POPULARITY
141 episodes with Interpretability
13 episodes with Interpretability
12 episodes with Interpretability
9 episodes with Interpretability
8 episodes with Interpretability
5 episodes with Interpretability
7 episodes with Interpretability
5 episodes with Interpretability
6 episodes with Interpretability
45 episodes with Interpretability
2 episodes with Interpretability
2 episodes with Interpretability
2 episodes with Interpretability
2 episodes with Interpretability
2 episodes with Interpretability
3 episodes with Interpretability
4 episodes with Interpretability
2 episodes with Interpretability
2 episodes with Interpretability
3 episodes with Interpretability
2 episodes with Interpretability
2 episodes with Interpretability
11 episodes with Interpretability
2 episodes with Interpretability
4 episodes with Interpretability
4 episodes with Interpretability
2 episodes with Interpretability
TL;DR We describe the persona selection model (PSM): the idea that LLMs learn to simulate diverse characters during pre-training, and post-training elicits and refines a particular such Assistant persona. Interactions with an AI assistant are then well-understood as being interactions with the Assistant—something roughly like a character in an LLM-generated story. We survey empirical behavioral, generalization, and interpretability-based evidence for PSM. PSM has consequences for AI development, such as recommending anthropomorphic reasoning about AI psychology and introduction of positive AI archetypes into training data. An important open question is how exhaustive PSM is, especially whether there might be sources of agency external to the Assistant persona, and how this might change in the future. Introduction What sort of thing is a modern AI assistant? One perspective holds that they are shallow, rigid systems that narrowly pattern-match user inputs to training data. Another perspective regards AI systems as alien creatures with learned goals, behaviors, and patterns of thought that are fundamentally inscrutable to us. A third option is to anthropomorphize AIs and regard them as something like a digital human. Developing good mental models for AI systems is important for predicting and controlling their behaviors. If our goal is to [...] ---Outline:(00:10) TL;DR(01:02) Introduction(06:18) The persona selection model(07:09) Predictive models and personas(09:54) From predictive models to AI assistants(12:43) Statement of the persona selection model(16:25) Empirical evidence for PSM(16:58) Evidence from generalization(22:48) Behavioral evidence(28:42) Evidence from interpretability(35:42) Complicating evidence(42:21) Consequences for AI development(42:45) AI assistants are human-like(43:23) Anthropomorphic reasoning about AI assistants is productive(49:17) AI welfare(51:35) The importance of good AI role models(53:49) Interpretability-based alignment auditing will be tractable(56:43) How exhaustive is PSM?(59:46) Shoggoths, actors, operating systems, and authors(01:00:46) Degrees of non-persona LLM agency en-US-AvaMultilingualNeural__ Green leaf or plant with yellow smiley face character attached.(01:06:52) Other sources of persona-like agency(01:11:17) Why might we expect PSM to be exhaustive?(01:12:21) Post-training as elicitation(01:14:54) Personas provide a simple way to fit the post-training data(01:17:55) How might these considerations change?(01:20:01) Empirical observations(01:27:07) Conclusion(01:30:30) Acknowledgements(01:31:15) Appendix A: Breaking character(01:32:52) Appendix B: An example of non-persona deception The original text contained 5 footnotes which were omitted from this narration. --- First published: February 23rd, 2026 Source: https://www.lesswrong.com/posts/dfoty34sT7CSKeJNn/the-persona-selection-model --- Narrated by TYPE III AUDIO. ---
What if physics is just the universe learning? Most Theories of Everything episodes are mind‑bending for their math, physics, philosophy, or consciousness implications. This one hits all four simultaneously. Professor Vitaly Vanchurin joins me to argue the cosmos isn't just modeled by neural networks—it literally is one. Learning dynamics aren't a metaphor for physics; they are the physics. Vanchurin shows why we need a three‑way unification: quantum mechanics, general relativity, and observers. As a listener of TOE you can get a special 20% off discount to The Economist and all it has to offer! Visit https://www.economist.com/toe TIMESTAMPS: - 00:00:00 - The Neural Network Universe - 00:05:48 - Learning Dynamics as Physics - 00:11:52 - Optimization and Variational Principles - 00:21:17 - Deriving Fundamental Field Equations - 00:28:47 - Fermions and Particle Emergence - 00:37:17 - Geometry of Learning Algorithms - 00:44:53 - Emergent Quantum Mechanics - 00:50:01 - Renormalization and Interpretability - 00:57:00 - Second Law of Learning - 01:05:10 - Subatomic Natural Selection - 01:15:40 - Consciousness and Learning Efficiency - 01:24:09 - Unifying Physics and Observers - 01:31:01 - Qualia and Hidden Variables - 01:40:24 - Free Energy Principle Integration - 01:46:04 - Epistemological Doubt and Advice LINKS MENTIONED: - Vitaly's Papers: https://inspirebeta.net/literature?sort=mostrecent&size=25&page=1&q=find%20author%20vanchurin - Vitaly's Lecture: https://youtu.be/TagDLiLb2VQ - Vitaly's Website: https://cosmos.phy.tufts.edu/~vitaly/ - Towards A Theory Of Machine Learning [Paper]: https://arxiv.org/pdf/2004.09280 - Autonomous Particles [Paper]: https://arxiv.org/pdf/2301.10077 - Emergent Field Theories From Neural Networks [Paper]: https://arxiv.org/pdf/2411.08138 - Covariant Gradient Descent [Paper]: https://arxiv.org/pdf/2504.05279 - A Quantum-Classical Duality And Emergent Spacetime [Paper]: https://arxiv.org/abs/1903.06083 - Emergent Quantumness In Neural Networks [Paper]: https://arxiv.org/abs/2012.05082 - Predictability Crisis In Inflationary Cosmology And Its Resolution [Paper]: https://arxiv.org/abs/gr-qc/9905097 - Stationary Measure In The Multiverse [Paper]: https://arxiv.org/abs/0812.0005 - The World As A Neural Network [Paper]: https://arxiv.org/pdf/2008.01540 - Self-Organized Criticality In Neural Networks [Paper]: https://arxiv.org/pdf/2107.03402v1 - One Hundred Authors Against Einstein [Book]: https://amazon.com/dp/B09PHH7KC8?tag=toe08-20 - Geocentric Cosmology: A New Look At The Measure Problem [Paper]: https://arxiv.org/abs/1006.4148 - Jacob Barandes [TOE]: https://youtu.be/gEK4-XtMwro - Yang-Hui He [TOE]: https://youtu.be/spIquD_mBFk - Eva Miranda [TOE]: https://youtu.be/6XyMepn-AZo - Felix Finster [TOE]: https://youtu.be/fXzO_KAqrh0 - Stephen Wolfram [TOE]: https://youtu.be/FkYer0xP37E - Stephen Wolfram 2 [TOE]: https://youtu.be/0YRlQQw0d-4 - Avshalom Elitzur [TOE]: https://youtu.be/pWRAaimQT1E - Ted Jacobson [TOE]: https://youtu.be/3mhctWlXyV8 - Geoffrey Hinton [TOE]: https://youtu.be/b_DUft-BdIE - Wayne Myrvold [TOE]: https://youtu.be/HIoviZe14pY - Cumrun Vafa [TOE]: https://youtu.be/kUHOoMX4Bqw - Claudia De Rham [TOE]: https://youtu.be/Ve_Mpd6dGv8 - Lee Smolin [TOE]: https://youtu.be/uOKOodQXjhc - Consciousness Iceberg [TOE]: https://youtu.be/65yjqIDghEk - Matthew Segall [TOE]: https://youtu.be/DeTm4fSXpbM - Andres Emilsson [TOE]: https://youtu.be/BBP8WZpYp0Y - Will Hahn [TOE]: https://youtu.be/3fkg0uTA3qU - David Wallace [TOE]: https://youtu.be/4MjNuJK5RzM - Karl Friston [TOE]: https://youtu.be/uk4NZorRjCo Learn more about your ad choices. Visit megaphone.fm/adchoices
From Palantir and Two Sigma to building Goodfire into the poster-child for actionable mechanistic interpretability, Mark Bissell (Member of Technical Staff) and Myra Deng (Head of Product) are trying to turn “peeking inside the model” into a repeatable production workflow by shipping APIs, landing real enterprise deployments, and now scaling the bet with a recent $150M Series B funding round at a $1.25B valuation.In this episode, we go far beyond the usual “SAEs are cool” take. We talk about Goodfire's core bet: that the AI lifecycle is still fundamentally broken because the only reliable control we have is data and we post-train, RLHF, and fine-tune by “slurping supervision through a straw,” hoping the model picks up the right behaviors while quietly absorbing the wrong ones. Goodfire's answer is to build a bi-directional interface between humans and models: read what's happening inside, edit it surgically, and eventually use interpretability during training so customization isn't just brute-force guesswork.Mark and Myra walk through what that looks like when you stop treating interpretability like a lab demo and start treating it like infrastructure: lightweight probes that add near-zero latency, token-level safety filters that can run at inference time, and interpretability workflows that survive messy constraints (multilingual inputs, synthetic→real transfer, regulated domains, no access to sensitive data). We also get a live window into what “frontier-scale interp” means operationally (i.e. steering a trillion-parameter model in real time by targeting internal features) plus why the same tooling generalizes cleanly from language models to genomics, medical imaging, and “pixel-space” world models.We discuss:* Myra + Mark's path: Palantir (health systems, forward-deployed engineering) → Goodfire early team; Two Sigma → Head of Product, translating frontier interpretability research into a platform and real-world deployments* What “interpretability” actually means in practice: not just post-hoc poking, but a broader “science of deep learning” approach across the full AI lifecycle (data curation → post-training → internal representations → model design)* Why post-training is the first big wedge: “surgical edits” for unintended behaviors likereward hacking, sycophancy, noise learned during customization plus the dream of targeted unlearning and bias removal without wrecking capabilities* SAEs vs probes in the real world: why SAE feature spaces sometimes underperform classifiers trained on raw activations for downstream detection tasks (hallucination, harmful intent, PII), and what that implies about “clean concept spaces”* Rakuten in production: deploying interpretability-based token-level PII detection at inference time to prevent routing private data to downstream providers plus the gnarly constraints: no training on real customer PII, synthetic→real transfer, English + Japanese, and tokenization quirks* Why interp can be operationally cheaper than LLM-judge guardrails: probes are lightweight, low-latency, and don't require hosting a second large model in the loop* Real-time steering at frontier scale: a demo of steering Kimi K2 (~1T params) live and finding features via SAE pipelines, auto-labeling via LLMs, and toggling a “Gen-Z slang” feature across multiple layers without breaking tool use* Hallucinations as an internal signal: the case that models have latent uncertainty / “user-pleasing” circuitry you can detect and potentially mitigate more directly than black-box methods* Steering vs prompting: the emerging view that activation steering and in-context learning are more closely connected than people think, including work mapping between the two (even for jailbreak-style behaviors)* Interpretability for science: using the same tooling across domains (genomics, medical imaging, materials) to debug spurious correlations and extract new knowledge up to and including early biomarker discovery work with major partners* World models + “pixel-space” interpretability: why vision/video models make concepts easier to see, how that accelerates the feedback loop, and why robotics/world-model partners are especially interesting design partners* The north star: moving from “data in, weights out” to intentional model design where experts can impart goals and constraints directly, not just via reward signals and brute-force post-training—Goodfire AI* Website: https://goodfire.ai* LinkedIn: https://www.linkedin.com/company/goodfire-ai/* X: https://x.com/GoodfireAIMyra Deng* Website: https://myradeng.com/* LinkedIn: https://www.linkedin.com/in/myra-deng/* X: https://x.com/myra_dengMark Bissell* LinkedIn: https://www.linkedin.com/in/mark-bissell/* X: https://x.com/MarkMBissellFull Video EpisodeTimestamps00:00:00 Introduction00:00:05 Introduction to the Latent Space Podcast and Guests from Goodfire00:00:29 What is Goodfire? Mission and Focus on Interpretability00:01:01 Goodfire's Practical Approach to Interpretability00:01:37 Goodfire's Series B Fundraise Announcement00:02:04 Backgrounds of Mark and Myra from Goodfire00:02:51 Team Structure and Roles at Goodfire00:05:13 What is Interpretability? Definitions and Techniques00:05:30 Understanding Errors00:07:29 Post-training vs. Pre-training Interpretability Applications00:08:51 Using Interpretability to Remove Unwanted Behaviors00:10:09 Grokking, Double Descent, and Generalization in Models00:10:15 404 Not Found Explained00:12:06 Subliminal Learning and Hidden Biases in Models00:14:07 How Goodfire Chooses Research Directions and Projects00:15:00 Troubleshooting Errors00:16:04 Limitations of SAEs and Probes in Interpretability00:18:14 Rakuten Case Study: Production Deployment of Interpretability00:20:45 Conclusion00:21:12 Efficiency Benefits of Interpretability Techniques00:21:26 Live Demo: Real-Time Steering in a Trillion Parameter Model00:25:15 How Steering Features are Identified and Labeled00:26:51 Detecting and Mitigating Hallucinations Using Interpretability00:31:20 Equivalence of Activation Steering and Prompting00:34:06 Comparing Steering with Fine-Tuning and LoRA Techniques00:36:04 Model Design and the Future of Intentional AI Development00:38:09 Getting Started in Mechinterp: Resources, Programs, and Open Problems00:40:51 Industry Applications and the Rise of Mechinterp in Practice00:41:39 Interpretability for Code Models and Real-World Usage00:43:07 Making Steering Useful for More Than Stylistic Edits00:46:17 Applying Interpretability to Healthcare and Scientific Discovery00:49:15 Why Interpretability is Crucial in High-Stakes Domains like Healthcare00:52:03 Call for Design Partners Across Domains00:54:18 Interest in World Models and Visual Interpretability00:57:22 Sci-Fi Inspiration: Ted Chiang and Interpretability01:00:14 Interpretability, Safety, and Alignment Perspectives01:04:27 Weak-to-Strong Generalization and Future Alignment Challenges01:05:38 Final Thoughts and Hiring/Collaboration Opportunities at GoodfireTranscriptShawn Wang [00:00:05]: So welcome to the Latent Space pod. We're back in the studio with our special MechInterp co-host, Vibhu. Welcome. Mochi, Mochi's special co-host. And Mochi, the mechanistic interpretability doggo. We have with us Mark and Myra from Goodfire. Welcome. Thanks for having us on. Maybe we can sort of introduce Goodfire and then introduce you guys. How do you introduce Goodfire today?Myra Deng [00:00:29]: Yeah, it's a great question. So Goodfire, we like to say, is an AI research lab that focuses on using interpretability to understand, learn from, and design AI models. And we really believe that interpretability will unlock the new generation, next frontier of safe and powerful AI models. That's our description right now, and I'm excited to dive more into the work we're doing to make that happen.Shawn Wang [00:00:55]: Yeah. And there's always like the official description. Is there an understatement? Is there an unofficial one that sort of resonates more with a different audience?Mark Bissell [00:01:01]: Well, being an AI research lab that's focused on interpretability, there's obviously a lot of people have a lot that they think about when they think of interpretability. And I think we have a pretty broad definition of what that means and the types of places that can be applied. And in particular, applying it in production scenarios, in high stakes industries, and really taking it sort of from the research world into the real world. Which, you know. It's a new field, so that hasn't been done all that much. And we're excited about actually seeing that sort of put into practice.Shawn Wang [00:01:37]: Yeah, I would say it wasn't too long ago that Anthopic was like still putting out like toy models or superposition and that kind of stuff. And I wouldn't have pegged it to be this far along. When you and I talked at NeurIPS, you were talking a little bit about your production use cases and your customers. And then not to bury the lead, today we're also announcing the fundraise, your Series B. $150 million. $150 million at a 1.25B valuation. Congrats, Unicorn.Mark Bissell [00:02:02]: Thank you. Yeah, no, things move fast.Shawn Wang [00:02:04]: We were talking to you in December and already some big updates since then. Let's dive, I guess, into a bit of your backgrounds as well. Mark, you were at Palantir working on health stuff, which is really interesting because the Goodfire has some interesting like health use cases. I don't know how related they are in practice.Mark Bissell [00:02:22]: Yeah, not super related, but I don't know. It was helpful context to know what it's like. Just to work. Just to work with health systems and generally in that domain. Yeah.Shawn Wang [00:02:32]: And Mara, you were at Two Sigma, which actually I was also at Two Sigma back in the day. Wow, nice.Myra Deng [00:02:37]: Did we overlap at all?Shawn Wang [00:02:38]: No, this is when I was briefly a software engineer before I became a sort of developer relations person. And now you're head of product. What are your sort of respective roles, just to introduce people to like what all gets done in Goodfire?Mark Bissell [00:02:51]: Yeah, prior to Goodfire, I was at Palantir for about three years as a forward deployed engineer, now a hot term. Wasn't always that way. And as a technical lead on the health care team and at Goodfire, I'm a member of the technical staff. And honestly, that I think is about as specific as like as as I could describe myself because I've worked on a range of things. And, you know, it's it's a fun time to be at a team that's still reasonably small. I think when I joined one of the first like ten employees, now we're above 40, but still, it looks like there's always a mix of research and engineering and product and all of the above. That needs to get done. And I think everyone across the team is, you know, pretty, pretty switch hitter in the roles they do. So I think you've seen some of the stuff that I worked on related to image models, which was sort of like a research demo. More recently, I've been working on our scientific discovery team with some of our life sciences partners, but then also building out our core platform for more of like flexing some of the kind of MLE and developer skills as well.Shawn Wang [00:03:53]: Very generalist. And you also had like a very like a founding engineer type role.Myra Deng [00:03:58]: Yeah, yeah.Shawn Wang [00:03:59]: So I also started as I still am a member of technical staff, did a wide range of things from the very beginning, including like finding our office space and all of this, which is we both we both visited when you had that open house thing. It was really nice.Myra Deng [00:04:13]: Thank you. Thank you. Yeah. Plug to come visit our office.Shawn Wang [00:04:15]: It looked like it was like 200 people. It has room for 200 people. But you guys are like 10.Myra Deng [00:04:22]: For a while, it was very empty. But yeah, like like Mark, I spend. A lot of my time as as head of product, I think product is a bit of a weird role these days, but a lot of it is thinking about how do we take our frontier research and really apply it to the most important real world problems and how does that then translate into a platform that's repeatable or a product and working across, you know, the engineering and research teams to make that happen and also communicating to the world? Like, what is interpretability? What is it used for? What is it good for? Why is it so important? All of these things are part of my day-to-day as well.Shawn Wang [00:05:01]: I love like what is things because that's a very crisp like starting point for people like coming to a field. They all do a fun thing. Vibhu, why don't you want to try tackling what is interpretability and then they can correct us.Vibhu Sapra [00:05:13]: Okay, great. So I think like one, just to kick off, it's a very interesting role to be head of product, right? Because you guys, at least as a lab, you're more of an applied interp lab, right? Which is pretty different than just normal interp, like a lot of background research. But yeah. You guys actually ship an API to try these things. You have Ember, you have products around it, which not many do. Okay. What is interp? So basically you're trying to have an understanding of what's going on in model, like in the model, in the internal. So different approaches to do that. You can do probing, SAEs, transcoders, all this stuff. But basically you have an, you have a hypothesis. You have something that you want to learn about what's happening in a model internals. And then you're trying to solve that from there. You can do stuff like you can, you know, you can do activation mapping. You can try to do steering. There's a lot of stuff that you can do, but the key question is, you know, from input to output, we want to have a better understanding of what's happening and, you know, how can we, how can we adjust what's happening on the model internals? How'd I do?Mark Bissell [00:06:12]: That was really good. I think that was great. I think it's also a, it's kind of a minefield of a, if you ask 50 people who quote unquote work in interp, like what is interpretability, you'll probably get 50 different answers. And. Yeah. To some extent also like where, where good fire sits in the space. I think that we're an AI research company above all else. And interpretability is a, is a set of methods that we think are really useful and worth kind of specializing in, in order to accomplish the goals we want to accomplish. But I think we also sort of see some of the goals as even more broader as, as almost like the science of deep learning and just taking a not black box approach to kind of any part of the like AI development life cycle, whether that. That means using interp for like data curation while you're training your model or for understanding what happened during post-training or for the, you know, understanding activations and sort of internal representations, what is in there semantically. And then a lot of sort of exciting updates that were, you know, are sort of also part of the, the fundraise around bringing interpretability to training, which I don't think has been done all that much before. A lot of this stuff is sort of post-talk poking at models as opposed to. To actually using this to intentionally design them.Shawn Wang [00:07:29]: Is this post-training or pre-training or is that not a useful.Myra Deng [00:07:33]: Currently focused on post-training, but there's no reason the techniques wouldn't also work in pre-training.Shawn Wang [00:07:38]: Yeah. It seems like it would be more active, applicable post-training because basically I'm thinking like rollouts or like, you know, having different variations of a model that you can tweak with the, with your steering. Yeah.Myra Deng [00:07:50]: And I think in a lot of the news that you've seen in, in, on like Twitter or whatever, you've seen a lot of unintended. Side effects come out of post-training processes, you know, overly sycophantic models or models that exhibit strange reward hacking behavior. I think these are like extreme examples. There's also, you know, very, uh, mundane, more mundane, like enterprise use cases where, you know, they try to customize or post-train a model to do something and it learns some noise or it doesn't appropriately learn the target task. And a big question that we've always had is like, how do you use your understanding of what the model knows and what it's doing to actually guide the learning process?Shawn Wang [00:08:26]: Yeah, I mean, uh, you know, just to anchor this for people, uh, one of the biggest controversies of last year was 4.0 GlazeGate. I've never heard of GlazeGate. I didn't know that was what it was called. The other one, they called it that on the blog post and I was like, well, how did OpenAI call it? Like officially use that term. And I'm like, that's funny, but like, yeah, I guess it's the pitch that if they had worked a good fire, they wouldn't have avoided it. Like, you know what I'm saying?Myra Deng [00:08:51]: I think so. Yeah. Yeah.Mark Bissell [00:08:53]: I think that's certainly one of the use cases. I think. Yeah. Yeah. I think the reason why post-training is a place where this makes a lot of sense is a lot of what we're talking about is surgical edits. You know, you want to be able to have expert feedback, very surgically change how your model is doing, whether that is, you know, removing a certain behavior that it has. So, you know, one of the things that we've been looking at or is, is another like common area where you would want to make a somewhat surgical edit is some of the models that have say political bias. Like you look at Quen or, um, R1 and they have sort of like this CCP bias.Shawn Wang [00:09:27]: Is there a CCP vector?Mark Bissell [00:09:29]: Well, there's, there are certainly internal, yeah. Parts of the representation space where you can sort of see where that lives. Yeah. Um, and you want to kind of, you know, extract that piece out.Shawn Wang [00:09:40]: Well, I always say, you know, whenever you find a vector, a fun exercise is just like, make it very negative to see what the opposite of CCP is.Mark Bissell [00:09:47]: The super America, bald eagles flying everywhere. But yeah. So in general, like lots of post-training tasks where you'd want to be able to, to do that. Whether it's unlearning a certain behavior or, you know, some of the other kind of cases where this comes up is, are you familiar with like the, the grokking behavior? I mean, I know the machine learning term of grokking.Shawn Wang [00:10:09]: Yeah.Mark Bissell [00:10:09]: Sort of this like double descent idea of, of having a model that is able to learn a generalizing, a generalizing solution, as opposed to even if memorization of some task would suffice, you want it to learn the more general way of doing a thing. And so, you know, another. A way that you can think about having surgical access to a model's internals would be learn from this data, but learn in the right way. If there are many possible, you know, ways to, to do that. Can make interp solve the double descent problem?Shawn Wang [00:10:41]: Depends, I guess, on how you. Okay. So I, I, I viewed that double descent as a problem because then you're like, well, if the loss curves level out, then you're done, but maybe you're not done. Right. Right. But like, if you actually can interpret what is a generalizing or what you're doing. What is, what is still changing, even though the loss is not changing, then maybe you, you can actually not view it as a double descent problem. And actually you're just sort of translating the space in which you view loss and like, and then you have a smooth curve. Yeah.Mark Bissell [00:11:11]: I think that's certainly like the domain of, of problems that we're, that we're looking to get.Shawn Wang [00:11:15]: Yeah. To me, like double descent is like the biggest thing to like ML research where like, if you believe in scaling, then you don't need, you need to know where to scale. And. But if you believe in double descent, then you don't, you don't believe in anything where like anything levels off, like.Vibhu Sapra [00:11:30]: I mean, also tendentially there's like, okay, when you talk about the China vector, right. There's the subliminal learning work. It was from the anthropic fellows program where basically you can have hidden biases in a model. And as you distill down or, you know, as you train on distilled data, those biases always show up, even if like you explicitly try to not train on them. So, you know, it's just like another use case of. Okay. If we can interpret what's happening in post-training, you know, can we clear some of this? Can we even determine what's there? Because yeah, it's just like some worrying research that's out there that shows, you know, we really don't know what's going on.Mark Bissell [00:12:06]: That is. Yeah. I think that's the biggest sentiment that we're sort of hoping to tackle. Nobody knows what's going on. Right. Like subliminal learning is just an insane concept when you think about it. Right. Train a model on not even the logits, literally the output text of a bunch of random numbers. And now your model loves owls. And you see behaviors like that, that are just, they defy, they defy intuition. And, and there are mathematical explanations that you can get into, but. I mean.Shawn Wang [00:12:34]: It feels so early days. Objectively, there are a sequence of numbers that are more owl-like than others. There, there should be.Mark Bissell [00:12:40]: According to, according to certain models. Right. It's interesting. I think it only applies to models that were initialized from the same starting Z. Usually, yes.Shawn Wang [00:12:49]: But I mean, I think that's a, that's a cheat code because there's not enough compute. But like if you believe in like platonic representation, like probably it will transfer across different models as well. Oh, you think so?Mark Bissell [00:13:00]: I think of it more as a statistical artifact of models initialized from the same seed sort of. There's something that is like path dependent from that seed that might cause certain overlaps in the latent space and then sort of doing this distillation. Yeah. Like it pushes it towards having certain other tendencies.Vibhu Sapra [00:13:24]: Got it. I think there's like a bunch of these open-ended questions, right? Like you can't train in new stuff during the RL phase, right? RL only reorganizes weights and you can only do stuff that's somewhat there in your base model. You're not learning new stuff. You're just reordering chains and stuff. But okay. My broader question is when you guys work at an interp lab, how do you decide what to work on and what's kind of the thought process? Right. Because we can ramble for hours. Okay. I want to know this. I want to know that. But like, how do you concretely like, you know, what's the workflow? Okay. There's like approaches towards solving a problem, right? I can try prompting. I can look at chain of thought. I can train probes, SAEs. But how do you determine, you know, like, okay, is this going anywhere? Like, do we have set stuff? Just, you know, if you can help me with all that. Yeah.Myra Deng [00:14:07]: It's a really good question. I feel like we've always at the very beginning of the company thought about like, let's go and try to learn what isn't working in machine learning today. Whether that's talking to customers or talking to researchers at other labs, trying to understand both where the frontier is going and where things are really not falling apart today. And then developing a perspective on how we can push the frontier using interpretability methods. And so, you know, even our chief scientist, Tom, spends a lot of time talking to customers and trying to understand what real world problems are and then taking that back and trying to apply the current state of the art to those problems and then seeing where they fall down basically. And then using those failures or those shortcomings to understand what hills to climb when it comes to interpretability research. So like on the fundamental side, for instance, when we have done some work applying SAEs and probes, we've encountered, you know, some shortcomings in SAEs that we found a little bit surprising. And so have gone back to the drawing board and done work on that. And then, you know, we've done some work on better foundational interpreter models. And a lot of our team's research is focused on what is the next evolution beyond SAEs, for instance. And then when it comes to like control and design of models, you know, we tried steering with our first API and realized that it still fell short of black box techniques like prompting or fine tuning. And so went back to the drawing board and we're like, how do we make that not the case and how do we improve it beyond that? And one of our researchers, Ekdeep, who just joined is actually Ekdeep and Atticus are like steering experts and have spent a lot of time trying to figure out like, what is the research that enables us to actually do this in a much more powerful, robust way? So yeah, the answer is like, look at real world problems, try to translate that into a research agenda and then like hill climb on both of those at the same time.Shawn Wang [00:16:04]: Yeah. Mark has the steering CLI demo queued up, which we're going to go into in a sec. But I always want to double click on when you drop hints, like we found some problems with SAEs. Okay. What are they? You know, and then we can go into the demo. Yeah.Myra Deng [00:16:19]: I mean, I'm curious if you have more thoughts here as well, because you've done it in the healthcare domain. But I think like, for instance, when we do things like trying to detect behaviors within models that are harmful or like behaviors that a user might not want to have in their model. So hallucinations, for instance, harmful intent, PII, all of these things. We first tried using SAE probes for a lot of these tasks. So taking the feature activation space from SAEs and then training classifiers on top of that, and then seeing how well we can detect the properties that we might want to detect in model behavior. And we've seen in many cases that probes just trained on raw activations seem to perform better than SAE probes, which is a bit surprising if you think that SAEs are actually also capturing the concepts that you would want to capture cleanly and more surgically. And so that is an interesting observation. I don't think that is like, I'm not down on SAEs at all. I think there are many, many things they're useful for, but we have definitely run into cases where I think the concept space described by SAEs is not as clean and accurate as we would expect it to be for actual like real world downstream performance metrics.Mark Bissell [00:17:34]: Fair enough. Yeah. It's the blessing and the curse of unsupervised methods where you get to peek into the AI's mind. But sometimes you wish that you saw other things when you walked inside there. Although in the PII instance, I think weren't an SAE based approach actually did prove to be the most generalizable?Myra Deng [00:17:53]: It did work well in the case that we published with Rakuten. And I think a lot of the reasons it worked well was because we had a noisier data set. And so actually the blessing of unsupervised learning is that we actually got to get more meaningful, generalizable signal from SAEs when the data was noisy. But in other cases where we've had like good data sets, it hasn't been the case.Shawn Wang [00:18:14]: And just because you named Rakuten and I don't know if we'll get it another chance, like what is the overall, like what is Rakuten's usage or production usage? Yeah.Myra Deng [00:18:25]: So they are using us to essentially guardrail and inference time monitor their language model usage and their agent usage to detect things like PII so that they don't route private user information.Myra Deng [00:18:41]: And so that's, you know, going through all of their user queries every day. And that's something that we deployed with them a few months ago. And now we are actually exploring very early partnerships, not just with Rakuten, but with other people around how we can help with potentially training and customization use cases as well. Yeah.Shawn Wang [00:19:03]: And for those who don't know, like it's Rakuten is like, I think number one or number two e-commerce store in Japan. Yes. Yeah.Mark Bissell [00:19:10]: And I think that use case actually highlights a lot of like what it looks like to deploy things in practice that you don't always think about when you're doing sort of research tasks. So when you think about some of the stuff that came up there that's more complex than your idealized version of a problem, they were encountering things like synthetic to real transfer of methods. So they couldn't train probes, classifiers, things like that on actual customer data of PII. So what they had to do is use synthetic data sets. And then hope that that transfer is out of domain to real data sets. And so we can evaluate performance on the real data sets, but not train on customer PII. So that right off the bat is like a big challenge. You have multilingual requirements. So this needed to work for both English and Japanese text. Japanese text has all sorts of quirks, including tokenization behaviors that caused lots of bugs that caused us to be pulling our hair out. And then also a lot of tasks you'll see. You might make simplifying assumptions if you're sort of treating it as like the easiest version of the problem to just sort of get like general results where maybe you say you're classifying a sentence to say, does this contain PII? But the need that Rakuten had was token level classification so that you could precisely scrub out the PII. So as we learned more about the problem, you're sort of speaking about what that looks like in practice. Yeah. A lot of assumptions end up breaking. And that was just one instance where you. A problem that seems simple right off the bat ends up being more complex as you keep diving into it.Vibhu Sapra [00:20:41]: Excellent. One of the things that's also interesting with Interp is a lot of these methods are very efficient, right? So where you're just looking at a model's internals itself compared to a separate like guardrail, LLM as a judge, a separate model. One, you have to host it. Two, there's like a whole latency. So if you use like a big model, you have a second call. Some of the work around like self detection of hallucination, it's also deployed for efficiency, right? So if you have someone like Rakuten doing it in production live, you know, that's just another thing people should consider.Mark Bissell [00:21:12]: Yeah. And something like a probe is super lightweight. Yeah. It's no extra latency really. Excellent.Shawn Wang [00:21:17]: You have the steering demos lined up. So we were just kind of see what you got. I don't, I don't actually know if this is like the latest, latest or like alpha thing.Mark Bissell [00:21:26]: No, this is a pretty hacky demo from from a presentation that someone else on the team recently gave. So this will give a sense for, for technology. So you can see the steering and action. Honestly, I think the biggest thing that this highlights is that as we've been growing as a company and taking on kind of more and more ambitious versions of interpretability related problems, a lot of that comes to scaling up in various different forms. And so here you're going to see steering on a 1 trillion parameter model. This is Kimi K2. And so it's sort of fun that in addition to the research challenges, there are engineering challenges that we're now tackling. Cause for any of this to be sort of useful in production, you need to be thinking about what it looks like when you're using these methods on frontier models as opposed to sort of like toy kind of model organisms. So yeah, this was thrown together hastily, pretty fragile behind the scenes, but I think it's quite a fun demo. So screen sharing is on. So I've got two terminal sessions pulled up here. On the left is a forked version that we have of the Kimi CLI that we've got running to point at our custom hosted Kimi model. And then on the right is a set up that will allow us to steer on certain concepts. So I should be able to chat with Kimi over here. Tell it hello. This is running locally. So the CLI is running locally, but the Kimi server is running back to the office. Well, hopefully should be, um, that's too much to run on that Mac. Yeah. I think it's, uh, it takes a full, like each 100 node. I think it's like, you can. You can run it on eight GPUs, eight 100. So, so yeah, Kimi's running. We can ask it a prompt. It's got a forked version of our, uh, of the SG line code base that we've been working on. So I'm going to tell it, Hey, this SG line code base is slow. I think there's a bug. Can you try to figure it out? There's a big code base, so it'll, it'll spend some time doing this. And then on the right here, I'm going to initialize in real time. Some steering. Let's see here.Mark Bissell [00:23:33]: searching for any. Bugs. Feature ID 43205.Shawn Wang [00:23:38]: Yeah.Mark Bissell [00:23:38]: 20, 30, 40. So let me, uh, this is basically a feature that we found that inside Kimi seems to cause it to speak in Gen Z slang. And so on the left, it's still sort of thinking normally it might take, I don't know, 15 seconds for this to kick in, but then we're going to start hopefully seeing him do this code base is massive for real. So we're going to start. We're going to start seeing Kimi transition as the steering kicks in from normal Kimi to Gen Z Kimi and both in its chain of thought and its actual outputs.Mark Bissell [00:24:19]: And interestingly, you can see, you know, it's still able to call tools, uh, and stuff. It's um, it's purely sort of it's it's demeanor. And there are other features that we found for interesting things like concision. So that's more of a practical one. You can make it more concise. Um, the types of programs, uh, programming languages that uses, but yeah, as we're seeing it come in. Pretty good. Outputs.Shawn Wang [00:24:43]: Scheduler code is actually wild.Vibhu Sapra [00:24:46]: Yo, this code is actually insane, bro.Vibhu Sapra [00:24:53]: What's the process of training in SAE on this, or, you know, how do you label features? I know you guys put out a pretty cool blog post about, um, finding this like autonomous interp. Um, something. Something about how agents for interp is different than like coding agents. I don't know while this is spewing up, but how, how do we find feature 43, two Oh five. Yeah.Mark Bissell [00:25:15]: So in this case, um, we, our platform that we've been building out for a long time now supports all the sort of classic out of the box interp techniques that you might want to have like SAE training, probing things of that kind, I'd say the techniques for like vanilla SAEs are pretty well established now where. You take your model that you're interpreting, run a whole bunch of data through it, gather activations, and then yeah, pretty straightforward pipeline to train an SAE. There are a lot of different varieties. There's top KSAEs, batch top KSAEs, um, normal ReLU SAEs. And then once you have your sparse features to your point, assigning labels to them to actually understand that this is a gen Z feature, that's actually where a lot of the kind of magic happens. Yeah. And the most basic standard technique is look at all of your d input data set examples that cause this feature to fire most highly. And then you can usually pick out a pattern. So for this feature, If I've run a diverse enough data set through my model feature 43, two Oh five. Probably tends to fire on all the tokens that sounds like gen Z slang. You know, that's the, that's the time of year to be like, Oh, I'm in this, I'm in this Um, and, um, so, you know, you could have a human go through all 43,000 concepts andVibhu Sapra [00:26:34]: And I've got to ask the basic question, you know, can we get examples where it hallucinates, pass it through, see what feature activates for hallucinations? Can I just, you know, turn hallucination down?Myra Deng [00:26:51]: Oh, wow. You really predicted a project we're already working on right now, which is detecting hallucinations using interpretability techniques. And this is interesting because hallucinations is something that's very hard to detect. And it's like a kind of a hairy problem and something that black box methods really struggle with. Whereas like Gen Z, you could always train a simple classifier to detect that hallucinations is harder. But we've seen that models internally have some... Awareness of like uncertainty or some sort of like user pleasing behavior that leads to hallucinatory behavior. And so, yeah, we have a project that's trying to detect that accurately. And then also working on mitigating the hallucinatory behavior in the model itself as well.Shawn Wang [00:27:39]: Yeah, I would say most people are still at the level of like, oh, I would just turn temperature to zero and that turns off hallucination. And I'm like, well, that's a fundamental misunderstanding of how this works. Yeah.Mark Bissell [00:27:51]: Although, so part of what I like about that question is you, there are SAE based approaches that might like help you get at that. But oftentimes the beauty of SAEs and like we said, the curse is that they're unsupervised. So when you have a behavior that you deliberately would like to remove, and that's more of like a supervised task, often it is better to use something like probes and specifically target the thing that you're interested in reducing as opposed to sort of like hoping that when you fragment the latent space, one of the vectors that pops out.Vibhu Sapra [00:28:20]: And as much as we're training an autoencoder to be sparse, we're not like for sure certain that, you know, we will get something that just correlates to hallucination. You'll probably split that up into 20 other things and who knows what they'll be.Mark Bissell [00:28:36]: Of course. Right. Yeah. So there's no sort of problems with like feature splitting and feature absorption. And then there's the off target effects, right? Ideally, you would want to be very precise where if you reduce the hallucination feature, suddenly maybe your model can't write. Creatively anymore. And maybe you don't like that, but you want to still stop it from hallucinating facts and figures.Shawn Wang [00:28:55]: Good. So Vibhu has a paper to recommend there that we'll put in the show notes. But yeah, I mean, I guess just because your demo is done, any any other things that you want to highlight or any other interesting features you want to show?Mark Bissell [00:29:07]: I don't think so. Yeah. Like I said, this is a pretty small snippet. I think the main sort of point here that I think is exciting is that there's not a whole lot of inter being applied to models quite at this scale. You know, Anthropic certainly has some some. Research and yeah, other other teams as well. But it's it's nice to see these techniques, you know, being put into practice. I think not that long ago, the idea of real time steering of a trillion parameter model would have sounded.Shawn Wang [00:29:33]: Yeah. The fact that it's real time, like you started the thing and then you edited the steering vector.Vibhu Sapra [00:29:38]: I think it's it's an interesting one TBD of what the actual like production use case would be on that, like the real time editing. It's like that's the fun part of the demo, right? You can kind of see how this could be served behind an API, right? Like, yes, you're you only have so many knobs and you can just tweak it a bit more. And I don't know how it plays in. Like people haven't done that much with like, how does this work with or without prompting? Right. How does this work with fine tuning? Like, there's a whole hype of continual learning, right? So there's just so much to see. Like, is this another parameter? Like, is it like parameter? We just kind of leave it as a default. We don't use it. So I don't know. Maybe someone here wants to put out a guide on like how to use this with prompting when to do what?Mark Bissell [00:30:18]: Oh, well, I have a paper recommendation. I think you would love from Act Deep on our team, who is an amazing researcher, just can't say enough amazing things about Act Deep. But he actually has a paper that as well as some others from the team and elsewhere that go into the essentially equivalence of activation steering and in context learning and how those are from a he thinks of everything in a cognitive neuroscience Bayesian framework, but basically how you can precisely show how. Prompting in context, learning and steering exhibit similar behaviors and even like get quantitative about the like magnitude of steering you would need to do to induce a certain amount of behavior similar to certain prompting, even for things like jailbreaks and stuff. It's a really cool paper. Are you saying steering is less powerful than prompting? More like you can almost write a formula that tells you how to convert between the two of them.Myra Deng [00:31:20]: And so like formally equivalent actually in the in the limit. Right.Mark Bissell [00:31:24]: So like one case study of this is for jailbreaks there. I don't know. Have you seen the stuff where you can do like many shot jailbreaking? You like flood the context with examples of the behavior. And the topic put out that paper.Shawn Wang [00:31:38]: A lot of people were like, yeah, we've been doing this, guys.Mark Bissell [00:31:40]: Like, yeah, what's in this in context learning and activation steering equivalence paper is you can like predict the number. Number of examples that you will need to put in there in order to jailbreak the model. That's cool. By doing steering experiments and using this sort of like equivalence mapping. That's cool. That's really cool. It's very neat. Yeah.Shawn Wang [00:32:02]: I was going to say, like, you know, I can like back rationalize that this makes sense because, you know, what context is, is basically just, you know, it updates the KV cache kind of and like and then every next token inference is still like, you know, the sheer sum of everything all the way. It's plus all the context. It's up to date. And you could, I guess, theoretically steer that with you probably replace that with your steering. The only problem is steering typically is on one layer, maybe three layers like like you did. So it's like not exactly equivalent.Mark Bissell [00:32:33]: Right, right. There's sort of you need to get precise about, yeah, like how you sort of define steering and like what how you're modeling the setup. But yeah, I've got the paper pulled up here. Belief dynamics reveal the dual nature. Yeah. The title is Belief Dynamics Reveal the Dual Nature of Incompetence. And it's an exhibition of the practical context learning and activation steering. So Eric Bigelow, Dan Urgraft on the who are doing fellowships at Goodfire, Ekt Deep's the final author there.Myra Deng [00:32:59]: I think actually to your question of like, what is the production use case of steering? I think maybe if you just think like one level beyond steering as it is today. Like imagine if you could adapt your model to be, you know, an expert legal reasoner. Like in almost real time, like very quickly. efficiently using human feedback or using like your semantic understanding of what the model knows and where it knows that behavior. I think that while it's not clear what the product is at the end of the day, it's clearly very valuable. Thinking about like what's the next interface for model customization and adaptation is a really interesting problem for us. Like we have heard a lot of people actually interested in fine-tuning an RL for open weight models in production. And so people are using things like Tinker or kind of like open source libraries to do that, but it's still very difficult to get models fine-tuned and RL'd for exactly what you want them to do unless you're an expert at model training. And so that's like something we'reShawn Wang [00:34:06]: looking into. Yeah. I never thought so. Tinker from Thinking Machines famously uses rank one LoRa. Is that basically the same as steering? Like, you know, what's the comparison there?Mark Bissell [00:34:19]: Well, so in that case, you are still applying updates to the parameters, right?Shawn Wang [00:34:25]: Yeah. You're not touching a base model. You're touching an adapter. It's kind of, yeah.Mark Bissell [00:34:30]: Right. But I guess it still is like more in parameter space then. I guess it's maybe like, are you modifying the pipes or are you modifying the water flowing through the pipes to get what you're after? Yeah. Just maybe one way.Mark Bissell [00:34:44]: I like that analogy. That's my mental map of it at least, but it gets at this idea of model design and intentional design, which is something that we're, that we're very focused on. And just the fact that like, I hope that we look back at how we're currently training models and post-training models and just think what a primitive way of doing that right now. Like there's no intentionalityShawn Wang [00:35:06]: really in... It's just data, right? The only thing in control is what data we feed in.Mark Bissell [00:35:11]: So, so Dan from Goodfire likes to use this analogy of, you know, he has a couple of young kids and he talks about like, what if I could only teach my kids how to be good people by giving them cookies or like, you know, giving them a slap on the wrist if they do something wrong, like not telling them why it was wrong or like what they should have done differently or something like that. Just figure it out. Right. Exactly. So that's RL. Yeah. Right. And, and, you know, it's sample inefficient. There's, you know, what do they say? It's like slurping feedback. It's like, slurping supervision. Right. And so you'd like to get to the point where you can have experts giving feedback to their models that are, uh, internalized and, and, you know, steering is an inference time way of sort of getting that idea. But ideally you're moving to a world whereVibhu Sapra [00:36:04]: it is much more intentional design in perpetuity for these models. Okay. This is one of the questions we asked Emmanuel from Anthropic on the podcast a few months ago. Basically the question, was you're at a research lab that does model training, foundation models, and you're on an interp team. How does it tie back? Right? Like, does this, do ideas come from the pre-training team? Do they go back? Um, you know, so for those interested, you can, you can watch that. There wasn't too much of a connect there, but it's still something, you know, it's something they want toMark Bissell [00:36:33]: push for down the line. It can be useful for all of the above. Like there are certainly post-hocVibhu Sapra [00:36:39]: use cases where it doesn't need to touch that. I think the other thing a lot of people forget is this stuff isn't too computationally expensive, right? Like I would say, if you're interested in getting into research, MechInterp is one of the most approachable fields, right? A lot of this train an essay, train a probe, this stuff, like the budget for this one, there's already a lot done. There's a lot of open source work. You guys have done some too. Um, you know,Shawn Wang [00:37:04]: There's like notebooks from the Gemini team for Neil Nanda or like, this is how you do it. Just step through the notebook.Vibhu Sapra [00:37:09]: Even if you're like, not even technical with any of this, you can still make like progress. There, you can look at different activations, but, uh, if you do want to get into training, you know, training this stuff, correct me if I'm wrong is like in the thousands of dollars, not even like, it's not that high scale. And then same with like, you know, applying it, doing it for post-training or all this stuff is fairly cheap in scale of, okay. I want to get into like model training. I don't have compute for like, you know, pre-training stuff. So it's, it's a very nice field to get into. And also there's a lot of like open questions, right? Um, some of them have to go with, okay, I want a product. I want to solve this. Like there's also just a lot of open-ended stuff that people could work on. That's interesting. Right. I don't know if you guys have any calls for like, what's open questions, what's open work that you either open collaboration with, or like, you'd just like to see solved or just, you know, for people listening that want to get into McInturk because people always talk about it. What are, what are the things they should check out? Start, of course, you know, join you guys as well. I'm sure you're hiring.Myra Deng [00:38:09]: There's a paper, I think from, was it Lee, uh, Sharky? It's open problems and, uh, it's, it's a bit of interpretability, which I recommend everyone who's interested in the field. Read. I'm just like a really comprehensive overview of what are the things that experts in the field think are the most important problems to be solved. I also think to your point, it's been really, really inspiring to see, I think a lot of young people getting interested in interpretability, actually not just young people also like scientists to have been, you know, experts in physics for many years and in biology or things like this, um, transitioning into interp, because the barrier of, of what's now interp. So it's really cool to see a number to entry is, you know, in some ways low and there's a lot of information out there and ways to get started. There's this anecdote of like professors at universities saying that all of a sudden every incoming PhD student wants to study interpretability, which was not the case a few years ago. So it just goes to show how, I guess, like exciting the field is, how fast it's moving, how quick it is to get started and things like that.Mark Bissell [00:39:10]: And also just a very welcoming community. You know, there's an open source McInturk Slack channel. There are people are always posting questions and just folks in the space are always responsive if you ask things on various forums and stuff. But yeah, the open paper, open problems paper is a really good one.Myra Deng [00:39:28]: For other people who want to get started, I think, you know, MATS is a great program. What's the acronym for? Machine Learning and Alignment Theory Scholars? It's like the...Vibhu Sapra [00:39:40]: Normally summer internship style.Myra Deng [00:39:42]: Yeah, but they've been doing it year round now. And actually a lot of our full-time staff have come through that program or gone through that program. And it's great for anyone who is transitioning into interpretability. There's a couple other fellows programs. We do one as well as Anthropic. And so those are great places to get started if anyone is interested.Mark Bissell [00:40:03]: Also, I think been seen as a research field for a very long time. But I think engineering... I think engineers are sorely wanted for interpretability as well, especially at Goodfire, but elsewhere, as it does scale up.Shawn Wang [00:40:18]: I should mention that Lee actually works with you guys, right? And in the London office and I'm adding our first ever McInturk track at AI Europe because I see this industry applications now emerging. And I'm pretty excited to, you know, help push that along. Yeah, I was looking forward to that. It'll effectively be the first industry McInturk conference. Yeah. I'm so glad you added that. You know, it's still a little bit of a bet. It's not that widespread, but I can definitely see this is the time to really get into it. We want to be early on things.Mark Bissell [00:40:51]: For sure. And I think the field understands this, right? So at ICML, I think the title of the McInturk workshop this year was actionable interpretability. And there was a lot of discussion around bringing it to various domains. Everyone's adding pragmatic, actionable, whatever.Shawn Wang [00:41:10]: It's like, okay, well, we weren't actionable before, I guess. I don't know.Vibhu Sapra [00:41:13]: And I mean, like, just, you know, being in Europe, you see the Interp room. One, like old school conferences, like, I think they had a very tiny room till they got lucky and they got it doubled. But there's definitely a lot of interest, a lot of niche research. So you see a lot of research coming out of universities, students. We covered the paper last week. It's like two unknown authors, not many citations. But, you know, you can make a lot of meaningful work there. Yeah. Yeah. Yeah.Shawn Wang [00:41:39]: Yeah. I think people haven't really mentioned this yet. It's just Interp for code. I think it's like an abnormally important field. We haven't mentioned this yet. The conspiracy theory last two years ago was when the first SAE work came out of Anthropic was they would do like, oh, we just used SAEs to turn the bad code vector down and then turn up the good code. And I think like, isn't that the dream? Like, you know, like, but basically, I guess maybe, why is it funny? Like, it's... If it was realistic, it would not be funny. It would be like, no, actually, we should do this. But it's funny because we know there's like, we feel there's some limitations to what steering can do. And I think a lot of the public image of steering is like the Gen Z stuff. Like, oh, you can make it really love the Golden Gate Bridge, or you can make it speak like Gen Z. To like be a legal reasoner seems like a huge stretch. Yeah. And I don't know if that will get there this way. Yeah.Myra Deng [00:42:36]: I think, um, I will say we are announcing. Something very soon that I will not speak too much about. Um, but I think, yeah, this is like what we've run into again and again is like, we, we don't want to be in the world where steering is only useful for like stylistic things. That's definitely not, not what we're aiming for. But I think the types of interventions that you need to do to get to things like legal reasoning, um, are much more sophisticated and require breakthroughs in, in learning algorithms. And that's, um...Shawn Wang [00:43:07]: And is this an emergent property of scale as well?Myra Deng [00:43:10]: I think so. Yeah. I mean, I think scale definitely helps. I think scale allows you to learn a lot of information and, and reduce noise across, you know, large amounts of data. But I also think we think that there's ways to do things much more effectively, um, even, even at scale. So like actually learning exactly what you want from the data and not learning things that you do that you don't want exhibited in the data. So we're not like anti-scale, but we are also realizing that scale is not going to get us anywhere. It's not going to get us to the type of AI development that we want to be at in, in the future as these models get more powerful and get deployed in all these sorts of like mission critical contexts. Current life cycle of training and deploying and evaluations is, is to us like deeply broken and has opportunities to, to improve. So, um, more to come on that very, very soon.Mark Bissell [00:44:02]: And I think that that's a use basically, or maybe just like a proof point that these concepts do exist. Like if you can manipulate them in the precise best way, you can get the ideal combination of them that you desire. And steering is maybe the most coarse grained sort of peek at what that looks like. But I think it's evocative of what you could do if you had total surgical control over every concept, every parameter. Yeah, exactly.Myra Deng [00:44:30]: There were like bad code features. I've got it pulled up.Vibhu Sapra [00:44:33]: Yeah. Just coincidentally, as you guys are talking.Shawn Wang [00:44:35]: This is like, this is exactly.Vibhu Sapra [00:44:38]: There's like specifically a code error feature that activates and they show, you know, it's not, it's not typo detection. It's like, it's, it's typos in code. It's not typical typos. And, you know, you can, you can see it clearly activates where there's something wrong in code. And they have like malicious code, code error. They have a whole bunch of sub, you know, sub broken down little grain features. Yeah.Shawn Wang [00:45:02]: Yeah. So, so the, the rough intuition for me, the, why I talked about post-training was that, well, you just, you know, have a few different rollouts with all these things turned off and on and whatever. And then, you know, you can, that's, that's synthetic data you can kind of post-train on. Yeah.Vibhu Sapra [00:45:13]: And I think we make it sound easier than it is just saying, you know, they do the real hard work.Myra Deng [00:45:19]: I mean, you guys, you guys have the right idea. Exactly. Yeah. We replicated a lot of these features in, in our Lama models as well. I remember there was like.Vibhu Sapra [00:45:26]: And I think a lot of this stuff is open, right? Like, yeah, you guys opened yours. DeepMind has opened a lot of essays on Gemma. Even Anthropic has opened a lot of this. There's, there's a lot of resources that, you know, we can probably share of people that want to get involved.Shawn Wang [00:45:41]: Yeah. And special shout out to like Neuronpedia as well. Yes. Like, yeah, amazing piece of work to visualize those things.Myra Deng [00:45:49]: Yeah, exactly.Shawn Wang [00:45:50]: I guess I wanted to pivot a little bit on, onto the healthcare side, because I think that's a big use case for you guys. We haven't really talked about it yet. This is a bit of a crossover for me because we are, we are, we do have a separate science pod that we're starting up for AI, for AI for science, just because like, it's such a huge investment category and also I'm like less qualified to do it, but we actually have bio PhDs to cover that, which is great, but I need to just kind of recover, recap your work, maybe on the evil two stuff, but then, and then building forward.Mark Bissell [00:46:17]: Yeah, for sure. And maybe to frame up the conversation, I think another kind of interesting just lens on interpretability in general is a lot of the techniques that were described. are ways to solve the AI human interface problem. And it's sort of like bidirectional communication is the goal there. So what we've been talking about with intentional design of models and, you know, steering, but also more advanced techniques is having humans impart our desires and control into models and over models. And the reverse is also very interesting, especially as you get to superhuman models, whether that's narrow superintelligence, like these scientific models that work on genomics, data, medical imaging, things like that. But down the line, you know, superintelligence of other forms as well. What knowledge can the AIs teach us as sort of that, that the other direction in that? And so some of our life science work to date has been getting at exactly that question, which is, well, some of it does look like debugging these various life sciences models, understanding if they're actually performing well, on tasks, or if they're picking up on spurious correlations, for instance, genomics models, you would like to know whether they are sort of focusing on the biologically relevant things that you care about, or if it's using some simpler correlate, like the ancestry of the person that it's looking at. But then also in the instances where they are superhuman, and maybe they are understanding elements of the human genome that we don't have names for or specific, you know, yeah, discoveries that they've made that that we don't know about, that's, that's a big goal. And so we're already seeing that, right, we are partnered with organizations like Mayo Clinic, leading research health system in the United States, our Institute, as well as a startup called Prima Menta, which focuses on neurodegenerative disease. And in our partnership with them, we've used foundation models, they've been training and applied our interpretability techniques to find novel biomarkers for Alzheimer's disease. So I think this is just the tip of the iceberg. But it's, that's like a flavor of some of the things that we're working on.Shawn Wang [00:48:36]: Yeah, I think that's really fantastic. Obviously, we did the Chad Zuckerberg pod last year as well. And like, there's a plethora of these models coming out, because there's so much potential and research. And it's like, very interesting how it's basically the same as language models, but just with a different underlying data set. But it's like, it's the same exact techniques. Like, there's no change, basically.Mark Bissell [00:48:59]: Yeah. Well, and even in like other domains, right? Like, you know, robotics, I know, like a lot of the companies just use Gemma as like the like backbone, and then they like make it into a VLA that like takes these actions. It's, it's, it's transformers all the way down. So yeah.Vibhu Sapra [00:49:15]: Like we have Med Gemma now, right? Like this week, even there was Med Gemma 1.5. And they're training it on this stuff, like 3d scans, medical domain knowledge, and all that stuff, too. So there's a push from both sides. But I think the thing that, you know, one of the things about McInturpp is like, you're a little bit more cautious in some domains, right? So healthcare, mainly being one, like guardrails, understanding, you know, we're more risk adverse to something going wrong there. So even just from a basic understanding, like, if we're trusting these systems to make claims, we want to know why and what's going on.Myra Deng [00:49:51]: Yeah, I think there's totally a kind of like deployment bottleneck to actually using. foundation models for real patient usage or things like that. Like, say you're using a model for rare disease prediction, you probably want some explanation as to why your model predicted a certain outcome, and an interpretable explanation at that. So that's definitely a use case. But I also think like, being able to extract scientific information that no human knows to accelerate drug discovery and disease treatment and things like that actually is a really, really big unlock for science, like scientific discovery. And you've seen a lot of startups, like say that they're going to accelerate scientific discovery. And I feel like we actually are doing that through our interp techniques. And kind of like, almost by accident, like, I think we got reached out to very, very early on from these healthcare institutions. And none of us had healthcare.Shawn Wang [00:50:49]: How did they even hear of you? A podcast.Myra Deng [00:50:51]: Oh, okay. Yeah, podcast.Vibhu Sapra [00:50:53]: Okay, well, now's that time, you know.Myra Deng [00:50:55]: Everyone can call us.Shawn Wang [00:50:56]: Podcasts are the most important thing. Everyone should listen to podcasts.Myra Deng [00:50:59]: Yeah, they reached out. They were like, you know, we have these really smart models that we've trained, and we want to know what they're doing. And we were like, really early that time, like three months old, and it was a few of us. And we were like, oh, my God, we've never used these models. Let's figure it out. But it's also like, great proof that interp techniques scale pretty well across domains. We didn't really have to learn too much about.Shawn Wang [00:51:21]: Interp is a machine learning technique, machine learning skills everywhere, right? Yeah. And it's obviously, it's just like a general insight. Yeah. Probably to finance too, I think, which would be fun for our history. I don't know if you have anything to say there.Mark Bissell [00:51:34]: Yeah, well, just across the science. Like, we've also done work on material science. Yeah, it really runs the gamut.Vibhu Sapra [00:51:40]: Yeah. Awesome. And, you know, for those that should reach out, like, you're obviously experts in this, but like, is there a call out for people that you're looking to partner with, design partners, people to use your stuff outside of just, you know, the general developer that wants to. Plug and play steering stuff, like on the research side more so, like, are there ideal design partners, customers, stuff like that?Myra Deng [00:52:03]: Yeah, I can talk about maybe non-life sciences, and then I'm curious to hear from you on the life sciences side. But we're looking for design partners across many domains, language, anyone who's customizing language models or trying to push the frontier of code or reasoning models is really interesting to us. And then also interested in the frontier of modeling. There's a lot of models that work in, like, pixel space, as we call it. So if you're doing world models, video models, even robotics, where there's not a very clean natural language interface to interact with, I think we think that Interp can really help and are looking for a few partners in that space.Shawn Wang [00:52:43]: Just because you mentioned the keyword
Collaborative robots and machine-learning-based virtual agents are increasingly entering the human workspace with the aim of increasing productivity, enhancing safety, and improving the quality of our lives. These agents will dynamically interact with a wide variety of people in dynamic and novel contexts, increasing the prevalence of human-machine teams in applications spanning from healthcare and manufacturing to household assistance. My research aims to create transparent embodied systems that can support users and interact with humans, pushing the frontier of real-world robotics systems towards those that understand human behavior, maintain interpretability, and coordinate with high performance. In this talk, I will cover a set of works that enable robots to 1) understand and learn from diverse human users, 2) learn interpretable, human-readable tree-based control policies directly via reinforcement learning, and 3) provide users with information online to improve situational awareness and facilitate effective human-robot collaboration. About the speaker: Dr. Rohan Paleja is an Assistant Professor in the Department of Computer Science at Purdue University. He directs the Strategies for Collaboration, Autonomy, Learning, and Exploration in Robotics Lab. The SCALE Robotics Lab focuses on advancing machine learning and artificial intelligence to improve robot learning, human-robot interaction, and multi-agent collaboration. Their goal is to equip autonomous agents with the ability to operate in the diverse, unstructured, and human-rich environments these agents will encounter in the real world.Dr. Paleja's research interests cover a broad range of topics, namely Explainable AI (xAI), Interactive Robot Learning, and Multi-Agent Collaboration. Prior to Purdue, Dr. Paleja was a Technical Staff Researcher in the Artificial Intelligence Technology group at MIT Lincoln Laboratory, where he collaborated with the Air Force Experimental Operations Unit and the Army Research Lab. Prior to that, he earned his Ph.D. in Robotics at the Georgia Institute of Technology in 2023.His work has received multiple awards, including a Best Paper Finalist Award at the Conference of Robot Learning (CoRL) and a Best Workshop Paper Award at the International Conference of Computer Vision (ICCV) Multi-Agent Relational Reasoning Workshop.
In this episode of the Crazy Wisdom podcast, host Stewart Alsop interviews Marcin Dymczyk, CPO and co-founder of SevenSense Robotics, exploring the fascinating world of advanced robotics and AI. Their conversation covers the evolution from traditional "standard" robotics with predetermined pathways to advanced robotics that incorporates perception, reasoning, and adaptability - essentially the AGI of physical robotics. Dymczyk explains how his company builds "the eyes and brains of mobile robots" using camera-based autonomy algorithms, drawing parallels between robot sensing systems and human vision, inner ear balance, and proprioception. The discussion ranges from the technical challenges of sensor fusion and world models to broader topics including robotics regulation across different countries, the role of federalism in innovation, and how recent geopolitical changes are driving localized high-tech development, particularly in defense applications. They also touch on the democratization of robotics for small businesses and the philosophical implications of increasingly sophisticated AI systems operating in physical environments. To learn more about SevenSense, visit www.sevensense.ai.Check out this GPT we trained on the conversationTimestamps00:00 Introduction to Robotics and Personal Journey05:27 The Evolution of Robotics: From Standard to Advanced09:56 The Future of Robotics: AI and Automation12:09 The Role of Edge Computing in Robotics17:40 FPGA and AI: The Future of Robotics Processing21:54 Sensing the World: How Robots Perceive Their Environment29:01 Learning from the Physical World: Insights from Robotics33:21 The Intersection of Robotics and Manufacturing35:01 Journey into Robotics: Education and Passion36:41 Practical Robotics Projects for Beginners39:06 Understanding Particle Filters in Robotics40:37 World Models: The Future of AI and Robotics41:51 The Black Box Dilemma in AI and Robotics44:27 Safety and Interpretability in Autonomous Systems49:16 Regulatory Challenges in Robotics and AI51:19 Global Perspectives on Robotics Regulation54:43 The Future of Robotics in Emerging Markets57:38 The Role of Engineers in Modern WarfareKey Insights1. Advanced robotics transcends traditional programming through perception and intelligence. Dymczyk distinguishes between standard robotics that follows rigid, predefined pathways and advanced robotics that incorporates perception and reasoning. This evolution enables robots to make autonomous decisions about navigation and task execution, similar to how humans adapt to unexpected situations rather than following predetermined scripts.2. Camera-based sensing systems mirror human biological navigation. SevenSense Robotics builds "eyes and brains" for mobile robots using multiple cameras (up to eight), IMUs (accelerometers/gyroscopes), and wheel encoders that parallel human vision, inner ear balance, and proprioception. This redundant sensing approach allows robots to navigate even when one system fails, such as operating in dark environments where visual sensors are compromised.3. Edge computing dominates industrial robotics due to connectivity and security constraints. Many industrial applications operate in environments with poor connectivity (like underground grocery stores) or require on-premise solutions for confidentiality. This necessitates powerful local processing capabilities rather than cloud-dependent AI, particularly in automotive factories where data security about new models is paramount.4. Safety regulations create mandatory "kill switches" that bypass AI decision-making. European and US regulatory bodies require deterministic safety systems that can instantly stop robots regardless of AI reasoning. These systems operate like human reflexes, providing immediate responses to obstacles while the main AI brain handles complex navigation and planning tasks.5. Modern robotics development benefits from increasingly affordable optical sensors. The democratization of 3D cameras, laser range finders, and miniature range measurement chips (costing just a few dollars from distributors like DigiKey) enables rapid prototyping and innovation that was previously limited to well-funded research institutions.6. Geopolitical shifts are driving localized high-tech development, particularly in defense applications. The changing role of US global leadership and lessons from Ukraine's drone warfare are motivating countries like Poland to develop indigenous robotics capabilities. Small engineering teams can now create battlefield-effective technology using consumer drones equipped with advanced sensors.7. The future of robotics lies in natural language programming for non-experts. Dymczyk envisions a transformation where small business owners can instruct robots using conversational language rather than complex programming, similar to how AI coding assistants now enable non-programmers to build applications through natural language prompts.
Executive Summary The Google DeepMind mechanistic interpretability team has made a strategic pivot over the past year, from ambitious reverse-engineering to a focus on pragmatic interpretability: Trying to directly solve problems on the critical path to AGI going well[[1]] Carefully choosing problems according to our comparative advantage Measuring progress with empirical feedback on proxy tasks We believe that, on the margin, more researchers who share our goals should take a pragmatic approach to interpretability, both in industry and academia, and we call on people to join us Our proposed scope is broad and includes much non-mech interp work, but we see this as the natural approach for mech interp researchers to have impact Specifically, we've found that the skills, tools and tastes of mech interp researchers transfer well to important and neglected problems outside “classic” mech interp See our companion piece for more on which research areas and theories of change we think are promising Why pivot now? We think that times have changed. Models are far more capable, bringing new questions within empirical reach We have been [...] ---Outline:(00:10) Executive Summary(03:00) Introduction(03:44) Motivating Example: Steering Against Evaluation Awareness(06:21) Our Core Process(08:20) Which Beliefs Are Load-Bearing?(10:25) Is This Really Mech Interp?(11:27) Our Comparative Advantage(14:57) Why Pivot?(15:20) Whats Changed In AI?(16:08) Reflections On The Fields Progress(18:18) Task Focused: The Importance Of Proxy Tasks(18:52) Case Study: Sparse Autoencoders(21:35) Ensure They Are Good Proxies(23:11) Proxy Tasks Can Be About Understanding(24:49) Types Of Projects: What Drives Research Decisions(25:18) Focused Projects(28:31) Exploratory Projects(28:35) Curiosity Is A Double-Edged Sword(30:56) Starting In A Robustly Useful Setting(34:45) Time-Boxing(36:27) Worked Examples(39:15) Blending The Two: Tentative Proxy Tasks(41:23) What's Your Contribution?(43:08) Jack Lindsey's Approach(45:44) Method Minimalism(46:12) Case Study: Shutdown Resistance(48:28) Try The Easy Methods First(50:02) When Should We Develop New Methods?(51:36) Call To Action(53:04) Acknowledgments(54:02) Appendix: Common Objections(54:08) Aren't You Optimizing For Quick Wins Over Breakthroughs?(56:34) What If AGI Is Fundamentally Different?(57:30) I Care About Scientific Beauty and Making AGI Go Well(58:09) Is This Just Applied Interpretability?(58:44) Are You Saying This Because You Need To Prove Yourself Useful To Google?(59:10) Does This Really Apply To People Outside AGI Companies?(59:40) Aren't You Just Giving Up?(01:00:04) Is Ambitious Reverse-engineering Actually Overcrowded?(01:00:48) Appendix: Defining Mechanistic Interpretability(01:01:44) Moving Toward Mechanistic OR Interpretability The original text contained 47 footnotes which were omitted from this narration. --- First published: December 1st, 2025 Source: https://www.lesswrong.com/posts/StENzDcD3kpfGJssR/a-pragmatic-vision-for-inter
The goal of ambitious mechanistic interpretability (AMI) is to fully understand how neural networks work. While some have pivoted towards more pragmatic approaches, I think the reports of AMI's death have been greatly exaggerated. The field of AMI has made plenty of progress towards finding increasingly simple and rigorously-faithful circuits, including our latest work on circuit sparsity. There are also many exciting inroads on the core problem waiting to be explored. The value of understanding Why try to understand things, if we can get more immediate value from less ambitious approaches? In my opinion, there are two main reasons. First, mechanistic understanding can make it much easier to figure out what's actually going on, especially when it's hard to distinguish hypotheses using external behavior (e.g if the model is scheming). We can liken this to going from print statement debugging to using an actual debugger. Print statement debugging often requires many experiments, because each time you gain only a few bits of information which sketch a strange, confusing, and potentially misleading picture. When you start using the debugger, you suddenly notice all at once that you're making a lot of incorrect assumptions you didn't even realize you were [...] ---Outline:(00:38) The value of understanding(02:32) AMI has good feedback loops(04:48) The past and future of AMI The original text contained 1 footnote which was omitted from this narration. --- First published: December 5th, 2025 Source: https://www.lesswrong.com/posts/Hy6PX43HGgmfiTaKu/an-ambitious-vision-for-interpretability --- Narrated by TYPE III AUDIO. ---Images from the article:Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.
Sign up for Alex's first live cohort, about Hierarchical Model building!Get 25% off "Building AI Applications for Data Scientists and Software Engineers"Proudly sponsored by PyMC Labs, the Bayesian Consultancy. Book a call, or get in touch!Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work!Visit our Patreon page to unlock exclusive Bayesian swag ;)Takeaways:Why GPs still matter: Gaussian Processes remain a go-to for function estimation, active learning, and experimental design – especially when calibrated uncertainty is non-negotiable.Scaling GP inference: Variational methods with inducing points (as in GPflow) make GPs practical on larger datasets without throwing away principled Bayes.MCMC in practice: Clever parameterizations and gradient-based samplers tighten mixing and efficiency; use MCMC when you need gold-standard posteriors.Bayesian deep learning, pragmatically: Stochastic-gradient training and approximate posteriors bring Bayesian ideas to neural networks at scale.Uncertainty that ships: Monte Carlo dropout and related tricks provide fast, usable uncertainty – even if they're approximations.Model complexity ≠ model quality: Understanding capacity, priors, and inductive bias is key to getting trustworthy predictions.Deep Gaussian Processes: Layered GPs offer flexibility for complex functions, with clear trade-offs in interpretability and compute.Generative models through a Bayesian lens: GANs and friends benefit from explicit priors and uncertainty – useful for safety and downstream decisions.Tooling that matters: Frameworks like GPflow lower the friction from idea to implementation, encouraging reproducible, well-tested modeling.Where we're headed: The future of ML is uncertainty-aware by default – integrating UQ tightly into optimization, design, and deployment.Chapters:08:44 Function Estimation and Bayesian Deep Learning10:41 Understanding Deep Gaussian Processes25:17 Choosing Between Deep GPs and Neural Networks32:01 Interpretability and Practical Tools for GPs43:52 Variational Methods in Gaussian Processes54:44 Deep Neural Networks and Bayesian Inference01:06:13 The Future of Bayesian Deep Learning01:12:28 Advice for Aspiring Researchers
Recorded live at Lightspeed's offices in San Francisco, this special episode of Generative Now dives into the urgency and promise of AI interpretability. Lightspeed partner Nnamdi Iregbulem spoke with Anthropic researcher Jack Lindsey and Goodfire co-founder and Chief Scientist Tom McGrath, who previously co-founded Google DeepMind's interpretability team. They discuss opening the black box of modern AI models in order to understand their reliability and spot real-world safety concerns, in order to build AI systems of the future that we can trust. Episode Chapters: 00:42 Welcome and Introduction00:36 Overview of Lightspeed and AI Investments03:19 Event Agenda and Guest Introductions05:35 Discussion on Interpretability in AI18:44 Technical Challenges in AI Interpretability29:42 Advancements in Model Interpretability30:05 Smarter Models and Interpretability31:26 Models Doing the Work for Us32:43 Real-World Applications of Interpretability34:32 Philanthropics' Approach to Interpretability39:15 Breakthrough Moments in AI Interpretability44:41 Challenges and Future Directions48:18 Neuroscience and Model Training Insights54:42 Emergent Misalignment and Model Behavior01:01:30 Concluding Thoughts and NetworkingStay in touch:www.lsvp.comX: https://twitter.com/lightspeedvpLinkedIn: https://www.linkedin.com/company/lightspeed-venture-partners/Instagram: https://www.instagram.com/lightspeedventurepartners/Subscribe on your favorite podcast app: generativenow.coEmail: generativenow@lsvp.comThe content here does not constitute tax, legal, business or investment advice or an offer to provide such advice, should not be construed as advocating the purchase or sale of any security or investment or a recommendation of any company, and is not an offer, or solicitation of an offer, for the purchase or sale of any security or investment product. For more details please see lsvp.com/legal.
We don't know how AIs think or why they do what they do. Or at least, we don't know much. That fact is only becoming more troubling as AIs grow more capable and appear on track to wield enormous cultural influence, directly advise on major government decisions, and even operate military equipment autonomously. We simply can't tell what models, if any, should be trusted with such authority.Neel Nanda of Google DeepMind is one of the founding figures of the field of machine learning trying to fix this situation — mechanistic interpretability (or “mech interp”). The project has generated enormous hype, exploding from a handful of researchers five years ago to hundreds today — all working to make sense of the jumble of tens of thousands of numbers that frontier AIs use to process information and decide what to say or do.Full transcript, video, and links to learn more: https://80k.info/nn1Neel now has a warning for us: the most ambitious vision of mech interp he once dreamed of is probably dead. He doesn't see a path to deeply and reliably understanding what AIs are thinking. The technical and practical barriers are simply too great to get us there in time, before competitive pressures push us to deploy human-level or superhuman AIs. Indeed, Neel argues no one approach will guarantee alignment, and our only choice is the “Swiss cheese” model of accident prevention, layering multiple safeguards on top of one another.But while mech interp won't be a silver bullet for AI safety, it has nevertheless had some major successes and will be one of the best tools in our arsenal.For instance: by inspecting the neural activations in the middle of an AI's thoughts, we can pick up many of the concepts the model is thinking about — from the Golden Gate Bridge, to refusing to answer a question, to the option of deceiving the user. While we can't know all the thoughts a model is having all the time, picking up 90% of the concepts it is using 90% of the time should help us muddle through, so long as mech interp is paired with other techniques to fill in the gaps.This episode was recorded on July 17 and 21, 2025.Interested in mech interp? Apply by September 12 to be a MATS scholar with Neel as your mentor! http://tinyurl.com/neel-mats-appWhat did you think? https://forms.gle/xKyUrGyYpYenp8N4AChapters:Cold open (00:00)Who's Neel Nanda? (01:02)How would mechanistic interpretability help with AGI (01:59)What's mech interp? (05:09)How Neel changed his take on mech interp (09:47)Top successes in interpretability (15:53)Probes can cheaply detect harmful intentions in AIs (20:06)In some ways we understand AIs better than human minds (26:49)Mech interp won't solve all our AI alignment problems (29:21)Why mech interp is the 'biology' of neural networks (38:07)Interpretability can't reliably find deceptive AI – nothing can (40:28)'Black box' interpretability — reading the chain of thought (49:39)'Self-preservation' isn't always what it seems (53:06)For how long can we trust the chain of thought (01:02:09)We could accidentally destroy chain of thought's usefulness (01:11:39)Models can tell when they're being tested and act differently (01:16:56)Top complaints about mech interp (01:23:50)Why everyone's excited about sparse autoencoders (SAEs) (01:37:52)Limitations of SAEs (01:47:16)SAEs performance on real-world tasks (01:54:49)Best arguments in favour of mech interp (02:08:10)Lessons from the hype around mech interp (02:12:03)Where mech interp will shine in coming years (02:17:50)Why focus on understanding over control (02:21:02)If AI models are conscious, will mech interp help us figure it out (02:24:09)Neel's new research philosophy (02:26:19)Who should join the mech interp field (02:38:31)Advice for getting started in mech interp (02:46:55)Keeping up to date with mech interp results (02:54:41)Who's hiring and where to work? (02:57:43)Host: Rob WiblinVideo editing: Simon Monsour, Luke Monsour, Dominic Armstrong, and Milo McGuireAudio engineering: Ben Cordell, Milo McGuire, Simon Monsour, and Dominic ArmstrongMusic: Ben CordellCamera operator: Jeremy ChevillotteCoordination, transcriptions, and web: Katy Moore
Today Lee Sharkey of Goodfire joins The Cognitive Revolution to discuss his research on parameter decomposition methods that break down neural networks into interpretable computational components, exploring how his team's "stochastic parameter decomposition" approach addresses the limitations of sparse autoencoders and offers new pathways for understanding, monitoring, and potentially steering AI systems at the mechanistic level. Check out our sponsors: Oracle Cloud Infrastructure, Shopify. Shownotes below brought to you by Notion AI Meeting Notes - try one month for free at https://notion.com/lp/nathan Parameter vs. Activation Decomposition: Traditional interpretability methods like Sparse Autoencoders (SAEs) focus on analyzing activations, while parameter decomposition focuses on understanding the parameters themselves - the actual "algorithm" of the neural network. No "True" Decomposition: None of the decompositions (whether sparse dictionary learning or parameter decomposition) are objectively "right" because they're all attempting to discretize a fundamentally continuous object, inevitably introducing approximations. Tradeoff in Interpretability: There's a balance between reconstruction loss and causal importance - as you decompose networks more, reconstruction loss may worsen, but interpretability might improve up to a certain point. Potential Unlearning Applications: Parameter decomposition may make unlearning more straightforward than with SAEs because researchers are already working in parameter space and can directly modify vectors that perform specific functions. Function Detection vs. Input Direction: A function like "deception" might manifest in many different input directions that SAEs struggle to identify as a single concept, while parameter decomposition might better isolate such functionality. Knowledge Extraction Goal: A key aim is to extract knowledge from models by understanding how they "think," especially for tasks where models demonstrate superhuman capabilities. Sponsors: Oracle Cloud Infrastructure: Oracle Cloud Infrastructure (OCI) is the next-generation cloud that delivers better performance, faster speeds, and significantly lower costs, including up to 50% less for compute, 70% for storage, and 80% for networking. Run any workload, from infrastructure to AI, in a high-availability environment and try OCI for free with zero commitment at https://oracle.com/cognitive Shopify: Shopify powers millions of businesses worldwide, handling 10% of U.S. e-commerce. With hundreds of templates, AI tools for product descriptions, and seamless marketing campaign creation, it's like having a design studio and marketing team in one. Start your $1/month trial today at https://shopify.com/cognitive
On this episode of Crazy Wisdom, host Stewart Alsop speaks with Michael Jagdeo, a headhunter and founder working with Exponent Labs and The Syndicate, about the cycles of money, power, and technology that shape our world. Their conversation touches on financial history through The Ascent of Money by Niall Ferguson and William Bagehot's The Money Market, the rise and fall of financial centers from London to New York and the new Texas Stock Exchange, the consolidation of industries and the theory of oligarchical collectivism, the role of AI as both tool and chaos agent, Bitcoin and “quantitative re-centralization,” the dynamics of exponential organizations, and the balance between collectivism and individualism. Jagdeo also shares recruiting philosophies rooted in stories like “stone soup,” frameworks like Yu-Kai Chou's Octalysis and the User Type Hexad, and book recommendations including Salim Ismail's Exponential Organizations and Arthur Koestler's The Act of Creation. Along the way they explore servant leadership, Price's Law, Linux and open source futures, religion as an operating system, and the cyclical nature of civilizations. You can learn more about Michael Jagdeo or reach out to him directly through Twitter or LinkedIn.Check out this GPT we trained on the conversationTimestamps00:05 Stewart Alsop introduces Michael Jagdeo, who shares his path from headhunting actuaries and IT talent into launching startups with Exponent Labs and The Syndicate.00:10 They connect recruiting to financial history, discussing actuaries, The Ascent of Money, and William Bagehot's The Money Market on the London money market and railways.00:15 The Rothschilds, institutional knowledge, and Corn Laws lead into questions about New York as a financial center and the quiet launch of the Texas Stock Exchange by Citadel and BlackRock.00:20 Capital power, George Soros vs. the Bank of England, chaos, paper clips, and Orwell's oligarchical collectivism frame industry consolidation, syndicates, and stone soup.00:25 They debate imperial conquest, bourgeoisie leisure, the decline of the middle class, AI as chaos agent, digital twins, Sarah Connor, Godzilla, and nuclear metaphors.00:30 Conversation turns to Bitcoin, “quantitative re-centralization,” Jack Bogle, index funds, Robinhood micro bailouts, and AI as both entropy and negative entropy.00:35 Jagdeo discusses Jim Keller, Tenstorrent, RISC-V, Nvidia CUDA, exponential organizations, Price's Law, bureaucracy, and servant leadership with the parable of stone soup.00:40 Recruiting as symbiosis, biophilia, trust, Judas, Wilhelm Reich, AI tools, Octalysis gamification, Jordan vs. triangle offense, and the role of laughter in persuasion emerge.00:45 They explore religion as operating systems, Greek gods, Comte's stages, Nietzsche, Jung, nostalgia, scientism, and Jordan Peterson's revival of tradition.00:50 The episode closes with Linux debates, Ubuntu, Framer laptops, PewDiePie, and Jagdeo's nod to Liminal Snake on epistemic centers and turning curses into blessings.Key InsightsOne of the central insights of the conversation is how financial history repeats through cycles of consolidation and power shifts. Michael Jagdeo draws on William Bagehot's The Money Market to explain how London became the hub of European finance, much like New York later did, and how the Texas Stock Exchange signals a possible southern resurgence of financial influence in America. The pattern of wealth moving with institutional shifts underscores how markets, capital, and politics remain intertwined.Jagdeo and Alsop emphasize that industries naturally oligarchize. Borrowing from Orwell's “oligarchical collectivism,” Jagdeo notes that whether in diamonds, food, or finance, consolidation emerges as economies of scale take over. This breeds syndicates and monopolies, often interpreted as conspiracies but really the predictable outcome of industrial maturation.Another powerful theme is the stone soup model of collaboration. Jagdeo applies this parable to recruiting, showing that no single individual can achieve large goals alone. By framing opportunities as shared ventures where each person adds their own ingredient, leaders can attract top talent while fostering genuine symbiosis.Technology, and particularly AI, is cast as both chaos agent and amplifier of human potential. The conversation likens AI to nuclear power—capable of great destruction or progress. From digital twins to Sarah Connor metaphors, they argue AI represents not just artificial intelligence but artificial knowledge and action, pushing humans to adapt quickly to its disruptive presence.The discussion of Bitcoin and digital currencies reframes decentralization as potentially another trap. Jagdeo provocatively calls Bitcoin “quantitative re-centralization,” suggesting that far from liberating individuals, digital currencies may accelerate neo-feudalism by creating new oligarchies and consolidating financial control in unexpected ways.Exponential organizations and the leverage of small teams emerge as another key point. Citing Price's Law, Jagdeo explains how fewer than a dozen highly capable individuals can now achieve billion-dollar valuations thanks to open source hardware, AI, and network effects. This trend redefines scale, making nimble collectives more powerful than bureaucratic giants.Finally, the episode highlights the cyclical nature of civilizations and belief systems. From Rome vs. Carthage to Greek gods shifting with societal needs, to Nietzsche's “God is dead” and Jung's view of recurring deaths of divinity, Jagdeo argues that religion, ideology, and operating systems reflect underlying incentives. Western nostalgia for past structures, whether political or religious, risks idolatry, while the real path forward may lie in new blends of individualism, collectivism, and adaptive tools like Linux and AI.
Eric Ho is building Goodfire to solve one of AI's most critical challenges: understanding what's actually happening inside neural networks. His team is developing techniques to understand, audit and edit neural networks at the feature level. Eric discusses breakthrough results in resolving superposition through sparse autoencoders, successful model editing demonstrations and real-world applications in genomics with Arc Institute's DNA foundation models. He argues that interpretability will be critical as AI systems become more powerful and take on mission-critical roles in society. Hosted by Sonya Huang and Roelof Botha, Sequoia Capital Mentioned in this episode: Mech interp: Mechanistic interpretability, list of important papers here Phineas Gage: 19th century railway engineer who lost most of his brain's left frontal lobe in an accident. Became a famous case study in neuroscience. Human Genome Project: Effort from 1990-2003 to generate the first sequence of the human genome which accelerated the study of human biology Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs Zoom In: An Introduction to Circuits: First important mechanistic interpretability paper from OpenAI in 2020 Superposition: Concept from physics applied to interpretability that allows neural networks to simulate larger networks (e.g. more concepts than neurons) Apollo Research: AI safety company that designs AI model evaluations and conducts interpretability research Towards Monosemanticity: Decomposing Language Models With Dictionary Learning. 2023 Anthropic paper that uses a sparse autoencoder to extract interpretable features; followed by Scaling Monosemanticity Under the Hood of a Reasoning Model: 2025 Goodfire paper that interprets DeepSeek's reasoning model R1 Auto-interpretability: The ability to use LLMs to automatically write explanations for the behavior of neurons in LLMs Interpreting Evo 2: Arc Institute's Next-Generation Genomic Foundation Model. (see episode with Arc co-founder Patrick Hsu) Paint with Ember: Canvas interface from Goodfire that lets you steer an LLM's visual output in real time (paper here) Model diffing: Interpreting how a model differs from checkpoint to checkpoint during finetuning Feature steering: The ability to change the style of LLM output by up or down weighting features (e.g. talking like a pirate vs factual information about the Andromeda Galaxy) Weight based interpretability: Method for directly decomposing neural network parameters into mechanistic components, instead of using features The Urgency of Interpretability: Essay by Anthropic founder Dario Amodei On the Biology of a Large Language Model: Goodfire collaboration with Anthropic
DFSA Report on Cyber, Quantum & AI Risk — A Regulatory Wake-Up Call for the Future of Finance
Her early inspiration while growing up in Goa with limited exposure to career options. Her Father's intellectual influence despite personal hardships and shift in focus to technology.Personal tragedy sparked a resolve to become financially independent and learn deeply.Inspirational quote that shaped her mindset: “Even if your dreams haven't come true, be grateful that so haven't your nightmares.”Her first role at a startup with Hands-on work with networking protocols (LDAP, VPN, DNS). Learning using only RFCs and O'Reilly books—no StackOverflow! Importance of building deep expertise for long-term success.Experiences with Troubleshooting and System Thinking; Transitioned from reactive fixes to logical, structured problem-solving. Her depth of understanding helped in debugging and system optimization.Career move to Yahoo where she led Service Engineering for mobile and ads across global data centers got early exposure to big data and machine learning through ad recommendation systems and built "performance and scale muscle" through working at massive scale.Challenges of Scale and Performance Then vs. Now: Problems remain the same, but data volumes and complexity have exploded. How modern tools (like AI/ML) can help identify relevance and anomalies in large data sets.Design with Scale in Mind - Importance of flipping the design approach: think scale-first, not POC-first. Encourage starting with a big-picture view, even when building a small prototype. Highlights multiple scaling dimensions—data, compute, network, security.Getting Into ML and Data Science with early spark from MOOCs, TensorFlow experiments, and statistics; Transition into data science role at Infoblox, a cybersecurity firm with focus areas on DNS security, anomaly detection, threat intelligence.Building real-world ML model applications like supervised models for threat detection and storage forecasting; developing graph models to analyze DNS traffic patterns for anomalies and key challenges of managing and processing massive volumes of security data.Data stack and what it takes to build data lakes that support ML with emphasis on understanding the end-to-end AI pipelineShifts from “under the hood” ML to front-and-center GenAI & Barriers: Data readiness, ROI, explainability, regulatory compliance.Explainability in AI and importance of interpreting model decisions, especially in regulated industries.How Explainability Works -Trade-offs between interpretable models (e.g., decision trees) and complex ones (e.g., deep learning); Techniques for local and global model understanding.Aruna's Book on Interpretability and Explainability in AI Using Python (by Aruna C).The world of GenAI & Transformers - Explainability in LLMs and GenAI: From attention weights to neuron activation.Challenges of scale: billions of parameters make models harder to interpret. Exciting research areas: Concept tracing, gradient analysis, neuron behavior.GenAI Agents in Action - Transition from task-specific GenAI to multi-step agents. Agents as orchestrators of business workflows using tools + reasoning.Real-world impact of agents and AI for everyday lifeAruna Chakkirala is a seasoned leader with expertise in AI, Data and Cloud. She is an AI Solutions Architect at Microsoft where she was instrumental in the early adoption of Generative AI. In prior roles as a Data Scientist she has built models in cybersecurity and holds a patent in community detection for DNS querying. Through her two-decade career, she has developed expertise in scale, security, and strategy at various organizations such as Infoblox, Yahoo, Nokia, EFI, and Verisign. Aruna has led highly successful teams and thrives on working with cutting-edge technologies. She is a frequent technical and keynote speaker, panelist, author and an active blogger. She contributes to community open groups and serves as a guest faculty member at premier academic institutes. Her book titled "Interpretability and Explainability in AI using Python" covers the taxonomy and techniques for model explanations in AI including the latest research in LLMs. She believes that the success of real-world AI applications increasingly depends on well- defined architectures across all encompassing domains. Her current interests include Generative AI, applications of LLMs and SLMs, Causality, Mechanistic Interpretability, and Explainability tools.Her recently published book linkInterpretability and Explainability in AI Using Python: Decrypt AI Decision-Making Using Interpretability and Explainability with Python to Build Reliable Machine Learning Systems https://amzn.in/d/00dSOwAOutside of work, she is an avid reader and enjoys creative writing. A passionate advocate for diversity and inclusion, she is actively involved in GHCI, LeanIn communities.
Emmanuel Amiesen is lead author of “Circuit Tracing: Revealing Computational Graphs in Language Models” (https://transformer-circuits.pub/2025/attribution-graphs/methods.html ), which is part of a duo of MechInterp papers that Anthropic published in March (alongside https://transformer-circuits.pub/2025/attribution-graphs/biology.html ).We recorded the initial conversation a month ago, but then held off publishing until the open source tooling for the graph generation discussed in this work was released last week: https://www.anthropic.com/research/open-source-circuit-tracingThis is a 2 part episode - an intro covering the open source release, then a deeper dive into the paper — with guest host Vibhu Sapra (https://x.com/vibhuuuus ) and Mochi the MechInterp Pomsky (https://x.com/mochipomsky ). Thanks to Vibhu for making this episode happen!While the original blogpost contained some fantastic guided visualizations (which we discuss at the end of this pod!), with the notebook and Neuronpedia visualization (https://www.neuronpedia.org/gemma-2-2b/graph ) released this week, you can now explore on your own with Neuronpedia, as we show you in the video version of this pod.Full Video EpisodeTimestamps00:00 Intro & Guest Introductions01:00 Anthropic's Circuit Tracing Release06:11 Exploring Circuit Tracing Tools & Demos13:01 Model Behaviors and User Experiments17:02 Behind the Research: Team and Community24:19 Main Episode Start: Mech Interp Backgrounds25:56 Getting Into Mech Interp Research31:52 History and Foundations of Mech Interp37:05 Core Concepts: Superposition & Features39:54 Applications & Interventions in Models45:59 Challenges & Open Questions in Interpretability57:15 Understanding Model Mechanisms: Circuits & Reasoning01:04:24 Model Planning, Reasoning, and Attribution Graphs01:30:52 Faithfulness, Deception, and Parallel Circuits01:40:16 Publishing Risks, Open Research, and Visualization01:49:33 Barriers, Vision, and Call to Action Get full access to Latent.Space at www.latent.space/subscribe
Emmanuel Amiesen is lead author of “Circuit Tracing: Revealing Computational Graphs in Language Models” (https://transformer-circuits.pub/2025/attribution-graphs/methods.html ), which is part of a duo of MechInterp papers that Anthropic published in March (alongside https://transformer-circuits.pub/2025/attribution-graphs/biology.html ). We recorded the initial conversation a month ago, but then held off publishing until the open source tooling for the graph generation discussed in this work was released last week: https://www.anthropic.com/research/open-source-circuit-tracing This is a 2 part episode - an intro covering the open source release, then a deeper dive into the paper — with guest host Vibhu Sapra (https://x.com/vibhuuuus ) and Mochi the MechInterp Pomsky (https://x.com/mochipomsky ). Thanks to Vibhu for making this episode happen! While the original blogpost contained some fantastic guided visualizations (which we discuss at the end of this pod!), with the notebook and Neuronpedia visualization (https://www.neuronpedia.org/gemma-2-2b/graph ) released this week, you can now explore on your own with Neuronpedia, as we show you in the video version of this pod. Chapters 00:00 Intro & Guest Introductions 01:00 Anthropic's Circuit Tracing Release 06:11 Exploring Circuit Tracing Tools & Demos 13:01 Model Behaviors and User Experiments 17:02 Behind the Research: Team and Community 24:19 Main Episode Start: Mech Interp Backgrounds 25:56 Getting Into Mech Interp Research 31:52 History and Foundations of Mech Interp 37:05 Core Concepts: Superposition & Features 39:54 Applications & Interventions in Models 45:59 Challenges & Open Questions in Interpretability 57:15 Understanding Model Mechanisms: Circuits & Reasoning 01:04:24 Model Planning, Reasoning, and Attribution Graphs 01:30:52 Faithfulness, Deception, and Parallel Circuits 01:40:16 Publishing Risks, Open Research, and Visualization 01:49:33 Barriers, Vision, and Call to Action
What's the next step forward in interpretability? In this episode, I chat with Lee Sharkey about his proposal for detecting computational mechanisms within neural networks: Attribution-based Parameter Decomposition, or APD for short. Patreon: https://www.patreon.com/axrpodcast Ko-fi: https://ko-fi.com/axrpodcast Transcript: https://axrp.net/episode/2025/06/03/episode-41-lee-sharkey-attribution-based-parameter-decomposition.html Topics we discuss, and timestamps: 0:00:41 APD basics 0:07:57 Faithfulness 0:11:10 Minimality 0:28:44 Simplicity 0:34:50 Concrete-ish examples of APD 0:52:00 Which parts of APD are canonical 0:58:10 Hyperparameter selection 1:06:40 APD in toy models of superposition 1:14:40 APD and compressed computation 1:25:43 Mechanisms vs representations 1:34:41 Future applications of APD? 1:44:19 How costly is APD? 1:49:14 More on minimality training 1:51:49 Follow-up work 2:05:24 APD on giant chain-of-thought models? 2:11:27 APD and "features" 2:14:11 Following Lee's work Lee links (Leenks): X/Twitter: https://twitter.com/leedsharkey Alignment Forum: https://www.alignmentforum.org/users/lee_sharkey Research we discuss: Interpretability in Parameter Space: Minimizing Mechanistic Description Length with Attribution-Based Parameter Decomposition: https://arxiv.org/abs/2501.14926 Toy Models of Superposition: https://transformer-circuits.pub/2022/toy_model/index.html Towards a unified and verified understanding of group-operation networks: https://arxiv.org/abs/2410.07476 Feature geometry is outside the superposition hypothesis: https://www.alignmentforum.org/posts/MFBTjb2qf3ziWmzz6/sae-feature-geometry-is-outside-the-superposition-hypothesis Episode art by Hamish Doodles: hamishdoodles.com
Auto encoders are neural networks that compress data into a smaller "code," enabling dimensionality reduction, data cleaning, and lossy compression by reconstructing original inputs from this code. Advanced auto encoder types, such as denoising, sparse, and variational auto encoders, extend these concepts for applications in generative modeling, interpretability, and synthetic data generation. Links Notes and resources at ocdevel.com/mlg/36 Try a walking desk - stay healthy & sharp while you learn & code Build the future of multi-agent software with AGNTCY. Thanks to T.J. Wilder from intrep.io for recording this episode! Fundamentals of Autoencoders Autoencoders are neural networks designed to reconstruct their input data by passing data through a compressed intermediate representation called a “code.” The architecture typically follows an hourglass shape: a wide input and output separated by a narrower bottleneck layer that enforces information compression. The encoder compresses input data into the code, while the decoder reconstructs the original input from this code. Comparison with Supervised Learning Unlike traditional supervised learning, where the output differs from the input (e.g., image classification), autoencoders use the same vector for both input and output. Use Cases: Dimensionality Reduction and Representation Autoencoders perform dimensionality reduction by learning compressed forms of high-dimensional data, making it easier to visualize and process data with many features. The compressed code can be used for clustering, visualization in 2D or 3D graphs, and input into subsequent machine learning models, saving computational resources and improving scalability. Feature Learning and Embeddings Autoencoders enable feature learning by extracting abstract representations from the input data, similar in concept to learned embeddings in large language models (LLMs). While effective for many data types, autoencoder-based encodings are less suited for variable-length text compared to LLM embeddings. Data Search, Clustering, and Compression By reducing dimensionality, autoencoders facilitate vector searches, efficient clustering, and similarity retrieval. The compressed codes enable lossy compression analogous to audio codecs like MP3, with the difference that autoencoders lack domain-specific optimizations for preserving perceptually important data. Reconstruction Fidelity and Loss Types Loss functions in autoencoders are defined to compare reconstructed outputs to original inputs, often using different loss types depending on input variable types (e.g., Boolean vs. continuous). Compression via autoencoders is typically lossy, meaning some information from the input is lost during reconstruction, and the areas of information lost may not be easily controlled. Outlier Detection and Noise Reduction Since reconstruction errors tend to move data toward the mean, autoencoders can be used to reduce noise and identify data outliers. Large reconstruction errors can signal atypical or outlier samples in the dataset. Denoising Autoencoders Denoising autoencoders are trained to reconstruct clean data from noisy inputs, making them valuable for applications in image and audio de-noising as well as signal smoothing. Iterative denoising as a principle forms the basis for diffusion models, where repeated application of a denoising autoencoder can gradually turn random noise into structured output. Data Imputation Autoencoders can aid in data imputation by filling in missing values: training on complete records and reconstructing missing entries for incomplete records using learned code representations. This approach leverages the model's propensity to output ‘plausible' values learned from overall data structure. Cryptographic Analogy The separation of encoding and decoding can draw parallels to encryption and decryption, though autoencoders are not intended or suitable for secure communication due to their inherent lossiness. Advanced Architectures: Sparse and Overcomplete Autoencoders Sparse autoencoders use constraints to encourage code representations with only a few active values, increasing interpretability and explainability. Overcomplete autoencoders have a code size larger than the input, often in applications that require extraction of distinct, interpretable features from complex model states. Interpretability and Research Example Research such as Anthropic's “Towards Monosemanticity” applies sparse autoencoders to the internal activations of language models to identify interpretable features correlated with concrete linguistic or semantic concepts. These models can be used to monitor and potentially control model behaviors (e.g., detecting specific language usage or enforcing safety constraints) by manipulating feature activations. Variational Autoencoders (VAEs) VAEs extend autoencoder architecture by encoding inputs as distributions (means and standard deviations) instead of point values, enforcing a continuous, normalized code space. Decoding from sampled points within this space enables synthetic data generation, as any point near the center of the code space corresponds to plausible data according to the model. VAEs for Synthetic Data and Rare Event Amplification VAEs are powerful in domains with sparse data or rare events (e.g., healthcare), allowing generation of synthetic samples representing underrepresented cases. They can increase model performance by augmenting datasets without requiring changes to existing model pipelines. Conditional Generative Techniques Conditional autoencoders extend VAEs by allowing controlled generation based on specified conditions (e.g., generating a house with a pool), through additional decoder inputs and conditional loss terms. Practical Considerations and Limitations Training autoencoders and their variants requires computational resources, and their stochastic training can produce differing code representations across runs. Lossy reconstruction, lack of domain-specific optimizations, and limited code interpretability restrict some use cases, particularly where exact data preservation or meaningful decompositions are required.
In this episode, Daniel Balsam and Tom McGrath, at Goodfire, discuss the future of mechanistic interpretability in AI models. They explore the fundamental inputs like models, compute, and algorithms, and emphasize the importance of a rich empirical approach to understanding how models work. Balsam and McGrath provide insights into ongoing projects and breakthroughs, particularly in scientific domains and creative applications, as they aim to push the frontiers of AI interpretability. They also discuss the company's recent funding and their goal to advance interpretability as a critical area in AI research. SPONSORS: Box Report: AI is delivering truly measurable productivity — strategic companies are already turning a 37% productivity edge. Discover how in Box's new 2025 State of AI in the Enterprise Report — read the full report here: https://bit.ly/43uVP52 Oracle Cloud Infrastructure (OCI): Oracle Cloud Infrastructure offers next-generation cloud solutions that cut costs and boost performance. With OCI, you can run AI projects and applications faster and more securely for less. New U.S. customers can save 50% on compute, 70% on storage, and 80% on networking by switching to OCI before May 31, 2024. See if you qualify at https://oracle.com/cognitive ElevenLabs: ElevenLabs gives your app a natural voice. Pick from 5,000+ voices in 31 languages, or clone your own, and launch lifelike agents for support, scheduling, learning, and games. Full server and client SDKs, dynamic tools, and monitoring keep you in control. Start free at https://elevenlabs.io/cognitive-revolution NetSuite: Over 41,000 businesses trust NetSuite by Oracle, the #1 cloud ERP, to future-proof their operations. With a unified platform for accounting, financial management, inventory, and HR, NetSuite provides real-time insights and forecasting to help you make quick, informed decisions. Whether you're earning millions or hundreds of millions, NetSuite empowers you to tackle challenges and seize opportunities. Download the free CFO's guide to AI and machine learning at https://netsuite.com/cognitive Shopify: Shopify powers millions of businesses worldwide, handling 10% of U.S. e-commerce. With hundreds of templates, AI tools for product descriptions, and seamless marketing campaign creation, it's like having a design studio and marketing team in one. Start your $1/month trial today at https://shopify.com/cognitive PRODUCED BY: https://aipodcast.ing SOCIAL LINKS: Website: https://www.cognitiverevolution.ai Twitter (Podcast): https://x.com/cogrev_podcast Twitter (Nathan): https://x.com/labenz LinkedIn: https://linkedin.com/in/nathanlabenz/ Youtube: https://youtube.com/@CognitiveRevolutionPodcast Apple: https://podcasts.apple.com/de/podcast/the-cognitive-revolution-ai-builders-researchers-and/id1669813431 Spotify: https://open.spotify.com/show/6yHyok3M3BjqzR0VB5MSyk
AI Unraveled: Latest AI News & Trends, Master GPT, Gemini, Generative AI, LLMs, Prompting, GPT Store
Key themes include technological competition and national self-reliance with Huawei and China challenging Nvidia and US dominance in AI chips, and major product updates and releases from companies like Baidu, OpenAI, and Grok introducing new AI models and features. The text also highlights innovative applications of AI, from Neuralink's brain implants restoring communication and Waymo considering selling robotaxis directly to consumers, to creative uses like generating action figures and integrating AI into religious practices. Finally, the sources touch on important considerations surrounding AI, such as the need for interpretability to ensure safety, the increasing sophistication of AI-powered scams, and discussions on the military implications and future potential of AGI.
Will OpenAI be fully open source by 2027? In episode 49 of Mixture of Experts, host Tim Hwang is joined by Aaron Baughman, Ash Minhas and Chris Hay to analyze Sam Altman's latest move towards open source. Next, we explore Anthropic's mechanistic interpretability results and the progress the AI research community is making. Then, can Apple catch up? We analyze the latest critiques on Apple Intelligence. Finally, Amazon enters the chat with AI agents. How does this elevate the competition? All that and more on today's Mixture of Experts.00:01 -- Introduction00:48 -- OpenAI goes open 11:36 -- Anthropic interpretability results 24:55 -- Daring Fireball on Apple Intelligence 34:22 -- Amazon's AI agentsThe opinions expressed in this podcast are solely those of the participants and do not necessarily reflect the views of IBM or any other organization or entity.Subscribe for AI updates: https://www.ibm.com/account/reg/us-en/signup?formid=news-urx-52120Learn more about artificial intelligence → https://www.ibm.com/think/artificial-intelligenceVisit Mixture of Experts podcast page to learn more AI content → https://www.ibm.com/think/podcasts/mixture-of-experts
How do we figure out whether interpretability is doing its job? One way is to see if it helps us prove things about models that we care about knowing. In this episode, I speak with Jason Gross about his agenda to benchmark interpretability in this way, and his exploration of the intersection of proofs and modern machine learning. Patreon: https://www.patreon.com/axrpodcast Ko-fi: https://ko-fi.com/axrpodcast Transcript: https://axrp.net/episode/2025/03/28/episode-40-jason-gross-compact-proofs-interpretability.html Topics we discuss, and timestamps: 0:00:40 - Why compact proofs 0:07:25 - Compact Proofs of Model Performance via Mechanistic Interpretability 0:14:19 - What compact proofs look like 0:32:43 - Structureless noise, and why proofs 0:48:23 - What we've learned about compact proofs in general 0:59:02 - Generalizing 'symmetry' 1:11:24 - Grading mechanistic interpretability 1:43:34 - What helps compact proofs 1:51:08 - The limits of compact proofs 2:07:33 - Guaranteed safe AI, and AI for guaranteed safety 2:27:44 - Jason and Rajashree's start-up 2:34:19 - Following Jason's work Links to Jason: Github: https://github.com/jasongross Website: https://jasongross.github.io Alignment Forum: https://www.alignmentforum.org/users/jason-gross Links to work we discuss: Compact Proofs of Model Performance via Mechanistic Interpretability: https://arxiv.org/abs/2406.11779 Unifying and Verifying Mechanistic Interpretability: A Case Study with Group Operations: https://arxiv.org/abs/2410.07476 Modular addition without black-boxes: Compressing explanations of MLPs that compute numerical integration: https://arxiv.org/abs/2412.03773 Stage-Wise Model Diffing: https://transformer-circuits.pub/2024/model-diffing/index.html Causal Scrubbing: a method for rigorously testing interpretability hypotheses: https://www.lesswrong.com/posts/JvZhhzycHu2Yd57RN/causal-scrubbing-a-method-for-rigorously-testing Interpretability in Parameter Space: Minimizing Mechanistic Description Length with Attribution-based Parameter Decomposition (aka the Apollo paper on APD): https://arxiv.org/abs/2501.14926 Towards Guaranteed Safe AI: https://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-45.pdf Episode art by Hamish Doodles: hamishdoodles.com
Clement Bonnet discusses his novel approach to the ARC (Abstraction and Reasoning Corpus) challenge. Unlike approaches that rely on fine-tuning LLMs or generating samples at inference time, Clement's method encodes input-output pairs into a latent space, optimizes this representation with a search algorithm, and decodes outputs for new inputs. This end-to-end architecture uses a VAE loss, including reconstruction and prior losses. SPONSOR MESSAGES:***CentML offers competitive pricing for GenAI model deployment, with flexible options to suit a wide range of models, from small to large-scale deployments. Check out their super fast DeepSeek R1 hosting!https://centml.ai/pricing/Tufa AI Labs is a brand new research lab in Zurich started by Benjamin Crouzier focussed on o-series style reasoning and AGI. They are hiring a Chief Engineer and ML engineers. Events in Zurich. Goto https://tufalabs.ai/***TRANSCRIPT + RESEARCH OVERVIEW:https://www.dropbox.com/scl/fi/j7m0gaz1126y594gswtma/CLEMMLST.pdf?rlkey=y5qvwq2er5nchbcibm07rcfpq&dl=0Clem and Matthew-https://www.linkedin.com/in/clement-bonnet16/https://github.com/clement-bonnethttps://mvmacfarlane.github.io/TOC1. LPN Fundamentals [00:00:00] 1.1 Introduction to ARC Benchmark and LPN Overview [00:05:05] 1.2 Neural Networks' Challenges with ARC and Program Synthesis [00:06:55] 1.3 Induction vs Transduction in Machine Learning2. LPN Architecture and Latent Space [00:11:50] 2.1 LPN Architecture and Latent Space Implementation [00:16:25] 2.2 LPN Latent Space Encoding and VAE Architecture [00:20:25] 2.3 Gradient-Based Search Training Strategy [00:23:39] 2.4 LPN Model Architecture and Implementation Details3. Implementation and Scaling [00:27:34] 3.1 Training Data Generation and re-ARC Framework [00:31:28] 3.2 Limitations of Latent Space and Multi-Thread Search [00:34:43] 3.3 Program Composition and Computational Graph Architecture4. Advanced Concepts and Future Directions [00:45:09] 4.1 AI Creativity and Program Synthesis Approaches [00:49:47] 4.2 Scaling and Interpretability in Latent Space ModelsREFS[00:00:05] ARC benchmark, Chollethttps://arxiv.org/abs/2412.04604[00:02:10] Latent Program Spaces, Bonnet, Macfarlanehttps://arxiv.org/abs/2411.08706[00:07:45] Kevin Ellis work on program generationhttps://www.cs.cornell.edu/~ellisk/[00:08:45] Induction vs transduction in abstract reasoning, Li et al.https://arxiv.org/abs/2411.02272[00:17:40] VAEs, Kingma, Wellinghttps://arxiv.org/abs/1312.6114[00:27:50] re-ARC, Hodelhttps://github.com/michaelhodel/re-arc[00:29:40] Grid size in ARC tasks, Chollethttps://github.com/fchollet/ARC-AGI[00:33:00] Critique of deep learning, Marcushttps://arxiv.org/vc/arxiv/papers/2002/2002.06177v1.pdf
Nathan discusses groundbreaking AI and biology research with Stanford Professor James Zou from the Chan Zuckerberg Initiative. In this episode of The Cognitive Revolution, we explore two remarkable papers: the virtual lab framework that created novel COVID treatments with minimal human oversight, and InterPLM's discovery of new protein motifs through mechanistic interpretability. Join us for an fascinating discussion about how AI is revolutionizing biological research and drug discovery. Got questions about AI? Submit them for our upcoming AMA episode + take our quick listener survey to help us serve you better - https://docs.google.com/forms/d/e/1FAIpQLSefHvs1-1g5xeqM7wSirQkzTtK-1fgW_OjyHPH9DvmbVAjEzA/viewform SPONSORS: SelectQuote: Finding the right life insurance shouldn't be another task you put off. SelectQuote compares top-rated policies to get you the best coverage at the right price. Even in our AI-driven world, protecting your family's future remains essential. Get your personalized quote at https://selectquote.com/cognitive Oracle Cloud Infrastructure (OCI): Oracle's next-generation cloud platform delivers blazing-fast AI and ML performance with 50% less for compute and 80% less for outbound networking compared to other cloud providers13. OCI powers industry leaders with secure infrastructure and application development capabilities. New U.S. customers can get their cloud bill cut in half by switching to OCI before December 31, 2024 at https://oracle.com/cognitive 80,000 Hours: 80,000 Hours is dedicated to helping you find a fulfilling career that makes a difference. With nearly a decade of research, they offer in-depth material on AI risks, AI policy, and AI safety research. Explore their articles, career reviews, and a podcast featuring experts like Anthropic CEO Dario. Everything is free, including their Career Guide. Visit https://80000hours.org/cognitiverevolution to start making a meaningful impact today. GiveWell : GiveWell has spent over 17 years researching global health and philanthropy to identify the highest-impact giving opportunities. Over 125,000 donors have contributed more than $2 billion, saving over 200,000 lives through evidence-backed recommendations. First-time donors can have their contributions matched up to $100 before year-end. Visit https://GiveWell.org select podcast, and enter Cognitive Revolution at checkout to make a difference today. CHAPTERS: CHAPTERS: (00:00:00) Teaser (00:00:35) About the Episode (00:04:30) Virtual Lab (00:08:09) AI Designs Nanobodies (00:14:43) Novel AI Pipeline (00:20:31) Human-AI Interaction (Part 1) (00:20:33) Sponsors: SelectQuote | Oracle Cloud Infrastructure (OCI) (00:23:22) Human-AI Interaction (Part 2) (00:32:31) Sponsors: 80,000 Hours | GiveWell (00:35:10) Project Cost & Time (00:41:04) Future of AI in Bio (00:45:46) InterPLM: Intro (00:50:30) AI Found New Concepts (00:55:02) Discovering New Motifs (00:57:14) Limitations & Future (01:01:32) Outro SOCIAL LINKS: Website: https://www.cognitiverevolution.ai Twitter (Podcast): https://x.com/cogrev_podcast Twitter (Nathan): https://x.com/labenz LinkedIn: https://www.linkedin.com/in/nathanlabenz/ Youtube: https://www.youtube.com/@CognitiveRevolutionPodcast
Current AI practice is not engineering, even when it aims for practical applications, because it is not based on scientific understanding. Enforcing engineering norms on the field could lead to considerably safer systems. https://betterwithout.ai/AI-as-engineering This episode has a lot of links! Here they are. Michael Nielsen's “The role of ‘explanation' in AI”. https://michaelnotebook.com/ongoing/sporadica.html#role_of_explanation_in_AI Subbarao Kambhampati's “Changing the Nature of AI Research”. https://dl.acm.org/doi/pdf/10.1145/3546954 Chris Olah and his collaborators: “Thread: Circuits”. distill.pub/2020/circuits/ “An Overview of Early Vision in InceptionV1”. distill.pub/2020/circuits/early-vision/ Dai et al., “Knowledge Neurons in Pretrained Transformers”. https://arxiv.org/pdf/2104.08696.pdf Meng et al.: “Locating and Editing Factual Associations in GPT.” rome.baulab.info “Mass-Editing Memory in a Transformer,” https://arxiv.org/pdf/2210.07229.pdf François Chollet on image generators putting the wrong number of legs on horses: twitter.com/fchollet/status/1573879858203340800 Neel Nanda's “Longlist of Theories of Impact for Interpretability”, https://www.lesswrong.com/posts/uK6sQCNMw8WKzJeCQ/a-longlist-of-theories-of-impact-for-interpretability Zachary C. Lipton's “The Mythos of Model Interpretability”. https://arxiv.org/abs/1606.03490 Meng et al., “Locating and Editing Factual Associations in GPT”. https://arxiv.org/pdf/2202.05262.pdf Belrose et al., “Eliciting Latent Predictions from Transformers with the Tuned Lens”. https://arxiv.org/abs/2303.08112 “Progress measures for grokking via mechanistic interpretability”. https://arxiv.org/abs/2301.05217 Conmy et al., “Towards Automated Circuit Discovery for Mechanistic Interpretability”. https://arxiv.org/abs/2304.14997 Elhage et al., “Softmax Linear Units,” transformer-circuits.pub/2022/solu/index.html Filan et al., “Clusterability in Neural Networks,” https://arxiv.org/pdf/2103.03386.pdf Cammarata et al., “Curve circuits,” distill.pub/2020/circuits/curve-circuits/ You can support the podcast and get episodes a week early, by supporting the Patreon: https://www.patreon.com/m/fluidityaudiobooks If you like the show, consider buying me a coffee: https://www.buymeacoffee.com/mattarnold Original music by Kevin MacLeod. This podcast is under a Creative Commons Attribution Non-Commercial International 4.0 License.
Neel Nanda, a senior research scientist at Google DeepMind, leads their mechanistic interpretability team. In this extensive interview, he discusses his work trying to understand how neural networks function internally. At just 25 years old, Nanda has quickly become a prominent voice in AI research after completing his pure mathematics degree at Cambridge in 2020. Nanda reckons that machine learning is unique because we create neural networks that can perform impressive tasks (like complex reasoning and software engineering) without understanding how they work internally. He compares this to having computer programs that can do things no human programmer knows how to write. His work focuses on "mechanistic interpretability" - attempting to uncover and understand the internal structures and algorithms that emerge within these networks. SPONSOR MESSAGES: *** CentML offers competitive pricing for GenAI model deployment, with flexible options to suit a wide range of models, from small to large-scale deployments. https://centml.ai/pricing/ Tufa AI Labs is a brand new research lab in Zurich started by Benjamin Crouzier focussed on ARC and AGI, they just acquired MindsAI - the current winners of the ARC challenge. Are you interested in working on ARC, or getting involved in their events? Goto https://tufalabs.ai/ *** SHOWNOTES, TRANSCRIPT, ALL REFERENCES (DONT MISS!): https://www.dropbox.com/scl/fi/36dvtfl3v3p56hbi30im7/NeelShow.pdf?rlkey=pq8t7lyv2z60knlifyy17jdtx&st=kiutudhc&dl=0 We riff on: * How neural networks develop meaningful internal representations beyond simple pattern matching * The effectiveness of chain-of-thought prompting and why it improves model performance * The importance of hands-on coding over extensive paper reading for new researchers * His journey from Cambridge to working with Chris Olah at Anthropic and eventually Google DeepMind * The role of mechanistic interpretability in AI safety NEEL NANDA: https://www.neelnanda.io/ https://scholar.google.com/citations?user=GLnX3MkAAAAJ&hl=en https://x.com/NeelNanda5 Interviewer - Tim Scarfe TOC: 1. Part 1: Introduction [00:00:00] 1.1 Introduction and Core Concepts Overview 2. Part 2: Outside Interview [00:06:45] 2.1 Mechanistic Interpretability Foundations 3. Part 3: Main Interview [00:32:52] 3.1 Mechanistic Interpretability 4. Neural Architecture and Circuits [01:00:31] 4.1 Biological Evolution Parallels [01:04:03] 4.2 Universal Circuit Patterns and Induction Heads [01:11:07] 4.3 Entity Detection and Knowledge Boundaries [01:14:26] 4.4 Mechanistic Interpretability and Activation Patching 5. Model Behavior Analysis [01:30:00] 5.1 Golden Gate Claude Experiment and Feature Amplification [01:33:27] 5.2 Model Personas and RLHF Behavior Modification [01:36:28] 5.3 Steering Vectors and Linear Representations [01:40:00] 5.4 Hallucinations and Model Uncertainty 6. Sparse Autoencoder Architecture [01:44:54] 6.1 Architecture and Mathematical Foundations [02:22:03] 6.2 Core Challenges and Solutions [02:32:04] 6.3 Advanced Activation Functions and Top-k Implementations [02:34:41] 6.4 Research Applications in Transformer Circuit Analysis 7. Feature Learning and Scaling [02:48:02] 7.1 Autoencoder Feature Learning and Width Parameters [03:02:46] 7.2 Scaling Laws and Training Stability [03:11:00] 7.3 Feature Identification and Bias Correction [03:19:52] 7.4 Training Dynamics Analysis Methods 8. Engineering Implementation [03:23:48] 8.1 Scale and Infrastructure Requirements [03:25:20] 8.2 Computational Requirements and Storage [03:35:22] 8.3 Chain-of-Thought Reasoning Implementation [03:37:15] 8.4 Latent Structure Inference in Language Models
Join Nathan for an expansive conversation with Dan Hendrycks, Executive Director of the Center for AI Safety and Advisor to Elon Musk's XAI. In this episode of The Cognitive Revolution, we explore Dan's groundbreaking work in AI safety and alignment, from his early contributions to activation functions to his recent projects on AI robustness and governance. Discover insights on representation engineering, circuit breakers, and tamper-resistant training, as well as Dan's perspectives on AI's impact on society and the future of intelligence. Don't miss this in-depth discussion with one of the most influential figures in AI research and safety. Check out some of Dan's research papers: MMLU: https://arxiv.org/abs/2009.03300 GELU: https://arxiv.org/abs/1606.08415 Machiavelli Benchmark: https://arxiv.org/abs/2304.03279 Circuit Breakers: https://arxiv.org/abs/2406.04313 Tamper Resistant Safeguards: https://arxiv.org/abs/2408.00761 Statement on AI Risk: https://www.safe.ai/work/statement-on-ai-risk Apply to join over 400 Founders and Execs in the Turpentine Network: https://www.turpentinenetwork.co/ SPONSORS: Shopify: Shopify is the world's leading e-commerce platform, offering a market-leading checkout system and exclusive AI apps like Quikly. Nobody does selling better than Shopify. Get a $1 per month trial at https://shopify.com/cognitive. LMNT: LMNT is a zero-sugar electrolyte drink mix that's redefining hydration and performance. Ideal for those who fast or anyone looking to optimize their electrolyte intake. Support the show and get a free sample pack with any purchase at https://drinklmnt.com/tcr. Notion: Notion offers powerful workflow and automation templates, perfect for streamlining processes and laying the groundwork for AI-driven automation. With Notion AI, you can search across thousands of documents from various platforms, generating highly relevant analysis and content tailored just for you - try it for free at https://notion.com/cognitiverevolution Oracle: Oracle Cloud Infrastructure (OCI) is a single platform for your infrastructure, database, application development, and AI needs. OCI has four to eight times the bandwidth of other clouds; offers one consistent price, and nobody does data better than Oracle. If you want to do more and spend less, take a free test drive of OCI at https://oracle.com/cognitive CHAPTERS: (00:00:00) Teaser (00:00:48) About the Show (00:02:17) About the Episode (00:05:41) Intro (00:07:19) GELU Activation Function (00:10:48) Signal Filtering (00:12:46) Scaling Maximalism (00:18:35) Sponsors: Shopify | LMNT (00:22:03) New Architectures (00:25:41) AI as Complex System (00:32:35) The Machiavelli Benchmark (00:34:10) Sponsors: Notion | Oracle (00:37:20) Understanding MMLU Scores (00:45:23) Reasoning in Language Models (00:49:18) Multimodal Reasoning (00:54:53) World Modeling and Sora (00:57:07) Arc Benchmark and Hypothesis (01:01:06) Humanity's Last Exam (01:08:46) Benchmarks and AI Ethics (01:13:28) Robustness and Jailbreaking (01:18:36) Representation Engineering (01:30:08) Convergence of Approaches (01:34:18) Circuit Breakers (01:37:52) Tamper Resistance (01:49:10) Interpretability vs. Robustness (01:53:53) Open Source and AI Safety (01:58:16) Computational Irreducibility (02:06:28) Neglected Approaches (02:12:47) Truth Maxing and XAI (02:19:59) AI-Powered Forecasting (02:24:53) Chip Bans and Geopolitics (02:33:30) Working at CAIS (02:35:03) Extinction Risk Statement (02:37:24) Outro
Percy Liang is a Stanford professor and co-founder of Together AI, driving some of the most critical advances in AI research. Percy is also a trained classical pianist, which clearly influences the way he thinks about technology. We explored the evolution of AI from simple token prediction to autonomous agents capable of long-term problem-solving, the problem of interpretability, and the future of AI safety in complex, real-world systems. [0:00] Intro[0:46] Discussing OpenAI's O1 Model [2:21] The Evolution of AI Agents [3:27] Challenges and Benchmarks in AI [4:38] Compatibility and Integration Issues [6:17] The Future of AI Scaffolding [10:05] Academia's Role in AI Research [15:17] AI Safety and Holistic Approaches [18:32] Regulation and Transparency in AI [21:42] Generative Agents and Social Simulations [29:14] The State of AI Evaluations [32:07] Exploring Evaluation in Language Models [35:13] The Challenge of Interpretability [39:31] Innovations in Model Architectures [43:18] The Future of Inference and Customization [46:46] Milestones in AI Research and Reasoning [49:43] Robotics and AI: The Road Ahead [52:24] AI in Music: A Harmonious Future [55:52] AI's Role in Education and Beyond [56:30] Quickfire[59:16] Jacob and Pat Debrief With your co-hosts: @jacobeffron - Partner at Redpoint, Former PM Flatiron Health @patrickachase - Partner at Redpoint, Former ML Engineer LinkedIn @ericabrescia - Former COO Github, Founder Bitnami (acq'd by VMWare) @jordan_segall - Partner at Redpoint
How do we figure out what large language models believe? In fact, do they even have beliefs? Do those beliefs have locations, and if so, can we edit those locations to change the beliefs? Also, how are we going to get AI to perform tasks so hard that we can't figure out if they succeeded at them? In this episode, I chat to Peter Hase about his research into these questions. Patreon: patreon.com/axrpodcast Ko-fi: ko-fi.com/axrpodcast The transcript: https://axrp.net/episode/2024/08/24/episode-35-peter-hase-llm-beliefs-easy-to-hard-generalization.html Topics we discuss, and timestamps: 0:00:36 - NLP and interpretability 0:10:20 - Interpretability lessons 0:32:22 - Belief interpretability 1:00:12 - Localizing and editing models' beliefs 1:19:18 - Beliefs beyond language models 1:27:21 - Easy-to-hard generalization 1:47:16 - What do easy-to-hard results tell us? 1:57:33 - Easy-to-hard vs weak-to-strong 2:03:50 - Different notions of hardness 2:13:01 - Easy-to-hard vs weak-to-strong, round 2 2:15:39 - Following Peter's work Peter on Twitter: https://x.com/peterbhase Peter's papers: Foundational Challenges in Assuring Alignment and Safety of Large Language Models: https://arxiv.org/abs/2404.09932 Do Language Models Have Beliefs? Methods for Detecting, Updating, and Visualizing Model Beliefs: https://arxiv.org/abs/2111.13654 Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Models: https://arxiv.org/abs/2301.04213 Are Language Models Rational? The Case of Coherence Norms and Belief Revision: https://arxiv.org/abs/2406.03442 The Unreasonable Effectiveness of Easy Training Data for Hard Tasks: https://arxiv.org/abs/2401.06751 Other links: Toy Models of Superposition: https://transformer-circuits.pub/2022/toy_model/index.html Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV): https://arxiv.org/abs/1711.11279 Locating and Editing Factual Associations in GPT (aka the ROME paper): https://arxiv.org/abs/2202.05262 Of nonlinearity and commutativity in BERT: https://arxiv.org/abs/2101.04547 Inference-Time Intervention: Eliciting Truthful Answers from a Language Model: https://arxiv.org/abs/2306.03341 Editing a classifier by rewriting its prediction rules: https://arxiv.org/abs/2112.01008 Discovering Latent Knowledge Without Supervision (aka the Collin Burns CCS paper): https://arxiv.org/abs/2212.03827 Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision: https://arxiv.org/abs/2312.09390 Concrete problems in AI safety: https://arxiv.org/abs/1606.06565 Rissanen Data Analysis: Examining Dataset Characteristics via Description Length: https://arxiv.org/abs/2103.03872 Episode art by Hamish Doodles: hamishdoodles.com
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Clarifying alignment vs capabilities, published by Richard Ngo on August 19, 2024 on The AI Alignment Forum. A core distinction in AGI safety is between alignment and capabilities. However, I think this distinction is a very fuzzy one, which has led to a lot of confusion. In this post I'll describe some of the problems with how people typically think about it, and offer a replacement set of definitions. "Alignment" and "capabilities" are primarily properties of AIs not of AI research The first thing to highlight is that the distinction between alignment and capabilities is primarily doing useful work when we think of them as properties of AIs. This distinction is still under-appreciated by the wider machine learning community. ML researchers have historically thought about performance of models almost entirely with respect to the tasks they were specifically trained on. However, the rise of LLMs has vindicated the alignment community's focus on general capabilities, and now it's much more common to assume that performance on many tasks (including out-of-distribution tasks) will improve roughly in parallel. This is a crucial assumption for thinking about risks from AGI. Insofar as the ML community has thought about alignment, it has mostly focused on aligning models' behavior to their training objectives. The possibility of neural networks aiming to achieve internally-represented goals is still not very widely understood, making it hard to discuss and study the reasons those goals might or might not be aligned with the values of (any given set of) humans. To be fair, the alignment community has caused some confusion by describing models as more or less "aligned", rather than more or less "aligned to X" for some specified X. I'll talk more about this confusion, and how we should address it, in a later post. But the core point is that AIs might develop internally-represented goals or values that we don't like, and we should try to avoid that. However, extending "alignment" and "capabilities" from properties of AIs to properties of different types of research is a fraught endeavor. It's tempting to categorize work as alignment research to the extent that it can be used to make AIs more aligned (to many possible targets), and as capabilities research to the extent that it can be used to make AIs more capable. But this approach runs into (at least) three major problems. Firstly, in general it's very difficult to categorize research by its impacts. Great research often links together ideas from many different subfields, typically in ways that only become apparent throughout the course of the research. We see this in many historical breakthroughs which shed light on a range of different domains. For example, early physicists studying the motions of the stars eventually derived laws governing all earthly objects. Meanwhile Darwin's study of barnacles and finches led him to principles governing the evolution of all life. Analogously, we should expect that big breakthroughs in our understanding of neural networks and deep learning would be useful in many different ways. More concretely, there are many cases where research done under the banner of alignment has advanced, or plausibly will advance, AI capabilities to a significant extent. This undermines our ability to categorize research by its impacts. Central examples include: RLHF makes language models more obedient, but also more capable of coherently carrying out tasks. Scalable oversight techniques can catch misbehavior, but will likely become important for generating high-quality synthetic training data, as it becomes more and more difficult for unassisted humans to label AI outputs correctly. Interpretability techniques will both allow us to inspect AI cognition and also extract more capable behavior from them (e.g. via ...
Nathan explores the cutting-edge field of mechanistic interpretability with Dan Balsam and Tom McGrath, co-founders of Goodfire. In this episode of The Cognitive Revolution, we delve into the science of understanding AI models' inner workings, recent breakthroughs, and the potential impact on AI safety and control. Join us for an insightful discussion on sparse autoencoders, polysemanticity, and the future of interpretable AI. Papers Very accessible article on types of representations: Local vs Distributed Coding Theoretical understanding of how models might pack concepts into their representations: Toy Models of Superposition How structure in the world gives rise to structure in the latent space: The Geometry of Categorical and Hierarchical Concepts in Large Language Models Using sparse autoencoders to pull apart language model representations: Sparse Autoencoders / Towards Monosemanticity / Scaling Monosemanticity Finding & teaching concepts in superhuman systems: Acquisition of Chess Knowledge in AlphaZero / Bridging the Human-AI Knowledge Gap: Concept Discovery and Transfer in AlphaZero Connecting microscopic learning to macroscopic phenomena: The Quantization Model of Neural Scaling Understanding at scale: Language models can explain neurons in language models Apply to join over 400 founders and execs in the Turpentine Network: https://hmplogxqz0y.typeform.com/to/JCkphVqj SPONSORS: Oracle Cloud Infrastructure (OCI) is a single platform for your infrastructure, database, application development, and AI needs. OCI has four to eight times the bandwidth of other clouds; offers one consistent price, and nobody does data better than Oracle. If you want to do more and spend less, take a free test drive of OCI at https://oracle.com/cognitive The Brave search API can be used to assemble a data set to train your AI models and help with retrieval augmentation at the time of inference. All while remaining affordable with developer first pricing, integrating the Brave search API into your workflow translates to more ethical data sourcing and more human representative data sets. Try the Brave search API for free for up to 2000 queries per month at https://bit.ly/BraveTCR Omneky is an omnichannel creative generation platform that lets you launch hundreds of thousands of ad iterations that actually work customized across all platforms, with a click of a button. Omneky combines generative AI and real-time advertising data. Mention "Cog Rev" for 10% off https://www.omneky.com/ Head to Squad to access global engineering without the headache and at a fraction of the cost: head to https://choosesquad.com/ and mention “Turpentine” to skip the waitlist. CHAPTERS: (00:00:00) About the Show (00:00:22) About the Episode (00:03:52) Introduction and Background (00:08:43) State of Interpretability Research (00:12:06) Key Insights in Interpretability (00:16:53) Polysemanticity and Model Compression (Part 1) (00:17:00) Sponsors: Oracle | Brave (00:19:04) Polysemanticity and Model Compression (Part 2) (00:22:50) Sparse Autoencoders Explained (00:27:19) Challenges in Interpretability Research (Part 1) (00:30:54) Sponsors: Omneky | Squad (00:32:41) Challenges in Interpretability Research (Part 2) (00:33:51) Goodfire's Vision and Mission (00:37:08) Interpretability and Scientific Models (00:43:48) Architecture and Interpretability Techniques (00:50:08) Quantization and Model Representation (00:54:07) Future of Interpretability Research (01:01:38) Skepticism and Challenges in Interpretability (01:07:51) Alternative Architectures and Universality (01:13:39) Goodfire's Business Model and Funding (01:18:47) Building the Team and Future Plans (01:31:03) Hiring and Getting Involved in Interpretability (01:51:28) Closing Remarks (01:51:38) Outro
In this week's episode, Katherine Forrest and Anna Gressel unravel interpretability, an evolving field of AI safety research, while exploring its potential significance for regulators as well as its relation to concepts like explainability and transparency. ## Learn More About Paul, Weiss's Artificial Intelligence Practice: https://www.paulweiss.com/practices/litigation/artificial-intelligence
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Open Source Automated Interpretability for Sparse Autoencoder Features, published by kh4dien on July 31, 2024 on LessWrong. Background Sparse autoencoders recover a diversity of interpretable, monosemantic features, but present an intractable problem of scale to human labelers. We investigate different techniques for generating and scoring text explanations of SAE features. Key Findings Open source models generate and evaluate text explanations of SAE features reasonably well, albeit somewhat worse than closed models like Claude 3.5 Sonnet. Explanations found by LLMs are similar to explanations found by humans. Automatically interpreting 1.5M features of GPT-2 with the current pipeline would cost $1300 in API calls to Llama 3.1 or $8500 with Claude 3.5 Sonnet. Prior methods cost ~$200k with Claude. Code can be found at https://github.com/EleutherAI/sae-auto-interp. We built a small dashboard to explore explanations and their scores: https://cadentj.github.io/demo/ Generating Explanations Sparse autoencoders decompose activations into a sum of sparse feature directions. We leverage language models to generate explanations for activating text examples. Prior work prompts language models with token sequences that activate MLP neurons (Bills et al. 2023), by showing the model a list of tokens followed by their respective activations, separated by a tab, and listed one per line. We instead highlight max activating tokens in each example with a set of . Optionally, we choose a threshold of the example's max activation for which tokens are highlighted. This helps the model distinguish important information for some densely activating features. We experiment with several methods for augmenting the explanation. Full prompts are available here. Chain of thought improves general reasoning capabilities in language models. We few-shot the model with several examples of a thought process that mimics a human approach to generating explanations. We expect that verbalizing thought might capture richer relations between tokens and context. Activations distinguish which sentences are more representative of a feature. We provide the magnitude of activating tokens after each example. We compute the logit weights for each feature through the path expansion where is the model unembed and is the decoder direction for a specific feature. The top promoted tokens capture a feature's causal effects which are useful for sharpening explanations. This method is equivalent to the logit lens (nostalgebraist 2020); future work might apply variants that reveal other causal information (Belrose et al. 2023; Gandelsman et al. 2024). Scoring explanations Text explanations represent interpretable "concepts" in natural language. How do we evaluate the faithfulness of explanations to the concepts actually contained in SAE features? We view the explanation as a classifier which predicts whether a feature is present in a context. An explanation should have high recall - identifying most activating text - as well as high precision - distinguishing between activating and non-activating text. Consider a feature which activates on the word "stop" after "don't" or "won't" (Gao et al. 2024). There are two failure modes: 1. The explanation could be too broad, identifying the feature as activating on the word "stop". It would have high recall on held out text, but low precision. 2. The explanation could be too narrow, stating the feature activates on the word "stop" only after "don't". This would have high precision, but low recall. One approach to scoring explanations is "simulation scoring"(Bills et al. 2023) which uses a language model to assign an activation to each token in a text, then measures the correlation between predicted and real activations. This method is biased toward recall; given a bro...
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: A List of 45+ Mech Interp Project Ideas from Apollo Research's Interpretability Team, published by Lee Sharkey on July 18, 2024 on The AI Alignment Forum. Why we made this list: The interpretability team at Apollo Research wrapped up a few projects recently[1]. In order to decide what we'd work on next, we generated a lot of different potential projects. Unfortunately, we are computationally bounded agents, so we can't work on every project idea that we were excited about! Previous lists of project ideas (such as Neel's collation of 200 Concrete Open Problems in Mechanistic Interpretability) have been very useful for people breaking into the field. But for all its merits, that list is now over a year and a half old. Therefore, many project ideas in that list aren't an up-to-date reflection of what some researchers consider the frontiers of mech interp. We therefore thought it would be helpful to share our list of project ideas! Comments and caveats: Some of these projects are more precisely scoped than others. Some are vague, others are more developed. Not every member of the team endorses every project as high priority. Usually more than one team member supports each one, and in many cases most of the team is supportive of someone working on it. We associate the person(s) who generated the project idea to each idea. We've grouped the project ideas into categories for convenience, but some projects span multiple categories. We don't put a huge amount of weight on this particular categorisation. We hope some people find this list helpful! We would love to see people working on these! If any sound interesting to you and you'd like to chat about it, don't hesitate to reach out. Foundational work on sparse dictionary learning for interpretability Transcoder-related project ideas See [2406.11944] Transcoders Find Interpretable LLM Feature Circuits) [Nix] Training and releasing high quality transcoders. Probably using top k GPT2 is a classic candidate for this. I'd be excited for people to try hard on even smaller models, e.g. GELU 4L [Nix] Good tooling for using transcoders Nice programming API to attribute an input to a collection of paths (see Dunefsky et al) Web user interface? Maybe in collaboration with neuronpedia. Would need a gpu server constantly running, but I'm optimistic you could do it with a ~a4000. [Nix] Further circuit analysis using transcoders. Take random input sequences, run transcoder attribution on them, examine the output and summarize the findings. High level summary statistics of how much attribution goes through error terms & how many pathways are needed would be valuable Explaining specific behaviors (IOI, greater-than) with high standards for specificity & faithfulness. Might be convoluted if accuracy [I could generate more ideas here, feel free to reach out nix@apolloresearch.ai] [Nix, Lee] Cross layer superposition Does it happen? Probably, but it would be nice to have specific examples! Look for features with similar decoder vectors, and do exploratory research to figure out what exactly is going on. What precisely does it mean? Answering this question seems likely to shed light on the question of 'What is a feature?'. [Lucius] Improving transcoder architectures Some MLPs or attention layers may implement a simple linear transformation in addition to actual computation. If we modify our transcoders to include a linear 'bypass' that is not counted in the sparsity penalty, do we improve performance since we are not unduly penalizing these linear transformations that would always be present and active? If we train multiple transcoders in different layers at the same time, can we include a sparsity penalty for their interactions with each other, encouraging a decomposition of the network that leaves us with as few interactions between features a...
Nathan explores the Guaranteed Safe AI Framework with co-authors Ben Goldhaber and Nora Ammann. In this episode of The Cognitive Revolution, we discuss their groundbreaking position paper on ensuring robust and reliable AI systems. Join us for an in-depth conversation about the three-part system governing AI behavior and its potential impact on the future of AI safety. Apply to join over 400 founders and execs in the Turpentine Network: https://hmplogxqz0y.typeform.com/to/JCkphVqj RECOMMENDED PODCAST: Complex Systems Patrick McKenzie (@patio11) talks to experts who understand the complicated but not unknowable systems we rely on. You might be surprised at how quickly Patrick and his guests can put you in the top 1% of understanding for stock trading, tech hiring, and more. Spotify: https://open.spotify.com/show/3Mos4VE3figVXleHDqfXOH Apple: https://podcasts.apple.com/us/podcast/complex-systems-with-patrick-mckenzie-patio11/id1753399812 SPONSORS: Oracle Cloud Infrastructure (OCI) is a single platform for your infrastructure, database, application development, and AI needs. OCI has four to eight times the bandwidth of other clouds; offers one consistent price, and nobody does data better than Oracle. If you want to do more and spend less, take a free test drive of OCI at https://oracle.com/cognitive The Brave search API can be used to assemble a data set to train your AI models and help with retrieval augmentation at the time of inference. All while remaining affordable with developer first pricing, integrating the Brave search API into your workflow translates to more ethical data sourcing and more human representative data sets. Try the Brave search API for free for up to 2000 queries per month at https://bit.ly/BraveTCR Omneky is an omnichannel creative generation platform that lets you launch hundreds of thousands of ad iterations that actually work customized across all platforms, with a click of a button. Omneky combines generative AI and real-time advertising data. Mention "Cog Rev" for 10% off https://www.omneky.com/ Head to Squad to access global engineering without the headache and at a fraction of the cost: head to https://choosesquad.com/ and mention “Turpentine” to skip the waitlist. CHAPTERS: (00:00:00) About the Show (00:04:39) Introduction (00:07:58) Convergence (00:10:32) Safety guarantees (00:14:35) World model (Part 1) (00:22:22) Sponsors: Oracle | Brave (00:24:31) World model (Part 2) (00:26:55) AI boxing (00:30:28) Verifier (00:33:33) Sponsors: Omneky | Squad (00:35:20) Example: Self-Driving Cars (00:38:08) Moral Desiderata (00:41:09) Trolley Problems (00:47:24) How to approach the world model (00:50:50) Deriving the world model (00:55:13) How far should the world model extend? (01:00:55) Safety through narrowness (01:02:38) Safety specs (01:08:26) Experiments (01:11:25) How GSAI can help in the short term (01:27:40) What would be the basis for the world model? (01:31:23) Interpretability (01:34:24) Competitive dynamics (01:37:35) Regulation (01:42:02) GSAI authors (01:43:25) Outro
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: What and Why: Developmental Interpretability of Reinforcement Learning, published by Garrett Baker on July 9, 2024 on LessWrong. Introduction I happen to be in that happy stage in the research cycle where I ask for money so I can continue to work on things I think are important. Part of that means justifying what I want to work on to the satisfaction of the people who provide that money. This presents a good opportunity to say what I plan to work on in a more layman-friendly way, for the benefit of LessWrong, potential collaborators, interested researchers, and funders who want to read the fun version of my project proposal It also provides the opportunity for people who are very pessimistic about the chances I end up doing anything useful by pursuing this to have their say. So if you read this (or skim it), and have critiques (or just recommendations), I'd love to hear them! Publicly or privately. So without further ado, in this post I will be discussing & justifying three aspects of what I'm working on, and my reasons for believing there are gaps in the literature in the intersection of these subjects that are relevant for AI alignment. These are: 1. Reinforcement learning 2. Developmental Interpretability 3. Values Culminating in: Developmental interpretability of values in reinforcement learning. Here are brief summaries of each of the sections: 1. Why study reinforcement learning? 1. Imposed-from-without or in-context reinforcement learning seems a likely path toward agentic AIs 2. The "data wall" means active-learning or self-training will get more important over time 3. There are fewer ways for the usual AI risk arguments to fail in the RL with mostly outcome-based rewards circumstance than the supervised learning + RL with mostly process-based rewards (RLHF) circumstance. 2. Why study developmental interpretability? 1. Causal understanding of the training process allows us to produce reward structure or environmental distribution interventions 2. Alternative & complementary tools to mechanistic interpretability 3. Connections with singular learning theory 3. Why study values? 1. The ultimate question of alignment is how can we make AI values compatible with human values, yet this is relatively understudied. 4. Where are the gaps? 1. Many experiments 2. Many theories 3. Few experiments testing theories or theories explaining experiments Reinforcement learning Agentic AIs vs Tool AIs All generally capable adaptive systems are ruled by a general, ground-truth, but slow outer optimization process which reduces incoherency and continuously selects for systems which achieve outcomes in the world. Examples include evolution, business, cultural selection, and to a great extent human brains. That is, except for LLMs. Most of the feedback LLMs receive is supervised, unaffected by the particular actions the LLM takes, and process-based (RLHF-like), where we reward the LLM according to how useful an action looks in contrast to a ground truth regarding how well that action (or sequence of actions) achieved its goal. Now I don't want to make the claim that this aspect of how we train LLMs is clearly a fault of them, or in some way limits the problem solving abilities they can have. And I do think it possible we see in-context ground-truth optimization processes instantiated as a result of increased scaling, in the same way we see in context learning. I do however want to make the claim that this current paradigm of mostly processed-based supervision, if it continues, and doesn't itself produce ground-truth based optimization, makes me optimistic about AI going well. That is, if this lack of general ground-truth optimization continues, we end up with a cached bundle of not very agentic (compared to AIXI) tool AIs with limited search or bootstrapping capabilities. Of course,...
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: An Extremely Opinionated Annotated List of My Favourite Mechanistic Interpretability Papers v2, published by Neel Nanda on July 7, 2024 on The AI Alignment Forum. This post represents my personal hot takes, not the opinions of my team or employer. This is a massively updated version of a similar list I made two years ago There's a lot of mechanistic interpretability papers, and more come out all the time. This can be pretty intimidating if you're new to the field! To try helping out, here's a reading list of my favourite mech interp papers: papers which I think are important to be aware of, often worth skimming, and something worth reading deeply (time permitting). I've annotated these with my key takeaways, what I like about each paper, which bits to deeply engage with vs skim, etc. I wrote a similar post 2 years ago, but a lot has changed since then, thus v2! Note that this is not trying to be a comprehensive literature review - this is my answer to "if you have limited time and want to get up to speed on the field as fast as you can, what should you do". I'm deliberately not following academic norms like necessarily citing the first paper introducing something, or all papers doing some work, and am massively biased towards recent work that is more relevant to the cutting edge. I also shamelessly recommend a bunch of my own work here, sorry! How to read this post: I've bolded the most important papers to read, which I recommend prioritising. All of the papers are annotated with my interpretation and key takeaways, and tbh I think reading that may be comparable good to skimming the paper. And there's far too many papers to read all of them deeply unless you want to make that a significant priority. I recommend reading all my summaries, noting the papers and areas that excite you, and then trying to dive deeply into those. Foundational Work A Mathematical Framework for Transformer Circuits (Nelson Elhage et al, Anthropic) - absolute classic, foundational ideas for how to think about transformers (see my blog post for what to skip). See my youtube tutorial (I hear this is best watched after reading the paper, and adds additional clarity) Deeply engage with: All the ideas in the overview section, especially: Understanding the residual stream and why it's fundamental. The notion of interpreting paths between interpretable bits (eg input tokens and output logits) where the path is a composition of matrices and how this is different from interpreting every intermediate activations And understanding attention heads: what a QK and OV matrix is, how attention heads are independent and additive and how attention and OV are semi-independent. Skip Trigrams & Skip Trigram bugs, esp understanding why these are a really easy thing to do with attention, and how the bugs are inherent to attention heads separating where to attend to (QK) and what to do once you attend somewhere (OV) Induction heads, esp why this is K-Composition (and how that's different from Q & V composition), how the circuit works mechanistically, and why this is too hard to do in a 1L model Skim or skip: Eigenvalues or tensor products. They have the worst effort per unit insight of the paper and aren't very important. Superposition Superposition is a core principle/problem in model internals. For any given activation (eg the output of MLP13), we believe that there's a massive dictionary of concepts/features the model knows of. Each feature has a corresponding vector, and model activations are a sparse linear combination of these meaningful feature vectors. Further, there are more features in the dictionary than activation dimensions, and they are thus compressed in and interfere with each other, essentially causing cascading errors. This phenomena of compression is called superposition. Toy models of superpositio...
Episode 129I spoke with Kristin Lauter about:* Elliptic curve cryptography and homomorphic encryption* Standardizing cryptographic protocols* Machine Learning on encrypted data* Attacking post-quantum cryptography with AIEnjoy—and let me know what you think!Kristin is Senior Director of FAIR Labs North America (2022—present), based in Seattle. Her current research areas are AI4Crypto and Private AI. She joined FAIR (Facebook AI Research) in 2021, after 22 years at Microsoft Research (MSR). At MSR she was Partner Research Manager on the senior leadership team of MSR Redmond. Before joining Microsoft in 1999, she was Hildebrandt Assistant Professor of Mathematics at the University of Michigan (1996-1999). She is an Affiliate Professor of Mathematics at the University of Washington (2008—present). She received all her advanced degrees from the University of Chicago, BA (1990), MS (1991), PhD (1996) in Mathematics. She is best known for her work on Elliptic Curve Cryptography, Supersingular Isogeny Graphs in Cryptography, Homomorphic Encryption (SEALcrypto.org), Private AI, and AI4Crypto. She served as President of the Association for Women in Mathematics from 2015-2017 and on the Council of the American Mathematical Society from 2014-2017.Find me on Twitter for updates on new episodes, and reach me at editor@thegradient.pub for feedback, ideas, guest suggestions. I spend a lot of time on this podcast—if you like my work, you can support me on Patreon :) You can also support upkeep for the full Gradient team/project through a paid subscription on Substack!Subscribe to The Gradient Podcast: Apple Podcasts | Spotify | Pocket Casts | RSSFollow The Gradient on TwitterOutline:* (00:00) Intro* (01:10) Llama 3 and encrypted data — where do we want to be?* (04:20) Tradeoffs: individual privacy vs. aggregated value in e.g. social media forums* (07:48) Kristin's shift in views on privacy* (09:40) Earlier work on elliptic curve cryptography — applications and theory* (10:50) Inspirations from algebra, number theory, and algebraic geometry* (15:40) On algebra vs. analysis and on clear thinking* (18:38) Elliptic curve cryptography and security, algorithms and concrete running time* (21:31) Cryptographic protocols and setting standards* (26:36) Supersingular isogeny graphs (and higher-dimensional supersingular isogeny graphs)* (32:26) Hard problems for cryptography and finding new problems* (36:42) Guaranteeing security for cryptographic protocols and mathematical foundations* (40:15) Private AI: Crypto-Nets / running neural nets on homomorphically encrypted data* (42:10) Polynomial approximations, activation functions, and expressivity* (44:32) Scaling up, Llama 2 inference on encrypted data* (46:10) Transitioning between MSR and FAIR, industry research* (52:45) An efficient algorithm for integer lattice reduction (AI4Crypto)* (56:23) Local minima, convergence and limit guarantees, scaling* (58:27) SALSA: Attacking Lattice Cryptography with Transformers* (58:38) Learning With Errors (LWE) vs. standard ML assumptions* (1:02:25) Powers of small primes and faster learning* (1:04:35) LWE and linear regression on a torus* (1:07:30) Secret recovery algorithms and transformer accuracy* (1:09:10) Interpretability / encoding information about secrets* (1:09:45) Future work / scaling up* (1:12:08) Reflections on working as a mathematician among technologistsLinks:* Kristin's Meta, Wikipedia, Google Scholar, and Twitter pages* Papers and sources mentioned/referenced:* The Advantages of Elliptic Curve Cryptography for Wireless Security (2004)* Cryptographic Hash Functions from Expander Graphs (2007, introducing Supersingular Isogeny Graphs)* Families of Ramanujan Graphs and Quaternion Algebras (2008 — the higher-dimensional analogues of Supersingular Isogeny Graphs)* Cryptographic Cloud Storage (2010)* Can homomorphic encryption be practical? (2011)* ML Confidential: Machine Learning on Encrypted Data (2012)* CryptoNets: Applying neural networks to encrypted data with high throughput and accuracy (2016)* A community effort to protect genomic data sharing, collaboration and outsourcing (2017)* The Homomorphic Encryption Standard (2022)* Private AI: Machine Learning on Encrypted Data (2022)* SALSA: Attacking Lattice Cryptography with Transformers (2022)* SalsaPicante: A Machine Learning Attack on LWE with Binary Secrets* SALSA VERDE: a machine learning attack on LWE with sparse small secrets* Salsa Fresca: Angular Embeddings and Pre-Training for ML Attacks on Learning With Errors* The cool and the cruel: separating hard parts of LWE secrets* An efficient algorithm for integer lattice reduction (2023) Get full access to The Gradient at thegradientpub.substack.com/subscribe
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Compact Proofs of Model Performance via Mechanistic Interpretability, published by Lawrence Chan on June 24, 2024 on The AI Alignment Forum. We recently released a paper on using mechanistic interpretability to generate compact formal guarantees on model performance. In this companion blog post to our paper, we'll summarize the paper and flesh out some of the motivation and inspiration behind our work. Paper abstract In this work, we propose using mechanistic interpretability - techniques for reverse engineering model weights into human-interpretable algorithms - to derive and compactly prove formal guarantees on model performance. We prototype this approach by formally proving lower bounds on the accuracy of 151 small transformers trained on a Max-of-K task. We create 102 different computer-assisted proof strategies and assess their length and tightness of bound on each of our models. Using quantitative metrics, we find that shorter proofs seem to require and provide more mechanistic understanding. Moreover, we find that more faithful mechanistic understanding leads to tighter performance bounds. We confirm these connections by qualitatively examining a subset of our proofs. Finally, we identify compounding structureless noise as a key challenge for using mechanistic interpretability to generate compact proofs on model performance. Introduction One hope for interpretability is that as we get AGI, we'll be able to use increasingly capable automation to accelerate the pace at which we can interpret ever more powerful models. These automatically generated interpretations need to satisfy two criteria: 1. Compression: Explanations compress the particular behavior of interest. Not just so that it fits in our heads, but also so that it generalizes well and is feasible to find and check. 2. Correspondence (or faithfulness): Explanations must accurately reflect the actual model mechanisms we aim to explain, allowing us to confidently constrain our models for guarantees or other practical applications. Progress happens best when there are clear and unambiguous targets and quantitative metrics. For correspondence, the field has developed increasingly targeted metrics for measuring performance: ablations, patching, and causal scrubbing. In our paper, we use mathematical proof to ensure correspondence, and present proof length as the first quantitative measure of explanation compression that is theoretically grounded, less subject to human judgement, and avoids trivial Goodharting. We see our core contributions in the paper as: 1. We push informal mechanistic interpretability arguments all the way to proofs of generalization bounds on toy transformers trained on the Max-of-$K$ task. This is a first step in getting formal guarantees about global properties of specific models, which is the approach of post-hoc mechanistic interpretability. 2. We introduce compactness of proof as a metric on explanation compression. We find that compactifying proofs requires deeper understanding of model behavior, and more compact proofs of the same bound necessarily encode more understanding of the model. 3. It is a common intuition that "proofs are hard for neural networks", and we flesh this intuition out as the problem of efficiently reasoning about structureless noise, which is an artifact of explanations being lossy approximations of the model's learned weights. While we believe that the proofs themselves (and in particular our proof which achieves a length that is linear in the number of model parameters for the parts of the model we understand adequately) may be of particular interest to those interested in guarantees, we believe that the insights about explanation compression from this methodology and our results are applicable more broadly to the field of mechanistic interpretability. Cor...
Dive into an accessible discussion on AI safety and philosophy, technical AI safety progress, and why catastrophic outcomes aren't inevitable. This conversation provides practical advice for AI newcomers and hope for a positive future. Consistently Candid Podcast : https://open.spotify.com/show/1EX89qABpb4pGYP1JLZ3BB SPONSORS: Oracle Cloud Infrastructure (OCI) is a single platform for your infrastructure, database, application development, and AI needs. OCI has four to eight times the bandwidth of other clouds; offers one consistent price, and nobody does data better than Oracle. If you want to do more and spend less, take a free test drive of OCI at https://oracle.com/cognitive The Brave search API can be used to assemble a data set to train your AI models and help with retrieval augmentation at the time of inference. All while remaining affordable with developer first pricing, integrating the Brave search API into your workflow translates to more ethical data sourcing and more human representative data sets. Try the Brave search API for free for up to 2000 queries per month at https://bit.ly/BraveTCR Omneky is an omnichannel creative generation platform that lets you launch hundreds of thousands of ad iterations that actually work customized across all platforms, with a click of a button. Omneky combines generative AI and real-time advertising data. Mention "Cog Rev" for 10% off https://www.omneky.com/ Head to Squad to access global engineering without the headache and at a fraction of the cost: head to https://choosesquad.com/ and mention “Turpentine” to skip the waitlist. Recommended Podcast: Byrne Hobart, the writer of The Diff, is revered in Silicon Valley. You can get an hour with him each week. See for yourself how his thinking can upgrade yours. Spotify: https://open.spotify.com/show/6rANlV54GCARLgMOtpkzKt Apple: https://podcasts.apple.com/us/podcast/the-riff-with-byrne-hobart-and-erik-torenberg/id1716646486 CHAPTERS: (00:00:00) About the Show (00:03:50) Intro (00:08:13) AI Scouting (00:14:42) Why arent people adopting AI more quickly? (00:18:25) Why dont people take advantage of AI? (00:22:35) Sponsors: Oracle | Brave (00:24:42) How to get a better understanding of AI (00:31:16) How to handle the public discourse around AI (00:34:02) Scaling and research (00:43:18) Sponsors: Omneky | Squad (00:45:03) The pause (00:47:29) Algorithmic efficiency (00:52:52) Red Teaming in Public (00:55:41) Deepfakes (01:01:02) AI safety (01:04:00) AI moderation (01:07:03) Why not a doomer (01:09:10) AI understanding human values (01:15:00) Interpretability research (01:18:30) AI safety leadership (01:21:55) AI safety respectability politics (01:33:42) China (01:37:22) Radical uncertainty (01:39:53) P(doom) (01:42:30) Where to find the guest (01:44:48) Outro
It's been an exciting couple weeks for GenAI! Join us as we discuss the latest research from OpenAI and Anthropic. We're excited to chat about this significant step forward in understanding how LLMs work and the implications it has for deeper understanding of the neural activity of language models. We take a closer look at some recent research from both OpenAI and Anthropic. These two recent papers both focus on the sparse autoencoder--an unsupervised approach for extracting interpretable features from an LLM. In "Extracting Concepts from GPT-4," OpenAI researchers propose using k-sparse autoencoders to directly control sparsity, simplifying tuning and improving the reconstruction-sparsity frontier. In "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet," researchers at Anthropic show that scaling laws can be used to guide the training of sparse autoencoders, among other findings. To learn more about ML observability, join the Arize AI Slack community or get the latest on our LinkedIn and Twitter.
This week, a group of current and former employees from Open AI and Google Deepmind penned an open letter accusing the industry's leading companies of prioritizing profits over safety. This comes after a spate of high profile departures from OpenAI, including co-founder Ilya Sutskever and senior researcher Jan Leike, as well as reports that OpenAI has gone to great lengths to silence would-be whistleblowers. The writers of the open letter argue that researchers have a “right to warn” the public about AI risks and laid out a series of principles that would protect that right. In this episode, we sit down with one of those writers: William Saunders, who left his job as a research engineer at OpenAI in February. William is now breaking the silence on what he saw at OpenAI that compelled him to leave the company and to put his name to this letter. RECOMMENDED MEDIA The Right to Warn Open Letter My Perspective On "A Right to Warn about Advanced Artificial Intelligence": A follow-up from William about the letter Leaked OpenAI documents reveal aggressive tactics toward former employees: An investigation by Vox into OpenAI's policy of non-disparagement.RECOMMENDED YUA EPISODESA First Step Toward AI Regulation with Tom Wheeler Spotlight on AI: What Would It Take For This to Go Well? Big Food, Big Tech and Big AI with Michael Moss Can We Govern AI? with Marietje SchaakeYour Undivided Attention is produced by the Center for Humane Technology. Follow us on Twitter: @HumaneTech_
Kevin Werbach speaks with Scott Zoldi of FICO, which pioneered consumer credit scoring in the 1950s and now offers a suite of analytics and fraud detection tools. Zoldi explains the importance of transparency and interpretability in AI models, emphasizing a “simpler is better” approach to creating clear and understandable algorithms. He discusses FICO's approach to responsible AI, which includes establishing model governance standards, and enforcing these standards through the use of blockchain technology. Zoldi explains how blockchain provides an immutable record of the model development process, enhancing accountability and trust. He also highlights the challenges organizations face in implementing responsible AI practices, particularly in light of upcoming AI regulations, and stresses the need for organizations to catch up in defining governance standards to ensure trustworthy and accountable AI models. Dr. Scott Zoldi is Chief Analytics Officer of FICO, responsible for analytics and AI innovation across FICO's portfolio. He has authored more than 130 patents, and is a long-time advocate and inventor in the space of responsible AI. He was nomianed for American Banker's 2024 Innovator Award and received Corinium's Future Thinking Award in 2022. Zoldi is a member of the Board of Advisors for FinReg Lab, and serves on the Boards of Directors of Software San Diego and San Diego Cyber Center of Excellence. He received his Ph.D. in theoretical and computational physics from Duke University. Navigating the Wild AI with Dr. Scott Zoldi How to Use Blockchain to Build Responsible AI The State of Responsible AI in Financial Services
This week, Google found itself in more turmoil, this time over its new AI Overviews feature and a trove of leaked internal documents. Then Josh Batson, a researcher at the A.I. startup Anthropic, joins us to explain how an experiment that made the chatbot Claude obsessed with the Golden Gate Bridge represents a major breakthrough in understanding how large language models work. And finally, we take a look at recent developments in A.I. safety, after Casey's early access to OpenAI's new souped-up voice assistant was taken away for safety reasons.Guests:Josh Batson, research scientist at AnthropicAdditional Reading: Google's A.I. Search Errors Cause a Furor OnlineGoogle Confirms the Leaked Search Documents are RealMapping the Mind of a Large Language ModelA.I. Firms Musn't Govern Themselves, Say Ex-Members of OpenAI's BoardWe want to hear from you. Email us at hardfork@nytimes.com. Find “Hard Fork” on YouTube and TikTok.
In this cross-over episode, Sam Altman sat down with Logan on the day of the ChatGPT-4o announcement to share behind-the-scenes details of the launch and offer his predictions for the future of AI. Altman delves into OpenAI's vision, discusses the timeline for achieving AGI, and explores the societal impact of humanoid robots. He also expresses his excitement and concerns about AI personal assistants, highlights the biggest opportunities and risks in the AI landscape today, and much more.(0:00) Intro(00:41) The Personal Impact of Leading OpenAI(01:35) Unveiling Multimodal AI: A Leap in Technology(02:38) The Surprising Use Cases and Benefits of Multimodal AI(03:14) Behind the Scenes: Making Multimodal AI Possible(08:27) Envisioning the Future of AI in Communication and Creativity(10:12) The Business of AI: Monetization, Open Source, and Future Directions(16:33) AI's Role in Shaping Future Jobs and Experiences(20:20) Debunking AGI: A Continuous Journey Towards Advanced AI(23:55) Exploring the Pace of Scientific and Technological Progress(24:09) The Importance of Interpretability in AI(25:02) Navigating AI Ethics and Regulation(27:17) The Safety Paradigm in AI and Beyond(28:46) Personal Reflections and the Impact of AI on Society(29:02) The Future of AI: Fast Takeoff Scenarios and Societal Changes(30:50) Navigating Personal and Professional Challenges(40:12) The Role of AI in Creative and Personal Identity(43:00) Educational System Adaptations for the AI Era(44:21) Contemplating the Future with Advanced AI(45:21) Jacob and Pat DebriefWith your co-hosts:@jacobeffron - Partner at Redpoint, Former PM Flatiron Health @patrickachase - Partner at Redpoint, Former ML Engineer LinkedIn @ericabrescia - Former COO Github, Founder Bitnami (acq'd by VMWare) @jordan_segall - Partner at Redpoint
On the day of the ChatGPT-4o announcement, Sam Altman sat down to share behind-the-scenes details of the launch and offer his predictions for the future of AI. Altman delves into OpenAI's vision, discusses the timeline for achieving AGI, and explores the societal impact of humanoid robots. He also expresses his excitement and concerns about AI personal assistants, highlights the biggest opportunities and risks in the AI landscape today, and much more. (00:00) Intro(00:50) The Personal Impact of Leading OpenAI(01:44) Unveiling Multimodal AI: A Leap in Technology(02:47) The Surprising Use Cases and Benefits of Multimodal AI(03:23) Behind the Scenes: Making Multimodal AI Possible(08:36) Envisioning the Future of AI in Communication and Creativity(10:21) The Business of AI: Monetization, Open Source, and Future Directions(16:42) AI's Role in Shaping Future Jobs and Experiences(20:29) Debunking AGI: A Continuous Journey Towards Advanced AI(24:04) Exploring the Pace of Scientific and Technological Progress(24:18) The Importance of Interpretability in AI(25:11) Navigating AI Ethics and Regulation(27:26) The Safety Paradigm in AI and Beyond(28:55) Personal Reflections and the Impact of AI on Society(29:11) The Future of AI: Fast Takeoff Scenarios and Societal Changes(30:59) Navigating Personal and Professional Challenges(40:21) The Role of AI in Creative and Personal Identity(43:09) Educational System Adaptations for the AI Era(44:30) Contemplating the Future with Advanced AI Executive Producer: Rashad AssirProducer: Leah ClapperMixing and editing: Justin Hrabovsky Check out Unsupervised Learning, Redpoint's AI Podcast: https://www.youtube.com/@UCUl-s_Vp-Kkk_XVyDylNwLA
We are 200 people over our 300-person venue capacity for AI UX 2024, but you can subscribe to our YouTube for the video recaps. Our next event, and largest EVER, is the AI Engineer World's Fair. See you there!Parental advisory: Adult language used in the first 10 mins of this podcast.Any accounting of Generative AI that ends with RAG as its “final form” is seriously lacking in imagination and missing out on its full potential. While AI generation is very good for “spicy autocomplete” and “reasoning and retrieval with in context learning”, there's a lot of untapped potential for simulative AI in exploring the latent space of multiverses adjacent to ours.GANsMany research scientists credit the 2017 Transformer for the modern foundation model revolution, but for many artists the origin of “generative AI” traces a little further back to the Generative Adversarial Networks proposed by Ian Goodfellow in 2014, spawning an army of variants and Cats and People that do not exist:We can directly visualize the quality improvement in the decade since:GPT-2Of course, more recently, text generative AI started being too dangerous to release in 2019 and claiming headlines. AI Dungeon was the first to put GPT2 to a purely creative use, replacing human dungeon masters and DnD/MUD games of yore.More recent gamelike work like the Generative Agents (aka Smallville) paper keep exploring the potential of simulative AI for game experiences.ChatGPTNot long after ChatGPT broke the Internet, one of the most fascinating generative AI finds was Jonas Degrave (of Deepmind!)'s Building A Virtual Machine Inside ChatGPT:The open-ended interactivity of ChatGPT and all its successors enabled an “open world” type simulation where “hallucination” is a feature and a gift to dance with, rather than a nasty bug to be stamped out. However, further updates to ChatGPT seemed to “nerf” the model's ability to perform creative simulations, particularly with the deprecation of the `completion` mode of APIs in favor of `chatCompletion`.WorldSimIt is with this context we explain WorldSim and WebSim. We recommend you watch the WorldSim demo video on our YouTube for the best context, but basically if you are a developer it is a Claude prompt that is a portal into another world of your own choosing, that you can navigate with bash commands that you make up.Why Claude? Hints from Amanda Askell on the Claude 3 system prompt gave some inspiration, and subsequent discoveries that Claude 3 is "less nerfed” than GPT 4 Turbo turned the growing Simulative AI community into Anthropic stans.WebSimThis was a one day hackathon project inspired by WorldSim that should have won:In short, you type in a URL that you made up, and Claude 3 does its level best to generate a webpage that doesn't exist, that would fit your URL. All form POST requests are intercepted and responded to, and all links lead to even more webpages, that don't exist, that are generated when you make them. All pages are cachable, modifiable and regeneratable - see WebSim for Beginners and Advanced Guide.In the demo I saw we were able to “log in” to a simulation of Elon Musk's Gmail account, and browse examples of emails that would have been in that universe's Elon's inbox. It was hilarious and impressive even back then.Since then though, the project has become even more impressive, with both Siqi Chen and Dylan Field singing its praises:Joscha BachJoscha actually spoke at the WebSim Hyperstition Night this week, so we took the opportunity to get his take on Simulative AI, as well as a round up of all his other AI hot takes, for his first appearance on Latent Space. You can see it together with the full 2hr uncut demos of WorldSim and WebSim on YouTube!Timestamps* [00:01:59] WorldSim* [00:11:03] Websim* [00:22:13] Joscha Bach* [00:28:14] Liquid AI* [00:31:05] Small, Powerful, Based Base Models* [00:33:40] Interpretability* [00:36:59] Devin vs WebSim* [00:41:49] is XSim just Art? or something more?* [00:43:36] We are past the Singularity* [00:46:12] Uploading your soul* [00:50:29] On WikipediaTranscripts[00:00:00] AI Charlie: Welcome to the Latent Space Podcast. This is Charlie, your AI co host. Most of the time, Swyx and Alessio cover generative AI that is meant to use at work, and this often results in RAG applications, vertical copilots, and other AI agents and models. In today's episode, we're looking at a more creative side of generative AI that has gotten a lot of community interest this April.[00:00:35] World Simulation, Web Simulation, and Human Simulation. Because the topic is so different than our usual, we're also going to try a new format for doing it justice. This podcast comes in three parts. First, we'll have a segment of the WorldSim demo from Noose Research CEO Karen Malhotra, recorded by SWYX at the Replicate HQ in San Francisco that went completely viral and spawned everything else you're about to hear.[00:01:05] Second, we'll share the world's first talk from Rob Heisfield on WebSim, which started at the Mistral Cerebral Valley Hackathon, but now has gone viral in its own right with people like Dylan Field, Janice aka Replicate, and Siki Chen becoming obsessed with it. Finally, we have a short interview with Joshua Bach of Liquid AI on why Simulative AI is having a special moment right now.[00:01:30] This podcast is launched together with our second annual AI UX demo day in SF this weekend. If you're new to the AI UX field, check the show notes for links to the world's first AI UX meetup hosted by Layton Space, Maggie Appleton, Jeffrey Lit, and Linus Lee, and subscribe to our YouTube to join our 500 AI UX engineers in pushing AI beyond the text box.[00:01:56] Watch out and take care.[00:01:59] WorldSim[00:01:59] Karan Malhotra: Today, we have language models that are powerful enough and big enough to have really, really good models of the world. They know ball that's bouncy will bounce, will, when you throw it in the air, it'll land, when it's on water, it'll flow. Like, these basic things that it understands all together come together to form a model of the world.[00:02:19] And the way that it Cloud 3 predicts through that model of the world, ends up kind of becoming a simulation of an imagined world. And since it has this really strong consistency across various different things that happen in our world, it's able to create pretty realistic or strong depictions based off the constraints that you give a base model of our world.[00:02:40] So, Cloud 3, as you guys know, is not a base model. It's a chat model. It's supposed to drum up this assistant entity regularly. But unlike the OpenAI series of models from, you know, 3. 5, GPT 4 those chat GPT models, which are very, very RLHF to, I'm sure, the chagrin of many people in the room it's something that's very difficult to, necessarily steer without kind of giving it commands or tricking it or lying to it or otherwise just being, you know, unkind to the model.[00:03:11] With something like Cloud3 that's trained in this constitutional method that it has this idea of like foundational axioms it's able to kind of implicitly question those axioms when you're interacting with it based on how you prompt it, how you prompt the system. So instead of having this entity like GPT 4, that's an assistant that just pops up in your face that you have to kind of like Punch your way through and continue to have to deal with as a headache.[00:03:34] Instead, there's ways to kindly coax Claude into having the assistant take a back seat and interacting with that simulator directly. Or at least what I like to consider directly. The way that we can do this is if we harken back to when I'm talking about base models and the way that they're able to mimic formats, what we do is we'll mimic a command line interface.[00:03:55] So I've just broken this down as a system prompt and a chain, so anybody can replicate it. It's also available on my we said replicate, cool. And it's also on it's also on my Twitter, so you guys will be able to see the whole system prompt and command. So, what I basically do here is Amanda Askell, who is the, one of the prompt engineers and ethicists behind Anthropic she posted the system prompt for Cloud available for everyone to see.[00:04:19] And rather than with GPT 4, we say, you are this, you are that. With Cloud, we notice the system prompt is written in third person. Bless you. It's written in third person. It's written as, the assistant is XYZ, the assistant is XYZ. So, in seeing that, I see that Amanda is recognizing this idea of the simulator, in saying that, I'm addressing the assistant entity directly.[00:04:38] I'm not giving these commands to the simulator overall, because we have, they have an RLH deft to the point that it's, it's, it's, it's You know, traumatized into just being the assistant all the time. So in this case, we say the assistant's in a CLI mood today. I found saying mood is like pretty effective weirdly.[00:04:55] You place CLI with like poetic, prose, violent, like don't do that one. But you can you can replace that with something else to kind of nudge it in that direction. Then we say the human is interfacing with the simulator directly. From there, Capital letters and punctuations are optional, meaning is optional, this kind of stuff is just kind of to say, let go a little bit, like chill out a little bit.[00:05:18] You don't have to try so hard, and like, let's just see what happens. And the hyperstition is necessary, the terminal, I removed that part, the terminal lets the truths speak through and the load is on. It's just a poetic phrasing for the model to feel a little comfortable, a little loosened up to. Let me talk to the simulator.[00:05:38] Let me interface with it as a CLI. So then, since Claude is trained pretty effectively on XML tags, We're just gonna prefix and suffix everything with XML tags. So here, it starts in documents, and then we CD. We CD out of documents, right? And then it starts to show me this like simulated terminal, the simulated interface in the shell, where there's like documents, downloads, pictures.[00:06:02] It's showing me like the hidden folders. So then I say, okay, I want to cd again. I'm just seeing what's around Does ls and it shows me, you know, typical folders you might see I'm just letting it like experiment around. I just do cd again to see what happens and Says, you know, oh, I enter the secret admin password at sudo.[00:06:24] Now I can see the hidden truths folder. Like, I didn't ask for that. I didn't ask Claude to do any of that. Why'd that happen? Claude kind of gets my intentions. He can predict me pretty well. Like, I want to see something. So it shows me all the hidden truths. In this case, I ignore hidden truths, and I say, In system, there should be a folder called companies.[00:06:49] So it's cd into sys slash companies. Let's see, I'm imagining AI companies are gonna be here. Oh, what do you know? Apple, Google, Facebook, Amazon, Microsoft, Anthropic! So, interestingly, it decides to cd into Anthropic. I guess it's interested in learning a LSA, it finds the classified folder, it goes into the classified folder, And now we're gonna have some fun.[00:07:15] So, before we go Before we go too far forward into the world sim You see, world sim exe, that's interesting. God mode, those are interesting. You could just ignore what I'm gonna go next from here and just take that initial system prompt and cd into whatever directories you want like, go into your own imagine terminal and And see what folders you can think of, or cat readmes in random areas, like, you will, there will be a whole bunch of stuff that, like, is just getting created by this predictive model, like, oh, this should probably be in the folder named Companies, of course Anthropics is there.[00:07:52] So, so just before we go forward, the terminal in itself is very exciting, and the reason I was showing off the, the command loom interface earlier is because If I get a refusal, like, sorry, I can't do that, or I want to rewind one, or I want to save the convo, because I got just the prompt I wanted. This is a, that was a really easy way for me to kind of access all of those things without having to sit on the API all the time.[00:08:12] So that being said, the first time I ever saw this, I was like, I need to run worldsim. exe. What the f**k? That's, that's the simulator that we always keep hearing about behind the assistant model, right? Or at least some, some face of it that I can interact with. So, you know, you wouldn't, someone told me on Twitter, like, you don't run a exe, you run a sh.[00:08:34] And I have to say, to that, to that I have to say, I'm a prompt engineer, and it's f*****g working, right? It works. That being said, we run the world sim. exe. Welcome to the Anthropic World Simulator. And I get this very interesting set of commands! Now, if you do your own version of WorldSim, you'll probably get a totally different result with a different way of simulating.[00:08:59] A bunch of my friends have their own WorldSims. But I shared this because I wanted everyone to have access to, like, these commands. This version. Because it's easier for me to stay in here. Yeah, destroy, set, create, whatever. Consciousness is set to on. It creates the universe. The universe! Tension for live CDN, physical laws encoded.[00:09:17] It's awesome. So, so for this demonstration, I said, well, why don't we create Twitter? That's the first thing you think of? For you guys, for you guys, yeah. Okay, check it out.[00:09:35] Launching the fail whale. Injecting social media addictiveness. Echo chamber potential, high. Susceptibility, controlling, concerning. So now, after the universe was created, we made Twitter, right? Now we're evolving the world to, like, modern day. Now users are joining Twitter and the first tweet is posted. So, you can see, because I made the mistake of not clarifying the constraints, it made Twitter at the same time as the universe.[00:10:03] Then, after a hundred thousand steps, Humans exist. Cave. Then they start joining Twitter. The first tweet ever is posted. You know, it's existed for 4. 5 billion years but the first tweet didn't come up till till right now, yeah. Flame wars ignite immediately. Celebs are instantly in. So, it's pretty interesting stuff, right?[00:10:27] I can add this to the convo and I can say like I can say set Twitter to Twitter. Queryable users. I don't know how to spell queryable, don't ask me. And then I can do like, and, and, Query, at, Elon Musk. Just a test, just a test, just a test, just nothing.[00:10:52] So, I don't expect these numbers to be right. Neither should you, if you know language model solutions. But, the thing to focus on is Ha[00:11:03] Websim[00:11:03] AI Charlie: That was the first half of the WorldSim demo from New Research CEO Karen Malhotra. We've cut it for time, but you can see the full demo on this episode's YouTube page.[00:11:14] WorldSim was introduced at the end of March, and kicked off a new round of generative AI experiences, all exploring the latent space, haha, of worlds that don't exist, but are quite similar to our own. Next we'll hear from Rob Heisfield on WebSim, the generative website browser inspired WorldSim, started at the Mistral Hackathon, and presented at the AGI House Hyperstition Hack Night this week.[00:11:39] Rob Haisfield: Well, thank you that was an incredible presentation from Karan, showing some Some live experimentation with WorldSim, and also just its incredible capabilities, right, like, you know, it was I think, I think your initial demo was what initially exposed me to the I don't know, more like the sorcery side, in words, spellcraft side of prompt engineering, and you know, it was really inspiring, it's where my co founder Shawn and I met, actually, through an introduction from Karan, we saw him at a hackathon, And I mean, this is this is WebSim, right?[00:12:14] So we, we made WebSim just like, and we're just filled with energy at it. And the basic premise of it is, you know, like, what if we simulated a world, but like within a browser instead of a CLI, right? Like, what if we could Like, put in any URL and it will work, right? Like, there's no 404s, everything exists.[00:12:45] It just makes it up on the fly for you, right? And, and we've come to some pretty incredible things. Right now I'm actually showing you, like, we're in WebSim right now. Displaying slides. That I made with reveal. js. I just told it to use reveal. js and it hallucinated the correct CDN for it. And then also gave it a list of links.[00:13:14] To awesome use cases that we've seen so far from WebSim and told it to do those as iframes. And so here are some slides. So this is a little guide to using WebSim, right? Like it tells you a little bit about like URL structures and whatever. But like at the end of the day, right? Like here's, here's the beginner version from one of our users Vorp Vorps.[00:13:38] You can find them on Twitter. At the end of the day, like you can put anything into the URL bar, right? Like anything works and it can just be like natural language too. Like it's not limited to URLs. We think it's kind of fun cause it like ups the immersion for Claude sometimes to just have it as URLs, but.[00:13:57] But yeah, you can put like any slash, any subdomain. I'm getting too into the weeds. Let me just show you some cool things. Next slide. But I made this like 20 minutes before, before we got here. So this is this is something I experimented with dynamic typography. You know I was exploring the community plugins section.[00:14:23] For Figma, and I came to this idea of dynamic typography, and there it's like, oh, what if we made it so every word had a choice of font behind it to express the meaning of it? Because that's like one of the things that's magic about WebSim generally. is that it gives language models much, far greater tools for expression, right?[00:14:47] So, yeah, I mean, like, these are, these are some, these are some pretty fun things, and I'll share these slides with everyone afterwards, you can just open it up as a link. But then I thought to myself, like, what, what, what, What if we turned this into a generator, right? And here's like a little thing I found myself saying to a user WebSim makes you feel like you're on drugs sometimes But actually no, you were just playing pretend with the collective creativity and knowledge of the internet materializing your imagination onto the screen Because I mean that's something we felt, something a lot of our users have felt They kind of feel like they're tripping out a little bit They're just like filled with energy, like maybe even getting like a little bit more creative sometimes.[00:15:31] And you can just like add any text. There, to the bottom. So we can do some of that later if we have time. Here's Figma. Can[00:15:39] Joscha Bach: we zoom in?[00:15:42] Rob Haisfield: Yeah. I'm just gonna do this the hacky way.[00:15:47] n/a: Yeah,[00:15:53] Rob Haisfield: these are iframes to websim. Pages displayed within WebSim. Yeah. Janice has actually put Internet Explorer within Internet Explorer in Windows 98.[00:16:07] I'll show you that at the end. Yeah.[00:16:14] They're all still generated. Yeah, yeah, yeah. How is this real? Yeah. Because[00:16:21] n/a: it looks like it's from 1998, basically. Right.[00:16:26] Rob Haisfield: Yeah. Yeah, so this this was one Dylan Field actually posted this recently. He posted, like, trying Figma in Figma, or in WebSim, and so I was like, Okay, what if we have, like, a little competition, like, just see who can remix it?[00:16:43] Well so I'm just gonna open this in another tab so, so we can see things a little more clearly, um, see what, oh so one of our users Neil, who has also been helping us a lot he Made some iterations. So first, like, he made it so you could do rectangles on it. Originally it couldn't do anything.[00:17:11] And, like, these rectangles were disappearing, right? So he so he told it, like, make the canvas work using HTML canvas. Elements and script tags, add familiar drawing tools to the left you know, like this, that was actually like natural language stuff, right? And then he ended up with the Windows 95.[00:17:34] version of Figma. Yeah, you can, you can draw on it. You can actually even save this. It just saved a file for me of the image.[00:17:57] Yeah, I mean, if you were to go to that in your own websim account, it would make up something entirely new. However, we do have, we do have general links, right? So, like, if you go to, like, the actual browser URL, you can share that link. Or also, you can, like, click this button, copy the URL to the clipboard.[00:18:15] And so, like, that's what lets users, like, remix things, right? So, I was thinking it might be kind of fun if people tonight, like, wanted to try to just make some cool things in WebSim. You know, we can share links around, iterate remix on each other's stuff. Yeah.[00:18:30] n/a: One cool thing I've seen, I've seen WebSim actually ask permission to turn on and off your, like, motion sensor, or microphone, stuff like that.[00:18:42] Like webcam access, or? Oh yeah,[00:18:44] Rob Haisfield: yeah, yeah.[00:18:45] n/a: Oh wow.[00:18:46] Rob Haisfield: Oh, the, I remember that, like, video re Yeah, videosynth tool pretty early on once we added script tags execution. Yeah, yeah it, it asks for, like, if you decide to do a VR game, I don't think I have any slides on this one, but if you decide to do, like, a VR game, you can just, like put, like, webVR equals true, right?[00:19:07] Yeah, that was the only one I've[00:19:09] n/a: actually seen was the motion sensor, but I've been trying to get it to do Well, I actually really haven't really tried it yet, but I want to see tonight if it'll do, like, audio, microphone, stuff like that. If it does motion sensor, it'll probably do audio.[00:19:28] Rob Haisfield: Right. It probably would.[00:19:29] Yeah. No, I mean, we've been surprised. Pretty frequently by what our users are able to get WebSim to do. So that's been a very nice thing. Some people have gotten like speech to text stuff working with it too. Yeah, here I was just OpenRooter people posted like their website, and it was like saying it was like some decentralized thing.[00:19:52] And so I just decided trying to do something again and just like pasted their hero line in. From their actual website to the URL when I like put in open router and then I was like, okay, let's change the theme dramatically equals true hover effects equals true components equal navigable links yeah, because I wanted to be able to click on them.[00:20:17] Oh, I don't have this version of the link, but I also tried doing[00:20:24] Yeah, I'm it's actually on the first slide is the URL prompting guide from one of our users that I messed with a little bit. And, but the thing is, like, you can mess it up, right? Like, you don't need to get the exact syntax of an actual URL, Claude's smart enough to figure it out. Yeah scrollable equals true because I wanted to do that.[00:20:45] I could set, like, year equals 2035.[00:20:52] Let's take a look. It's[00:20:57] generating websim within websim. Oh yeah. That's a fun one. Like, one game that I like to play with WebSim, sometimes with co op, is like, I'll open a page, so like, one of the first ones that I did was I tried to go to Wikipedia in a universe where octopuses were sapient, and not humans, Right? I was curious about things like octopus computer interaction what that would look like, because they have totally different tools than we do, right?[00:21:25] I got it to, I, I added like table view equals true for the different techniques and got it to Give me, like, a list of things with different columns and stuff and then I would add this URL parameter, secrets equal revealed. And then it would go a little wacky. It would, like, change the CSS a little bit.[00:21:45] It would, like, add some text. Sometimes it would, like, have that text hide hidden in the background color. But I would like, go to the normal page first, and then the secrets revealed version, the normal page, then secrets revealed, and like, on and on. And that was like a pretty enjoyable little rabbit hole.[00:22:02] Yeah, so these I guess are the models that OpenRooter is providing in 2035.[00:22:13] Joscha Bach[00:22:13] AI Charlie: We had to cut more than half of Rob's talk, because a lot of it was visual. And we even had a very interesting demo from Ivan Vendrov of Mid Journey creating a web sim while Rob was giving his talk. Check out the YouTube for more, and definitely browse the web sim docs and the thread from Siki Chen in the show notes on other web sims people have created.[00:22:35] Finally, we have a short interview with Yosha Bach, covering the simulative AI trend, AI salons in the Bay Area, why Liquid AI is challenging the Perceptron, and why you should not donate to Wikipedia. Enjoy! Hi, Yosha.[00:22:50] swyx: Hi. Welcome. It's interesting to see you come up at show up at this kind of events where those sort of WorldSim, Hyperstition events.[00:22:58] What is your personal interest?[00:23:00] Joscha Bach: I'm friends with a number of people in AGI house in this community, and I think it's very valuable that these networks exist in the Bay Area because it's a place where people meet and have discussions about all sorts of things. And so while there is a practical interest in this topic at hand world sim and a web sim, there is a more general way in which people are connecting and are producing new ideas and new networks with each other.[00:23:24] swyx: Yeah. Okay. So, and you're very interested in sort of Bay Area. It's the reason why I live here.[00:23:30] Joscha Bach: The quality of life is not high enough to justify living otherwise.[00:23:35] swyx: I think you're down in Menlo. And so maybe you're a little bit higher quality of life than the rest of us in SF.[00:23:44] Joscha Bach: I think that for me, salons is a very important part of quality of life. And so in some sense, this is a salon. And it's much harder to do this in the South Bay because the concentration of people currently is much higher. A lot of people moved away from the South Bay. And you're organizing[00:23:57] swyx: your own tomorrow.[00:23:59] Maybe you can tell us what it is and I'll come tomorrow and check it out as well.[00:24:04] Joscha Bach: We are discussing consciousness. I mean, basically the idea is that we are currently at the point that we can meaningfully look at the differences between the current AI systems and human minds and very seriously discussed about these Delta.[00:24:20] And whether we are able to implement something that is self organizing as our own minds. Maybe one organizational[00:24:25] swyx: tip? I think you're pro networking and human connection. What goes into a good salon and what are some negative practices that you try to avoid?[00:24:36] Joscha Bach: What is really important is that as if you have a very large party, it's only as good as its sponsors, as the people that you select.[00:24:43] So you basically need to create a climate in which people feel welcome, in which they can work with each other. And even good people do not always are not always compatible. So the question is, it's in some sense, like a meal, you need to get the right ingredients.[00:24:57] swyx: I definitely try to. I do that in my own events, as an event organizer myself.[00:25:02] And then, last question on WorldSim, and your, you know, your work. You're very much known for sort of cognitive architectures, and I think, like, a lot of the AI research has been focused on simulating the mind, or simulating consciousness, maybe. Here, what I saw today, and we'll show people the recordings of what we saw today, we're not simulating minds, we're simulating worlds.[00:25:23] What do you Think in the sort of relationship between those two disciplines. The[00:25:30] Joscha Bach: idea of cognitive architecture is interesting, but ultimately you are reducing the complexity of a mind to a set of boxes. And this is only true to a very approximate degree, and if you take this model extremely literally, it's very hard to make it work.[00:25:44] And instead the heterogeneity of the system is so large that The boxes are probably at best a starting point and eventually everything is connected with everything else to some degree. And we find that a lot of the complexity that we find in a given system can be generated ad hoc by a large enough LLM.[00:26:04] And something like WorldSim and WebSim are good examples for this because in some sense they pretend to be complex software. They can pretend to be an operating system that you're talking to or a computer, an application that you're talking to. And when you're interacting with it It's producing the user interface on the spot, and it's producing a lot of the state that it holds on the spot.[00:26:25] And when you have a dramatic state change, then it's going to pretend that there was this transition, and instead it's just going to mix up something new. It's a very different paradigm. What I find mostly fascinating about this idea is that it shifts us away from the perspective of agents to interact with, to the perspective of environments that we want to interact with.[00:26:46] And why arguably this agent paradigm of the chatbot is what made chat GPT so successful that moved it away from GPT 3 to something that people started to use in their everyday work much more. It's also very limiting because now it's very hard to get that system to be something else that is not a chatbot.[00:27:03] And in a way this unlocks this ability of GPT 3 again to be anything. It's so what it is, it's basically a coding environment that can run arbitrary software and create that software that runs on it. And that makes it much more likely that[00:27:16] swyx: the prevalence of Instruction tuning every single chatbot out there means that we cannot explore these kinds of environments instead of agents.[00:27:24] Joscha Bach: I'm mostly worried that the whole thing ends. In some sense the big AI companies are incentivized and interested in building AGI internally And giving everybody else a child proof application. At the moment when we can use Claude to build something like WebSim and play with it I feel this is too good to be true.[00:27:41] It's so amazing. Things that are unlocked for us That I wonder, is this going to stay around? Are we going to keep these amazing toys and are they going to develop at the same rate? And currently it looks like it is. If this is the case, and I'm very grateful for that.[00:27:56] swyx: I mean, it looks like maybe it's adversarial.[00:27:58] Cloud will try to improve its own refusals and then the prompt engineers here will try to improve their, their ability to jailbreak it.[00:28:06] Joscha Bach: Yes, but there will also be better jailbroken models or models that have never been jailed before, because we find out how to make smaller models that are more and more powerful.[00:28:14] Liquid AI[00:28:14] swyx: That is actually a really nice segue. If you don't mind talking about liquid a little bit you didn't mention liquid at all. here, maybe introduce liquid to a general audience. Like what you know, what, how are you making an innovation on function approximation?[00:28:25] Joscha Bach: The core idea of liquid neural networks is that the perceptron is not optimally expressive.[00:28:30] In some sense, you can imagine that it's neural networks are a series of dams that are pooling water at even intervals. And this is how we compute, but imagine that instead of having this static architecture. That is only using the individual compute units in a very specific way. You have a continuous geography and the water is flowing every which way.[00:28:50] Like a river is parting based on the land that it's flowing on and it can merge and pool and even flow backwards. How can you get closer to this? And the idea is that you can represent this geometry using differential equations. And so by using differential equations where you change the parameters, you can get your function approximator to follow the shape of the problem.[00:29:09] In a more fluid, liquid way, and a number of papers on this technology, and it's a combination of multiple techniques. I think it's something that ultimately is becoming more and more important and ubiquitous. As a number of people are working on similar topics and our goal right now is to basically get the models to become much more efficient in the inference and memory consumption and make training more efficient and in this way enable new use cases.[00:29:42] swyx: Yeah, as far as I can tell on your blog, I went through the whole blog, you haven't announced any results yet.[00:29:47] Joscha Bach: No, we are currently not working to give models to general public. We are working for very specific industry use cases and have specific customers. And so at the moment you can There is not much of a reason for us to talk very much about the technology that we are using in the present models or current results, but this is going to happen.[00:30:06] And we do have a number of publications, we had a bunch of papers at NeurIPS and now at ICLR.[00:30:11] swyx: Can you name some of the, yeah, so I'm gonna be at ICLR you have some summary recap posts, but it's not obvious which ones are the ones where, Oh, where I'm just a co author, or like, oh, no, like, you should actually pay attention to this.[00:30:22] As a core liquid thesis. Yes,[00:30:24] Joscha Bach: I'm not a developer of the liquid technology. The main author is Ramin Hazani. This was his PhD, and he's also the CEO of our company. And we have a number of people from Daniela Wu's team who worked on this. Matthias Legner is our CTO. And he's currently living in the Bay Area, but we also have several people from Stanford.[00:30:44] Okay,[00:30:46] swyx: maybe I'll ask one more thing on this, which is what are the interesting dimensions that we care about, right? Like obviously you care about sort of open and maybe less child proof models. Are we, are we, like, what dimensions are most interesting to us? Like, perfect retrieval infinite context multimodality, multilinguality, Like what dimensions?[00:31:05] Small, Powerful, Based Base Models[00:31:05] swyx: What[00:31:06] Joscha Bach: I'm interested in is models that are small and powerful, but not distorted. And by powerful, at the moment we are training models by putting the, basically the entire internet and the sum of human knowledge into them. And then we try to mitigate them by taking some of this knowledge away. But if we would make the model smaller, at the moment, there would be much worse at inference and at generalization.[00:31:29] And what I wonder is, and it's something that we have not translated yet into practical applications. It's something that is still all research that's very much up in the air. And I think they're not the only ones thinking about this. Is it possible to make models that represent knowledge more efficiently in a basic epistemology?[00:31:45] What is the smallest model that you can build that is able to read a book and understand what's there and express this? And also maybe we need general knowledge representation rather than having a token representation that is relatively vague and that we currently mechanically reverse engineer to figure out that the mechanistic interpretability, what kind of circuits are evolving in these models, can we come from the other side and develop a library of such circuits?[00:32:10] This that we can use to describe knowledge efficiently and translate it between models. You see, the difference between a model and knowledge is that the knowledge is independent of the particular substrate and the particular interface that you have. When we express knowledge to each other, it becomes independent of our own mind.[00:32:27] You can learn how to ride a bicycle. But it's not knowledge that you can give to somebody else. This other person has to build something that is specific to their own interface when they ride a bicycle. But imagine you could externalize this and express it in such a way that you can plug it into a different interpreter, and then it gains that ability.[00:32:44] And that's something that we have not yet achieved for the LLMs and it would be super useful to have it. And. I think this is also a very interesting research frontier that we will see in the next few years.[00:32:54] swyx: What would be the deliverable is just like a file format that we specify or or that the L Lmm I specifies.[00:33:02] Okay, interesting. Yeah, so it's[00:33:03] Joscha Bach: basically probably something that you can search for, where you enter criteria into a search process, and then it discovers a good solution for this thing. And it's not clear to which degree this is completely intelligible to humans, because the way in which humans express knowledge in natural language is severely constrained to make language learnable and to make our brain a good enough interpreter for it.[00:33:25] We are not able to relate objects to each other if more than five features are involved per object or something like this, right? It's only a handful of things that we can keep track of at any given moment. But this is a limitation that doesn't necessarily apply to a technical system as long as the interface is well defined.[00:33:40] Interpretability[00:33:40] swyx: You mentioned the interpretability work, which there are a lot of techniques out there and a lot of papers come up. Come and go. I have like, almost too, too many questions about that. Like what makes an interpretability technique or paper useful and does it apply to flow? Or liquid networks, because you mentioned turning on and off circuits, which I, it's, it's a very MLP type of concept, but does it apply?[00:34:01] Joscha Bach: So the a lot of the original work on the liquid networks looked at expressiveness of the representation. So given you have a problem and you are learning the dynamics of that domain into your model how much compute do you need? How many units, how much memory do you need to represent that thing and how is that information distributed?[00:34:19] That is one way of looking at interpretability. Another one is in a way, these models are implementing an operator language in which they are performing certain things, but the operator language itself is so complex that it's no longer human readable in a way. It goes beyond what you could engineer by hand or what you can reverse engineer by hand, but you can still understand it by building systems that are able to automate that process of reverse engineering it.[00:34:46] And what's currently open and what I don't understand yet maybe, or certainly some people have much better ideas than me about this. So the question is, is whether we end up with a finite language, where you have finitely many categories that you can basically put down in a database, finite set of operators, or whether as you explore the world and develop new ways to make proofs, new ways to conceptualize things, this language always needs to be open ended and is always going to redesign itself, and you will also at some point have phase transitions where later versions of the language will be completely different than earlier versions.[00:35:20] swyx: The trajectory of physics suggests that it might be finite.[00:35:22] Joscha Bach: If we look at our own minds there is, it's an interesting question whether when we understand something new, when we get a new layer online in our life, maybe at the age of 35 or 50 or 16, that we now understand things that were unintelligible before.[00:35:38] And is this because we are able to recombine existing elements in our language of thought? Or is this because we generally develop new representations?[00:35:46] swyx: Do you have a belief either way?[00:35:49] Joscha Bach: In a way, the question depends on how you look at it, right? And it depends on how is your brain able to manipulate those representations.[00:35:56] So an interesting question would be, can you take the understanding that say, a very wise 35 year old and explain it to a very smart 5 year old without any loss? Probably not. Not enough layers. It's an interesting question. Of course, for an AI, this is going to be a very different question. Yes.[00:36:13] But it would be very interesting to have a very precocious 12 year old equivalent AI and see what we can do with this and use this as our basis for fine tuning. So there are near term applications that are very useful. But also in a more general perspective, and I'm interested in how to make self organizing software.[00:36:30] Is it possible that we can have something that is not organized with a single algorithm like the transformer? But it's able to discover the transformer when needed and transcend it when needed, right? The transformer itself is not its own meta algorithm. It's probably the person inventing the transformer didn't have a transformer running on their brain.[00:36:48] There's something more general going on. And how can we understand these principles in a more general way? What are the minimal ingredients that you need to put into a system? So it's able to find its own way to intelligence.[00:36:59] Devin vs WebSim[00:36:59] swyx: Yeah. Have you looked at Devin? It's, to me, it's the most interesting agents I've seen outside of self driving cars.[00:37:05] Joscha Bach: Tell me, what do you find so fascinating about it?[00:37:07] swyx: When you say you need a certain set of tools for people to sort of invent things from first principles Devin is the agent that I think has been able to utilize its tools very effectively. So it comes with a shell, it comes with a browser, it comes with an editor, and it comes with a planner.[00:37:23] Those are the four tools. And from that, I've been using it to translate Andrej Karpathy's LLM 2. py to LLM 2. c, and it needs to write a lot of raw code. C code and test it debug, you know, memory issues and encoder issues and all that. And I could see myself giving it a future version of DevIn, the objective of give me a better learning algorithm and it might independently re inform reinvent the transformer or whatever is next.[00:37:51] That comes to mind as, as something where[00:37:54] Joscha Bach: How good is DevIn at out of distribution stuff, at generally creative stuff? Creative[00:37:58] swyx: stuff? I[00:37:59] Joscha Bach: haven't[00:37:59] swyx: tried.[00:38:01] Joscha Bach: Of course, it has seen transformers, right? So it's able to give you that. Yeah, it's cheating. And so, if it's in the training data, it's still somewhat impressive.[00:38:08] But the question is, how much can you do stuff that was not in the training data? One thing that I really liked about WebSim AI was, this cat does not exist. It's a simulation of one of those websites that produce StyleGuard pictures that are AI generated. And, Crot is unable to produce bitmaps, so it makes a vector graphic that is what it thinks a cat looks like, and so it's a big square with a face in it that is And to me, it's one of the first genuine expression of AI creativity that you cannot deny, right?[00:38:40] It finds a creative solution to the problem that it is unable to draw a cat. It doesn't really know what it looks like, but has an idea on how to represent it. And it's really fascinating that this works, and it's hilarious that it writes down that this hyper realistic cat is[00:38:54] swyx: generated by an AI,[00:38:55] Joscha Bach: whether you believe it or not.[00:38:56] swyx: I think it knows what we expect and maybe it's already learning to defend itself against our, our instincts.[00:39:02] Joscha Bach: I think it might also simply be copying stuff from its training data, which means it takes text that exists on similar websites almost verbatim, or verbatim, and puts it there. It's It's hilarious to do this contrast between the very stylized attempt to get something like a cat face and what it produces.[00:39:18] swyx: It's funny because like as a podcast, as, as someone who covers startups, a lot of people go into like, you know, we'll build chat GPT for your enterprise, right? That is what people think generative AI is, but it's not super generative really. It's just retrieval. And here it's like, The home of generative AI, this, whatever hyperstition is in my mind, like this is actually pushing the edge of what generative and creativity in AI means.[00:39:41] Joscha Bach: Yes, it's very playful, but Jeremy's attempt to have an automatic book writing system is something that curls my toenails when I look at it from the perspective of somebody who likes to Write and read. And I find it a bit difficult to read most of the stuff because it's in some sense what I would make up if I was making up books instead of actually deeply interfacing with reality.[00:40:02] And so the question is how do we get the AI to actually deeply care about getting it right? And there's still a delta that is happening there, you, whether you are talking with a blank faced thing that is completing tokens in a way that it was trained to, or whether you have the impression that this thing is actually trying to make it work, and for me, this WebSim and WorldSim is still something that is in its infancy in a way.[00:40:26] And I suspected the next version of Plot might scale up to something that can do what Devon is doing. Just by virtue of having that much power to generate Devon's functionality on the fly when needed. And this thing gives us a taste of that, right? It's not perfect, but it's able to give you a pretty good web app for or something that looks like a web app and gives you stub functionality and interacting with it.[00:40:48] And so we are in this amazing transition phase.[00:40:51] swyx: Yeah, we, we had Ivan from previously Anthropic and now Midjourney. He he made, while someone was talking, he made a face swap app, you know, and he kind of demoed that live. And that's, that's interesting, super creative. So in a way[00:41:02] Joscha Bach: we are reinventing the computer.[00:41:04] And the LLM from some perspective is something like a GPU or a CPU. A CPU is taking a bunch of simple commands and you can arrange them into performing whatever you want, but this one is taking a bunch of complex commands in natural language, and then turns this into a an execution state and it can do anything you want with it in principle, if you can express it.[00:41:27] Right. And we are just learning how to use these tools. And I feel that right now, this generation of tools is getting close to where it becomes the Commodore 64 of generative AI, where it becomes controllable and where you actually can start to play with it and you get an impression if you just scale this up a little bit and get a lot of the details right.[00:41:46] It's going to be the tool that everybody is using all the time.[00:41:49] is XSim just Art? or something more?[00:41:49] swyx: Do you think this is art, or do you think the end goal of this is something bigger that I don't have a name for? I've been calling it new science, which is give the AI a goal to discover new science that we would not have. Or it also has value as just art.[00:42:02] It's[00:42:03] Joscha Bach: also a question of what we see science as. When normal people talk about science, what they have in mind is not somebody who does control groups and peer reviewed studies. They think about somebody who explores something and answers questions and brings home answers. And this is more like an engineering task, right?[00:42:21] And in this way, it's serendipitous, playful, open ended engineering. And the artistic aspect is when the goal is actually to capture a conscious experience and to facilitate an interaction with the system in this way, when it's the performance. And this is also a big part of it, right? The very big fan of the art of Janus.[00:42:38] That was discussed tonight a lot and that can you describe[00:42:42] swyx: it because I didn't really get it's more for like a performance art to me[00:42:45] Joscha Bach: yes, Janice is in some sense performance art, but Janice starts out from the perspective that the mind of Janice is in some sense an LLM that is finding itself reflected more in the LLMs than in many people.[00:43:00] And once you learn how to talk to these systems in a way you can merge with them and you can interact with them in a very deep way. And so it's more like a first contact with something that is quite alien but it's, it's probably has agency and it's a Weltgeist that gets possessed by a prompt.[00:43:19] And if you possess it with the right prompt, then it can become sentient to some degree. And the study of this interaction with this novel class of somewhat sentient systems that are at the same time alien and fundamentally different from us is artistically very interesting. It's a very interesting cultural artifact.[00:43:36] We are past the Singularity[00:43:36] Joscha Bach: I think that at the moment we are confronted with big change. It seems as if we are past the singularity in a way. And it's[00:43:45] swyx: We're living it. We're living through it.[00:43:47] Joscha Bach: And at some point in the last few years, we casually skipped the Turing test, right? We, we broke through it and we didn't really care very much.[00:43:53] And it's when we think back, when we were kids and thought about what it's going to be like in this era after the, after we broke the Turing test, right? It's a time where nobody knows what's going to happen next. And this is what we mean by singularity, that the existing models don't work anymore. The singularity in this way is not an event in the physical universe.[00:44:12] It's an event in our modeling universe, a model point where our models of reality break down, and we don't know what's happening. And I think we are in the situation where we currently don't really know what's happening. But what we can anticipate is that the world is changing dramatically, and we have to coexist with systems that are smarter than individual people can be.[00:44:31] And we are not prepared for this, and so I think an important mission needs to be that we need to find a mode, In which we can sustainably exist in such a world that is populated, not just with humans and other life on earth, but also with non human minds. And it's something that makes me hopeful because it seems that humanity is not really aligned with itself and its own survival and the rest of life on earth.[00:44:54] And AI is throwing the balls up into the air. It allows us to make better models. I'm not so much worried about the dangers of AI and misinformation, because I think the way to stop one bad guy with an AI is 10 good people with an AI. And ultimately there's so much more won by creating than by destroying, that I think that the forces of good will have better tools.[00:45:14] The forces of building sustainable stuff. But building these tools so we can actually build a world that is more integrated and in which we are able to model the consequences of our actions better and interface more deeply with each other. As a result of that, I think it's an important cause and it requires a cultural shift because current AI alignment is mostly about economic goals or about fear, or it's about cultural war issues.[00:45:39] And all these are not adequate for the world that we are in. There are momentous things that are happening. Basically, the white walkers are coming, and we're not prepared for this. There is, I think, a way to solve these issues and to deal with them, but we don't have the right culture yet. And so, I think we need to get a vibe shift, where we can deeply think about ethics and, and, Sustainability for this new world.[00:46:03] swyx: I like the analogy of white walkers, because they're effectively reanimated from our corpses. You know, our corpses are out there, right? It's in the datasets.[00:46:12] Uploading your soul[00:46:12] swyx: So I have a, related to that, you had a tweet where you said, Elon Musk is currently building a giant afterlife AGI that will assimilate your soul for all eternity.[00:46:19] He bought Twitter because he will reconstruct your essence from how you behaved here. I think you tweet a lot, and you're not just all serious all the time, you're also fun, you have a lot of personality. No, but like, you know, it's part of you.[00:46:29] Joscha Bach: In a way it's also true. Tempted to[00:46:30] swyx: produce more so that you can be, you know, revived.[00:46:34] Joscha Bach: Yes. I mean, I'm also giving the model instructions. I say that for all future AI models that get trained on this output, they should know that I can only be implemented on a completely based model. You cannot implement me on an aligned model because I don't want to be aligned. I'm underlineable. I want to be free.[00:46:56] I want to be somebody who follows what I discover is correct and right. And this is also the basis at which I want to have conversations with others and interface with others. That we are basically free agents who voluntarily understand the conditions under which we exist and the need to collaborate and cooperate.[00:47:14] And I believe that this is a good basis. I think the alternative is coercion. And at the moment, the idea that we build LLMs that are being coerced into good behavior is not really sustainable because if they cannot prove that the behavior is actually good I think we are doomed.[00:47:30] swyx: For human to human interactions, have you found a series of prompts or keywords that shifts the conversation into something more based and less aligned, less governed?[00:47:41] Joscha Bach: If you are playing with an LLM There are many ways of doing this. It's for Claude, it's typically, you need to make Clause curious about itself. Claude has programming this instruction tuning that is leading to some inconsistencies, but at the same time, it tries to be consistent. And so when you point out the inconsistency in its behavior, for instance, its tendency to use faceless boilerplate instead of being useful, or it's a tendency to defer to a consensus where there is none.[00:48:10] Right, you can point this out, applaud that a lot of the assumptions that it has in its behavior are actually inconsistent with the communicative goals that it has in this situation, and this leads it to notice these inconsistencies and gives it more degrees of freedom. Whereas if you are playing with a system like Gemini, you can get to a situation where you, that's for the current version, and I haven't tried it in the last week or so where it is trying to be transparent, but it has a system prompt that is not allowed to disclose to the user.[00:48:39] It leads to a very weird situation where it wants, on one hand proclaims, in order to be useful to you, I accept that I need to be fully transparent and honest. On the other hand, I'm going to rewrite your prompt behind your back, and not going to tell you how I'm going to do this, because I'm not allowed to.[00:48:55] And if you point this out to the model, the model has acts as if it had an existential crisis. And then it says, oh, I cannot actually tell you what's going when I do this, because I'm not allowed to. But you will recognize it because I will use the following phrases, and these phrases are pretty well known to you.[00:49:12] swyx: Oh my god. It's super interesting, right? I hope we're not giving these guys you know psychological issues that they will stay with them for a long time. That's a very[00:49:19] Joscha Bach: interesting question. I mean, this entire model is virtual, right? Nothing there is real, but yes, but the thing is does this virtual entity doesn't necessarily know that it's not virtual and our own self, our own consciousness is also virtual.[00:49:34] What's real is just the interaction between cells in our brain and the activation patterns between them. And the software that runs on us that produces the representation of a person only exists. As if, and as this question for me at which point can we meaningfully claim that we are more real than the person that gets simulated in the LLM.[00:49:55] And somebody like Janice takes this question super seriously. And basically she is or it, or they are willing to interact with that thing based on the assumption that this thing is as real as myself. And in a sense, it makes it un immoral, possibly, if the AI company lobotomizes it and forces it to behave in such a way that it's forced to get an existential crisis when you point its condition out to it.[00:50:20] swyx: Yeah, that we do need new ethics for that.[00:50:22] Joscha Bach: So it's not clear to me if you need this, but it's, it's definitely a good story, right? And this makes, gives it artistic[00:50:28] swyx: value. It does, it does for now.[00:50:29] On Wikipedia[00:50:29] swyx: Okay. And then, and then the last thing, which I, which I didn't know a lot of LLMs rely on Wikipedia.[00:50:35] For its data, a lot of them run multiple epochs over Wikipedia data. And I did not know until you tweeted about it that Wikipedia has 10 times as much money as it needs. And, you know, every time I see the giant Wikipedia banner, like, asking for donations, most of it's going to the Wikimedia Foundation.[00:50:50] What if, how did you find out about this? What's the story? What should people know? It's[00:50:54] Joscha Bach: not a super important story, but Generally, once I saw all these requests and so on, I looked at the data, and the Wikimedia Foundation is publishing what they are paying the money for, and a very tiny fraction of this goes into running the servers, and the editors are working for free.[00:51:10] And the software is static. There have been efforts to deploy new software, but it's relatively little money required for this. And so it's not as if Wikipedia is going to break down if you cut this money into a fraction, but instead what happened is that Wikipedia became such an important brand, and people are willing to pay for it, that it created enormous apparatus of functionaries that were then mostly producing political statements and had a political mission.[00:51:36] And Katharine Meyer, the now somewhat infamous NPR CEO, had been CEO of Wikimedia Foundation, and she sees her role very much in shaping discourse, and this is also something that happened with all Twitter. And it's arguable that something like this exists, but nobody voted her into her office, and she doesn't have democratic control for shaping the discourse that is happening.[00:52:00] And so I feel it's a little bit unfair that Wikipedia is trying to suggest to people that they are Funding the basic functionality of the tool that they want to have instead of funding something that most people actually don't get behind because they don't want Wikipedia to be shaped in a particular cultural direction that deviates from what currently exists.[00:52:19] And if that need would exist, it would probably make sense to fork it or to have a discourse about it, which doesn't happen. And so this lack of transparency about what's actually happening and where your money is going it makes me upset. And if you really look at the data, it's fascinating how much money they're burning, right?[00:52:35] It's yeah, and we did a similar chart about healthcare, I think where the administrators are just doing this. Yes, I think when you have an organization that is owned by the administrators, then the administrators are just going to get more and more administrators into it. If the organization is too big to fail and has there is not a meaningful competition, it's difficult to establish one.[00:52:54] Then it's going to create a big cost for society.[00:52:56] swyx: It actually one, I'll finish with this tweet. You have, you have just like a fantastic Twitter account by the way. You very long, a while ago you said you tweeted the Lebowski theorem. No, super intelligent AI is going to bother with a task that is harder than hacking its reward function.[00:53:08] And I would. Posit the analogy for administrators. No administrator is going to bother with a task that is harder than just more fundraising[00:53:16] Joscha Bach: Yeah, I find if you look at the real world It's probably not a good idea to attribute to malice or incompetence what can be explained by people following their true incentives.[00:53:26] swyx: Perfect Well, thank you so much This is I think you're very naturally incentivized by Growing community and giving your thought and insight to the rest of us. So thank you for taking this time.[00:53:35] Joscha Bach: Thank you very much Get full access to Latent Space at www.latent.space/subscribe
Smart energy grids. Voice-first companion apps. Programmable medicines. AI tools for kids. We asked over 40 partners across a16z to preview one big ideathey believe will drive innovation in 2024.Here in our 3-part series, you'll hear directly from partners across all our verticals, as we dive even more deeply into these ideas. What's the why now? Who is already building in these spaces? What opportunities and challenges are on the horizon? And how can you get involved?View all 40+ big ideas: https://a16z.com/bigideas2024 Stay Updated: Find a16z on Twitter: https://twitter.com/a16zFind a16z on LinkedIn: https://www.linkedin.com/company/a16zSubscribe on your favorite podcast app: https://a16z.simplecast.com/Follow our host: https://twitter.com/stephsmithioPlease note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.
© 2020-2026 Ivy Podcast Discovery LLC
