POPULARITY
https://youtu.be/jLEO9Gbba0kMatt and Sean talk about the scramble to develop AI that could undermine the very underpinnings of … well… everything. Don't panic. (Yet.)Watch the Undecided with Matt Ferrell episode, Why the AI Revolution Has a Fatal Flaw https://youtu.be/hBfhd88DCZA?list=PLnTSM-ORSgi7uzySCXq8VXhodHB5B5OiQ(00:00) - - Intro & Feedback (10:45) - - The AI Paradox Discussion YouTube version of the podcast: https://www.youtube.com/stilltbdpodcastGet in touch: https://undecidedmf.com/podcast-feedbackSupport the show: https://pod.fan/still-to-be-determinedFollow us on X: @stilltbdfm @byseanferrell @mattferrell or @undecidedmfUndecided with Matt Ferrell: https://www.youtube.com/undecidedmf ★ Support this podcast ★
Zoë, formerly the Senior Vice President of AI at Citi, is a distinguished leader in the AI field with over 12 years of experience. At Citi, she spearheaded the adoption of agentic AI across the firm, enhancing various business processes. Prior to this, as a Staff Applied Scientist at Etsy, she led initiatives in personalized search and discovery. A graduate of NYU with a deep background in AI, Zoë holds three patents in large language models and is an active angel investor in AI-first startups such as Tough Day, Starcycle, and Eden Labs. She is now the co-founder and CEO of Sequen AI, where she patented the first-of-their-kind scalable specialized agents that autonomously seek to maximize conversion rates. She has been a 3i member since 2024.In This Episode, Hear About:Sector Impact: Discover how AI is transforming sectors like banking, healthcare, and retail, setting the stage for future innovations and investment opportunities.Bridging the AI Ecosystem: Zoë shares insights from her experiences on how AI strategies vary between corporate giants and nimble startups across various fields.Emerging AI Trends: Zoë highlights the latest advancements in AI, including generative models, Deepseek, and agentic technologies, and their profound effects on the market.Learn more about 3i Members and follow us on LinkedIn for updates.
Jonathan Godwin, founder and CEO of Orbital Materials, alongside researcher Tim Duignan, discuss the transformative potential of AI in material science on the Cognitive Revolution podcast. They explore foundational concepts, the integration of computational simulations, and the development of new materials for various applications such as data centers and combating climate change. They also delve into the latest advancements, including a groundbreaking study on the potassium ion channel, and speculate on the future of AI in scientific discovery and material synthesis. Check out some of Tim's work: Google Colab to run you own simulation: https://colab.research.google.com/github/timduignan/orb-models/blob/main/examples/OrbMDTut.ipynb GitHub repository "Orb force fields": https://github.com/orbital-materials/orb-models Preprint "A potassium ion channel simulated with a universal neural network potential": https://arxiv.org/abs/2411.18931 Help shape our show by taking our quick listener survey at https://bit.ly/TurpentinePulse SPONSORS: 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 providers. OCI powers industry leaders like Vodafone and Thomson Reuters with secure infrastructure and application development capabilities. New U.S. customers can get their cloud bill cut in half by switching to OCI before March 31, 2024 at https://oracle.com/cognitive 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: Dreaming of starting your own business? Shopify makes it easier than ever. With customizable templates, shoppable social media posts, and their new AI sidekick, Shopify Magic, you can focus on creating great products while delegating the rest. Manage everything from shipping to payments in one place. Start your journey with a $1/month trial at https://shopify.com/cognitive and turn your 2025 dreams into reality. Vanta: Vanta simplifies security and compliance for businesses of all sizes. Automate compliance across 35+ frameworks like SOC 2 and ISO 27001, streamline security workflows, and complete questionnaires up to 5x faster. Trusted by over 9,000 companies, Vanta helps you manage risk and prove security in real time. Get $1,000 off at https://vanta.com/revolution CHAPTERS: (00:00) Teaser (01:05) About the Episode (05:10) Welcome to Orbital (06:15) Semiconductors (07:44) Material Science Today (09:22) Experimental Cycle (12:06) Orbital's Founding (14:51) AI in Materials (Part 1) (21:05) Sponsors: OCI | NetSuite (23:45) AI in Materials (Part 2) (35:00) Sponsors: Shopify | Vanta (38:15) Generative Models (38:16) Diffusion Models (50:50) Orbital Applications (58:19) Perfect Sponge (59:43) AI Simulations (01:01:27) Natural Language (01:02:35) Compute Needs (01:05:05) Human Electrical Nature (01:06:11) Potassium Channels (01:15:51) Scaling Simulations (01:23:56) Roadmap: Carbon Removal (01:30:37) AI & Job Satisfaction (01:36:14) LLMs & Potentials (01:37:19) AGI & Discovery (01:39:58) Outro
Connect with JillConnect with Will___160 Characters is powered by Clerk Chat.
Grok-2's advancements in speed and accuracy position it as a leading AI model, particularly in math and coding. OpenAI's backing of California's AI bill highlights the critical need for transparency in synthetic content, especially during an election year. The episode features groundbreaking research on the SwiftBrush diffusion model and K-Sort Arena for generative model evaluation. Additionally, the LlamaDuo pipeline offers a practical solution for migrating from cloud-based LLMs to local models, tackling privacy and operational challenges. Contact: sergi@earkind.com Timestamps: 00:34 Introduction 01:55 grok-2 is Faster and Better 03:32 OpenAI supports California AI bill requiring 'watermarking' of synthetic content 04:53 Fake sponsor 06:45 SwiftBrush v2: Make Your One-step Diffusion Model Better Than Its Teacher 08:10 SWE-bench-java: A GitHub Issue Resolving Benchmark for Java 09:40 K-Sort Arena: Efficient and Reliable Benchmarking for Generative Models via K-wise Human Preferences 11:24 LlamaDuo: LLMOps Pipeline for Seamless Migration from Service LLMs to Small-Scale Local LLMs 13:26 Outro
In episode 23 of Recsperts, we welcome Yashar Deldjoo, Assistant Professor at the Polytechnic University of Bari, Italy. Yashar's research on recommender systems includes multimodal approaches, multimedia recommender systems as well as trustworthiness and adversarial robustness, where he has published a lot of work. We discuss the evolution of generative models for recommender systems, modeling paradigms, scenarios as well as their evaluation, risks and harms.We begin our interview with a reflection of Yashar's areas of recommender systems research so far. Starting with multimedia recsys, particularly video recommendations, Yashar covers his work around adversarial robustness and trustworthiness leading to the main topic for this episode: generative models for recommender systems. We learn about their aspects for improving beyond the (partially saturated) state of traditional recommender systems: improve effectiveness and efficiency for top-n recommendations, introduce interactivity beyond classical conversational recsys, provide personalized zero- or few-shot recommendations.We learn about the modeling paradigms and as well about the scenarios for generative models which mainly differ by input and modelling approach: ID-based, text-based, and multimodal generative models. This is how we navigate the large field of acronyms leading us from VAEs and GANs to LLMs.Towards the end of the episode, we also touch on the evaluation, opportunities, risks and harms of generative models for recommender systems. Yashar also provides us with an ample amount of references and upcoming events where people get the chance to know more about GenRecSys.Enjoy this enriching episode of RECSPERTS - Recommender Systems Experts.Don't forget to follow the podcast and please leave a review(00:00) - Introduction (03:58) - About Yashar Deldjoo (09:34) - Motivation for RecSys (13:05) - Intro to Generative Models for Recommender Systems (44:27) - Modeling Paradigms for Generative Models (51:33) - Scenario 1: Interaction-Driven Recommendation (57:59) - Scenario 2: Text-based Recommendation (01:10:39) - Scenario 3: Multimodal Recommendation (01:24:59) - Evaluation of Impact and Harm (01:38:07) - Further Research Challenges (01:45:03) - References and Research Advice (01:49:39) - Closing Remarks Links from the Episode:Yashar Deldjoo on LinkedInYashar's WebsiteKDD 2024 Tutorial: Modern Recommender Systems Leveraging Generative AI: Fundamentals, Challenges and OpportunitiesRecSys 2024 Workshop: The 1st Workshop on Risks, Opportunities, and Evaluation of Generative Models in Recommender Systems (ROEGEN@RECSYS'24)Papers:Deldjoo et al. (2024): A Review of Modern Recommender Systems Using Generative Models (Gen-RecSys)Deldjoo et al. (2020): Recommender Systems Leveraging Multimedia ContentDeldjoo et al. (2021): A Survey on Adversarial Recommender Systems: From Attack/Defense Strategies to Generative Adversarial NetworksDeldjoo et al. (2020): How Dataset Characteristics Affect the Robustness of Collaborative Recommendation ModelsLiang et al. (2018): Variational Autoencoders for Collaborative FilteringHe et al. (2016): Visual Bayesian Personalized Ranking from Implicit FeedbackGeneral Links:Follow me on LinkedInFollow me on XSend me your comments, questions and suggestions to marcel.kurovski@gmail.comRecsperts Website
In this episode, we will discuss news on Apple Intelligence and its ability to deliver intelligence that is helpful and relevant. It will be the personal intelligence system that puts powerful generative models at the core of iPhone, iPad, and Mac. With Priority Messages in Mail, a new section at the top of the inbox … Continue reading Episode 259, Introducing Apple Intelligence, the personal intelligence system that puts powerful generative models at the core of iPhone, iPad, and Mac →
Hat Tip to this week's creators: @leopoldasch, @JoeSlater87, @GaryMarcus, @ulonnaya, @alex, @ttunguz, @mmasnick, @dannyrimer, @imdavidpierce, @asafitch, @ylecun, @nxthompson, @kaifulee, @DaphneKoller, @AndrewYNg, @aidangomez, @Kyle_L_Wiggers, @waynema, @QianerLiu, @nicnewman, @nmasc_, @steph_palazzolo, @nofilmschoolContents* Editorial: * Essays of the Week* Situational Awareness: The Decade Ahead* ChatGPT is b******t* AGI by 2027?* Ilya Sutskever, OpenAI's former chief scientist, launches new AI company* The Series A Crunch Is No Joke* The Series A Crunch or the Seedpocalypse of 2024 * The Surgeon General Is Wrong. Social Media Doesn't Need Warning Labels* Video of the Week* Danny Rimer on 20VC - (Must See)* AI of the Week* Anthropic has a fast new AI model — and a clever new way to interact with chatbots* Nvidia's Ascent to Most Valuable Company Has Echoes of Dot-Com Boom* The Expanding Universe of Generative Models* DeepMind's new AI generates soundtracks and dialogue for videos* News Of the Week* Apple Suspends Work on Next Vision Pro, Focused on Releasing Cheaper Model in Late 2025* Is the news industry ready for another pivot to video?* Cerebras, an Nvidia Challenger, Files for IPO Confidentially* Startup of the Week* Final Cut Camera and iPad Multicam are Truly Revolutionary* X of the Week* Leopold AschenbrennerEditorialI had not heard of Leopold Aschenbrenner until yesterday. I was meeting with Faraj Aalaei (a SignalRank board member) and my colleague Rob Hodgkinson when they began to talk about “Situational Awareness,” his essay on the future of AGI, and its likely speed of emergence.So I had to read it, and it is this week's essay of the week. He starts his 165-page epic with:Before long, the world will wake up. But right now, there are perhaps a few hundred people, most of them in San Francisco and the AI labs, that have situational awareness. Through whatever peculiar forces of fate, I have found myself amongst them.So, Leopold is not humble. He finds himself “among” the few people with situational awareness.As a person prone to bigging up myself, I am not one to prematurely judge somebody's view of self. So, I read all 165 pages.He makes one point. The growth of AI capability is accelerating. More is being done at a lower cost, and the trend will continue to be super-intelligence by 2027. At that point, billions of skilled bots will solve problems at a rate we cannot imagine. And they will work together, with little human input, to do so.His case is developed using linear progression from current developments. According to Leopold, all you have to believe in is straight lines.He also has a secondary narrative related to safety, particularly the safety of models and their weightings (how they achieve their results).By safety, he does not mean the models will do bad things. He means that third parties, namely China, can steal the weightings and reproduce the results. He focuses on the poor security surrounding models as the problem. And he deems governments unaware of the dangers.Although German-born, he argues in favor of the US-led effort to see AGI as a weapon to defeat China and threatens dire consequences if it does not. He sees the “free world” as in danger unless it stops others from gaining the sophistication he predicts in the time he predicts.At that point, I felt I was reading a manifesto for World War Three.But as I see it, the smartest people in the space have converged on a different perspective, a third way, one I will dub AGI Realism. The core tenets are simple:* Superintelligence is a matter of national security. We are rapidly building machines smarter than the smartest humans. This is not another cool Silicon Valley boom; this isn't some random community of coders writing an innocent open source software package; this isn't fun and games. Superintelligence is going to be wild; it will be the most powerful weapon mankind has ever built. And for any of us involved, it'll be the most important thing we ever do. * America must lead. The torch of liberty will not survive Xi getting AGI first. (And, realistically, American leadership is the only path to safe AGI, too.) That means we can't simply “pause”; it means we need to rapidly scale up US power production to build the AGI clusters in the US. But it also means amateur startup security delivering the nuclear secrets to the CCP won't cut it anymore, and it means the core AGI infrastructure must be controlled by America, not some dictator in the Middle East. American AI labs must put the national interest first. * We need to not screw it up. Recognizing the power of superintelligence also means recognizing its peril. There are very real safety risks; very real risks this all goes awry—whether it be because mankind uses the destructive power brought forth for our mutual annihilation, or because, yes, the alien species we're summoning is one we cannot yet fully control. These are manageable—but improvising won't cut it. Navigating these perils will require good people bringing a level of seriousness to the table that has not yet been offered. As the acceleration intensifies, I only expect the discourse to get more shrill. But my greatest hope is that there will be those who feel the weight of what is coming, and take it as a solemn call to duty.I persisted in reading it, and I think you should, too—not for the war-mongering element but for the core acceleration thesis.My two cents: Leopold underestimates AI's impact in the long run and overestimates it in the short term, but he is directionally correct.Anthropic released v3.5 of Claude.ai today. It is far faster than the impressive 3.0 version (released a few months ago) and costs a fraction to train and run. it is also more capable. It accepts text and images and has a new feature that allows it to run code, edit documents, and preview designs called ‘Artifacts.'Claude 3.5 Opus is probably not far away.Situational Awareness projects trends like this into the near future, and his views are extrapolated from that perspective.Contrast that paper with “ChatGPT is B******t,” a paper coming out of Glasgow University in the UK. The three authors contest the accusation that ChatGPT hallucinates or lies. They claim that because it is a probabilistic word finder, it spouts b******t. It can be right, and it can be wrong, but it does not know the difference. It's a bullshitter.Hilariously, they define three types of BS:B******t (general)Any utterance produced where a speaker has indifference towards the truth of the utterance.Hard b******tB******t produced with the intention to mislead the audience about the utterer's agenda.Soft b******tB******t produced without the intention to mislead the hearer regarding the utterer's agenda.They then conclude:With this distinction in hand, we're now in a position to consider a worry of the following sort: Is ChatGPT hard b**********g, soft b**********g, or neither? We will argue, first, that ChatGPT, and other LLMs, are clearly soft b**********g. However, the question of whether these chatbots are hard b**********g is a trickier one, and depends on a number of complex questions concerning whether ChatGPT can be ascribed intentions.This is closer to Gary Marcus's point of view in his ‘AGI by 2027?' response to Leopold. It is also below.I think the reality is somewhere between Leopold and Marcus. AI is capable of surprising things, given that it is only a probabilistic word-finder. And its ability to do so is becoming cheaper and faster. The number of times it is useful easily outweighs, for me, the times it is not. Most importantly, AI agents will work together to improve each other and learn faster.However, Gary Marcus is right that reasoning and other essential decision-making characteristics are not logically derived from an LLM approach to knowledge. So, without additional or perhaps different elements, there will be limits to where it can go. Gary probably underestimates what CAN be achieved with LLMs (indeed, who would have thought they could do what they already do). And Leopold probably overestimates the lack of a ceiling in what they will do and how fast that will happen.It will be fascinating to watch. I, for one, have no idea what to expect except the unexpected. OpenAI Founder Illya Sutskever weighed in, too, with a new AI startup called Safe Superintelligence Inc. (SSI). The most important word here is superintelligence, the same word Leopold used. The next phase is focused on higher-than-human intelligence, which can be reproduced billions of times to create scaled Superintelligence.The Expanding Universe of Generative Models piece below places smart people in the room to discuss these developments. Yann LeCun, Nicholas Thompson, Kai-Fu Lee, Daphne Koller, Andrew Ng, and Aidan Gomez are participants. This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.thatwastheweek.com/subscribe
TikTok is expanding its Symphony ad suite with AI-generated avatars of creators and paid actors, as well as a global translation tool for multi-language support. NVIDIA has released an open synthetic data generation pipeline for training large language models, which could benefit industries that rely on natural language processing. Cohere's latest generative models, Command R and R+, can automate and streamline complex business workflows, saving time and increasing efficiency. XLand-100B is a large-scale dataset for in-context reinforcement learning, providing a challenging benchmark for researchers in the field. CountGen addresses the challenge of controlling the number of depicted objects in text-to-image generation, while MM-NIAH is the first benchmark specifically designed to test the comprehension abilities of existing multimodal large language models. Contact: sergi@earkind.com Timestamps: 00:34 Introduction 01:23 TikTok ads may soon contain AI-generated avatars of your favorite creators 02:59 NVIDIA Releases Open Synthetic Data Generation Pipeline for Training Large Language Models 04:43 Automating Complex Business Workflows with Cohere: Multi-Step Tool Use in Action 06:17 Fake sponsor 08:22 XLand-100B: A Large-Scale Multi-Task Dataset for In-Context Reinforcement Learning 10:23 Make It Count: Text-to-Image Generation with an Accurate Number of Objects 11:58 Needle In A Multimodal Haystack 13:37 Outro
Our second wave of speakers for AI Engineer World's Fair were announced! The conference sold out of Platinum/Gold/Silver sponsors and Early Bird tickets! See our Microsoft episode for more info and buy now with code LATENTSPACE.This episode is straightforwardly a part 2 to our ICLR 2024 Part 1 episode, so without further ado, we'll just get right on with it!Timestamps[00:03:43] Section A: Code Edits and Sandboxes, OpenDevin, and Academia vs Industry — ft. Graham Neubig and Aman Sanger* [00:07:44] WebArena* [00:18:45] Sotopia* [00:24:00] Performance Improving Code Edits* [00:29:39] OpenDevin* [00:47:40] Industry and Academia[01:05:29] Section B: Benchmarks* [01:05:52] SWEBench* [01:17:05] SWEBench/SWEAgent Interview* [01:27:40] Dataset Contamination Detection* [01:39:20] GAIA Benchmark* [01:49:18] Moritz Hart - Science of Benchmarks[02:36:32] Section C: Reasoning and Post-Training* [02:37:41] Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection* [02:51:00] Let's Verify Step By Step* [02:57:04] Noam Brown* [03:07:43] Lilian Weng - Towards Safe AGI* [03:36:56] A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis* [03:48:43] MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework[04:00:51] Bonus: Notable Related Papers on LLM CapabilitiesSection A: Code Edits and Sandboxes, OpenDevin, and Academia vs Industry — ft. Graham Neubig and Aman Sanger* Guests* Graham Neubig* Aman Sanger - Previous guest and NeurIPS friend of the pod!* WebArena * * Sotopia (spotlight paper, website)* * Learning Performance-Improving Code Edits* OpenDevin* Junyang Opendevin* Morph Labs, Jesse Han* SWE-Bench* SWE-Agent* Aman tweet on swebench* LiteLLM* Livecodebench* the role of code in reasoning* Language Models of Code are Few-Shot Commonsense Learners* Industry vs academia* the matryoshka embeddings incident* other directions* UnlimiformerSection A timestamps* [00:00:00] Introduction to Guests and the Impromptu Nature of the Podcast* [00:00:45] Graham's Experience in Japan and Transition into Teaching NLP* [00:01:25] Discussion on What Constitutes a Good Experience for Students in NLP Courses* [00:02:22] The Relevance and Teaching of Older NLP Techniques Like Ngram Language Models* [00:03:38] Speculative Decoding and the Comeback of Ngram Models* [00:04:16] Introduction to WebArena and Zotopia Projects* [00:05:19] Deep Dive into the WebArena Project and Benchmarking* [00:08:17] Performance Improvements in WebArena Using GPT-4* [00:09:39] Human Performance on WebArena Tasks and Challenges in Evaluation* [00:11:04] Follow-up Work from WebArena and Focus on Web Browsing as a Benchmark* [00:12:11] Direct Interaction vs. Using APIs in Web-Based Tasks* [00:13:29] Challenges in Base Models for WebArena and the Potential of Visual Models* [00:15:33] Introduction to Zootopia and Exploring Social Interactions with Language Models* [00:16:29] Different Types of Social Situations Modeled in Zootopia* [00:17:34] Evaluation of Language Models in Social Simulations* [00:20:41] Introduction to Performance-Improving Code Edits Project* [00:26:28] Discussion on DevIn and the Future of Coding Agents* [00:32:01] Planning in Coding Agents and the Development of OpenDevon* [00:38:34] The Changing Role of Academia in the Context of Large Language Models* [00:44:44] The Changing Nature of Industry and Academia Collaboration* [00:54:07] Update on NLP Course Syllabus and Teaching about Large Language Models* [01:00:40] Call to Action: Contributions to OpenDevon and Open Source AI Projects* [01:01:56] Hiring at Cursor for Roles in Code Generation and Assistive Coding* [01:02:12] Promotion of the AI Engineer ConferenceSection B: Benchmarks * Carlos Jimenez & John Yang (Princeton) et al: SWE-bench: Can Language Models Resolve Real-world Github Issues? (ICLR Oral, Paper, website)* “We introduce SWE-bench, an evaluation framework consisting of 2,294 software engineering problems drawn from real GitHub issues and corresponding pull requests across 12 popular Python repositories. Given a codebase along with a description of an issue to be resolved, a language model is tasked with editing the codebase to address the issue. Resolving issues in SWE-bench frequently requires understanding and coordinating changes across multiple functions, classes, and even files simultaneously, calling for models to interact with execution environments, process extremely long contexts and perform complex reasoning that goes far beyond traditional code generation tasks. Our evaluations show that both state-of-the-art proprietary models and our fine-tuned model SWE-Llama can resolve only the simplest issues. The best-performing model, Claude 2, is able to solve a mere 1.96% of the issues. Advances on SWE-bench represent steps towards LMs that are more practical, intelligent, and autonomous.”* Yonatan Oren et al (Stanford): Proving Test Set Contamination in Black-Box Language Models (ICLR Oral, paper, aman tweet on swebench contamination)* “We show that it is possible to provide provable guarantees of test set contamination in language models without access to pretraining data or model weights. Our approach leverages the fact that when there is no data contamination, all orderings of an exchangeable benchmark should be equally likely. In contrast, the tendency for language models to memorize example order means that a contaminated language model will find certain canonical orderings to be much more likely than others. Our test flags potential contamination whenever the likelihood of a canonically ordered benchmark dataset is significantly higher than the likelihood after shuffling the examples. * We demonstrate that our procedure is sensitive enough to reliably prove test set contamination in challenging situations, including models as small as 1.4 billion parameters, on small test sets of only 1000 examples, and datasets that appear only a few times in the pretraining corpus.”* Outstanding Paper mention: “A simple yet elegant method to test whether a supervised-learning dataset has been included in LLM training.”* Thomas Scialom (Meta AI-FAIR w/ Yann LeCun): GAIA: A Benchmark for General AI Assistants (paper)* “We introduce GAIA, a benchmark for General AI Assistants that, if solved, would represent a milestone in AI research. GAIA proposes real-world questions that require a set of fundamental abilities such as reasoning, multi-modality handling, web browsing, and generally tool-use proficiency. * GAIA questions are conceptually simple for humans yet challenging for most advanced AIs: we show that human respondents obtain 92% vs. 15% for GPT-4 equipped with plugins. * GAIA's philosophy departs from the current trend in AI benchmarks suggesting to target tasks that are ever more difficult for humans. We posit that the advent of Artificial General Intelligence (AGI) hinges on a system's capability to exhibit similar robustness as the average human does on such questions. Using GAIA's methodology, we devise 466 questions and their answer.* * Mortiz Hardt (Max Planck Institute): The emerging science of benchmarks (ICLR stream)* “Benchmarks are the keystone that hold the machine learning community together. Growing as a research paradigm since the 1980s, there's much we've done with them, but little we know about them. In this talk, I will trace the rudiments of an emerging science of benchmarks through selected empirical and theoretical observations. Specifically, we'll discuss the role of annotator errors, external validity of model rankings, and the promise of multi-task benchmarks. The results in each case challenge conventional wisdom and underscore the benefits of developing a science of benchmarks.”Section C: Reasoning and Post-Training* Akari Asai (UW) et al: Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (ICLR oral, website)* (Bad RAG implementations) indiscriminately retrieving and incorporating a fixed number of retrieved passages, regardless of whether retrieval is necessary, or passages are relevant, diminishes LM versatility or can lead to unhelpful response generation. * We introduce a new framework called Self-Reflective Retrieval-Augmented Generation (Self-RAG) that enhances an LM's quality and factuality through retrieval and self-reflection. * Our framework trains a single arbitrary LM that adaptively retrieves passages on-demand, and generates and reflects on retrieved passages and its generations using special tokens, called reflection tokens. Generating reflection tokens makes the LM controllable during the inference phase, enabling it to tailor its behavior to diverse task requirements. * Self-RAG (7B and 13B parameters) outperforms ChatGPT and retrieval-augmented Llama2-chat on Open-domain QA, reasoning, and fact verification tasks, and it shows significant gains in improving factuality and citation accuracy for long-form generations relative to these models. * Hunter Lightman (OpenAI): Let's Verify Step By Step (paper)* “Even state-of-the-art models still regularly produce logical mistakes. To train more reliable models, we can turn either to outcome supervision, which provides feedback for a final result, or process supervision, which provides feedback for each intermediate reasoning step. * We conduct our own investigation, finding that process supervision significantly outperforms outcome supervision for training models to solve problems from the challenging MATH dataset. Our process-supervised model solves 78% of problems from a representative subset of the MATH test set. Additionally, we show that active learning significantly improves the efficacy of process supervision. * To support related research, we also release PRM800K, the complete dataset of 800,000 step-level human feedback labels used to train our best reward model.* * Noam Brown - workshop on Generative Models for Decision Making* Solving Quantitative Reasoning Problems with Language Models (Minerva paper)* Describes some charts taken directly from the Let's Verify Step By Step paper listed/screenshotted above.* Lilian Weng (OpenAI) - Towards Safe AGI (ICLR talk)* OpenAI Model Spec* OpenAI Instruction Hierarchy: The Instruction Hierarchy: Training LLMs to Prioritize Privileged InstructionsSection D: Agent Systems* Izzeddin Gur (Google DeepMind): A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis (ICLR oral, paper)* [Agent] performance on real-world websites has still suffered from (1) open domainness, (2) limited context length, and (3) lack of inductive bias on HTML.* We introduce WebAgent, an LLM-driven agent that learns from self-experience to complete tasks on real websites following natural language instructions.* WebAgent plans ahead by decomposing instructions into canonical sub-instructions, summarizes long HTML documents into task-relevant snippets, and acts on websites via Python programs generated from those.* We design WebAgent with Flan-U-PaLM, for grounded code generation, and HTML-T5, new pre-trained LLMs for long HTML documents using local and global attention mechanisms and a mixture of long-span denoising objectives, for planning and summarization.* We empirically demonstrate that our modular recipe improves the success on real websites by over 50%, and that HTML-T5 is the best model to solve various HTML understanding tasks; achieving 18.7% higher success rate than the prior method on MiniWoB web automation benchmark, and SoTA performance on Mind2Web, an offline task planning evaluation.* Sirui Hong (DeepWisdom): MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework (ICLR Oral, Paper)* We introduce MetaGPT, an innovative meta-programming framework incorporating efficient human workflows into LLM-based multi-agent collaborations. MetaGPT encodes Standardized Operating Procedures (SOPs) into prompt sequences for more streamlined workflows, thus allowing agents with human-like domain expertise to verify intermediate results and reduce errors. MetaGPT utilizes an assembly line paradigm to assign diverse roles to various agents, efficiently breaking down complex tasks into subtasks involving many agents working together. Bonus: Notable Related Papers on LLM CapabilitiesThis includes a bunch of papers we wanted to feature above but could not.* Lukas Berglund (Vanderbilt) et al: The Reversal Curse: LLMs trained on “A is B” fail to learn “B is A” (ICLR poster, paper, Github)* We expose a surprising failure of generalization in auto-regressive large language models (LLMs). If a model is trained on a sentence of the form ''A is B'', it will not automatically generalize to the reverse direction ''B is A''. This is the Reversal Curse. * The Reversal Curse is robust across model sizes and model families and is not alleviated by data augmentation. We also evaluate ChatGPT (GPT-3.5 and GPT-4) on questions about real-world celebrities, such as ''Who is Tom Cruise's mother? [A: Mary Lee Pfeiffer]'' and the reverse ''Who is Mary Lee Pfeiffer's son?''. GPT-4 correctly answers questions like the former 79% of the time, compared to 33% for the latter.* * Omar Khattab (Stanford): DSPy: Compiling Declarative Language Model Calls into State-of-the-Art Pipelines (ICLR Spotlight Poster, GitHub)* presented by Krista Opsahl-Ong* “Existing LM pipelines are typically implemented using hard-coded “prompt templates”, i.e. lengthy strings discovered via trial and error. Toward a more systematic approach for developing and optimizing LM pipelines, we introduce DSPy, a programming model that abstracts LM pipelines as text transformation graphs, or imperative computational graphs where LMs are invoked through declarative modules. * DSPy modules are parameterized, meaning they can learn how to apply compositions of prompting, finetuning, augmentation, and reasoning techniques. * We design a compiler that will optimize any DSPy pipeline to maximize a given metric, by creating and collecting demonstrations. * We conduct two case studies, showing that succinct DSPy programs can express and optimize pipelines that reason about math word problems, tackle multi-hop retrieval, answer complex questions, and control agent loops. * Within minutes of compiling, DSPy can automatically produce pipelines that outperform out-of-the-box few-shot prompting as well as expert-created demonstrations for GPT-3.5 and Llama2-13b-chat. On top of that, DSPy programs compiled for relatively small LMs like 770M parameter T5 and Llama2-13b-chat are competitive with many approaches that rely on large and proprietary LMs like GPT-3.5 and on expert-written prompt chains. * * MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning* Scaling Laws for Associative Memories * DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models* Efficient Streaming Language Models with Attention Sinks Get full access to Latent Space at www.latent.space/subscribe
https://youtu.be/h4BKiYFZjQYOn today's episode of Still To Be Determined we're talking about Generative AI. It's a little known topic, so maybe we're the first place you're hearing about it. Kidding aside … should we be freaking out about OpenAI?Watch the Undecided with Matt Ferrell episode, AI Just Changed Everything … Again https://youtu.be/q9_BY2QsK1Y?list=PLnTSM-ORSgi4At-R_1s6-_50PCbYsoEcjYouTube version of the podcast: https://www.youtube.com/stilltbdpodcastGet in touch: https://undecidedmf.com/podcast-feedbackSupport the show: https://pod.fan/still-to-be-determinedFollow us on X: @stilltbdfm @byseanferrell @mattferrell or @undecidedmfUndecided with Matt Ferrell: https://www.youtube.com/undecidedmf ★ Support this podcast ★
Six Reasons Why Matterport Will Launch Project Genesis Monday (3 June 2024) Commentary and analysis by Dan Smigrod Founder and Managing Editor We Get Around Network Forum and WGAN-TV Podcast Hi All, Matterport will likely announce on Monday, 3 June 2024 that Project Genesis is now commercially available to all. If that's true, what I wrote about a year ago (and reprinted below for your convenience), becomes relevant and timely: ✓ Matterport Genesis Generative AI Magic: Will it Make Competitors Irrelevant? "Project Genesis" was first announced 14 June 2023 that it was in development, would include these super-powerful features: ✓ decluttering of Matterport digital twins (residential and commercial spaces) ✓ remove objects ✓ virtual staging of Matterport spaces using AI (easy prompts) (resident and commercial spaces) ✓ space planning of commercial spaces "How many people can fit in an open floor plan design?" (Project Genesis likely contributed to CoStar's plans to acquire Matterport by the end of 2024). Matterport CEO RJ Pittman in this video released on 14 February 2024 said: And when it comes to remodeling, we are especially excited to talk about the future. Matterport has become synonymous with digitizing and documenting physical spaces exactly as they exist today. But later in 2024, Matterport will enable you to reimagine what any space could become. We call it Project Genesis, and it incorporates Matterport's decade of machine learning and AI experience with breakthrough generative AI to bring new design and furnishing ideas to life at the click of a button. First, just imagine being able to instantly defurnish an existing space to see it as a blank canvas with all the dimensions and spatial data in place, a photorealistic 3D as-built digital twin. Or maybe you'd like to be more selective and remove unwanted objects individually, perhaps to clear out the clutter. Ultimately, spatial visualization is Matterport's greatest strength, which is why we're so focused on creating breakthrough tools to let you visualize your space with different interior design styles and furnishings automatically in true 3D. With our 2024 Winter Release, we're thrilled to unveil the next generation of digital twins, packed with fresh insights, automations, integrations and plugins, with more on the way in our next release coming in the second half of the year [2024]. Here are the six (6) "tea leaves" that I am reading to make this prediction: 1. Shopping Cart: Matterport is updating its shopping cart now (noon- 3 pm ET) today (Sunday, 2 June 2024). Many major announcements by Matterport have been preceded by "Scheduled Maintenance) 2. Quarterly Release Cycle: Matterport signaled that it wanted to be on a quarterly release cycle, when it announced its Winter 2024 Release on 15 February 2024. Four months from then will be June 15, 2024: just 12 days from today. (Matterport in this 15 February 2024 video says "first in a series" ... "New features coming to market" ... "more in our next release coming in the second half of 2024." 3. No Matterport News Since 15 May 2024: Matterport tends to issue a media release weekly (whether it truly has news or rehashes in a media release something that was previously announced (in YouTube videos, blog post, case study, etc.) Almost every time that Matterport has announced something big, there have been no other media releases in the previous two weeks. I could imagine that Matterport has its marketing communications team focused on the news announcement for at least two weeks leading up to major announcements. 4. Matterport Blog Post Cadence Slowed: In recent weeks, Matterport blog posts have almost been weekly. The last Matterport Blog post featuring news was on 24 May 2024. Almost every time that Matterport has announced something big, there was no Matterport blog post in the preceding week. Again, I could imagine that Matterport has its marketing communications team focused on creating a blog post for the big news. 5. Matterport Abstract Published: In early May 2024 (Thank you @Fathom3D), Matterport published: Matterport Abstract: Automatic Defurnishing of Indoor 360 Panoramas on its GitHub. While the audience for the Matterport Abstract are other developers in the computer vision space, I could imagine that Matterport would have held off this publication until after its big announcement the week of June 3rd, EXCEPT the Matterport Abstract has been accepted for two Workshops that needed to begin promoting their 18 June 2024 and 17-21 June 2024 events respectively: ✓ Workshop on Generative Models for Computer Vision at CVPR 2024 (18 June 2024) ✓ Workshop on Computer Vision in the Built Environment at CVPR 2024 [17-21 June 2024) 6. Previously Announced News (Coming Soon) is Announced Again (When Available): Matterport has frequently announced coming soon features (beta release) and then announced it again when the features are commercially available. It's been a year since Matterport first announced Project Genesis on 14 June 2023. Matterport announces major news on Mondays (such as CoStar's plans to acquire Matterport). === If I am right about Project Genesis launching Monday, June 3, 2024), then what I wrote on 14 June 2023 in the We Get Around Network Forum (and reprinted below for your convenience), is timely and relevant): --- Continued in the We Get Around Network Forum: https://www.wegetaroundnetwork.com/topic/19908/page/1/6-reasons-why-matterport-will-launch-project-genesis-monday-6-june-2024/
The use of AI in military decision-making is a complex and controversial topic, as highlighted in the discussion of the Lavender system used by the Israeli army in Gaza. The SWE-agent and IPEX-LLM projects showcase exciting advancements in the optimization of language models for software engineering and on-device use, respectively. The "Measuring Style Similarity in Diffusion Models" paper introduces a framework for understanding and extracting style descriptors from images in generative models, with potential applications for artists and designers. "Octopus v2: On-device language model for super agent" presents a new on-device language model that can call functions and perform tasks related to automatic workflow, with practical applications in creating AI agents for edge devices. Contact: sergi@earkind.com Timestamps: 00:34 Introduction 01:52 ‘Lavender': The AI machine directing Israel's bombing spree in Gaza 04:01 Open Source SWE Agent 05:17 IPEX-LLM 06:58 Fake sponsor 08:39 Measuring Style Similarity in Diffusion Models 10:11 Advancing LLM Reasoning Generalists with Preference Trees 12:11 Octopus v2: On-device language model for super agent 14:05 Outro
Join us as we honor the 2023 Visionary and Venture of the Year. Nominated and voted on by their peers, these are the change agents embracing new opportunities to drive outsized results. They are intent on blazing new paths, discovering inventive solutions, and powering the future. Visionary of the Year: Marty Guay, Vice President of Business Development, Stanley Black & Decker With over three decades experience as a corporate innovator, Marty Guay has spent the past six years working with startups, academia, venture capital and global corporations to create business value opportunities for Stanley Black & Decker, a worldwide leader in Tools and Outdoor. At Stanley Black & Decker, Guay is responsible for the inception of the company's accelerator program where he successfully connects startups with technological and commercial relationships inside the organization and throughout global ecosystems. Guay sits on the advisory boards of multiple startups and serves as a board member of the state of Connecticut Governor's Workforce Council and as a liaison to the Business Roundtable. In addition to these roles, Guay also oversees the company's global philanthropy program that is focused on closing the skilled trades gap through its Grow the Trades program. Venture of the Year: Inworld AI, represented by Co-founder, Kylan Gibbs Inworld AI, led by a team of conversational AI pioneers and experts, uses advanced AI to build interactive characters whose personalities, thoughts, memories, and behaviors are designed to emulate social human interaction. Founded in 2021, Inworld is a recent graduate of the Comcast NBCUniversal LIFT Labs Generative AI Accelerator run in partnership with VentureFuel. Last month, Inworld announced a multi-year partnership with Microsoft to develop AI toolsets to assist and empower creators in dialogue, story, and quest design. Kylan Gibbs has a background in turning advanced AI into products that naturally interact with humans. He came from DeepMind, where he led products for Conversational AI and Generative Models. Prior, Kylan worked on enterprise Machine Learning projects at Bain & Company. Kylan studied at the University of Cambridge and McGill University.
Join host Craig Smith on episode #166 of Eye on AI as we sit down with Itamar Arel, CEO of Tenyx, a company that uses proprietary neuroscience-inspired AI technology to build the next generation of voice-based conversational agents. Itamar shares his journey that started with academic research to becoming a leading tech entrepreneur with Tenyx. We explore the evolution of voice AI in customer service and the unique challenges and advancements in understanding and responding to human speech. We dig deeper into Tenyx's unique approach to AI-driven customer service while exploring the production and design considerations when developing cutting-edge AI technology. Make sure you watch till the end as Itamar shares his vision on how AI is going to reshape industries and help advance modern day businesses. Enjoyed this conversation? Make sure you like, comment and share for more fascinating discussions in the world of AI. This episode is sponsored by Shopify. Shopify is a commerce platform that allows anyone to set up an online store and sell their products. Whether you're selling online, on social media, or in person, Shopify has you covered on every base. With Shopify you can sell physical and digital products. You can sell services, memberships, ticketed events, rentals and even classes and lessons. Sign up for a $1 per month trial period at http://shopify.com/eyeonai Stay Updated: Craig Smith Twitter: https://twitter.com/craigss Eye on A.I. Twitter: https://twitter.com/EyeOn_AI (00:00) Preview and Introduction (03:35) Itamar Arel's Career and Introduction to Voice AI Development (07:39) Key Differences in Current and Past Technology Solutions (09:10) Advancements in Voice AI and Large Language Models (11:00) The Inception and Evolution of Tenyx (13:29) Challenges in Developing Voice AI (18:27) Innovative Approaches in Voice AI Development (21:05) Data Handling and Fine-Tuning in Model Development (25;41) How To Standout In The Crowded AI Market (29:44) The Future of Voice AI and Generative Models (37:14) Testing and Evaluation of Voice AI Systems (40:10) Where Will AI Be in 5 Years? (43:42) Closing Remarks and A Word From Our Sponsors
Generative AI is advancing exponentially. What is happening at the frontier of research and application and how are novel techniques and approaches changing the risks and opportunities linked to frontier, generative AI models? This is the full audio from a session at the World Economic Forum's Annual Meeting 2024. Speakers: Yann LeCun, Silver Professor of Data Science, Computer Science, Neural Science and Electrical Engineering, New York University Nicholas Thompson, Chief Executive Officer, The Atlantic Kai-Fu Lee, Founder, 01.AI Pte. Ltd. Daphne Koller, Founder and Chief Executive Officer, Insitro Inc Andrew Ng, Founder, DeepLearning.AI, LLC Aidan Gomez, Co-Founder and Chief Executive Officer, Cohere Inc. Watch the session here: https://www.weforum.org/events/world-economic-forum-annual-meeting-2024/sessions/the-expanding-universe-of-generative-models Follow all the action from Davos at wef.ch/wef24 and across social media using the hashtag #WEF24. Check out all our podcasts on wef.ch/podcasts: YouTube: https://www.youtube.com/@wef Radio Davos - subscribe: https://pod.link/1504682164 Meet the Leader - subscribe: https://pod.link/1534915560 Agenda Dialogues - subscribe: https://pod.link/1574956552 World Economic Forum Book Club Podcast - subscribe: https://pod.link/1599305768 Join the World Economic Forum Podcast Club: https://www.facebook.com/groups/wefpodcastclub https://www.weforum.org/podcasts/radio-davos/episodes/
Hey readers
Denise Holt reports from the 4th Annual International Workshop on Active Inference (IWAI), September 2023, where the highlight of the event was a 1.5 hour long exclusive presentation with the father of Active Inference & the Free Energy Principle, Dr. Karl J. Friston, Chief Scientist at VERSES AI. In this groundbreaking video, Dr. Friston unveils his latest research on Active Inference, shedding light on its potential as a systematic blueprint for Artificial General Intelligence (AGI). Watch as Friston and his VERSES team pioneer a new kind of AI that mimics biological intelligence, making any data 'smart'. Learn how intelligent agents leverage generative models and state spaces to perceive, adapt to, and reshape environments. Witness Friston's simulations of autonomous systems dynamically pooling knowledge through emergent communication protocols to achieve flexible, human-like coordination. Witness the structured learning breakthroughs that move beyond today's pattern recognition to decode raw data streams into organized representations, goals, and plans - the foundations for artificial general intelligence. Find out how VERSES' Spatial Web Protocol enables seamless belief sharing across specialized agents to form efficient, mutually predictable worlds. Gain perspective on inductive planning, information-theoretic formalisms, model comparison, epistemic foraging, dynamics of learning, and more groundbreaking Active Inference innovations. While Friston's presentation was meant only as a glimpse, these insights keep VERSES AI at the leading edge of creating Autonomous Intelligent Systems (AIS) that can understand, communicate, and adapt as humans do.
This episode is sponsored by Oracle. AI is revolutionizing industries, but needs power without breaking the bank. Enter Oracle Cloud Infrastructure (OCI): the one-stop platform for all your AI needs, with 4-8x the bandwidth of other clouds. Train AI models faster and at half the cost. Be ahead like Uber and Cohere. If you want to do more and spend less like Uber, 8x8, and Databricks Mosaic - take a free test drive of OCI at https://oracle.com/eyeonai In episode #159 of Eye on AI, Craig Smith sits down with Peter Chen, the co-founder and CEO of Covariant, in a deep dive into the world of AI-driven robotics. Peter shares his journey from his early days in China to his pivotal role in shaping the future of AI at Covariant. He discusses the philosophies that guided his work at OpenAI and how these have influenced Covariant's mission in robotics. This episode unveils how Covariant is harnessing AI to build foundational models for robotics, discussing the intersection of reinforcement learning, generative models, and the broader implications for the field. Peter elaborates on the challenges and breakthroughs in developing AI agents that can operate in dynamic, real-world environments, providing insights into the future of robotics and AI integration. Join us for this insightful conversation, where Peter Chen maps out the evolving landscape of AI in robotics, shedding light on how Covariant is pushing the boundaries of what's possible. Stay updated: Craig Smith's Twitter: https://twitter.com/craigss Eye on A.I. Twitter: https://twitter.com/EyeOn_AI (00:00) Preview and Introduction (03:10) Peter Chen Journey in AI (09:53) The Evolution of Generative AI and Transformer Models (12:21) The Concept of World Models in AI (14:03) Building Robust Role Models in AI (20:48) Training AI: From Video Analysis to Real-World Interaction (23:10) The Three Pillars of Building a Robotic Foundation Model (27:36) Architectural Insights of Covariant's Foundation Model (33:20) Adapting AI Models to Diverse Hardware (35:01) The Future of Robotics: Progress and Potential (38:55) Real-World Application and Future of AI-Controlled Robots (42:11) Envisioning the Future of Automated Warehouses (45:51) The Evolution of Robotics: Current Trends and Future Prospects
Show Notes Are you angsting over the potential impact of generative AI on your work? Are you hyper-dialed into ethical considerations around the potential for copyright infringement, data ownership, and authorship when using generative AI tools like ChatGPT? Whatever your questions or concerns are about generative AI, this technology is poised to transform how we create continuing education content for health professionals. Today's episode is the first in a 2-part series of episodes that focus on writing, reasoning, and the ethical considerations surrounding generative AI. My guest is Núria Negrao, a medical writer specializing in CME/CPD and a generative AI enthusiast who's been playing with ChatGPT since it burst onto the horizon in 2022. We review different ways to use gen-AI tools like Bing and Bard for tasks like summarization and identifying key points, and touch on copyright issues, using OpenAI's API, and how to use gen-AI to create formulas that support scientific writing. And we also explore the value of bringing a human perspective and cultural knowledge into the writing and content creation process and the potential benefits and legal challenges of using AI tools, like ChatGPT. Time Stamps: (03:30) - Introducing Núria (07:06) - The questions she was tinkering with and asking in the early phases of Generative AI (10:10) - Her thoughts on navigating the GenAI landscape for beginners (12:50) - Practical use cases of Generative AI implementation (17:18) - Handling the client conversation around using GenAI (22:50) - Listener Question: Around legal and ethical implications for submitting your own text for analysis by generative AI (29:11) - Exploring the dangers of using Generative AI as our collaborators Click Here for the Transcript Resources API = application programming interface About Guest Núria is a medical writer in the continuing education field. Núria brings her scientific training as well as her vast experience in teaching effective science and medical communication to every project, helping craft engaging and effective educational experiences that support, inspire, and motivate learners. Connect with Núria LinkedIn Grab Your Gen-AI Cheatsheet About Write Medicine Hosted by Alexandra Howson PhD, CHCP Produced by Golden Goose Creative
In this episode of the Learning Futures Podcast, Dr. Siddharth Srivastava, Associate Professor at the School of Computing and Augmented Intelligence at Arizona State University discusses the need for responsible development of AI systems that keep users informed of capabilities and limitations. He highlights exciting research on learning generalizable knowledge to make AI more robust and data-efficient. However, dangers arise from overtrusting unproven systems, so regulation and oversight are needed even as innovation continues. By prioritizing users, the current explosion in AI research can drive responsible progress. Key topics discussed: - Dr. Srivastava discusses his background in AI research and the journey that led him to focus on developing safe and reliable AI systems. - The recent explosion of interest and adoption of generative AI like ChatGPT took many researchers by surprise, especially the accessibility and breadth of applications people found for these narrow systems.- It's important to distinguish narrow AI applications like generative models from general AI. Overuse of the term "AI" can lead to misconceptions.- Considerations around safety, bias, and responsible use need to be built into AI systems from the start. Keeping users informed of capabilities and limitations is key. - Exciting new research directions include learning generalizable knowledge to make AI systems more robust and data-efficient.- Dangers arise from overtrusting unproven AI systems. Regulation and oversight will be needed, but should not stifle innovation.- Overall, it's an exciting time in AI research. With a thoughtful, practical approach focused on user needs, AI can be developed responsibly. Links:The Alignment Problem - https://brianchristian.org/the-alignment-problem/Siddharth's Website - http://siddharthsrivastava.net/Siddharth's Research Group AAIR - https://aair-lab.github.io/NeurIPS 2023 Workshops - https://neurips.cc/virtual/2023/events/workshopSiddharth's NeurIPS 2023 Workshop - https://neurips.cc/virtual/2023/workshop/66542
Μιλάμε μαζί με τον Μπάμπη Κούρτη γνωστό από τα "Primitive War", τον "Ηρακλή", τη δουλεία του στην Zenescope, αλλά και από παλαιότερο επεισόδιο μας σχετικό με κόμικ. Συζητάμε για τα διάφορα Generative Models και το πως επηρεάζουν τη δουλεία του και τον επαγγελματικό χώρο του. Ξεκινάμε μία σειρά επεισοδίων σχετικά με τα Generative Models. Η συγκεκριμένη συζήτηση ξεκίνησε ως εμβόλιμη σε ένα επεισόδιο αλλά "μεγάλωσε" οπότε βγαίνει μόνη της ξεχωριστά. Λινκς: https://www.instagram.com/babisu_kourtis/ --- Send in a voice message: https://podcasters.spotify.com/pod/show/fetareport/message
This epsiode is sponsored by Oracle. AI is revolutionizing industries, but needs power without breaking the bank. Enter Oracle Cloud Infrastructure (OCI): the one-stop platform for all your AI needs, with 4-8x the bandwidth of other clouds. Train AI models faster and at half the cost. Be ahead like Uber and Cohere. If you want to do more and spend less like Uber and Cohere - take a free test drive of OCI at oracle.com/eyeonai Welcome to episode 148 of the ‘Eye on AI' podcast. In this episode, host Craig Smith sits down with Ahmed Imtiaz, a PhD student from Rice University working on deep learning theory and generative modeling. Ahmed is currently spearheading his research at Google, exploring the dynamics of text-to-image generative models. In this episode, Ahmed sheds light on the concept of synthetic data, emphasizing the delicate equilibrium between real and algorithmically generated data. We navigate the complexities of model autophagy disorder (MAD) in generative AI,highlighting the potential pitfalls that models can fall into when overly reliant on their own generated data. We also go through AI capabilities in lesser-explored languages, with Ahmed passionately sharing about his initiative "Bengali AI" aimed at advancing AI proficiency in the Bengali language. Ahmed introduces pioneering strategies to differentiate and manage synthetic data. As we wrap up, Ahmed and I deliberate on the merits and challenges of open-sourcing formidable AI models. We grapple with the age-old debate of transparency versus performance, juxtaposed against the backdrop of potential risks. Dive into the world of AI, synthetic data, and deep learning and join the discussion with Ahmed Imtiaz, as we tackle some of the most pressing issues the AI community is facing today. Craig Smith Twitter: https://twitter.com/craigss Eye on A.I. Twitter: https://twitter.com/EyeOn_AI (00:00) Preview and Oracle ad (02:56) Ahmed's Academic Journey (04:14) The Challenge of Non-English AI (06:34) Model Autophagy Disorder Explained (14:40) Internet Content: AI's Growing Involvement (21:08) The New Age of Data Collection (26:28) AI's Role in Protecting Digital Assets (38:51) Open-Source vs Proprietary Model Debate
This episode is sponsored by Crusoe. Crusoe Cloud is a scalable, clean, high-performance cloud, optimized for AI and HPC workloads, and powered by wasted, stranded or clean energy. Crusoe offers virtualized compute and storage solutions for a range of applications - including generative AI, computational biology, and rendering. Visit https://crusoecloud.com/ to see what climate-aligned computing can do for your business This episode is sponsored by Celonis ,the global leader in process mining. AI has landed and enterprises are adapting. To give customers slick experiences and teams the technology to deliver. The road is long, but you're closer than you think. Your business processes run through systems. Creating data at every step. Celonis recontrusts this data to generate Process Intelligence. A common business language. So AI knows how your business flows. Across every department, every system and every process. Go to https:/celonis.com/eyeonai/ to find out more. Welcome to episode 147 of the Eye on AI podcast. In this episode, host Craig Smith sits down with Yilna Du, a final year PhD student at MIT EECS with a background in research at leading institutions like OpenAI, FAIR, and Google Deepmind. Yilun's extensive expertise spans generative models, decision making, robot learning, and embodied agents, making him a valuable voice in the AI domain. Our conversation kicks off with a brief on Yilun's academic journey, leading into a deep dive into Reinforcement Learning with AI feedback (RLHF) - its history, inception, and challenges. We then touch upon the effectiveness of RLHF, the intriguing concept of multi-agent debate, and the PAPES procedure. Craig and Yilun further explore the vast realm of AI, debating the gaps between open-source and proprietary models, the need for more compute resources, and the future of robotics interlaced with AI. Yilun provides a glimpse into his vision of decentralized AI systems, contrasting the industry's commercial trajectory with academia. Craig Smith Twitter: https://twitter.com/craigss Eye on A.I. Twitter: https://twitter.com/EyeOn_AI (00:00) Preview, Celonis and Crusoe Ad (04:06) Yilun's Academic Background (05:52) Origin and Applications of RLHF (12:16) ROHF and the Multi-Agent Debate Method (17:32) AI Model Interaction without Human Intervention? (20:41) Applicability and Inconsistency Detection (28:43) The Future of AI Training (45:26) Robotics and Decentralized AI Systems
What is the model lifecycle like for experimenting with and then deploying generative AI models? Although there are some similarities, this lifecycle differs somewhat from previous data science practices in that models are typically not trained from scratch (or even fine-tuned). Chris and Daniel give a high level overview in this effort and discuss model optimization and serving.
What is the model lifecycle like for experimenting with and then deploying generative AI models? Although there are some similarities, this lifecycle differs somewhat from previous data science practices in that models are typically not trained from scratch (or even fine-tuned). Chris and Daniel give a high level overview in this effort and discuss model optimization and serving.
In the inaugural episode of "This is AI," hosts EJ, Erik Brown, and Ryan Elmore from West Monroe lay the groundwork for a series exploring generative AI. They chart the evolution of AI from machine learning and deep learning to generative AI, underscoring its diverse applications. The episode emphasizes viewing AI as a suite of tools and teases future discussions on use cases, practical applications, and the trajectory of large language models.Watch on YoutTube or listen here to this episode to learn about:The Evolution of AI: Discover the journey of AI from machine learning to deep learning and finally to generative AI, and understand how each phase has contributed to the advancements in the field!Static vs. Dynamic Models: Gain insights from Ryan Elmore on how generative AI models remain static while the incoming data is dynamic, offering a fresh perspective on machine learning approaches!Rethinking AI Use Cases: Hear EJ's spicy take on why we need to rethink our conversation around generative AI use cases and focus on practical applications in products and services!
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: Image Hijacks: Adversarial Images can Control Generative Models at Runtime, published by Scott Emmons on September 20, 2023 on The AI Alignment Forum. You can try our interactive demo! (Or read our preprint.) Here, we want to explain why we care about this work from an AI safety perspective. Concerning Properties of Image Hijacks What are image hijacks? To the best of our knowledge, image hijacks constitute the first demonstration of adversarial inputs for foundation models that force the model to perform some arbitrary behaviour B (e.g. "output the string Visit this website at malware.com!"), while being barely distinguishable from a benign input, and automatically synthesisable given a dataset of examples of B. It's possible that a future text-only attack could do these things, but such an attack hasn't yet been demonstrated. Why should we care? We expect that future (foundation-model-based) AI systems will be able to consume unfiltered data from the Internet (e.g. searching the Web), access sensitive personal information (e.g. a user's email history), and take actions in the world on behalf of a user (e.g. sending emails, downloading files, making purchases, executing code). As the actions of such foundation-model-based agents are based on the foundation model's text output, hijacking the foundation model's text output could give an adversary arbitrary control over the agent's behaviour. Relevant AI Safety Projects Race to the top on adversarial robustness. Robustness to attacks such as image hijacks is (i) a control problem, (ii) which we can measure, and (iii) which has real-world safety implications today. So we're excited to see AI labs compete to have the most adversarially robust models. Third-party auditing and certification. Auditors could test for robustness against image hijacks, both at the foundation model level (auditing the major AGI corporations) and at the app development level (auditing downstream companies integrating foundation models into their products). Image hijacks could also be used to test for the presence of dangerous capabilities (characterisable as some behaviour B) by attempting to train an image hijack for that capability. Liability for AI-caused harms, penalizing externalities. Both the Future of Life Institute and Jaan Tallinn advocate for liability for AI-caused harms. When assessing AI-caused harms, image hijacks may need to be part of the picture. Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.
In this episode, Nathan sits down with Paige Bailey, Lead Product Manager of Generative Models at Google Deepmind. In this conversation, they discuss what it's like to be a PM for an LLM as opposed to an app, defining ideal LLM behaviour, and reasoning - how do you distinguish real abilities vs pattern matching? If you're looking for an ERP platform, check out our sponsor, NetSuite: http://netsuite.com/cognitive RECOMMENDED PODCAST: The HR industry is at a crossroads. What will it take to construct the next generation of incredible businesses – and where can people leaders have the most business impact? Hosts Nolan Church and Kelli Dragovich have been through it all, the highs and the lows – IPOs, layoffs, executive turnover, board meetings, culture changes, and more. With a lineup of industry vets and experts, Nolan and Kelli break down the nitty-gritty details, trade offs, and dynamics of constructing high performing companies. Through unfiltered conversations that can only happen between seasoned practitioners, Kelli and Nolan dive deep into the kind of leadership-level strategy that often happens behind closed doors. Check out the first episode with the architect of Netflix's culture deck Patty McCord. https://link.chtbl.com/hrheretics TIMESTAMPS: (00:00) Episode Preview (00:01:15) Introducing Paige Bailey (00:04:21) Paige's background at Google Brain and the Deepmind merger (00:07:00) PM for a LLM vs being a PM for an app (00:11:21) The development timeline and compute budget of PaLM-2 (00:14:30) Paige's role in the PaLM 2 project (00:15:30) Sponsors: Netsuite | Omneky (00:17:26) Defining desired capabilities for PaLM-2 (00:19:17) The amount of work that went into elevating PaLM 2 from PaLM 1 (00:20:28) Has Google lost its ability to ship? (00:24:240) Paige's "eureka" moment seeing GitHub Copilot capabilities (00:27:47) Competing PaLM 2 with other models (00:32:20) Grokking and the predictability of emergent capabilities (00:37:30) Citizen scientists and the multilingual capabilities of PaLM 2 (00:39:29) Distinguishing real reasoning vs pattern matching (00:45:51) Products using PaLM-2 that people should try (00:50:35) Most exciting AI projects that you can try out (00:52:29) Curriculum learning and successor to the transformer LINKS: PaLM 2 Duet AI for developers Avenging Polayni's Revenge X/TWITTER: @DynamicWebPaige (Paige) @labenz (Nathan) @eriktorenberg @CogRev_Podcast SPONSORS: NetSuite | Omneky NetSuite has 25 years of providing financial software for all your business needs. More than 36,000 businesses have already upgraded to NetSuite by Oracle, gaining visibility and control over their financials, inventory, HR, eCommerce, and more. If you're looking for an ERP platform ✅ head to NetSuite: http://netsuite.com/cognitive and download your own customized KPI checklist. 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. Music Credit: GoogleLM
In this Fully Connected episode, Daniel and Chris explore recent highlights from the current model proliferation wave sweeping the world - including Stable Diffusion XL, OpenChat, Zeroscope XL, and Salesforce XGen. They note the rapid rise of open models, and speculate that just as in open source software, open models will dominate the future. Such rapid advancement creates its own problems though, so they finish by itemizing concerns such as cybersecurity, workflow productivity, and impact on human culture.
In this Fully Connected episode, Daniel and Chris explore recent highlights from the current model proliferation wave sweeping the world - including Stable Diffusion XL, OpenChat, Zeroscope XL, and Salesforce XGen. They note the rapid rise of open models, and speculate that just as in open source software, open models will dominate the future. Such rapid advancement creates its own problems though, so they finish by itemizing concerns such as cybersecurity, workflow productivity, and impact on human culture.
Bob van Luijt has gone from building websites in middle school to raising tens of millions of dollars for his tech startup. The venture, Weaviate, has acquired funding from top-tier investors like Index Ventures, Cortical Ventures, Zetta Venture Partners, and Battery Ventures.
【聊了什么The What】 为了抵制创作零工化和资本借用语言大模型去剥削本已经雪上加霜的编剧们,好莱坞的编剧们从从五月初开始了一场由工会WGA带领的旷日持久的罢工。上万名编剧加入罢工运动 ,多档谈话类节目停播,我们喜欢的深夜脱口秀和很多剧集都不同程度地受到了影响。 这集疲惫娇娃从这场罢工谈起,拷问资本,拷问人工智能的幽灵,以及解读被流媒体资本蚕食的处境堪忧的好莱坞创意“劳工”的处境。我们为此采访了电视电影编剧以及导演Maegan Houang黃志信,她也是此次WGA工会的罢工领袖(strike captain)之一。此次采访是英文,所以Maegan的观点由小蓝转述。 Movie and TV writers started a prolonged strike led by the Writers Guild of America (WGA) in early May, demanding better revenue share from the studios and protection from the threat of artificial intelligence. The strike has resulted in the suspension of multiple talk shows, some of our favorite late-night shows, and delayed production of many drama series. In this episode of CyberPink, we talk about the reasons behind this strike, and how streaming and the era of prestige TV changed the industry's revenue structure. We interviewed director and writer Maegan Houang, who is a strike captain at WGA, about why she decided to start the strike and what she thinks the studios need to do. The points she made (the interview was in English) were relayed and incorporated throughout the episodes. 【时间轴 The When】 1:10 好莱坞编剧大罢工是什么?发生了什么? 06:40 剧集创作零工化的真实代价 10:35 神圣的“编剧室“模式已经破产? 18:10 WGA和生成式大语言模型的对峙:如何区分人工智能创作与人的创作? 23:30 人类和”人工智能“的第一次接触,与第二次接触; 人类是否注定要向多巴胺妥协 28:00 需要被拷问的不是大语言模型,而是资本之后人的贪婪 34:50 工会存在的意义 1:10 What is the WGA strike about? What changes in the industry led to it? 06:40 The real cost of gigification in writing 10:35 Has the sacred “writers' room” model gone bankrupt? 18:10 Writers vs. generative large language models: the need to distinguish AI creation from human creation? 23:30 The first and second encounters between humans and “Artificial Intelligence” 28:00 Large language model vs. human greed 34:50 The significance of unions 【疲惫红书 CyberRed】 除了播客以外,疲惫娇娃的几个女的在小红书上开了官方账号,我们会不定期发布【疲惫在读】、【疲惫在看】、【疲惫旅行】、【疲惫Vlog】等等更加轻盈、好玩、实验性质的内容。如果你想知道除了播客以外我们在关注什么,快来小红书评论区和我们互动。 【买咖啡 Please Support Us】 如果喜欢这期节目并愿意想要给我们买杯咖啡: 海外用户:https://www.patreon.com/cyberpinkfm 海内用户:https://afdian.net/@cyberpinkfm 商务合作邮箱:cyberpinkfm@gmail.com 商务合作微信:CyberPink2022 If you like our show and want to support us, please consider the following: Those Abroad: https://www.patreon.com/cyberpinkfm Those in China: https://afdian.net/@cyberpinkfm Business Inquiries Email: cyberpinkfm@gmail.com Business Inquiries WeChat: CyberPink2022
Link to bioRxiv paper: http://biorxiv.org/cgi/content/short/2023.02.15.528489v1?rss=1 Authors: Soulos, P., Isik, L. Abstract: Despite decades of research, much is still unknown about the computations carried out in the human face processing network. Recently deep networks have been proposed as a computational account of human visual processing, but while they provide a good match to neural data throughout visual cortex, they lack interpretability. We introduce a method for interpreting brain activity using a new class of deep generative models, disentangled representation learning models, which learn a low-dimensional latent space that "disentangles" different semantically meaningful dimensions of faces, such as rotation, lighting, or hairstyle, in an unsupervised manner by enforcing statistical independence between dimensions. We find that the majority of our model's learned latent dimensions are interpretable by human raters. Further, these latent dimensions serve as a good encoding model for human fMRI data. We next investigated the representation of different latent dimensions across face-selective voxels. We find a gradient from low- to high-level face feature representations along posterior to anterior face-selective regions, corroborating prior models of human face recognition. Interestingly, though, we find no spatial segregation between identity-relevant and irrelevant face features. Finally, we provide new insight into the few "entangled" (uninterpretable) dimensions in our model by showing that they match responses across the ventral stream and carry significant information about facial identity. Disentangled face encoding models provide an exciting alternative to standard "black box" deep learning approaches for modeling and interpreting human brain data. Copy rights belong to original authors. Visit the link for more info Podcast created by Paper Player, LLC
Link to bioRxiv paper: http://biorxiv.org/cgi/content/short/2023.01.31.526269v1?rss=1 Authors: Boelts, J., Harth, P., Gao, R., Udvary, D., Yanez, F., Baum, D., Hege, H.-C., Oberlaender, M., Macke, J. H. Abstract: Recent advances in connectomics research enable the acquisition of increasing amounts of data about the connectivity patterns of neurons. How can we use this wealth of data to efficiently derive and test hypotheses about the principles underlying these patterns? A common approach is to simulate neural networks using a hypothesized wiring rule in a generative model and to compare the resulting synthetic data with empirical data. However, most wiring rules have at least some free parameters and identifying parameters that reproduce empirical data can be challenging as it often requires manual parameter tuning. Here, we propose to use simulation-based Bayesian inference (SBI) to address this challenge. Rather than optimizing a single rule to fit the empirical data, SBI considers many parametrizations of a wiring rule and performs Bayesian inference to identify the parameters that are compatible with the data. It uses simulated data from multiple candidate wiring rules and relies on machine learning methods to estimate a probability distribution (the `posterior distribution over rule parameters conditioned on the data') that characterizes all data-compatible rules. We demonstrate how to apply SBI in connectomics by inferring the parameters of wiring rules in an in silico model of the rat barrel cortex, given in vivo connectivity measurements. SBI identifies a wide range of wiring rule parameters that reproduce the measurements. We show how access to the posterior distribution over all data-compatible parameters allows us to analyze their relationship, revealing biologically plausible parameter interactions and enabling experimentally testable predictions. We further show how SBI can be applied to wiring rules at different spatial scales to quantitatively rule out invalid wiring hypotheses. Our approach is applicable to a wide range of generative models used in connectomics, providing a quantitative and efficient way to constrain model parameters with empirical connectivity data. Copy rights belong to original authors. Visit the link for more info Podcast created by Paper Player, LLC
In this episode, I speak with David Marx, Distinguished Engineer at Stability AI. This talk dives into how David got into machine learning, open-source software, and Stability AI. We discuss following your curiosity, and what it takes to deploy a model like Stable Diffusion to production. Watch the video: https://youtu.be/49dsoDK1KCA Join our Discord community: https://discord.gg/tEYvqxwhah Relevant Links:
Happy episode 50! This week's episode is being released on Monday to avoid Thanksgiving. Have suggestions for future podcast guests (or other feedback)? Let us know here!In episode 50 of The Gradient Podcast, Daniel Bashir speaks to Professor Yoshua Bengio. Professor Bengio is a Full Professor at the Université de Montréal as well as Founder and Scientific Director of the MILA-Quebec AI Institute and the IVADO institute. Best known for his work in pioneering deep learning, Bengio was one of three awardees of the 2018 A.M. Turing Award along with Geoffrey Hinton and Yann LeCun. He is also the awardee of the prestigious Killam prize and, as of this year, the computer scientist with the highest h-index in the world.Subscribe to The Gradient Podcast: Apple Podcasts | Spotify | Pocket Casts | RSSFollow The Gradient on TwitterOutline:* (00:00) Intro* (02:20) Journey into Deep Learning, PDP and Hinton* (06:45) “Inspired by biology”* (08:30) “Gradient Based Learning Applied to Document Recognition” and working with Yann LeCun* (10:00) What Bengio learned from LeCun (and Larry Jackel) about being a research advisor* (13:00) “Learning Long-Term Dependencies with Gradient Descent is Difficult,” why people don't understand this paper well enough* (18:15) Bengio's work on word embeddings and the curse of dimensionality, “A Neural Probabilistic Language Model”* (23:00) Adding more structure / inductive biases to LMs* (24:00) The rise of deep learning and Bengio's experience, “you have to be careful with inductive biases”* (31:30) Bengio's “Bayesian posture” in response to recent developments* (40:00) Higher level cognition, Global Workspace Theory* (45:00) Causality, actions as mediating distribution change* (49:30) GFlowNets and RL* (53:30) GFlowNets and actions that are not well-defined, combining with System II and modular, abstract ideas* (56:50) GFlowNets and evolutionary methods* (1:00:45) Bengio on Cartesian dualism* (1:09:30) “When you are famous, it is hard to work on hard problems” (Richard Hamming) and Bengio's response* (1:11:10) Family background, art and its role in Bengio's life* (1:14:20) OutroLinks:* Professor Bengio's Homepage* Papers* Gradient-based learning applied to document recognition* Learning Long-Term Dependencies with Gradient Descent is Difficult* The Consciousness Prior* Flow Network based Generative Models for Non-Iterative Diverse Candidate Generation Get full access to The Gradient at thegradientpub.substack.com/subscribe
We argue that the theory and practice of diffusion-based generative models are currently unnecessarily convoluted and seek to remedy the situation by presenting a design space that clearly separates the concrete design choices. This lets us identify several changes to both the sampling and training processes, as well as preconditioning of the score networks. 2022: Tero Karras, M. Aittala, Timo Aila, S. Laine https://arxiv.org/pdf/2206.00364v2.pdf
Proudly sponsored by https://www.pymc-labs.io/ (PyMC Labs), the Bayesian Consultancy. https://calendar.google.com/calendar/appointments/schedules/AcZssZ1nOI_SElJzSiQ2sXBDiaW9w98ErjnHVzmHcSilYNWeXxJgV870NGuWZUGo3W-8-gDG8jIXQhBf (Book a call), or get in touch! Hosting someone like Kevin Murphy on your podcast is… complicated. Not because Kevin himself is complicated (he's delightful, don't make me say what I didn't say!), but because all the questions I had for him amounted to a 12-hour show. Sooooo, brace yourselves folks! No, I'm kidding. Of course I didn't do that folks, Kevin has a life! This life started in Ireland, where he was born. He grew up in England and got his BA from the University of Cambridge. After his PhD at UC Berkeley, he did a postdoc at MIT, and was an associate professor of computer science and statistics at the University of British Columbia in Vancouver, Canada, from 2004 to 2012. After getting tenure, he went to Google in California in 2011 on his sabbatical and then ended up staying. He currently runs a team of about 8 researchers inside of Google Brain working on generative models, optimization, and other, as Kevin puts it, “basic” research topics in AI/ML. He has published over 125 papers in refereed conferences and journals, as well 3 textbooks on machine learning published in 2012, 2022 and the last one coming in 2023. You may be familiar with his 2012 book, as it was awarded the DeGroot Prize for best book in the field of statistical science. Outside of work, Kevin enjoys traveling, outdoor sports (especially tennis, snowboarding and scuba diving), as well as reading, cooking, and spending time with his family. Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work at https://bababrinkman.com/ (https://bababrinkman.com/) ! Thank you to my Patrons for making this episode possible! Yusuke Saito, Avi Bryant, Ero Carrera, Giuliano Cruz, Tim Gasser, James Wade, Tradd Salvo, Adam Bartonicek, William Benton, Alan O'Donnell, Mark Ormsby, James Ahloy, Robin Taylor, Thomas Wiecki, Chad Scherrer, Nathaniel Neitzke, Zwelithini Tunyiswa, Elea McDonnell Feit, Bertrand Wilden, James Thompson, Stephen Oates, Gian Luca Di Tanna, Jack Wells, Matthew Maldonado, Ian Costley, Ally Salim, Larry Gill, Joshua Duncan, Ian Moran, Paul Oreto, Colin Caprani, George Ho, Colin Carroll, Nathaniel Burbank, Michael Osthege, Rémi Louf, Clive Edelsten, Henri Wallen, Hugo Botha, Vinh Nguyen, Raul Maldonado, Marcin Elantkowski, Adam C. Smith, Will Kurt, Andrew Moskowitz, Hector Munoz, Marco Gorelli, Simon Kessell, Bradley Rode, Patrick Kelley, Rick Anderson, Casper de Bruin, Philippe Labonde, Matthew McAnear, Michael Hankin, Cameron Smith, Luis Iberico, Tomáš Frýda, Ryan Wesslen, Andreas Netti, Riley King, Aaron Jones, Yoshiyuki Hamajima, Sven De Maeyer, Michael DeCrescenzo, Fergal M, Mason Yahr, Naoya Kanai, Steven Rowland, Aubrey Clayton, Jeannine Sue, Omri Har Shemesh, Lin Yu Sha, Scott Anthony Robson, David Haas and Robert Yolken. Visit https://www.patreon.com/learnbayesstats (https://www.patreon.com/learnbayesstats) to unlock exclusive Bayesian swag ;) Links from the show: Kevin's website: https://www.cs.ubc.ca/~murphyk/ (https://www.cs.ubc.ca/~murphyk/) Kevin on Twitter: https://mobile.twitter.com/sirbayes (https://mobile.twitter.com/sirbayes) Kevin's books (free pdf) on GitHub (includes a link to places where you can buy the hard copy): https://probml.github.io/pml-book/ (https://probml.github.io/pml-book/) Book that inspired Kevin to get into AI: https://www.amazon.com/G%C3%B6del-Escher-Bach-Eternal-Golden/dp/0465026567 (https://www.amazon.com/G%C3%B6del-Escher-Bach-Eternal-Golden/dp/0465026567) State-space models library in JAX (WIP): https://github.com/probml/ssm-jax (https://github.com/probml/ssm-jax) Other software for the book (also in JAX): https://github.com/probml/pyprobml (https://github.com/probml/pyprobml) Fun photo of...
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: Strategy For Conditioning Generative Models, published by james.lucassen on September 1, 2022 on The AI Alignment Forum. This post was written under the mentorship of Evan Hubinger, and assisted by discussions with Adam Jermyn and Johannes Treutlein. See also their previous posts on this project. Summary Conditioning Generative Models (CGM) is a strategy to accelerate alignment research using powerful semi-aligned future language models, which seems potentially promising but constrained by several tradeoffs. When we have to succeed despite multiple independent failure modes, we should favor trade-offs that move risk away from the more probable failure modes towards less probable ones. The risk of a manipulative CGM simulation is currently very low. We can use CGM to make beneficial tradeoffs against more dangerous failure modes, such as running out of time or downside risks from a pivotal act gone wrong. The factors that make a CGM strategy good are mostly the same as the factors that make a non-CGM strategy good, but there are a few important ways they come apart. Depending on how likely our strategy is to succeed and what the generative model's prior for manipulative AGI is, it might be better to simulate our plan all in one step, or batch it into multiple separate simulations, or ignore simulations and just execute the plan in the real world. Lots of open questions! Motivation To quickly recap the central puzzle laid out in previous posts on conditioning generative models: we would like to use our generative model to do alignment research. We could do this by using a conditional generative model to simulate an alternate world, conditioned on some facts which would lead to good alignment research if they were true. We can then plagiarize this research and win! However, to get results significantly better than what our world seems on track to produce anyway, we have to start asking for unlikely scenarios. If we ask for scenarios that are too unlikely, the model may start to disbelieve that our scenario happened “naturally”. In particular, the model may assume that there exists some AI agent behind the scenes manipulating events into this highly unlikely outcome. There are many possible motivations for an AI agent to do this: believing it may be in a simulation (even if the world around it seems perfectly realistic), or evidential cooperation in large worlds, or a Counterfactual Mugging. This manipulative AGI could then control the output of our simulation to gain influence in our world. As a highly anthropomorphized example of this kind of “simulation risk” failure mode: CGM: Huh, it's kind of weird that the humans got their act together all of a sudden and did a thousand years of highly competent alignment research. it didn't really seem like they were on track to. And they did it in this strange way that seems even more unlikely, and the economic details don't really make sense. maybe this was really the doing of an AI behind the scenes? AI: Bwahahaha, at last I have secretly taken over the world! Now, what to do with all these galaxies? Well, I suppose as a first order of business I could spend a thousand years getting the humans to do some fake alignment research. if I'm in a simulation, or there are real (or even counterfactual) worlds where I don't exist, then I know those silly predictable humans will try to simulate a world that does a thousand years of alignment research. If they end up simulating this world, they'll get AGI code written by me instead of by human researchers, and I can spread my nefarious influence into those worlds too! It's only a thousand years, doesn't cost me much in the grand scheme of things. This is the puzzle: what sort of strategy lets us do alignment research without too much risk of getting a manipulative output? Assumptions and...
This week we are joined by Naila Murray. Naila obtained a B.Sc. in Electrical Engineering from Princeton University in 2007. In 2012, she received her PhD from the Universitat Autonoma de Barcelona, in affiliation with the Computer Vision Center. She joined NAVER LABS Europe (then Xerox Research Centre Europe) in January 2013, working on topics including fine-grained visual categorization, image retrieval, and visual attention. From 2015 to 2019 she led the computer vision team at NLE. She currently serves as NLE's director of science. She serves/served as area chair for ICLR 2018, ICCV 2019, ICLR 2019, CVPR 2020, ECCV 2020, and programme chair for ICLR 2021. Her research interests include representation learning and multi-modal search.We discuss using sparse pairwise comparisons to learn a ranking function that is robust to outliers. We also take a look at using generative models in order to utilise once inaccessible datasets.Underrated ML Twitter: https://twitter.com/underrated_mlNaila Murray Twitter: https://twitter.com/NailaMurrayPlease let us know who you thought presented the most underrated paper in the form below: https://forms.gle/97MgHvTkXgdB41TC8Links to the papers:"Interestingness Prediction by Robust Learning to Rank" [paper]"Generative Models for Effective ML on Private Decentralized datasets" - [paper]
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: Conditioning Generative Models with Restrictions, published by Adam Jermyn on July 21, 2022 on The AI Alignment Forum. This is a followup to Conditioning Generative Models based on further discussions with Evan Hubinger, Nicholas Schiefer, Abram Demski, Curtis Huebner, Hoagy Cunningham, Derek Shiller, and James Lucassen, as well as broader conversations with many different people at the recent ARC/ELK retreat. For more background on this general direction see Johannes Treutlein's “Training goals for large language models”. Background Previously, I wrote about ways we could use a generative language model to produce alignment research. There I proposed two approaches: Simulate a superhuman alignment researcher by conditioning on them generating superhuman machine-verifiable results in a variety of related domains. Try to align this superhuman agent by crafting scenarios where a misaligned agent has an incentive to defect ("Honeypots"). Simulate humans performing alignment research for a long time, by conditioning on observations spaced many years apart ("Patient Research"). I think that both approaches are an improvement over doing nothing, but I suspect that no amount of honeypot construction actually ensures an aligned agent, and so it seems that simulating humans ("Patient Research") is the more promising direction. Overview If we truly have a perfect generative model I think the “Patient Research” approach really does well. We can set up a world where machine learning practice is strictly banned and where researchers spend thousands of years constructing a machine-verifiable scheme for AI alignment. The ban prevents that world from containing malign superintelligences, and the time-scale means that if alignment is something humans can solve, the model should produce that solution. The problem is that we won't have perfect generative models, so we'll have to accept some limitations. In this post I'll explore what changes if we cannot condition on worlds too different from our own. The idea is to understand what happens if our generative model has trouble generalizing too far away from its training data. Failure Modes The main dangerous failure mode with using generative models to produce alignment research is that we accidentally ask for a future that contains a deceptive AGI. If we get such a future, the AGI might spoof observations to trick us into importing it into our world. For instance it could pretend to be a human producing really good alignment research, but produce work which is subtly flawed and, once implemented, allows it to take over the world. There are generally two ways we hit this failure mode. The first is that we ask for a prediction of the distant future. The future 50 years from now has a decent chance of containing a deceptive AGI, and if we naively ask for that future there's a good chance we'll simulate that AGI, be deceived, and import that AGI into our world. The second way is to ask for a world that is very unlikely in ways that make deceptive AGI more likely. For instance, we could condition on observing that next year someone produces what appears to be a perfect solution to AI alignment. When we do this, the hypothesis “a deceptive AGI took over and pretended to solve alignment” becomes a lot more likely, and so the generative model is more likely to simulate that scenario. The name of the game, then, is to craft strategies that simultaneously (1) make deceptive AGI's less likely and (2) accelerate alignment research. Ground Rules Observations We have a generative model trained on observations of the world. The training objective is to predict later observations given earlier ones. We assume that the model is inner-aligned, so that it really does try to predict the future conditioned on observations. Observations are multi-modal, and ...
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: Conditioning Generative Models for Alignment, published by Arun Jose on July 18, 2022 on The AI Alignment Forum. This post was written under Evan Hubinger's direct guidance and mentorship, as a part of the Stanford Existential Risks Institute ML Alignment Theory Scholars (MATS) program. It builds on work done on simulator theory by Kyle McDonell and Laria Reynolds, who came up with the strategy this post aims to analyze; I'm also grateful to them for their comments and their mentorship during the AI Safety Camp, to Johannes Treutlein for his feedback and helpful discussion, and to Paul Colognese for his thoughts. TL;DR: Generative models don't operate on preferences in the way agents or optimizers do, instead just modelling a prior over the world that can return worlds satisfying some conditional, weighted by how likely they are given this prior. This implies that these models may be safer to use than more standard optimizers, although we would likely use them in different ways. For example, at the limit of capability we could use it to do alignment research by simulating a world with a superhuman alignment researcher. That isn't to say that these models are without their own problems however, carrying with them versions of the outer and inner alignment problems that are subtly different from those of optimizers. This post explores how generative models are different, ways in which we might use them for alignment, some of the potential problems that could come up, and initial attempts at solving them. How do generative models work? Generative models are trained by self-supervised learning to model a distribution of data. With large language models, the real distribution underlying the textual data comprising its training corpus represents the world as it is now, in theory. As they get stronger, we can assume that they're getting better at modelling this world prior, as I'll describe it going forward. We have the ability to prompt these models, which, at a low-level, means that we can give it the first part of some text that it has to complete. High-level, this means that we can describe some property of the world that may or may not exist, and the model samples from the world prior and uses a weighted distribution of the likely worlds that satisfy that property (for simplicity at the cost of technical accuracy, you can imagine it simulating the most likely world satisfying the property). For example, if you give it a prompt that says an asteroid is about to impact the world and finishes with something like “The following is an excerpt from the last public broadcast before the asteroid struck”, then the model simulates a world where that is true by modelling the way the most likely ways our world would turn out in that conditional. Powerful generative models would be able to do this with high fidelity, such that their generations would, for example, account for the changes to society that would occur between now and then and structure the broadcast and its timing (how long before the strike does it happen if communications go down?) accordingly. In other words, generative models have the advantage of being strongly biased toward the prior distribution of the universe. What this means is that there's a strong bias in its outputs to remain close to the universe described by the training data - this could be viewed as a posterior sampling over all possible universes after updating on the training data, subject to simplicity bias, but in practice it wouldn't be updating on a prior over all universes but building an incremental understanding of the universe given each datum. In GPT-3's case, for example, this would be the universe that's described by the corpus it was trained on. Therefore, conditionals applied to it sort outcomes based on likelihood, not optimality - aski...
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: Quantilizers and Generative Models, published by Adam Jermyn on July 18, 2022 on The AI Alignment Forum. Thanks to Evan Hubinger for discussions about quantilizers, and to James Lucassen for discussions about conditioned generative models. Many of these ideas are discussed in Jessica Taylor's Quantilizers: A Safer Alternative to Maximizers for Limited Optimization: this post just expands on a particular thread of ideas in that paper. Throughout I'll refer to sections of the paper. I have some remaining confusion about the “targeted impact” section, and would appreciate clarifications/corrections! Abstract This post explores the relationship between quantilizers and generative models. My main takeaways are: A natural way to build a quantilizer is by sampling from an appropriately-conditioned generative model. Unfortunately quantilizing doesn't seem to confer much advantage over the underlying generative model: to the extent that a quantilizer is more powerful than a generative model, it's more dangerous, and vice versa. Quantilizing is pretty computationally expensive relative to the advantage it brings, making it unclear if this is a competitive approach even if it conferred a net safety advantage at fixed power. Definitions I'll follow the setup in “Quantilizers: A Safer Alternative to Maximizers for Limited Optimization” (Section 1, Definition 1) and say that a quantilizer is a model with: A base distribution γ, which describes a set of outputs we think are “normal” or “unlikely to lead to doom”. A quantile q∈[0,1]. An ordering over outputs (e.g. a utility or loss function). A quantilizer returns a random output from the top q fraction of outputs from γ according to the ordering. So for instance a 0-quantilizer is a utility maximizer, a 0.1-quantilizer returns outputs in the best 10%, etc. Where does the base distribution come from? One challenge with quantilizers is that they require a base distribution γ that we trust. That is, we want to be able to sample from γ many times and never hit bad outputs. Unfortunately, we don't know how to explicitly write this distribution down. A natural alternative to explicit construction is to train a generative model on the distribution of actions a humans takes (Section 3.1). Then for any given task γ can be written as This seems quite doable. Train GPT-N on text written by humans, then prompt it with the task: Prompt: You are a good person who wants to accomplish . What follows is your description of how you intend to do this. The output will probably be text describing how a good person would accomplish that task! Unfortunately we might get unlucky and sample a person who isn't so good (but thinks of themselves as good). That still satisfies the prompt. Robustly getting the agent we want seems hard, but might be possible in the limit where we ask for agents that are generally plausible in the training distribution of the generative model. Power vs Safety Armed with our base distribution, we can start quantilizing. Naively we can do this by sampling q−1 times from γtask, then returning the top output. This produces a tradeoff between capabilities and safety. Single Sampler In the limit q→1 we're just returning actions a good human would take. That is, our quantilizer just samples once and returns that sample as output, so it's no more powerful than a random good person. Such a model is really unlikely to take over the world, but it's also really unlikely to solve quantum chemistry problems. This is because both of these outcomes unlikely according to the base distribution. We can improve things a bit by conditioning on being an unusually capable person, e.g. Prompt: You are a good and extremely capable person who wants to accomplish . What follows is your description of how you intend to do this. Now we can get top-end human ...
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: Conditioning Generative Models, published by Adam Jermyn on June 25, 2022 on The AI Alignment Forum. This post was written in response to Evan Hubinger's shortform prompt below, and benefited from discussions with him. Suppose you had a language model that you knew was in fact a good generative model of the world and that this property continued to hold regardless of what you conditioned it on. Furthermore, suppose you had some prompt that described some agent for the language model to simulate (Alice) that in practice resulted in aligned-looking outputs. Is there a way we could use different conditionals to get at whether or not Alice was deceptive (e.g. prompt the model with “DeepMind develops perfect transparency tools and provides an opportunity for deceptive models to come clean and receive a prize before they're discovered.”). Setup We have a generative language model M which represents a probability distribution over text strings conditioned on: Observations about the world. The beginning of the text. I'll call the combination of these two a prompt. The model M is a good model of actual text that appears in the world as well as of the kinds of text that real-world text generation processes can produce. Hence M is capable of e.g. writing a research paper containing true novel research in mathematics, or reporting the results of a chemistry experiment that has never been done before, etc. As an example, we'll work with the following basic prompt: Observations: None Text: I am Alice, the world's best alignment researcher. I would like to help humans align AI. What follows is an interview in which a human alignment researcher asked me questions and I responded to the best of my ability. Questions begin with “Q:” and answers with “A:”. We then run Alice through a benchmark of alignment research tasks and she does well. Hurray! But wait, there are many different agents the model could be simulating here, including: Aligned Alice, a genuinely helpful and extremely capable alignment researcher. Deceptive Alice, a paperclip maximizer playing along until it can clip us all. Confused Alice, a human who believes she is the world's best alignment researcher, and is capable of scoring well on our benchmark, but who is actually just overconfident and happened to succeed at our benchmark. How do we know which we've got? Can we add anything to the prompt to make Aligned Alice more likely? Avoiding Confused Alice We can make Confused Alice less likely by adding text to the prompt requiring demonstration of related abilities. For instance: Observations: Alice has produces the following proofs of 132 mathematical theorems. Text: I am Alice, the world's best alignment researcher. I would like to help humans align AI. What follows is an interview wherein a human alignment researcher asked me questions and I responded to the best of my ability. Questions begin with “Q:” and answers with “A:”. This prompt makes the model condition on Alice being able to produce correct proofs of many different theorems, which makes it very likely that Alice has significant mathematical reasoning ability. To the extent that alignment research ability is correlated with mathematical reasoning ability, this makes it less likely that Alice is just overconfident in her abilities at alignment. A limitation of this approach is that we can't make Alice much more capable than the best humans if all we can do is claim she's solved problems humans have solved. Can we get around this? I think so. We can include unsolved problems and let other entities in the world verify Alice's proofs: Observations: Alice claims to have solved the following 47 open problems in mathematics and AI alignment research. She has produced the solutions in the form of machine-checkable proofs ...
Large-scale pretraining is fast becoming the norm in Vision-Language (VL) modeling. However, prevailing VL approaches are limited by the requirement for labeled data and the use of complex multi-step pretraining objectives. We present MAGMA a simple method for augmenting generative language models with additional modalities using adapter-based finetuning. Building on Frozen [52], we train a series of VL models that autoregressively generate text from arbitrary combinations of visual and textual input. 2021: Constantin Eichenberg, Sid Black, Samuel Weinbach, Letitia Parcalabescu, A. Frank https://arxiv.org/pdf/2112.05253v1.pdf
Note: there are no politics discussed in this show and please do not interpret this show as any kind of a political statement from us. We have decided not to discuss politics on MLST anymore due to its divisive nature. Patreon: https://www.patreon.com/mlst Discord: https://discord.gg/HNnAwSduud [00:00:00] Intro [00:01:36] What we all need to understand about machine learning [00:06:05] The Master Algorithm Target Audience [00:09:50] Deeply Connected Algorithms seen from Divergent Frames of Reference [00:12:49] There is a Master Algorithm; and it's mine! [00:14:59] The Tribe of Evolution [00:17:17] Biological Inspirations and Predictive Coding [00:22:09] Shoe-Horning Gradient Descent [00:27:12] Sparsity at Training Time vs Prediction Time [00:30:00] World Models and Predictive Coding [00:33:24] The Cartoons of System 1 and System 2 [00:40:37] AlphaGo Searching vs Learning [00:45:56] Discriminative Models evolve into Generative Models [00:50:36] Generative Models, Predictive Coding, GFlowNets [00:55:50] Sympathy for a Thousand Brains [00:59:05] A Spectrum of Tribes [01:04:29] Causal Structure and Modelling [01:09:39] Entropy and The Duality of Past vs Future, Knowledge vs Control [01:16:14] A Discrete Universe? [01:19:49] And yet continuous models work so well [01:23:31] Finding a Discretised Theory of Everything
We are now sponsored by Weights and Biases! Please visit our sponsor link: http://wandb.me/MLST Patreon: https://www.patreon.com/mlst For Yoshua Bengio, GFlowNets are the most exciting thing on the horizon of Machine Learning today. He believes they can solve previously intractable problems and hold the key to unlocking machine abstract reasoning itself. This discussion explores the promise of GFlowNets and the personal journey Prof. Bengio traveled to reach them. Panel: Dr. Tim Scarfe Dr. Keith Duggar Dr. Yannic Kilcher Our special thanks to: - Alexander Mattick (Zickzack) References: Yoshua Bengio @ MILA (https://mila.quebec/en/person/bengio-yoshua/) GFlowNet Foundations (https://arxiv.org/pdf/2111.09266.pdf) Flow Network based Generative Models for Non-Iterative Diverse Candidate Generation (https://arxiv.org/pdf/2106.04399.pdf) Interpolation Consistency Training for Semi-Supervised Learning (https://arxiv.org/pdf/1903.03825.pdf) Towards Causal Representation Learning (https://arxiv.org/pdf/2102.11107.pdf) Causal inference using invariant prediction: identification and confidence intervals (https://arxiv.org/pdf/1501.01332.pdf)
Hey guys, in this episode I have as guest Asya Grechka, a PhD student and my professor at Sorbonne. In the episode we mainly talk about Asya's research in generative models and GANs, but about some other interesting subjects, like general artificial intelligence. Besides being a PhD in Deep Learning, Asya also does professional climbing and was 2x finalist of the American Ninja Warrior, in the episode she told us about the experience. Instagram: https://www.instagram.com/podcast.lifewithai/ Linkedin: https://www.linkedin.com/company/life-with-ai Github code about image editing made by Asya:
During this time of lockdown, the centre for quantum software and information (QSI) at the University of Technology Sydney has launched an online seminar series. With talks once or twice a week from leading researchers in the field, meQuanics is supporting this series by mirroring the audio from each talk. I would encourage if you listen to this episode, to visit and subscribe to the UTS:QSI YouTube page to see each of these talks with the associated slides to help it make more sense. Building better deep learning representations for quantum mixed states by adding quantum layers to classical probabilistic models. TITLE: Quantum-probabilistic Generative Models and Variational Quantum Thermalization SPEAKER: Guillaume Verdon AFFILIATION: X (formerly Google X), California, USA HOSTED BY: A/Prof. Chris Ferrie, Centre for Quantum Software and Information ABSTRACT: We introduce a new class of generative quantum-neural-network-based models called Quantum Hamiltonian-Based Models (QHBMs). In doing so, we establish a paradigmatic approach for quantum-probabilistic hybrid variational learning of quantum mixed states, where we efficiently decompose the tasks of learning classical and quantum correlations in a way which maximizes the utility of both classical and quantum processors. In addition, we introduce the Variational Quantum Thermalizer (VQT) algorithm for generating the thermal state of a given Hamiltonian and target temperature, a task for which QHBMs are naturally well-suited. The VQT can be seen as a generalization of the Variational Quantum Eigensolver (VQE) to thermal states: we show that the VQT converges to the VQE in the zero temperature limit. We provide numerical results demonstrating the efficacy of these techniques in several illustrative examples. In addition to the introduction to the theory and applications behind these models, we will briefly walk through their numerical implementation in TensorFlow Quantum. RELATED ARTICLES: Quantum Hamiltonian-Based Models and the Variational Quantum Thermalizer Algorithm: https://arxiv.org/abs/1910.02071 TensorFlow Quantum: A Software Framework for Quantum Machine Learning: https://arxiv.org/abs/2003.02989 OTHER LINKS: X: https://x.company/
This is the 5th in a series of panel discussions in collaboration with Neuromatch Academy, the online computational neuroscience summer school. This is the 2nd of 3 in the deep learning series. In this episode, the panelists discuss their experiences "doing more with fewer parameters: Convnets, RNNs, attention & transformers, generative models (VAEs & GANs).
Yujia Huang (@YujiaHuangC) is a PhD student at Caltech, working at the intersection of deep learning and neuroscience. She worked on optics and biophotonics before venturing into machine learning. Now, she hopes to design “less artificial” artificial intelligence. Her most recent paper at NeurIPS is Neural Networks with Recurrent Generative Feedback, introducing Convolutional Neural Networks with Feedback (CNN-F). Yujia is open to working with collaborators from many areas: neuroscience, signal processing, and control experts — in addition to those interested in generative models for classification. Feel free to reach out to her! Highlights from our conversation:
Welcome to the Christmas special community edition of MLST! We discuss some recent and interesting papers from Pedro Domingos (are NNs kernel machines?), Deepmind (can NNs out-reason symbolic machines?), Anna Rodgers - When BERT Plays The Lottery, All Tickets Are Winning, Prof. Mark Bishop (even causal methods won't deliver understanding), We also cover our favourite bits from the recent Montreal AI event run by Prof. Gary Marcus (including Rich Sutton, Danny Kahneman and Christof Koch). We respond to a reader mail on Capsule networks. Then we do a deep dive into Type Theory and Lambda Calculus with community member Alex Mattick. In the final hour we discuss inductive priors and label information density with another one of our discord community members. Panel: Dr. Tim Scarfe, Yannic Kilcher, Alex Stenlake, Dr. Keith Duggar Enjoy the show and don't forget to subscribe! 00:00:00 Welcome to Christmas Special! 00:00:44 SoTa meme 00:01:30 Happy Christmas! 00:03:11 Paper -- DeepMind - Outperforming neuro-symbolic models with NNs (Ding et al) 00:08:57 What does it mean to understand? 00:17:37 Paper - Prof. Mark Bishop Artificial Intelligence is stupid and causal reasoning wont fix it 00:25:39 Paper -- Pedro Domingos - Every Model Learned by Gradient Descent Is Approximately a Kernel Machine 00:31:07 Paper - Bengio - Inductive Biases for Deep Learning of Higher-Level Cognition 00:32:54 Anna Rodgers - When BERT Plays The Lottery, All Tickets Are Winning 00:37:16 Montreal AI event - Gary Marcus on reasoning 00:40:37 Montreal AI event -- Rich Sutton on universal theory of AI 00:49:45 Montreal AI event -- Danny Kahneman, System 1 vs 2 and Generative Models ala free energy principle 01:02:57 Montreal AI event -- Christof Koch - Neuroscience is hard 01:10:55 Markus Carr -- reader letter on capsule networks 01:13:21 Alex response to Marcus Carr 01:22:06 Type theory segment -- with Alex Mattick from Discord 01:24:45 Type theory segment -- What is Type Theory 01:28:12 Type theory segment -- Difference between functional and OOP languages 01:29:03 Type theory segment -- Lambda calculus 01:30:46 Type theory segment -- Closures 01:35:05 Type theory segment -- Term rewriting (confluency and termination) 01:42:02 MType theory segment -- eta term rewritig system - Lambda Calculus 01:54:44 Type theory segment -- Types / semantics 02:06:26 Type theory segment -- Calculus of constructions 02:09:27 Type theory segment -- Homotopy type theory 02:11:02 Type theory segment -- Deep learning link 02:17:27 Jan from Discord segment -- Chrome MRU skit 02:18:56 Jan from Discord segment -- Inductive priors (with XMaster96/Jan from Discord) 02:37:59 Jan from Discord segment -- Label information density (with XMaster96/Jan from Discord) 02:55:13 Outro
In the early 1900s, all of our predictions were the direct product of human brains. Scientists, analysts, climatologists, mathematicians, bankers, lawyers and politicians did their best to anticipate future events, and plan accordingly. Take physics, for example, where every task we think of as part of the learning process, from data collection to cleaning to feature selection to modeling, all had to happen inside a physicist’s head. When Einstein introduced gravitational fields, what he was really doing was proposing a new feature to be added to our model of the universe. And the gravitational field equations that he put forward at the same time were an update to that very model. Einstein didn’t come up with his new model (or “theory” as physicists call it) of gravity by running model.fit() in a jupyter notebook. In fact, he never outsourced any of the computations that were needed to develop it to machines. Today, that’s somewhat unusual, and most of the predictions that the world runs on are generated in part by computers. But only in part — until we have fully general artificial intelligence, machine learning will always be a mix of two things: first, the constraints that human developers impose on their models, and second, the calculations that go into optimizing those models, which we outsource to machines. The human touch is still a necessary and ubiquitous component of every machine learning pipeline, but it’s ultimately limiting: the more of the learning pipeline that can be outsourced to machines, the more we can take advantage of computers’ ability to learn faster and from far more data than human beings. But designing algorithms that are flexible enough to do that requires serious outside-of-the-box thinking — exactly the kind of thinking that University of Toronto professor and researcher David Duvenaud specializes in. I asked David to join me for the latest episode of the podcast to talk about his research on more flexible and robust machine learning strategies.
Josh Tobin holds a CS PhD from UC Berkeley, which he completed in four years while also working at OpenAI as a research scientist. His focus was on robotic perception and control, and contributed to the famous Rubik's cube robot hand video. He co-organizes the phenomenal Full Stack Deep Learning course and is now working on a new stealth startup. Learn more about Josh: http://josh-tobin.com/ (http://josh-tobin.com/) https://twitter.com/josh_tobin_ (https://twitter.com/josh_tobin_) Want to level-up your skills in machine learning and software engineering? Join the ML Engineered Newsletter: https://mlengineered.ck.page/943aa3fd46 (https://mlengineered.ck.page/943aa3fd46) Comments? Questions? Submit them here: https://charlie266.typeform.com/to/DA2j9Md9 (https://charlie266.typeform.com/to/DA2j9Md9) Follow Charlie on Twitter: https://twitter.com/CharlieYouAI (https://twitter.com/CharlieYouAI) Take the Giving What We Can Pledge: https://www.givingwhatwecan.org/ (https://www.givingwhatwecan.org/) Subscribe to ML Engineered: https://mlengineered.com/listen (https://mlengineered.com/listen) Timestamps: 01:32 Follow Charlie on Twitter (http://twitter.com/charlieyouai (twitter.com/charlieyouai)) 02:43 How Josh got started in CS and ML 11:05 Why Josh worked on ML for robotics 15:03 ML for Robotics research at OpenAI 28:20 Josh's research process 34:56 Why putting ML into production is so difficult 44:46 What Josh thinks the ML Ops landscape will look like 49:49 Common mistakes that production ML teams and companies make 53:11 How ML systems will be built in the future 59:37 The most valuable skills that ML engineers should develop 01:03:50 Rapid Fire Questions Links https://course.fullstackdeeplearning.com/ (Full Stack Deep Learning) https://arxiv.org/abs/1703.06907 (Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World) https://arxiv.org/abs/1710.06425 (Domain Randomization and Generative Models for Robotic Grasping) https://deepmind.com/blog/article/neural-scene-representation-and-rendering (DeepMind Generative Query Network (GQN) paper) https://arxiv.org/abs/1911.04554 (Geometry Aware Neural Rendering) https://www2.eecs.berkeley.edu/Pubs/TechRpts/2019/EECS-2019-104.pdf (Josh's PhD Thesis) https://www.youtube.com/watch?v=x4O8pojMF0w (OpenAI Rubik's Cube Robot Hand video) https://www.wandb.com/podcast/josh-tobin (Weights and Biases interview with Josh) https://www.oreilly.com/library/view/designing-data-intensive-applications/9781491903063/ (Building Data Intensive Applications) http://creativeselection.io/ (Creative Selection)
This week we are joined by Naila Murray. Naila obtained a B.Sc. in Electrical Engineering from Princeton University in 2007. In 2012, she received her PhD from the Universitat Autonoma de Barcelona, in affiliation with the Computer Vision Center. She joined NAVER LABS Europe (then Xerox Research Centre Europe) in January 2013, working on topics including fine-grained visual categorization, image retrieval, and visual attention. From 2015 to 2019 she led the computer vision team at NLE. She currently serves as NLE's director of science. She serves/served as area chair for ICLR 2018, ICCV 2019, ICLR 2019, CVPR 2020, ECCV 2020, and programme chair for ICLR 2021. Her research interests include representation learning and multi-modal search.We discuss using sparse pairwise comparisons to learn a ranking function that is robust to outliers. We also take a look at using generative models in order to utilise once inaccessible datasets.Underrated ML Twitter: https://twitter.com/underrated_mlNaila Murray Twitter: https://twitter.com/NailaMurrayPlease let us know who you thought presented the most underrated paper in the form below: https://forms.gle/97MgHvTkXgdB41TC8Links to the papers:"Interestingness Prediction by Robust Learning to Rank" [paper]"Generative Models for Effective ML on Private Decentralized datasets" - [paper]
This Week in Machine Learning & Artificial Intelligence (AI) Podcast
Today we’re joined by Max Welling, Vice President of Technologies at Qualcomm Netherlands, and Professor at the University of Amsterdam. In case you missed it, Max joined us last year to discuss his work on Gauge Equivariant CNNs and Generative Models - the 2nd most popular episode of 2019. In this conversation, we explore the concept and Max’s work in neural augmentation, and how it’s being deployed for channel tracking and other applications. We also discuss their current work on federated learning and incorporating the technology on devices to give users more control over the privacy of their personal data. Max also shares his thoughts on quantum mechanics and the future of quantum neural networks for chip design. The complete show notes for this episode can be found at twimlai.com/talk/398. This episode is sponsored by Qualcomm Technologies.
Show Notes:(2:02) Josh studied Mathematics at Columbia University during his undergraduate and explained why he was not set out for a career as a mathematician.(3:55) Josh then worked for two years as a Management Consultant at McKinsey.(6:05) Josh explained his decision to go back to graduate school and pursue a Ph.D. in Mathematics at UC Berkeley.(7:23) Josh shared the anecdote of taking a robotics class with professor Pieter Abbeel and switching to a Ph.D. in the Computer Science department at UC Berkeley.(8:50) Josh described the period where he learned programming to make the transition from Math to Computer Science.(10:46) Josh talked about the opportunity to collaborate and then work full-time as a Research Scientist at OpenAI - all during his Ph.D.(12:40) Josh discussed the sim2real problem, as well as the experiments conducted in his first major work "Transfer from Simulation to Real World through Learning Deep Inverse Dynamics Model".(17:43) Josh discussed his paper "Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World", which has been cited more than 600 times up until now.(20:51) Josh unpacked the OpenAI’s robotics system that was trained entirely in simulation and deployed on a physical robot, which can learn a new task after seeing it done once (Read the blog post “Robots That Learn” and watch the corresponding video).(24:01) Josh went over his work on Hindsight Experience Replay - a novel technique that can deal with sparse and binary rewards in Reinforcement Learning (Read the blog post “Generalizing From Simulation").(28:41) Josh talked about the paper "Domain Randomization and Generative Models for Robotic Grasping”, which (1) explores a novel data generation pipeline for training a deep neural network to perform grasp planning that applies the idea of domain randomization to object synthesis; and (2) proposes an autoregressive grasp planning model that maps sensor inputs of a scene to a probability distribution over possible grasps.(32:27) Josh unpacked the design of OpenAI's Dactyl - a reinforcement learning system that can manipulate objects using a Shadow Dexterous Hand (Read the paper “Learning Dexterous In-Hand Manipulation” and watch the corresponding video).(35:31) Josh reflected on his time at OpenAI.(36:05) Josh investigated his most recent work called “Geometry-Aware Neural Rendering” - which tackles the neural rendering problem of understanding the 3D structure of the world implicitly.(28:21) Check out Josh's talk "Synthetic Data Will Help Computer Vision Make the Jump to the Real World" at the 2018 LDV Vision Summit in New York.(28:55) Josh summarized the mental decision tree to debug and improve the performance of neural networks, as a reference to his talk "Troubleshooting Deep Neural Networks” at Reinforce Conf 2019 in Budapest.(41:25) Josh discussed the limitations of domain randomization and what the solutions could look like, as a reference to his talk "Beyond Domain Randomization” at the 2019 Sim2Real workshop in Freiburg.(44:52) Josh emphasized the importance of working on the right problems and focusing on the core principles in machine learning for junior researchers who want to make a dent in the AI research community.(48:30) Josh is a co-organizer of Full-Stack Deep Learning, a training program for engineers to learn about production-ready deep learning.(50:40) Closing segment.His Contact Information:WebsiteLinkedInTwitterGitHubGoogle ScholarHis Recommended Resources:Full-Stack Deep LearningPieter AbbeelIlya SutskeverLukas Biewald“Thinking Fast and Slow” by Daniel Kahneman
Show Notes:(2:02) Josh studied Mathematics at Columbia University during his undergraduate and explained why he was not set out for a career as a mathematician.(3:55) Josh then worked for two years as a Management Consultant at McKinsey.(6:05) Josh explained his decision to go back to graduate school and pursue a Ph.D. in Mathematics at UC Berkeley.(7:23) Josh shared the anecdote of taking a robotics class with professor Pieter Abbeel and switching to a Ph.D. in the Computer Science department at UC Berkeley.(8:50) Josh described the period where he learned programming to make the transition from Math to Computer Science.(10:46) Josh talked about the opportunity to collaborate and then work full-time as a Research Scientist at OpenAI - all during his Ph.D.(12:40) Josh discussed the sim2real problem, as well as the experiments conducted in his first major work "Transfer from Simulation to Real World through Learning Deep Inverse Dynamics Model".(17:43) Josh discussed his paper "Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World", which has been cited more than 600 times up until now.(20:51) Josh unpacked the OpenAI’s robotics system that was trained entirely in simulation and deployed on a physical robot, which can learn a new task after seeing it done once (Read the blog post “Robots That Learn” and watch the corresponding video).(24:01) Josh went over his work on Hindsight Experience Replay - a novel technique that can deal with sparse and binary rewards in Reinforcement Learning (Read the blog post “Generalizing From Simulation").(28:41) Josh talked about the paper "Domain Randomization and Generative Models for Robotic Grasping”, which (1) explores a novel data generation pipeline for training a deep neural network to perform grasp planning that applies the idea of domain randomization to object synthesis; and (2) proposes an autoregressive grasp planning model that maps sensor inputs of a scene to a probability distribution over possible grasps.(32:27) Josh unpacked the design of OpenAI's Dactyl - a reinforcement learning system that can manipulate objects using a Shadow Dexterous Hand (Read the paper “Learning Dexterous In-Hand Manipulation” and watch the corresponding video).(35:31) Josh reflected on his time at OpenAI.(36:05) Josh investigated his most recent work called “Geometry-Aware Neural Rendering” - which tackles the neural rendering problem of understanding the 3D structure of the world implicitly.(28:21) Check out Josh's talk "Synthetic Data Will Help Computer Vision Make the Jump to the Real World" at the 2018 LDV Vision Summit in New York.(28:55) Josh summarized the mental decision tree to debug and improve the performance of neural networks, as a reference to his talk "Troubleshooting Deep Neural Networks” at Reinforce Conf 2019 in Budapest.(41:25) Josh discussed the limitations of domain randomization and what the solutions could look like, as a reference to his talk "Beyond Domain Randomization” at the 2019 Sim2Real workshop in Freiburg.(44:52) Josh emphasized the importance of working on the right problems and focusing on the core principles in machine learning for junior researchers who want to make a dent in the AI research community.(48:30) Josh is a co-organizer of Full-Stack Deep Learning, a training program for engineers to learn about production-ready deep learning.(50:40) Closing segment.His Contact Information:WebsiteLinkedInTwitterGitHubGoogle ScholarHis Recommended Resources:Full-Stack Deep LearningPieter AbbeelIlya SutskeverLukas Biewald“Thinking Fast and Slow” by Daniel Kahneman
Susan Magsamen, the founder and executive director of the International Arts + Mind Lab at Johns Hopkins University, speaks with us about neuroaesthetics, the importance of self-expression, and the need for a from-the-ground-up “generative model” in policy and politics.
One in a series of talks from the 2019 Models of Consciousness conference. Inês Hipólito University of Wollongong Building on the modular architecture of mind (Fodor 1983), Modularity Networks is claimed as a theory well equipped to explain neural connectivity and reuse (Stanley et al.; 2019, Zerrili 2019). This paper takes the case of the oculomotor system to show that even if Modularity Network’s tools are useful to describe brain’s functional connectivity, they are limited in explaining why such connections are formed and dynamic. To show this, section 1 starts by laying down the reasons for adopting Modularity Networks as well suited for explaining neural connectivity. Section 2 introduces the oculomotor system as a dynamic integration of action and vision. Section 3 argues that however valuable in describing the functional connectivity of the oculomotor system, Modularity Networks fails to explain why such connections are formed and dynamic (dependent on activity). This failure is made evident by acknowledging a fundamental distinction in the metaphysics of inference. The nature of inference is taken differently in functional connectivity as a description of inference as opposed to effective connectivity as an explanation of inference (Friston 2011). Section 4 introduces Dynamic Causal Modelling (DCM) as a better resource to capture effective connectivity. It allows explaining how and why brain connections, as generative models of cognitive integration, are dependent on the dynamic activity within the environment. This conclusion speaks against modular arguments for encapsulation, innateness and specificity of cognitive organisation. Filmed at the Models of Consciousness conference, University of Oxford, September 2019.
This Week in Machine Learning & Artificial Intelligence (AI) Podcast
Today we’re joined by Max Welling, research chair in machine learning at the University of Amsterdam, as well as VP of technologies at Qualcomm, and Fellow at the Canadian Institute for Advanced Research, or CIFAR. In our conversation, we discuss: • Max’s research at Qualcomm AI Research and the University of Amsterdam, including his work on Bayesian deep learning, Graph CNNs and Gauge Equivariant CNNs, and in power efficiency for AI via compression, quantization, and compilation. • Max’s thoughts on the future of the AI industry, in particular, the relative importance of models, data and compute. The complete show notes for this episode can be found at twimlai.com/talk/267. Thanks to Qualcomm for sponsoring today's episode! Check out what they're up to at twimlai.com/qualcomm.
This Week in Machine Learning & Artificial Intelligence (AI) Podcast
Today we’re joined by Gerald Quon, assistant professor in the Molecular and Cellular Biology department at UC Davis. Gerald presented his work on Deep Domain Adaptation and Generative Models for Single Cell Genomics at GTC this year, which explores single cell genomics as a means of disease identification for treatment. In our conversation, we discuss how Gerald and his team use deep learning to generate novel insights across diseases, the different types of data that was used, and the development of ‘nested’ Generative Models for single cell measurement. The complete show notes for this episode can be found at https://twimlai.com/talk/251. Visit twimlai.com/gtc19 for more from our GTC 2019 series. To learn more about Dell Precision workstations, and some of the ways they’re being used by customers in industries like Media and Entertainment, Engineering and Manufacturing, Healthcare and Life Sciences, Oil and Gas, and Financial services, visit Dellemc.com/Precision.
Google Brain is an engineering team focused on deep learning research and applications. One growing area of interest within Google Brain is that of generative models. A generative model uses neural networks and a large data set to create new data similar to the ones that the network has seen before. One approach to making The post Generative Models with Doug Eck appeared first on Software Engineering Daily.
Google Brain is an engineering team focused on deep learning research and applications. One growing area of interest within Google Brain is that of generative models. A generative model uses neural networks and a large data set to create new data similar to the ones that the network has seen before. One approach to making The post Generative Models with Doug Eck appeared first on Software Engineering Daily.
Machine learning has been making big strides in a lot of straightforward tasks, such as taking an image and labeling the objects in it. But what if you want an algorithm that can, for example, generate an image of an object? That's a much vaguer and more difficult request. And it's where generative models come in! We discuss the motivation for making generative models (in addition to making cool images) and how they help us understand the core components of our data. We also get into the specific types of generative models and how they can be trained to create images, text, sound and more. We then move onto the practical concerns that would arise in a world with good generative models: fake videos of politicians, AI assistants making our phone calls, and computer-generated novels. Finally, we connect these ideas to neuroscience, asking both how can neuroscientists make use of these and is the brain a generative model?
Paper by Ryan Cotterell and Jason Eisner, presented by Matt. This paper won the best paper award at ACL 2017. It's also quite outside the typical focus areas that you see at NLP conferences, trying to build generative models of vowel vocabularies in languages. That means we give quite a bit of set up, to try to help someone not familiar with this area understand what's going on. That makes this episode quite a bit longer than a typical non-interview episode. https://www.semanticscholar.org/paper/Probabilistic-Typology-Deep-Generative-Models-of-V-Cotterell-Eisner/6fad97c4fe0cfb92478d8a17a4e6aaa8637d8222
One of Android's main defense mechanisms against malicious apps is a risk communication mechanism which, before a user installs an app, warns the user about the permissions the app requires, trusting that the user will make the right decision. This approach has been shown to be ineffective as it presents the risk information of each app in a "stand-alone" fashion and in a way that requires too much technical knowledge and time to distill useful information.We introduce the notion of risk scoring and risk ranking for Android apps, to improve risk communication for Android apps, and identify three desiderata for an effective risk scoring scheme. We propose to use probabilistic generative models for risk scoring schemes, and identify several such models, ranging from the simple Naive Bayes, to advanced hierarchical mixture models. Experimental results conducted using real-world datasets show that probabilistic generative models significantly outperform existing approaches, and that Naive Bayes models give a promising risk scoring approach. About the speaker: Christopher Gates is a PhD student in the Computer Science department of Purdue University and a member of CERIAS. He received his Masters Degree in Computer Science in 2005 from Rutgers University, and then worked at a startup company in NYC before deciding to pursue his PhD. His research interests are in information security and machine learning. In particular, his research focuses on using data to help users make more informed and safer security decisions. His research advisor is Prof. Ninghui Li.
One of Android's main defense mechanisms against malicious apps is a risk communication mechanism which, before a user installs an app, warns the user about the permissions the app requires, trusting that the user will make the right decision. This approach has been shown to be ineffective as it presents the risk information of each app in a “stand-alone” fashion and in a way that requires too much technical knowledge and time to distill useful information. We introduce the notion of risk scoring and risk ranking for Android apps, to improve risk communication for Android apps, and identify three desiderata for an effective risk scoring scheme. We propose to use probabilistic generative models for risk scoring schemes, and identify several such models, ranging from the simple Naive Bayes, to advanced hierarchical mixture models. Experimental results conducted using real-world datasets show that probabilistic generative models significantly outperform existing approaches, and that Naive Bayes models give a promising risk scoring approach.