Podcasts about oom

Live Slow Ride Fast Podcast

Play Episode Listen Later Jun 17, 2024 31:42

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: Questionable Narratives of "Situational Awareness", published by fergusq on June 17, 2024 on The Effective Altruism Forum. Introduction This is a response to the Situational Awareness essay series by Leopold Aschenbrenner. As a disclaimer, I am an AI pessimist, meaning that I don't believe there is evidence for AGI appearing any time soon. I do also believe that even if you are an AI optimist, you should view Aschenbrenner's text critically, as it contains numerous flawed arguments and questionable narratives, which I will go through in this post. The text has numerous dubious technical claims and flawed arguments, including misleading statements regarding RLHF[1], uncited claims of human intelligence[2], use of made-up units such as OOM[3] without any serious technical argumentation, use of made-up charts that extrapolate these made-up units, claims that current models could be "unhobbled"[4], and baseless claims such as that current AI is at the level of a preschooler or a high school student[5]. I have given some thoughts on these in the footnotes, although I don't consider myself the best person to criticize them. Instead, I will be focusing more on the narrative structure of the text, which I think is more important than the technical part. After reading this text, it gave me heavy propaganda-vibes, as if it were a political piece that tries to construct a narrative that aims to support certain political goals. Its technical argumentation is secondary to creating a compelling narrative (or a group of narratives). I will first go through the two most problematic narratives, the conspiracy-esque and US-centric narratives. Then, I will talk a bit about the technological narrative, which is the main narrative of the text. I stress that I don't necessarily believe that there is any malign intent behind these narratives, or that Aschenbrenner is trying to intentionally mislead people with them. However, I believe they should be pointed out, as I think these narratives are harmful to the AI safety community. The concepts of AGI and intelligence explosion are outlandish and suspicious to people not accepting them. Using narratives often utilized by bad-faith actors makes it easier for readers to just discard what is being said. Conspiracy narratives The text opens with a description of how the writer is part of a very small group of enlightened people who have learned the truth: 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. [...] Perhaps they will be an odd footnote in history, or perhaps they will go down in history like Szilard and Oppenheimer and Teller. If they are seeing the future even close to correctly, we are in for a wild ride. This invokes a conspiracy theory narrative that the world is "asleep" and must "wake up", and only a small group of conspirators and enlightened individuals know what is really going on. This is then compared to real-life "conspiracies" such as the Manhattan project to draw credibility for such narratives while ignoring the clear differences to them, such that the Manhattan project was a highly-organized goal-directed attempt to construct a weapon, which is not remotely similar to the decentralized actors currently developing AI systems. Later in the text, a hypothetical "AGI Manhattan Project" is described, further trying to frame the current AI discussion as being similar to the discussion that happened the couple of years before the Manhattan project in real life. Again, this ignores the fact that AI is being researched by thousands of people across the world, both in universities and in companies, and it has clear commercial value, wh...

ai san francisco manhattan speech narrative ea oppenheimer questionable teller agi situational awareness rationalist oom rlhf szilard

"De etappe van zaterdag - daar zal Pogi nog een keer uithalen"

Play Episode Listen Later May 24, 2024 35:07

Laurens en Stefan gaan verder. Lau vanuit Oklahoma, Stefan vanuit de Bajes. Giro mijmeringen en gravelgedoe alom. Natuurlijk komt de val van Ivar aan bod, en de kleine verhalen uit deze Giro: die van de Oom van Georg, over de DNF van Piccolo en de vele redenen waarom Tim Merlier een baas is.En hoe zit het nou met die trui van Karsten? Je hoort het allemaal, in de Live Slow Ride Fast podcastNordVPN proberen? Ga naar nordvpn.com/ridefast voor een kletser van een korting + vier maanden extra gratis! Risicovrij, snel en veilig. En: 30 dagen geld-terug-garantie. Dus ga naar nordvpn.com/ridefast

oklahoma ga acast zo daar giro georg dus neem piccolo natuurlijk karsten meld laurens dnf lau ivar keer etappe zaterdag oom tim merlier live slow ride fast

LW - Scaling of AI training runs will slow down after GPT-5 by Maxime Riché

Play Episode Listen Later Apr 26, 2024 6:25

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: Scaling of AI training runs will slow down after GPT-5, published by Maxime Riché on April 26, 2024 on LessWrong. My credence: 33% confidence in the claim that the growth in the number of GPUs used for training SOTA AI will slow down significantly directly after GPT-5. It is not higher because of (1) decentralized training is possible, and (2) GPT-5 may be able to increase hardware efficiency significantly, (3) GPT-5 may be smaller than assumed in this post, (4) race dynamics. TLDR: Because of a bottleneck in energy access to data centers and the need to build OOM larger data centers. Update: See Vladimir_Nesov's comment below for why this claim is likely wrong, since decentralized training seems to be solved. The reasoning behind the claim: Current large data centers consume around 100 MW of power, while a single nuclear power plant generates 1GW. The largest seems to consume 150 MW. An A100 GPU uses 250W, and around 1kW with overheard. B200 GPUs, uses ~1kW without overhead. Thus a 1MW data center can support maximum 1k to 2k GPUs. GPT-4 used something like 15k to 25k GPUs to train, thus around 15 to 25MW. Large data centers are around 10-100 MW. This is likely one of the reason why top AI labs are mostly only using ~ GPT-4 level of FLOPS to train new models. GPT-5 will mark the end of the fast scaling of training runs. A 10-fold increase in the number of GPUs above GPT-5 would require a 1 to 2.5 GW data center, which doesn't exist and would take years to build, OR would require decentralized training using several data centers. Thus GPT-5 is expected to mark a significant slowdown in scaling runs. The power consumption required to continue scaling at the current rate is becoming unsustainable, as it would require the equivalent of multiple nuclear power plants. I think this is basically what Sam Altman, Elon Musk and Mark Zuckerberg are saying in public interviews. The main focus to increase capabilities will be one more time on improving software efficiency. In the next few years, investment will also focus on scaling at inference time and decentralized training using several data centers. If GPT-5 doesn't unlock research capabilities, then after GPT-5, scaling capabilities will slow down for some time towards historical rates, with most gains coming from software improvements, a bit from hardware improvement, and significantly less than currently from scaling spending. Scaling GPUs will be slowed down by regulations on lands, energy production, and build time. Training data centers may be located and built in low-regulation countries. E.g., the Middle East for cheap land, fast construction, low regulation, and cheap energy, thus maybe explaining some talks with Middle East investors. Unrelated to the claim: Hopefully, GPT-5 is still insufficient for self-improvement: Research has pretty long horizon tasks that may require several OOM more compute. More accurate world models may be necessary for longer horizon tasks and especially for research (hopefully requiring the use of compute-inefficient real, non-noisy data, e.g., real video). "Hopefully", moving to above human level requires RL. "Hopefully", RL training to finetune agents is still several OOM less efficient than pretraining and/or is currently too noisy to improve the world model (this is different than simply shaping propensities) and doesn't work in the end. Guessing that GPT-5 will be at expert human level on short horizon tasks but not on long horizon tasks nor on doing research (improving SOTA), and we can't scale as fast as currently above that. How big is that effect going to be? Using values from: https://epochai.org/blog/the-longest-training-run, we have estimates that in a year, the effective compute is increased by: Software efficiency: x1.7/year (1 OOM in 3.9 y) Hardware efficiency: x1.3/year ...

AF - Improving SAE's by Sqrt()-ing L1 & Removing Lowest Activating Features by Logan Riggs Smith

Play Episode Listen Later Mar 15, 2024 7:23

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: Improving SAE's by Sqrt()-ing L1 & Removing Lowest Activating Features, published by Logan Riggs Smith on March 15, 2024 on The AI Alignment Forum. TL;DR We achieve better SAE performance by: Removing the lowest activating features Replacing the L1(feature_activations) penalty function with L1(sqrt(feature_activations)) with 'better' meaning: we can reconstruct the original LLM activations w/ lower MSE & with fewer features/datapoint. As a sneak peak (the graph should make more sense as we build up to it, don't worry!): Now in more details: Sparse Autoencoders (SAEs) reconstruct each datapoint in [layer 3's residual stream activations of Pythia-70M-deduped] using a certain amount of features (this is the L0-norm of the hidden activation in the SAE). Typically the higher activations are interpretable & the lowest of activations non-interpretable. Here is a feature that activates mostly on apostrophe (removing it also makes it worse at predicting "s"). The lower activations are conceptually similar, but then we have a huge amount of tokens that are something else. From a datapoint viewpoint, there's a similar story: given a specific datapoint, the top activation features make a lot of sense, but the lowest ones don't (ie if 20 features activate that reconstruct a specific datapoint, the top ~5 features make a decent amount of sense & the lower 15 make less and less sense) Are these low-activating features actually important for downstream performance (eg CE)? Or are they modeling noise in the underlying LLM (which is why we see conceptually similar datapoints in lower activation points)? Ablating Lowest Features There are a few different ways to remove the "lowest" feature activations. Dataset View: Lowest k-features per datapoint Feature View: Features have different activation values. Some are an OOM larger than others on average, so we can set feature specific thresholds. Percentage of max activation - remove all feature activations that are < [10%] of max activation for that feature Quantile - Remove all features in the [10th] percentile activations for each feature Global Threshold - Let's treat all features the same. Set all feature activations less than [0.1] to 0. It turns out that the simple global threshold performs the best: [Note: "CE" refers to the CE when you replace [layer 3 residual stream]'s activations with the reconstruction from the SAE. Ultimately we want the original model's CE with the smallest amount of feature's per datapoint (L0 norm).] You can halve the L0 w/ a small (~0.08) increase in CE. Sadly, there is an increase in both MSE & CE. If MSE was higher & CE stayed the same, then that supports the hypothesis that the SAE is modeling noise at lower activations (ie noise that's important for MSE/reconstruction but not for CE/downstream performance). But these lower activations are important for both MSE & CE similarly. For completion sake, here's a messy graph w/ all 4 methods: [Note: this was run on a different SAE than the other images] There may be a more sophisticated methods that take into account feature-information (such as whether it's an outlier feature or feature frequency), but we'll be sticking w/ the global threshold for the rest of the post. Sweeping Across SAE's with Different L0's You can get widly different L0's by just sweeping the weight on the L1 penalty term where increasing the L0 increases reconstruction but at the cost of more, potentially polysemantic, features per datapoint. Does the above phenomona extend to SAE's w/ different L0's? Looks like it does & the models seems to follow a pareto frontier. Using L1(sqrt(feature_activation)) @Lucia Quirke trained SAE's with L1(sqrt(feature_activations)) (this punishes smaller activations more & larger activations less) and anecdotally noticed less of these smaller, unintepreta...

speech improving ea activating lowest tl llm riggs percentage l1 sae mse rationalist oom

Repairing Damage to Others, Keeping Training Playful, and Red-Lining When it Counts | ASK ME ANYTHING

training damage counts ask me anything repairing new order playful lining ryan michler oom iron council

Play Episode Listen Later Mar 13, 2024 66:35

In this week's ASK ME ANYTHING, Ryan Michler and Kipp Sorensen take on your questions from the Iron Council and Order of Man Facebook Group. Hit Ryan up on Instagram at @ryanmichler and share what's working in your life. ⠀ SHOW HIGHLIGHTS (0:00) Episode Intro (20:54) What are the differences between the periods of pushing hard towards your goals to the point of overextending oneself and experiencing mental burnout? (36:48) If someone had started listening to the OOM pod years ago but stopped, should they pick back up where they left off, or start with the most recent episode? (39:46) What is the process you'd recommend to someone choosing their team on The Iron Council? (42:38) When do you know when it's time to let go of a friend? (49:00) Am I sending my kids the wrong impression about physical presentation when I'm less careful about it at home? (54:15) How do we begin to repair and own up to the mistakes and repercussions of them that we might have made in regards to our family? Order of Man Merchandise. Pick yours up today! Get your signed copy of Ryan's latest book, The Masculinity Manifesto Want maximum health, wealth, relationships, and abundance in your life? Sign up for our free course, 30 Days to Battle Ready ⠀ Download the NEW Order of Man Twelve-Week Battle Planner App and maximize your week.

game confidence european golf penge dp world tour walker cup challenge tour oom theopen

Play Episode Listen Later Feb 24, 2024 33:18

Introducing Five and Nine Live!Following up on Episode 4.07. Awakening the Healer, we present our very first live podcast recording — a special episode for Season 4 recorded live at the Center for the Enlightenment Arts in Bushwick, Brooklyn, at Gathering Future Ancestors: a meditative experience of immersive storytelling, participatory art, and sound — exploring water as an entryway to our shared humanity.In this discussion, host AX Mina and guest Helen Banach talk about multihyphenate life, the practice of healing and the difference between holding space and holding pain. They close with a tarot reading for the collective using the Mixed Signals Tarot deck designed by M Eifler, with insights on how to navigate grief, cultivate care and tend to the fire within all of us to build differently in this time of deep crisis.The Season 4 finale will be coming out in the next few weeks. In the meantime, check out our most recent episodes with the 2023 residents of One of Many Studio's Wadi Rum residency:* Azad's Kite: Art in a Time of Crisis* Learning to Trust the Gut* It's More Clear at the Bottom* PLUS a bonus minisode of our Embodied Water PracticeTarot Cards Discussed* Five of Cups* Seven of Cups* Eight of Wands* Queen of Swords (aka Maker of Swords in the Mixed Signals Deck)Images of the cards are available at thisisfiveandnine.com.Season 4 is co-presented with One of Many Studio (OoM), an experience design studio at the intersection of healing and justice work. We envision a movement of Future Ancestors built through transformative, intentionally curated experiences. Connect with OoM on Substack.Five and Nine is a podcast and newsletter at the crossroads of magic, work and economic justice. We publish “moonthly” — every new moon

time trust awakening healers maker substack swords centering bushwick azad wadi rum oom healing gifts

EP71 - Marco Penge | CONFIDENCE is KEY to my game!

onepodlife

Play Episode Listen Later Jan 22, 2024 43:31

Marco Penge has rising to the top of the European golf pyramid by winning the Challenge Tour season-long OOM, but it hasn't been easy. After finding success in his early years, Marco almost made the Walker Cup team but with some poor results he started to suffer on and off the course. Marco opens up about the struggles with his mental health and the actions he took to become a DP World Tour player. As always, Marco also shares what he thinks are the top three things any aspiring golfer should be doing. Getting the insight from someone who has made it to the top is also valuable and Marco definitely gave a good account of what it takes. Like ONEPODLIFE? Share it with your friends and leave a review on whatever platform you are listening from, I would love to hear your thoughts. FOLLOW my journey on Instagram @oneputtlife Learn more about me by visiting oneputtlife.com

Play Episode Listen Later Jan 17, 2024 15:46

Symbolic language is at the heart of magic and also at the heart of human expression. In this minisode, Five and Nine joins Adelle Lin, an artist and technologist who explores the poetry of water and the Arabic, Chinese and Hebrew languages in the desert. We talk about time, displacement, and what it feels like to roll a 160 foot scroll down the side of a hill. Also: Ana offers a reflection on deep time and where water comes from.Five and Nine Season 4 is about crisis. It's our first season recorded on location, with the majority of our episodes produced in the beautiful Wadi Rum desert in Jordan, thanks to a special artist residency hosted by One of Many Studio (OoM).Music performed by Hashim Bin Muaitiq, recorded live in Wadi Rum.Season 4 is co-presented with One of Many Studio (OoM), an experience design & consulting studio connecting people with what it means to be a Future Ancestor. OoM works at the intersection of immersive experiences and social change, elevating critical conversations to transform the way we relate to ourselves, each other, and our legacies.Five and Nine is a podcast and newsletter at the crossroads of magic, work and economic justice. We publish “moonthly” — every new moon

music water chinese poetry hebrew arabic scroll symbolic wadi rum oom written language

water practice transformation slowing down embodied aliveness oom

Play Episode Listen Later Dec 21, 2023 4:24

It's the still of winter here in the northern hemisphere as the solstice arrives, and we just have a few more episodes of Season 4 left.In this bonus minisode, Season 4 co-hosts Britt Pham and Nour Batyne of One of Many Studio offer a meditation that helps us face the climate crisis by honoring our interconnectedness.“We are gathering among unceded waters,” they offer. “These waters hold the legacy and the souls of ancestors who stewarded them with humanity in the face of violence and inhumanity. We honor them.”Related episodes:*

LW - Originality vs. Correctness by alkjash

Play Episode Listen Later Dec 6, 2023 34:12

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: Originality vs. Correctness, published by alkjash on December 6, 2023 on LessWrong. I talk with Alkjash about valuing original thinking vs getting things right. We discuss a few main threads: What are the benefits of epistemic specialisation? What about generalism? How much of the action in an actual human mind is in tweaking your distribution over hypotheses and how much over making sure you're considering good hypotheses? If your hope is to slot into an epistemic process that figures out what's correct in part by you coming up with novel ideas, will processes that are out to get you make you waste your life? Intellectual generals vs supersoldiers Over time I've noticed that I care less and less about epistemic rationality - i.e. being correct - and more and more about being original. Of course the final goal is to produce thoughts that are original AND correct, but I find the originality condition more stringent and worth optimizing for. This might be a feature of working in mathematics where verifying correctness is cheap and reliable. Huh that feels like an interesting take. I don't have a super strong take on originality vs. correctness, but I do think I live my life with more of a "if you don't understand the big picture and your environment well, you'll get got, and also the most important things are 10000x more important than the median important thing, so you really need to be able to notice those opportunities, which requires an accurate map of how things work in-aggregate". Which like, isn't in direct conflict with what you are saying, though maybe is. I think I have two big sets of considerations that make me hesitant to optimize for originality over correctness (and also a bunch the other way around, but I'll argue for one side here first): The world itself is really heavy-tailed and having a good understanding of how most of the world works, while sacrificing deeper understanding of how a narrower slicer of the world works, seems worth it since behind every part of reality that you haven't considered, a crucial consideration might lurk that completely shifts what you want to be doing with your life The obvious example from an LW perspective is encountering the arguments for AI Risk vs. not and some related considerations around "living in the most important century". But also broader things like encountering the tools of proof and empirical science and learning how to program. The world is adversarial in the sense that if you are smart and competent, there are large numbers of people and institutions optimizing to get you to do things that are advantageous to them, ignoring your personal interests. Most smart people "get got" and end up orienting their lives around some random thing they don't even care about that much, because they've gotten their OODA loop captured by some social environment that makes it hard for them to understand what is going on or learn much about what they actually want to do with their lives. I think navigating an adversarial environment like this requires situational awareness and broad maps of the world, and prioritizing originality over correctness IMO makes one substantially more susceptible to a large set of attacks. Some quick gut reactions that I'll reflect/expand on: I think the world is not as heavy-tailed for most human utility functions as you claim. Revealed preferences suggest that saving the world is probably within an OOM as good (to me and most other people) as living ten years longer, or something like this. Same with the difference between $1m and $1b. One of the core heuristics I have is that your perspective (which is one that seems predominant on LW) is one of very low trust in "the intellectual community," leading to every individual doing all the computations from the ground up for themselves. It feels to...

speech revealed ea intellectual imo originality lw ooda correctness rationalist oom lesswrong

Plan Your Week: Brainstorming

focus plan prioritize brainstorming brainstorm oom

Play Episode Listen Later Oct 22, 2023 36:56

A new coaching series with an episode dropping every Sunday! Join the OOM listener community & get the forms by joining as a member here

How to Overcome Loneliness, How to Get Buy-In From Others, and Working Through and Emotional Affair | ASK ME ANYTHING

overcome loneliness ask me anything battle ready emotional affair oom iron council

Play Episode Listen Later Sep 13, 2023 43:34

In this week's ASK ME ANYTHING, Kipp Sorensen and Sean Villalovos are together “in-person” for a rare occasion to take on your questions from the Iron Council. Hit Ryan up on Instagram at @ryanmichler and share what's working in your life. ⠀ SHOW HIGHLIGHTS (0:00) Episode intro⠀ (4:00) What things would you recommend men do when they are feeling lonely? (11:30) How has 9/11 impacted you and how do you think it has impacted the country? (16:50) As men, should we share our most intimate struggles with our significant others? (22:50) Do either of you experience a shame spiral? (26:00) How can I start a fundraising operation without taking time away from my work or other obligations? (30:00) I am curious about your opinion on how to get people to buy into a mission? (34:50) When you started OOM, how long was it until you created a steady revenue stream? (39:30) What can I do to help gain trust back after my wife engaged in an emotional affair? Subscribe to the Order of Man YouTube Channel Battle Planners are back in stock. Pick yours up today! Get your signed copy of Ryan's new book, The Masculinity Manifesto For more information on the Iron Council brotherhood. Want maximum health, wealth, relationships, and abundance in your life? Sign up for our free course, 30 Days to Battle Ready

My Fave Travel Apps!

Play Episode Listen Later Sep 12, 2023 48:00

Would love to hear yours too! Join the conversation as a OOM-member by clicking the Buy Me A Coffee link here

fave travel apps oom

4 Steps To Help Kids Organize

organize memberships help kids oom

Play Episode Listen Later Sep 9, 2023 43:07

I'm adapting my 4 step organization process to help you help your kids organize their spaces. Along with some practical, tried & true tips! Join the conversation in the OOM member-only Facebook group by clicking on Memberships on my Buy Me A Coffee page here

Do These 3 Things Daily

Play Episode Listen Later Aug 28, 2023 33:00

“When life tosses you lemons make lemonade”—“They” say.

oom

Quantity vs Quality In Business

Play Episode Listen Later Aug 21, 2023 44:12

Inspired by an interesting phone call I received from someone selling me an expensive coaching program. Knowing your reason for doing business while being equally aware of which season you're in is essential in knowing how to navigate your business! Here are some tips on how to not allow others to influence you away from your goals. Also, I would

quantity oom

EA - XPT forecasts on (some) Direct Approach model inputs by Forecasting Research Institute

model speech figures ea forecasting tai forecasts inputs research institute epoch xpt rationalist oom direct approach josh rosenberg

Play Episode Listen Later Aug 21, 2023 33:50

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: XPT forecasts on (some) Direct Approach model inputs, published by Forecasting Research Institute on August 21, 2023 on The Effective Altruism Forum. This post was co-authored by the Forecasting Research Institute and Rose Hadshar. Thanks to Josh Rosenberg for managing this work, Zachary Jacobs and Molly Hickman for the underlying data analysis, Kayla Gamin for fact-checking and copy-editing, and the whole FRI XPT team for all their work on this project. Special thanks to staff at Epoch for their feedback and advice. Summary Superforecaster and expert forecasts from the Existential Risk Persuasion Tournament (XPT) differ substantially from Epoch's default Direct Approach model inputs on algorithmic progress and investment: InputEpoch (default)XPT superforecasterXPT expertNotesBaseline growth rate in algorithmic progress (OOM/year)0.21-0.650.09-0.20.15-0.23Current spending ($, millions)$60$35$60Yearly growth in spending (%)34%-91.4%6.40%-11%5.7%-19.5% Epoch: 80% confidence interval (CI) XPT: 90% CI, based on 2024-2030 forecasts Epoch: 2023 estimate XPT: 2024 median forecast Epoch: 80% CI XPT: 90% CI, based on 2024-2050 forecasts Note that there are no XPT forecasts relating to other inputs to the Direct Approach model, most notably the compute requirements parameters. Taking the Direct Approach model as given and using relevant XPT forecasts as inputs where possible leads to substantial differences in model output: OutputEpoch default inputsXPT superforecaster inputsXPT expert inputsMedian TAI arrival yearProbability of TAI by 2050Probability of TAI by 2070Probability of TAI by 2100 2036 2065 2052 70% 38% 49% 76% 53% 65% 80% 66% 74% Note that regeneration affects model outputs, so these results can't be replicated directly, and the TAI probabilities presented here differ slightly from those in Epoch's blog post. Figures given here are the average of 5 regenerations. Epoch is drawing on recent research which was not available at the time the XPT forecasters made their forecasts (the XPT closed in October 2022). Most of the difference in outputs comes down to differences in forecasts on baseline growth rate in algorithmic progress and yearly growth in spending, where XPT forecasts differ radically from the Epoch default inputs (which extrapolate historical trends). XPT forecasters' all-things-considered transformative artificial intelligence (TAI) timelines are much longer than those which the Direct Approach model outputs using XPT inputs: Source of 2070 forecastXPT superforecasterXPT expertDirect Approach model53%65%XPT postmortem survey question on probability of TAI by 20703.75%16% If you buy the assumptions of the Direct Approach model, and XPT forecasts on relevant inputs, this pushes timelines out by two to three decades compared with the default Epoch inputs. However, it still implies TAI by 2070. It seems very likely that XPT forecasters would not buy the assumptions of the Direct Approach model: their explicitly stated probabilities on TAI by 2070 are

Listen At The End Of Your Week

Play Episode Listen Later Aug 18, 2023 21:07

Pat yourself on the back my friend! You made it to the end of your week! So what's next?? Listen to this episode! Don't forget to check out the options INCLUDING the OOM listener membership! Click here for my Buy Me A Coffee page to join https://www.buymeacoffee.com/oomama

oom

10 Organizing Products Worth The Money

Play Episode Listen Later Aug 11, 2023 37:08

***I uploaded a video in the member-only group showing you what my closet looks like!*** What?! Another episode from me in one week?! Yes!! I'm catching up to give you more advice & insight into some fantastic organization products well worth the spend. Join the convo on the OOM member-only Facebook group & share a photo of your FAVE products! Click on the link below to support my show by buying me a cup of coffee (I'll give you a shout-out on the show!) The link is where you can also join as a member for only $9 per month! https://www.buymeacoffee.com/oomama

money products organizing fave oom

Episode 10: His Best Self

Chromythica

Play Episode Listen Later Jul 25, 2023 160:29

In our tenth episode, Professor Z and Ember hang out in the woods, make a new friend, and learn about why the forest is in a little bit of a state these days. Back at the Palace, Temerity and Oom ponder the fate of King DTF's forced guests and wind up tackling a particularly thorny problem: getting DTF to be his best self. Content notices for this episode include: references to and portrayals of ableism; alcohol use; portrayals of incarceration; strong language.This episode was recorded and produced on unceded Lisjan Ohlone land. We give our great respect to the Lisjan Ohlone and the Indigenous peoples of all the places that you, our audience, are watching and listening from. The lands that we are on have always been and will always be Indigenous lands. For more information about how you can support Lisjan Ohlone land reclamation, please visit https://sogoreate-landtrust.org. SAG-AFTRA Foundation: https://members.sagfoundation.org/donate Entertainment Community Fund: https://entertainmentcommunity.org Team Members: Justin Brown (Ember): Artistic DirectorFairuz Rougeaux (Temerity Vane): Producer & Creative DirectorAlex Rudy (Professor Z): Technical Director Esther Wallace (GM): Creative Director, Lorekeeper, and Executive Producer David Yamashiro (Oom Gildrose): Production AdvisorChromythica is a member of the Rainbow Roll Network, an event-focused, cooperative network for LGBTQIA-led, creator-owned Actual Play shows. To learn more about Rainbow Roll Network shows, please visit rainbowroll.network. Our character art and logos, credits and break music, and some theme music are by Justin Brown.Other music in this episode is sourced from zapsplat.com and the Descript media assets catalog.Chromythica uses trademarks and/or copyrights owned by Paizo Inc., used under Paizo's Community Use Policy (paizo.com/communityuse). We are expressly prohibited from charging you to use or access this content. Chromythica is not published, endorsed, or specifically approved by Paizo. For more information about Paizo Inc. and Paizo products, visit paizo.com.

indigenous dungeons and dragons palace lgbtqia best self pathfinder ttrpg descript paizo dtf justin brown entertainment community fund oom paizo inc temerity

EA - Who's right about inputs to the biological anchors model? by rosehadshar

Play Episode Listen Later Jul 24, 2023 11:42

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: Who's right about inputs to the biological anchors model?, published by rosehadshar on July 24, 2023 on The Effective Altruism Forum. In this post, I compared forecasts from Ajeya Cotra and from forecasters in the Existential Risk Persuasion Tournament (XPT) relating to some of the inputs to Cotra's biological anchors model. Here, I give my personal take on which of those forecasts seem more plausible. Note that: I'm only considering the inputs to the bio anchors model which we have XPT forecasts for. This notably excludes the 2020 training requirements distribution, which is a very important driver of model outputs. My take is based on considering the explicit arguments that Cotra and the XPT forecasters gave, rather than on independent research. My take is subjective. I've been working with the Forecasting Research Institute (who ran the XPT) since November 2022, and this is a potential source of bias. I'm publishing this post in a personal capacity and it hasn't gone through FRI's review process. I originally wrote this early in 2023. I've tried to update it as new information came out, but I likely haven't done a comprehensive job of this. To recap, here are the relevant forecasts: See workings here and here. The 'most aggressive' and 'most conservative' forecasts can be considered equivalent to 90% confidence intervals for the median estimate. Hardware For FLOP/$ in 2025, I think both Cotra and the XPGT forecasters are wrong, but Cotra will prove more right. Epoch's current estimate of highest GPU price-performance is 4.2e18 FLOP per $. They also find a trend in GPU price-performance of 0.1 OOM/year for state of the art GPUs. So I'll extrapolate 4.2e18 to 5.97E+18. For compute price halving time to 2100, I think it depends how likely you think it is that novel technologies like optical computing will reduce compute prices in future. This is the main argument Cotra puts forward for expecting such low prices. It's an argument made in XPT too, but less weight is put on it. Counterarguments given in XPT: fundamental physical limits, progress getting harder, rare materials capping how much prices can drop, catastrophe/extinction, optimisation shifting to memory architectures. Cotra mentions some but not all of these (she doesn't mention rare materials or memory architectures). Cotra flags that she thinks after 2040 her forecasts on this are pretty unreliable. But, because of how wrong their 2024 and 2030 forecasts seem to be, I'm not inclined to put much weight on XPT forecasts here either. I'll go with the most aggressive XPT figure, which is close to Cotra's. I don't have an inside view on the likelihood of novel technologies causing further price drops. Note that the disagreement about compute price halving times drives a lot of the difference in model output. Willingness to spend On the most expensive training run by 2025, I think Cotra is a bit too aggressive and XPT forecasters are much too conservative. In 2022, Cotra updated downwards a bit on the likelihood of a $1bn training run by 2025. There isn't much time left for Cotra to be right. Cotra was predicting $20m by the end of 2020, and $80m by the end of 2021. GPT-3 was $4.6m in 2020. If you buy that unreleased proprietary models are likely to be 2-8x more expensive than public ones (which Cotra argues), that XPT forecasters missed this consideration, and that GPT-3 isn't proprietary and/or unreleased (flagging because I'm unsure what Cotra actually means by proprietary/unreleased), then this could be consistent with Cotra's forecasts. Epoch estimates that GPT-4 cost $50m to train at some point in 2022. Again, this could be in line with Cotra's predictions. More importantly, GPT-4 costs make XPT forecasters look quite wrong already - their 2024 prediction was surpassed in 2022. This is especially striking i...

model speech ea flop biological gpt willingness gpu anchors gpus inputs epoch xpt rationalist oom ajeya cotra

LW - BCIs and the ecosystem of modular minds by beren

Play Episode Listen Later Jul 22, 2023 23:07

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: BCIs and the ecosystem of modular minds, published by beren on July 21, 2023 on LessWrong. Crossposted from my personal blog. Epistemic status: Much more speculative than previous posts but points towards an aspect of the future that is becoming clearer which I think is underappreciated at present. If you are interested in any of these thoughts please reach out. For many years, the primary AI risk model was one of rapid take-off (FOOM) of a single AI entering a recursive self-improvement loop and becoming utterly dominant over humanity. There were lots of debates about whether this 'fast-takeoff' model was correct or whether instead we would enter a slow-takeoff regime. In my opinion, the evidence is pretty definitive that at the moment we are entering a slow-takeoff regime, and arguably have been in it for the last few years (historically takeoff might be dated to the release of GPT-3). The last few years have undoubtedly been years of scaling monolithic very large models. The primary mechanism of improvement has been increasing the size of a monolithic general model. We have discovered that a single large model can outperform many small, specialized models on a wide variety of tasks. This trend is especially strong for language models. We also see a similar trend in image models and other modalities where large transformer or diffusion architectures work extremely well and scaling them up in both parameter size and data leads to large and predictable gains. However, soon this scaling era will necessarily come to an end temporarily. This is necessary because the size of training runs and models is rapidly exceeding what companies can realistically spend on compute (and what NVIDIA can produce). GPT-4 training cost at least 100m. It is likely that GPT-5, or a successor run in the next few years will cost >1B. At this scale, only megacap tech companies can afford another OOM and beyond that there is only powerful nation-states, which seem to be years away. Other modalities such as visual and audio have several more OOMs of scaling to go yet but if the demand is there they can also be expended in a few years. More broadly, scaling up model training is now a firmly understood process and has moved from a science to engineering and there now exist battle-tested libraries (both internal to companies and somewhat open-source) which allow for large scale training runs to be primarily bottlenecked by hardware and not by sorting out the software and parallelism stack. Beyond a-priori considerations, there are also some direct signals. Sam Altman recently said that scaling will not be the primary mechanism for improvement in the future. Other researchers have expressed similar views. Of course scaling will continue well into the future, and there are also many low hanging fruit in efficiency improvements to be made, both in terms of parameter efficiency and data efficiency. However, if we do not reach AGI in the next few years, then it seems increasingly likely that we will not reach AGI in the near-future simply by scaling. If this is true, we will move into a slow takeoff world. AI technology will still improve, but will become much more democratized and distributed than at present. Many companies will catch up to the technological frontier and foundation model inference and even training will increasingly become a commodity. More and more of the economy will be slowly automated, although there will be a lot of lag here simply due to the large amount of low-hanging fruit, the need for maturity of the underlying software stack and business models, and simply that things progress slowly in the real world. AI progress will look a lot more like electrification (as argued by Scott Alexander) than like nuclear weapons or some other decisive technological breakthrough. What will be...

ai speech minds ea nvidia ecosystem gpt 1b sam altman agi modular scott alexander beren epistemic rationalist oom lesswrong foom

☀️ 4.00. Trailer: Live from the Wadi Rum Desert

covid-19 spotify google apple desert recommend happy solstice wadi rum oom

Play Episode Listen Later Jun 22, 2023 3:13

Happy Solstice, everyone.Five and Nine Season 4 is about crisis. It's our first field studio season, with the majority of our episodes recorded from the beautiful Wadi Rum desert in Jordan, thanks to a special artist residency hosted by One of Many Studio (OoM).We ask our guests what's changed since the beginning of the COVID-19 emergency, and how to think about our relationship to the world, the earth and the people around us. We talk global tourism, the quiet of the pandemic, the art of healing, and the world of international development.Five and Nine is pleased to present this season with OoM, an experience design & consulting studio connecting people with what it means to be a Future Ancestor. Stay tuned for Episode 1 in the coming weeks. Subscribe at thisisfiveandnine.com and on Apple, Spotify, Google and Instagram.Enjoying the show? You can support us in three ways:* Become a paid subscriber now for just $6 per month and get access to our paid programming. This podcast is always free, but paid subscribers will get access to special content, like discounts on our classes with The Shipman Agency, free listens of our standalone meditations, and a monthly tarotscope we produce in partnership with Ignota Books.* Recommend this show to others. Do you know anyone who you think might enjoy this podcast? Send them a link. Ask them to tune in. You can send them snippets of our shows on Instagram, at @fiveandnine_podcast.* Leave us a review on Apple or Spotify. Reviews help bring visibility and credibility to indie podcasts like ours and help people know what to expect when tuning in. This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit fiveandnine.substack.com/subscribe

AF - PaLM-2 & GPT-4 in "Extrapolating GPT-N performance" by Lukas Finnveden

Play Episode Listen Later May 30, 2023 11:50

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: PaLM-2 & GPT-4 in "Extrapolating GPT-N performance", published by Lukas Finnveden on May 30, 2023 on The AI Alignment Forum. Two and a half years ago, I wrote Extrapolating GPT-N performance, trying to predict how fast scaled-up models would improve on a few benchmarks. One year ago, I added PaLM to the graphs. Another spring has come and gone, and there are new models to add to the graphs: PaLM-2 and GPT-4. (Though I only know GPT-4's performance on a small handful of benchmarks.) Converting to Chinchilla scaling laws In previous iterations of the graph, the x-position represented the loss on GPT-3's validation set, and the x-axis was annotated with estimates of size+data that you'd need to achieve that loss according to the Kaplan scaling laws. (When adding PaLM to the graph, I estimated its loss using those same Kaplan scaling laws.) In these new iterations, the x-position instead represents an estimate of (reducible) loss according to the Chinchilla scaling laws. Even without adding any new data-points, this predicts faster progress, since the Chinchilla scaling laws describes how to get better performance for less compute. The appendix describes how I estimate Chinchilla reducible loss for GPT-3 and PaLM-1. Briefly: For the GPT-3 data points, I convert from loss reported in the GPT-3 paper, to the minimum of parameters and tokens you'd need to achieve that loss according to Kaplan scaling laws, and then plug those numbers of parameters and tokens into the Chinchilla loss function. For PaLM-1, I straightforwardly put its parameter- and token-count into the Chinchilla loss function. To start off, let's look at a graph with only GPT-3 and PaLM-1, with a Chinchilla x-axis. Here's a quick explainer of how to read the graphs (the original post contains more details). Each dot represents a particular model's performance on a particular category of benchmarks (taken from papers about GPT-3 and PaLM). Color represents benchmark; y-position represents benchmark performance (normalized between random and my guess of maximum possible performance). The x-axis labels are all using the Chinchilla scaling laws to predict reducible loss-per-token, number of parameters, number of tokens, and total FLOP (if language models at that loss were trained Chinchilla-optimally). Compare to the last graph in this comment, which is the same with a Kaplan x-axis. Some things worth noting: PaLM is now ~0.5 OOM of compute less far along the x-axis. This corresponds to the fact that you could get PaLM for cheaper if you used optimal parameter- and data-scaling. The smaller GPT-3 models are farther to the right on the x-axis. I think this is mainly because the x-axis in my previous post had a different interpretation. The overall effect is that the data points get compressed together, and the slope becomes steeper. Previously, the black "Average" sigmoid reached 90% at ~1e28 FLOP. Now it looks like it reaches 90% at ~5e26 FLOP. Let's move on to PaLM-2. If you want to guess whether PaLM-2 and GPT-4 will underperform or outperform extrapolations, now might be a good time to think about that. PaLM-2 If this CNBC leak is to be trusted, PaLM-2 uses 340B parameters and is trained on 3.6T tokens. That's more parameters and less tokens than is recommended by the Chinchilla training laws. Possible explanations include: The model isn't dense. Perhaps it implements some type of mixture-of-experts situation that means that its effective parameter-count is smaller. It's trained Chinchilla-optimally for multiple epochs on a 3.6T token dataset. The leak is wrong. If we assume that the leak isn't too wrong, I think that fairly safe bounds for PaLM-2's Chinchilla-equivalent compute is: It's as good as a dense Chinchilla-optimal model trained on just 3.6T tokens, i.e. one with 3.6T/20=180B parameters. This would ...

performance color speech average cnbc ea compare flop palm converting gpt kaplan chinchillas 340b rationalist 6t extrapolating oom finnveden

LW - PaLM-2 & GPT-4 in "Extrapolating GPT-N performance" by Lukas Finnveden

california europe los angeles wine spread fahrenheit linear biggest problem amy lee oom

Play Episode Listen Later May 30, 2023 11:49

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: PaLM-2 & GPT-4 in "Extrapolating GPT-N performance", published by Lukas Finnveden on May 30, 2023 on LessWrong. Two and a half years ago, I wrote Extrapolating GPT-N performance, trying to predict how fast scaled-up models would improve on a few benchmarks. One year ago, I added PaLM to the graphs. Another spring has come and gone, and there are new models to add to the graphs: PaLM-2 and GPT-4. (Though I only know GPT-4's performance on a small handful of benchmarks.) Converting to Chinchilla scaling laws In previous iterations of the graph, the x-position represented the loss on GPT-3's validation set, and the x-axis was annotated with estimates of size+data that you'd need to achieve that loss according to the Kaplan scaling laws. (When adding PaLM to the graph, I estimated its loss using those same Kaplan scaling laws.) In these new iterations, the x-position instead represents an estimate of (reducible) loss according to the Chinchilla scaling laws. Even without adding any new data-points, this predicts faster progress, since the Chinchilla scaling laws describes how to get better performance for less compute. The appendix describes how I estimate Chinchilla reducible loss for GPT-3 and PaLM-1. Briefly: For the GPT-3 data points, I convert from loss reported in the GPT-3 paper, to the minimum of parameters and tokens you'd need to achieve that loss according to Kaplan scaling laws, and then plug those numbers of parameters and tokens into the Chinchilla loss function. For PaLM-1, I straightforwardly put its parameter- and token-count into the Chinchilla loss function. To start off, let's look at a graph with only GPT-3 and PaLM-1, with a Chinchilla x-axis. Here's a quick explainer of how to read the graphs (the original post contains more details). Each dot represents a particular model's performance on a particular category of benchmarks (taken from papers about GPT-3 and PaLM). Color represents benchmark; y-position represents benchmark performance (normalized between random and my guess of maximum possible performance). The x-axis labels are all using the Chinchilla scaling laws to predict reducible loss-per-token, number of parameters, number of tokens, and total FLOP (if language models at that loss were trained Chinchilla-optimally). Compare to the last graph in this comment, which is the same with a Kaplan x-axis. Some things worth noting: PaLM is now ~0.5 OOM of compute less far along the x-axis. This corresponds to the fact that you could get PaLM for cheaper if you used optimal parameter- and data-scaling. The smaller GPT-3 models are farther to the right on the x-axis. I think this is mainly because the x-axis in my previous post had a different interpretation. The overall effect is that the data points get compressed together, and the slope becomes steeper. Previously, the black "Average" sigmoid reached 90% at ~1e28 FLOP. Now it looks like it reaches 90% at ~5e26 FLOP. Let's move on to PaLM-2. If you want to guess whether PaLM-2 and GPT-4 will underperform or outperform extrapolations, now might be a good time to think about that. PaLM-2 If this CNBC leak is to be trusted, PaLM-2 uses 340B parameters and is trained on 3.6T tokens. That's more parameters and less tokens than is recommended by the Chinchilla training laws. Possible explanations include: The model isn't dense. Perhaps it implements some type of mixture-of-experts situation that means that its effective parameter-count is smaller. It's trained Chinchilla-optimally for multiple epochs on a 3.6T token dataset. The leak is wrong. If we assume that the leak isn't too wrong, I think that fairly safe bounds for PaLM-2's Chinchilla-equivalent compute is: It's as good as a dense Chinchilla-optimal model trained on just 3.6T tokens, i.e. one with 3.6T/20=180B parameters. This would make it 6180e...

performance color speech average cnbc ea compare flop palm converting gpt kaplan chinchillas 340b rationalist 6t extrapolating oom lesswrong finnveden

Amy Lee - Solving Wine's Biggest Problem

Organic Wine Podcast

Play Episode Listen Later May 22, 2023 63:22

This is a special episode. More than an episode, it's a direct request to all of you listening right now. Here's the request: let's solve the glass bottle problem right now. If you've been listening to this podcast recently, the name OOM should be familiar to you. They're a sponsor of this podcast, and they are a company based here in my fair city that is tackling the bottle re-use challenge head on. They have begun collecting, de-labeling, cleaning and sanitizing wine bottles to re-sell. They've encountered some problems that they can't solve on their own… they need you. Or really, we all need each other. As you listen to this conversation with OOM co-founder Amy Lee, you'll see what I mean. Amy wants OOM to help eliminate single use packaging across all industries. The scope of this conversation is mainly focused on California, but this is a conversation that needs to happen and is happening everywhere. The reason I wanted to get this conversation out to you is because any of us trying to do this anywhere will encounter the same problems, and sharing these problems and their potential solutions as a global community of winemakers and wine lovers will move all of these efforts forward toward solutions much more quickly. The main issues come down to two things that all of us listening can help make happen: first, we need to use label materials that can be removed without chemical processes, and second, we need to agree on just a handful of standard bottle shapes and colors that we all use if we buy new glass. Why do we need to do this? Why is this conversation not only important, but urgent? Because glass is far and away the biggest source of emissions for the wine industry, and re-using bottles can drastically reduce the emissions associated with producing and using new glass. Also, most wine bottles do not get recycled in the US. Those of you listening in Europe do much better with your recycling, but in the US we recycle less than 31% of our wine bottles. And the bad news about recycling glass is that it produces a lot of emissions to heat glass to close to 3000 degrees Fahrenheit so that it can be re-molded. My hope is that those of you listening now can choose to alter your bottle and label purchasing behavior immediately to begin to facilitate a transition to a re-use system. If you're not a wine producer, tell your favorite producers about this opportunity. Let them know you'd like them to embrace these bottling choices and that you'd not only be okay with it, you'd love it. If you're a wine maker, get everyone at your custom crush onto the same bottles and labels. Spread this podcast and this message to everyone you know in wine. Because it will take all of us, and we'll need to work with the glass producers too. I was at a local wine fair yesterday here in Los Angeles for natural wine producers. I think every producer and supporter there was philosophically receptive to this kind of change, but what was lacking was a moment in the center of that event where someone called everyone in attendance to attention and rallied us all together as community of like-minded individuals who have a lot of power to make that change happen, and appeal to us to take action to make this happen. This is that appeal. And if you are hosting or organizing an event or know someone who is, please consider structuring that moment into your festival. Whether it's to instigate action to create a bottle re-use program, or a three-minute appeal to make any other change happen that we desperately need to make, I'm beginning to feel like these festivals are missed opportunities to do something important. We have linear systems in place right now. Linear systems can only exist if we assume the earth's resources are infinite, if we assume that we can continue to take without giving back. We all know this assumption is tragically wrong – linear systems all have dead ends, and so it's time to set up a new circular system based on the assumption that our world and its resources are precious and finite and require us to give back on the same level at which we take. This conversation is about how we start to do that. Resources, bottle skus, and label specs for a re-use system at: OrganicWinePodcast.com OOM.earth/owp To help make this positive change happen, please join our patreon community.

LW - Contra Yudkowsky on AI Doom by jacob cannell

Play Episode Listen Later Apr 24, 2023 13:05

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: Contra Yudkowsky on AI Doom, published by jacob cannell on April 24, 2023 on LessWrong. Eliezer Yudkowsky predicts doom from AI: that humanity faces likely extinction in the near future (years or decades) from a rogue unaligned superintelligent AI system. Moreover he predicts that this is the default outcome, and AI alignment is so incredibly difficult that even he failed to solve it. EY is an entertaining and skilled writer, but do not confuse rhetorical writing talent for depth and breadth of technical knowledge. I do not have EY's talents there, or Scott Alexander's poetic powers of prose. My skill points instead have gone near exclusively towards extensive study of neuroscience, deep learning, and graphics/GPU programming. More than most, I actually have the depth and breadth of technical knowledge necessary to evaluate these claims in detail. I have evaluated this model in detail and found it substantially incorrect and in fact brazenly naively overconfident. Intro Even though the central prediction of the doom model is necessarily un-observable for anthropic reasons, alternative models (such as my own, or moravec's, or hanson's) have already made substantially better predictions, such that EY's doom model has low posterior probability. EY has espoused this doom model for over a decade, and hasn't updated it much from what I can tell. Here is the classic doom model as I understand it, starting first with key background assumptions: The brain inefficiency assumption: The human brain is inefficient in multiple dimensions/ways/metrics that translate into intelligence per dollar; inefficient as a hardware platform in key metrics such as thermodynamic efficiency. The mind inefficiency or human incompetence assumption: In terms of software he describes the brain as an inefficient complex "kludgy mess of spaghetti-code". He derived these insights from the influential evolved modularity hypothesis as popularized in ev pysch by Tooby and Cosmides. He boo-hooed neural networks, and in fact actively bet against them in actions by hiring researchers trained in abstract math/philosophy, ignoring neuroscience and early DL, etc. The more room at the bottom assumption: Naturally dovetailing with points 1 and 2, EY confidently predicts there is enormous room for further hardware improvement, especially through strong drexlerian nanotech. The alien mindspace assumption: EY claims human mindspace is an incredibly narrow twisty complex target to hit, whereas the space of AI mindspace is vast, and AI designs will be something like random rolls from this vast alien mindspace resulting in an incredibly low probability of hitting the narrow human target. Doom naturally follows from these assumptions: Sometime in the near future some team discovers the hidden keys of intelligence and creates a human-level AGI which then rewrites its own source code, initiating a self improvement recursion which quickly results in the AGI developing strong nanotech and killing all humans within a matter of days or even hours. If assumptions 1 and 2 don't hold (relative to 3) then there is little to no room for recursive self improvement. If assumption 4 is completely wrong then the default outcome is not doom regardless. Every one of his key assumptions is mostly wrong, as I and others predicted well in advance. EY seems to have been systematically overconfident as an early futurist, and then perhaps updated later to avoid specific predictions, but without much updating his mental models (specifically his nanotech-woo model, as we will see). Brain Hardware Efficiency EY correctly recognizes that thermodynamic efficiency is a key metric for computation/intelligence, and he confidently, brazenly claims (as of late 2021), that the brain is about 6 OOM from thermodynamic efficiency limits: Which brings me to...

ai speech doom naturally ea ey dl gpu agi scott alexander eliezer yudkowsky rationalist cannell oom yudkowsky lesswrong tooby cosmides

LW - The Brain is Not Close to Thermodynamic Limits on Computation by DaemonicSigil

Play Episode Listen Later Apr 24, 2023 8:10

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: The Brain is Not Close to Thermodynamic Limits on Computation, published by DaemonicSigil on April 24, 2023 on LessWrong. Introduction This post is written as a response to jacob_cannel's recent post Contra Yudkowsky on AI Doom. He writes: EY correctly recognizes that thermodynamic efficiency is a key metric for computation/intelligence, and he confidently, brazenly claims (as of late 2021), that the brain is about 6 OOM from thermodynamic efficiency limits EY is just completely out of his depth here: he doesn't seem to understand how the Landauer limit actually works, doesn't seem to understand that synapses are analog MACs which minimally require OOMs more energy than simple binary switches, doesn't seem to understand that interconnect dominates energy usage regardless, etc. Most of Jacob's analysis for brain efficiency is contained in this post: Brain Efficiency: Much More than You Wanted to Know. I believe this analysis is flawed with respect to the thermodynamic energy efficiency of the brain. That's the scope of this post: I will respond to Jacob's claims about thermodynamic limits on brain energy efficiency. Other constraints are out of scope, as is a discussion of the rest of the analysis in Brain Efficiency. The Landauer limit Just to review quickly, the Landauer limit says that erasing 1 bit of information has an energy cost of kTlog2. This energy must be dissipated as heat into the environment. Here k is Boltzmann's constant, while T is the temperature of the environment. At room temperature, this is about 0.02 eV. Erasing a bit is something that you have to do quite often in many types of computations, and the more bit erasures your computation needs, the more energy it costs to do that computation. (To give a general sense of how many erasures are needed to do a given amount of computation: If we add n-bit numbers a and b to get a+bmod2n, and then throw away the original values of a and b, that costs n bit erasures. I.e. the energy cost is nkTlog2.) Extra reliability costs? Brain Efficiency claims that the energy dissipation required to erase a bit becomes many times larger when we try to erase the bit reliably. The key transition error probability α is constrained by the bit energy: α=e−EbkBT Here's a range of bit energies and corresponding minimal room temp switch error rates (in electronvolts): α=0.49, Eb=0.02eV α=0.01, Eb=0.1eV α=10−25, Eb=1eV This adds a factor of about 50 to the energy cost of erasing a bit, so this would be quite significant if true. To back up this claim, Jacob cites this paper by Michael P. Frank. The relevant equation is pulled from section 2. However, in that entire section, Frank is temporarily assuming that the energy used to represent the bit internally is entirely dissipated when it comes time for the bit to be erased. Dissipating that entire energy is not required by the laws of physics, however. Frank himself explicitly mentions this in the paper (see section 3): The energy used to represent the bit can be partially recovered when erasing it. Only kTlog2 must actually be dissipated when erasing a bit, even if we ask for very high reliability. (I originally became suspicious of Jacob's numbers here based on a direct calculation. Details in this comment for those interested.) Analog signals? Quoting Brain Efficiency: Analog operations are implemented by a large number of quantal/binary carrier units; with the binary precision equivalent to the signal to noise ratio where the noise follows a binomial distribution. Because of this analog representation, Jacob estimates about 6000 eV required to do the equivalent of an 8 bit multiplication. However, the laws of physics don't require us to do our floating point operations in analog. "are implemented" does not imply "have to be implemented". Digital multiplication of two 8 bit ...

digital brain speech limits ev ea ey analog macs eb erasing computation thermodynamics boltzmann rationalist oom landauer lesswrong dissipating

Ben Sims 'Run It Red' 098

Run it Red with Ben Sims

Play Episode Listen Later Apr 13, 2023 120:01

Run it Red's latest drop is here! full tracklist below... If you're enjoying the show, and have the means, drop a donation to some worthy charities in the link below. Stay Safe, Sims x Charity Link: biglink.to/Charities Subscribe: >>> fanlink.to/runitred Spotify Playlist: bit.ly/RUNITREDSPOTIFY 1. Mal Hombre - Sequence. Voyager 2. Cosmic Xplorer - Loophole. Fenomenos 3. Mike Storm - Returning Data. Voyager 4. We Call It Voight-Kampff - Largo. International Day Off 5. FLAWS - Structure. Knowledge Imprint 6. RNGD - Teta. Edit Select 7. MasCon - Enzym. Profano 8. Jeroen Search - Nerve Stimulation. Lazy Reflex Complex 9. Jonas Kopp - Elliptical Movement. Prophet 10. RNGD - Vice. Edit Select 11. MasCon - Eisbein. MORD 12. Linear Phase - SP-17. Edit Select 13. Jonas Kopp - Wave Movement. Prophet 14. Alexander Johansson & Mattias Fridell - Trappan. Lomsk 15. RNGD - Un Dub. Edit Select 16. Phara - Evoke. Token 17. Fireground - Gaze. Tresor 18. Elyas & DJ Sack - Solved. Seclusion 19. ANNE - Hideous. Soma 20. ARKVS - Killing The Vibe. TH Tar Hallow 21. Jeroen Search - Motor Nerve. Lazy Reflex Complex 22. Anti Patriot - Hypocrisy. Sonic Mind 23. Max Watts - Dusk (Alex Falk Remix). Faith Beat 24..Mark Broom - Tube. Rekids 25. Developer - Throne To Throne. Modular 26. Alexander Johansson & Mattias Fridell - Mitt i Prick. Lomsk 27. Truncate - Untitled Tool. Common Sense 28. MarAxe - Warp Drive. Mind Burn 29. Stroef - Outer Courts. Unreleased 30. Mark Broom_Hardgroove 4 Life. Unreleased 31. Alexa Strange - Now. Unreleased 32. ANNĒ - Blind. Mutual Rytm 33. Daryl Stay - Recoil. PLOY 34. Zadig - Tape 11. Construct Re-Form 35. RNGD - Rojo Amargo. Edit Select 36. ANNE - Cardigan. Soma 37. Fadi Mohem - Relic. Spandau20 38. Rebbeca Goldberg - Automated (Mark Broom Remix). U-Trax 39. Hemissi - Shyd. Symbolism 40. Temudo - You Spelled Corn Wrong. Blueprint 41. Temudo - Lea & Blanche. Blueprint 42. A Thousand Details - John the Dentist. Oom 43. Get Cosy - Vottak. Bunkaball 44. Zadig - Tape 8. Construct Re-Form 45. Samuel L Session & Van Czar - Morning Debris. Symbolism 46. Orlando Voorn (feat. Emil & Boo Williams) - 909 (OV Remix). Heist 47. Anti Patriot - Newspeak. Sonic Mind 48. Ritzi Lee - Pointwise Iteration. Symbolism 49. Vohkinne - Seek Refuge. Emphatic 50. Ritzi Lee - S-Waves. Symbolism 51. Viels & Pyramidal Decode - Entropia Variabile. MORD 52. Ana Rs - From The Brink. Symbolism 53. PWCCA - Overall Function. Devotion 54. ROMPHY - Origins. Room 55. Alexander Johansson & Mattias Fridell - Tonkonst. Lomsk 56. Ritzi Lee - P-Waves. Symbolism 57. Dimi Angelis - Illusory Correlation. S.Lab LTD 58. Nørbak - Thibaud Said It's Sharp. Hayes 59. Ribé - Séquito. Semantica 60. Jonas Kopp- Circular Movement. Prophet 61. PWCCA - Encoded Chain. Newrhythmic 62. Trolley Route - Artificial Materials. Semantica 63. A4 - Sounds Of The Earth. Voyager 64. PWCCA - Functional Analysis. Devotion 65. Ana Rs - The Shadow Between Us. Symbolism 66. Tetelepta - 20+20. Indigo Aera 67. Trolley Route - Overlapping Shapes. Semantica 68. Trolley Route - Possessing The Organic. Semantica 69. ASA 808 - Soma 70. Photonz - Orpheus (Kara Konchar Remix). One Eyed Jacks 71. Peverelist - Pulse III. Livity Sound 72. DJ Prime Cuts - Kunta Kinte. The Trilogy Tapes

LW - The surprising parameter efficiency of vision models by beren

Play Episode Listen Later Apr 8, 2023 7:15

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: The surprising parameter efficiency of vision models, published by beren on April 8, 2023 on LessWrong. Crossposted from my personal blog. Epistemic status: This is a short post meant to highlight something I do not yet understand and therefore a potential issue with my models. I would also be interested to hear if anybody else has a good model of this. Why do vision (and audio) models work so well despite being so small? State of the art models like stable diffusion and midjourney work exceptionally well, generating near-photorealistic art and images and give users a fair degree of controllability over their generations. I would estimate with a fair degree of confidence that the capabilities of these models probably surpass the mental imagery abilities of almost all humans (they definitely surpass mine and a number of people I have talked to). However, these models are also super small in terms of parameters. The original stable diffusion is only 890M parameters. In terms of dataset size, image models are at a rough equality with humans. The stable diffusion dataset is 2 billion images. Assuming that you see 10 images per second every second you are awake and that you are awake 18 hours a day, you can observe 230 million images per year and so get the same data input as stable diffusion after 10 years. Of course, the images you see are much more redundant and we made some highly aggressive assumptions but after a human lifetime being in the same OOM as a SOTA image model is not insane. On the other hand, the hundreds of billions to trillions of tokens fed to LLMs is orders of magnitude beyond what humans could ever experience. A similar surprising smallness occurs in audio models. OpenAI's Whisper can do almost flawless audio transcription (including multilingual translation!) with just 1.6B parameters. Let's contrast this to the brain. Previously, I estimated that we should expect the visual cortex to have on the order of 100B parameters, if not more. The auditory cortex should be of roughly the same order of magnitude, but slightly smaller than the visual cortex. That is two orders of magnitude larger than state of the art DL models in these modalities. This contrasts with state of the art language models which appear to be approximately equal to the brain in parameter count and abilities. Small (1-10B) language models are clearly inferior to the brain at producing valid text and completions as well as standard question answering and factual recall tasks. Human parity in factual knowledge is reached somewhere between GPT-2 and GPT-3. Human language abilities are still not entirely surpassed with GPT-3 (175B parameters) or GPT-4 (presumably significantly larger). This puts large language models within approximately the same order of magnitude as the human linguistic cortex. What could be the reasons for this discrepancy? Off the top of my head I can think of a number which are below (and ranked by rough intuitive plausibility), and it would be interesting to try to investigate these further. Also, if anybody has ideas or evidence either way please send me a message. 1.) The visual cortex vs image models is not a fair comparison. The brain does lots of stuff image generation models can't do such as parse and render very complex visual scenes, deals with saccades and having two eyes, and, crucially, handle video data and moving stimuli. We haven't fully cracked video yet and it is plausible that to do so existing vision models require an OOM or two more of scale. 2.) There are specific inefficiencies in the brain's processing of images that image models skip which do not apply to language models. One very obvious example of this is convolutions. While CNNs have convolutional filters which are applied to all tiles of the image individually, the brain cannot do this an...

vision state human speech surprising models whispers efficiency ea openai gpt assuming dl 6b parameter sota 100b beren epistemic rationalist oom lesswrong

LW - The 0.2 OOMs/year target by Cleo Nardo

Play Episode Listen Later Mar 31, 2023 7:55

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: The 0.2 OOMs/year target, published by Cleo Nardo on March 30, 2023 on LessWrong. Paris Climate Accords In the early 21st century, the climate movement converged around a "2°C target", shown in Article 2(1)(a) of the Paris Climate Accords: The 2°C target helps facilitate coordination between nations, organisations, and individuals. It provided a clear, measurable goal. It provided a sense of urgency and severity. It promoted a sense of shared responsibility. It helped to align efforts across different stakeholders. It created a shared understanding of what success would look like. The AI governance community should converge around a similar target. 0.2 OOMs/year target In this article, I propose a target of 0.2 OOMs/year. OOM stands for "orders of magnitude", and corresponds to a ten-fold increase. This corresponds to a 58% year-on-year growth. I do not propose any specific policy for achieving the 0.2 OOMs/year target, because the purpose of the target is to unify stakeholders even if they support different policies. I do not propose any specific justification for the 0.2 OOMs/year target, because the purpose of the target is to unify stakeholders even if they have different justifications. Here is the statement: "Humanity — which includes all nations, organisations, and individuals — should limit the growth rate of machine learning training runs from 2020 until 2050 to below 0.2 OOMs/year." The statement is intentionally ambiguous about how to measure "the growth rate of machine learning training runs". I suspect that a good proxy metric would be the effective training footprint (defined below) but I don't think the proxy metric should be included in the statement of the target itself. Effective training footprint What is the effective training footprint? The effective training footprint, measured in FLOPs, is one proxy metric for the growth rate of machine learning training runs. The footprint of a model is defined, with caveats, as the total number of FLOPs used to train the model since initialisation. Caveats: A randomly initialised model has a footprint of 0 FLOPs. If the model is trained from a randomly initialised model using SGD or a variant, then its footprint is the total number of FLOPs used in the training process. If a pre-trained base model is used for the initialisation of another training process (such as unsupervised learning, supervised learning, fine-tuning, or reinforcement learning), then the footprint of the resulting model will include the footprint of the pre-trained model. If multiple models are composed to form a single cohesive model, then the footprint of the resulting model is the sum of the footprints of each component model. If there is a major algorithmic innovation which divides by a factor of r the FLOPs required to train a model to a particular score on downstream tasks, then the footprint of models trained with that innovation is multiplied by the same factor r. This list of caveats to the definition of Effective Training Footprint is non-exhaustive. Future consultations may yield additional caveats, or replace Effective Training Footprint with an entirely different proxy metric. Fixing the y-axis According to the 0.2 OOMs/year target, there cannot exist an ML model during the year (2022+x) with a footprint exceeding f(x), where f(x+1)=100.2×f(x). That means that log10f(x)=(0.2x+a) FLOPs for some fixed constant a. If we consult EpochAI's plot of compute training runs during the large-scale era of ML, we see that footprints have been growing with approximately 0.5 OOMs/year. We can use this trend to fix the value of A. In 2022, the footprint was approximately 1.0e+24. Therefore a=24. In other words, log10f(x)=0.2x+24. I have used 2022 as an anchor. If I had used 2016 instead, then the 0.2 OOMs/yr target would've been stricter. If I...

ai future target humanity speech fixing ea ml flops nardo rationalist sgd oom lesswrong

#61 Addicted to Acronyms (Rerun)

Streetwise Hebrew

Play Episode Listen Later Feb 28, 2023 8:15

We love using ראשי תיבות - acronyms - in modern Hebrew. We take the initials and between the last two letters we add inverted commas (two apostrophes) to show that it's an acronym rather than an ordinary word. Sometimes the transformation from word to acronym is so extreme that some words even end up changing their gender! New Words and Expressions: Drishat shalom chama – דְּרִישַׁת שָׁלוֹם חַמָה Timsor drishat shalom la-mishpacha – תִּמְסוֹר דְּרִישַׁת שָׁלוֹם לַמִשְׁפָּחָה Moser dash – מוֹסֵר דַּ”ש Moser dash la-chaverim – מוֹסֵר דַּ”ש לַחֲבֵרִים Timsor/timseri/timseroo le-Chayim dash – תִּמְסוֹר/תִּמְסְרִי/תִּמְסְרוּ לְחַיִים דָּ”ש Halevay ve-ha-sofash ha-ze lo yigamer af paam – הַלְוַואי וְהַסּוֹפָ”ש הַזֶּה לֹא יִיגַמֵר אַף פַּעַם Dash ham – דַּ”ש חַם Sofash, sof shavu'a – סוֹפָ”ש, סוֹף שָבוּעַ Sof ha-shavu'a – סוֹף הַשַבוּעַ She-yihye achla sofash – שֶיִּהְיֶה אַחְלָה סוֹפָ”ש Galey tsahal – גָּלֵי צַהַ”ל Tsahal, tsva ha-hagana leisrael – צַהַ”ל, צְבַא הָהַגּנָה לְיִשְרָאֵל Tsava – צָבָא Gal, galim, galey – גַּל, גַּלִּים, גַּלֵי Hu ba-tsava – הוּא בַּצָּבָא Mankal, menahel klali – מַנְכָּ”ל, מְנָהֵל כְּלָלִי Samankal, sgan mankal – סַמַנְכָֹּ”ל, סְגַן מנכ”ל Samankal ha-chevra – סַמַנְכָֹּ”ל הַחֶבְרָה Hool – חוּ”ל Az histovavta lecha be-hool – אָז הִסְתּוֹבַבְתָ לְךָ בְּחוּ”ל Rosh ha-memshala amar she-hu ba la-oom kedey lehagid et ha'emet – רֹאשׁ הַמֶּמְשָלָה אָמַר שֶׁהוּא בָּא לָאוּ”ם כְּדֵי לְהָגִיד אֶת הָאֶמֶת Oom, Oomot Me'uchadot – אוּ”ם, אוּמוֹת מְאוּחָדוֹת Be-derech klal – בְּדֶרֶך כְּלָל, בד”כ Tel Aviv – ת”א Playlist and Clips: Yardena arazi & Lahakat Ha-nachal – Drishat Shalom (Lyrics) Boaz Shar'abi & Matti Caspi – Shalom Aleychem (Lyrics) The Ultras – Sofshavu'a (Lyrics)

israel hebrew addicted lyrics sof rerun hu acronyms moser rosh new words galey tsahal oom chayim

Be Cringey! 100th Episode

Play Episode Listen Later Jan 21, 2023 55:48

I learned that sometimes being “cringey” can be okay…and that's what I've learned with the One Organized Mama podcast! On this episode I share some other “A-Ha!” moments that I've learned from this podcast and how you can apply to your life too. Also, don't forget to give me feedback on my listener form here: https://docs.google.com/forms/d/e/1FAIpQLScKcNcTdXft54x2TIP057xr_Y8IGBjH3mFW_jI8DaTRGZOhdA/viewform Also, interested in buying me a cup of coffee or joining the OOM membership? Here's the link: https://www.buymeacoffee.com/oomama

cringey oom

Overcoming Alcoholism, Regaining Control of Emotions, and Manipulative Communication | ASK ME ANYTHING

man overcoming emotions alcoholism ask me anything new order regaining manipulative ryan michler oom iron council

Play Episode Listen Later Jan 4, 2023 60:45

In this week's ASK ME ANYTHING, Ryan Michler and Kipp Sorensen take on your questions from the Iron Council - the exclusive brotherhood of the Order of Man movement. Hit Ryan up on Instagram at @ryanmichler and share what's working in your life. ⠀ SHOW HIGHLIGHTS (0:00) Show kick-off⠀ (7:00) What is your greatest achievement of 2022 and what did you fail to achieve? (26:00) Do you believe that perfection is unattainable as a martial artist? (33:20) What advice would you give for someone trying to rebuild financially after a rough year in 2022? (38:20) What is one thing that guys should be paying attention to in 2023, but are most likely not? (40:00) Do you have any advice on when to say no versus when to enjoy the moment with loved ones? (45:50) What are you going to do to make the Iron Council better in 2023? (47:30) Were there any moments in 2022 where you were reactive and wish you had been proactive? What lessons did you learn from the experience? (53:50) As the OOM has grown, has it become what you had envisioned or has that vision evolved? Get your signed copy of Ryan's new book, The Masculinity Manifesto For more information on the Iron Council brotherhood. Want maximum health, wealth, relationships, and abundance in your life? Sign up for our free course, 30 Days to Battle Ready ⠀ Download the NEW Order of Man Twelve-Week Battle Planner App and maximize your week.

Making Sense of Data with Harry Perks

Screaming in the Cloud

Play Episode Listen Later Dec 8, 2022 30:48

About HarryHarry has worked at Sysdig for over 6 years, helping organizations mature their journey to cloud native. He's witnessed the evolution of bare metal, VMs, and finally Kubernetes establish itself as the de-facto for container orchestration. He is part of the product team building Sysdig's troubleshooting and cost offering, helping customers increase their confidence operating and managing Kubernetes.Previously, Harry ran, and later sold, a cloud hosting provider where he was working hands on with systems administration. He studied information security and lives in the UK.Links Referenced:Sysdig: https://sysdig.com/ TranscriptAnnouncer: Hello, and welcome to Screaming in the Cloud with your host, Chief Cloud Economist at The Duckbill Group, Corey Quinn. This weekly show features conversations with people doing interesting work in the world of cloud, thoughtful commentary on the state of the technical world, and ridiculous titles for which Corey refuses to apologize. This is Screaming in the Cloud.Corey: This episode is brought to us by our friends at Pinecone. They believe that all anyone really wants is to be understood, and that includes your users. AI models combined with the Pinecone vector database let your applications understand and act on what your users want… without making them spell it out. Make your search application find results by meaning instead of just keywords, your personalization system make picks based on relevance instead of just tags, and your security applications match threats by resemblance instead of just regular expressions. Pinecone provides the cloud infrastructure that makes this easy, fast, and scalable. Thanks to my friends at Pinecone for sponsoring this episode. Visit Pinecone.io to understand more.Corey: This episode is brought to you in part by our friends at Veeam. Do you care about backups? Of course you don't. Nobody cares about backups. Stop lying to yourselves! You care about restores, usually right after you didn't care enough about backups. If you're tired of the vulnerabilities, costs, and slow recoveries when using snapshots to restore your data, assuming you even have them at all living in AWS-land, there is an alternative for you. Check out Veeam, that's V-E-E-A-M for secure, zero-fuss AWS backup that won't leave you high and dry when it's time to restore. Stop taking chances with your data. Talk to Veeam. My thanks to them for sponsoring this ridiculous podcast.Corey: Welcome to Screaming in the Cloud. I'm Corey Quinn. This promoted episode has been brought to us by our friends at Sysdig, and they have sent one of their principal product managers to suffer my slings and arrows. Please welcome Harry Perks.Harry: Hey, Corey, thanks for hosting me. Good to meet you.Corey: An absolute pleasure and thanks for basically being willing to suffer all of the various nonsense about to throw your direction. Let's start with origin stories; I find that those tend to wind up resonating the most. Back when I first noticed Sysdig coming into the market, because it was just launching at that point, it seemed like it was a… we'll call it an innovative approach to observability, though I don't recall that we use the term observability back then. It more or less took a look at whatever an application was doing almost at a system call level and tracing what was going on as those requests worked on an individual system, and then providing those in a variety of different forms to reason about. Is that directionally correct as far as the origin story goes, where my misremembering an evening event I went to what feels like half a lifetime ago?Harry: I'd say the latter, but just because it's a funnier answer. But that's correct. So, Sysdig was created by Loris Degioanni, one of the founders of Wireshark. And when containers and Kubernetes was being incepted, you know, it kind of created this problem where you kind of lacked visibility into what's going on inside these opaque boxes, right? These black boxes which are containers.So, we started using system calls as a source of truth for… I don't want to say observability, but observability, and using those system calls to essentially see what's going on inside containers from the outside. And leveraging system calls, we were able to pull up metrics, such as what are the golden signals of applications running in containers, network traffic. So, it's a very simple way to instrument applications. And that was really how monitoring started. And then Sysdig kind of morphed into a security product.Corey: What was it that drove that transformation? Because generally speaking, when you have a product that's in a particular space that's aimed at a particular niche pivots into something that feels as orthogonal as security don't tend to be something that you see all that often. What did you folks see that wound up driving that change?Harry: The same challenges that were being presented by containers and microservices for monitoring were the same challenges for security. So, for runtime security, it was very difficult for our customers to be able to understand what the heck is going on inside the container. Is a crypto miner being spun up? Is there malicious activity going on? So, it made logical sense to use that same data source - system calls - to understand the monitoring and the security posture of applications.Corey: One of the big challenges out there is that security tends to be one of those pervasive things—I would argue that observability does too—where once you have a position of being able to see what is going on inside of an environment and be able to reason about it. And this goes double for inside of containers, which from a cloud provider perspective, at least seems to be, “Oh, yeah, just give us the containers, we don't care what's going on inside, so we're never going to ask, notice, or care.” And being able to bridge between that lack of visibility between—from the outside of container land and inside of container land has been a perennial problem. There are security implications, there are cost implications, there are observability challenges to be sure, and of course, reliability concerns that flow directly from that, which is, I think, most people, at least historically, contextualize observability. It's a fancy word to describe is the site about to fall over and crash into the sea. At least in my experience. Is that your definition of observability, or if I basically been hijacked by a number of vendors who have decided to relabel what they'd been doing for 15 years as observability?Harry: [laugh]. I think observability is one of those things that is down to interpretation depending on what is the most recent vendor you've been speaking with. But to me, observability is: am I happy? Am I sad? Are my applications happy? Are they sad?Am I able to complete business-critical transactions that keep me online, and keep me afloat? So, it's really as simple as that. There are different ways to implement observability, but it's really, you know, you can't improve the performance, and you can't improve the security posture of things, you can't see, right? So, how do I make sure I can see everything? And what do I do with that data is really what observability means to me.Corey: The entire observability space across the board is really one of those areas that is defined, on some level, by outliers within it. It's easy to wind up saying that any given observability tool will—oh, it alerts you when your application breaks. The problem is that the interesting stuff is often found in the margins, in the outlier products that wind up emerging from it. What is the specific area of that space where Sysdig tends to shine the most?Harry: Yeah, so you're right. The outliers typically cause problems and often you don't know what you don't know. And I think if you look at Kubernetes specifically, there is a whole bunch of new problems and challenges and things that you need to be looking at that didn't exist five to ten years ago, right? There are new things that can break. You know, you've got a pod that's stuck in a CrashLoopBackOff.And hey, I'm a developer who's running my application on Kubernetes. I've got this pod in a CrashLoopBackOff. I don't know what that means. And then suddenly I'm being expected to alert on these problems. Well, how can I alert on things that I didn't even know were a problem?So, one of the things that Sysdig is doing on the observability side is we're looking at all of this data and we're actually presenting opinionated views that help customers make sense of that data. Almost like, you know, I could present this data and give it to my grandma, and she would say, “Oh, yeah, okay. You've got these pods in CrashLoopBackoff you've got these pods that are being CPU throttled. Hey, you know, I didn't know I had to worry about CPU limits, or, you know, memory limits and now I'm suffering, kind of, OOM kills.” So, I think one of the things that's quite unique about Sysdig on the monitoring side that a lot of customers are getting value from is kind of demystifying some of those challenges and making a lot of that data actionable.Corey: At the time of this recording, I've not yet bothered to run Kubernetes in anger by which I, of course, mean production. My production environment is of course called ‘Anger' similarly to the way that my staging environment is called ‘Theory' because things work in theory, but not in production. That is going to be changing in the first quarter of next year, give or take. The challenge with that, though, is that so much has changed—we'll say—since the evolution of Kubernetes into something that is mainstream production in most shops. I stopped working in production environments before that switch really happened, so I'm still at a relatively amateurish level of understanding around a lot of these things.I'm still thinking about old-school problems, like, “Okay, how big do I make each one of the nodes in my Kubernetes cluster?” Yeah, if I get big systems, it's likelier that there will be economies of scale that start factoring in fewer nodes to manage, but it does increase the blast radius if one of those nodes gets affected by something that takes it offline for a while. I'm still at the very early stages of trying to wrap my head around the nuances of running these things in a production environment. Cost is, of course, a separate argument. My clients run it everywhere and I can reason about it surprisingly well for something that is not lending itself to easy understanding it by any sense of the word and you almost have to intuit its existence just by looking at the AWS bill.Harry: No, I like your observations. And I think the last part there around costs is something that I'm seeing a lot in the industry and in our customers is, okay, suddenly, you know, I've got a great monitoring posture, or observability posture, whatever that really means. I've got a great security posture. As customers are maturing in their journey to Kubernetes, suddenly there are a bunch of questions that are being asked from atop—and we've kind of seen this internally—such as, “Hey, what is the ROI of each customer?”Or, “What is the ROI of a specific product line or feature that we deliver to our customers?”And we couldn't answer those problems. And we couldn't answer those problems because we're running a bunch of applications and software on Kubernetes and when we receive our billing reports from the multiple different cloud providers we use— Azure, AWS, and GCP—we just received a big fat bill that was compute, and we were unable to kind of break that down by the different teams and business units, which is a real problem. And one of the problems that we really wanted to start solving, both for internal uses, but also for our customers, as well.Corey: Yeah, when you have a customer coming in, the easy part of the equation is well how much revenue are we getting from a customer? Well, that's easy enough to just wind up polling your finance group and, “Yeah, how much have they paid us this year?” “Great. Good to know.” Then it gets really confusing over on the cost side because it gets into a unit economic model that I think most shops don't have a particularly advanced understanding of.If we have another hundred customers sign up this month, what will it cost us to service them? And what are the variables that change those numbers? It really gets into a fascinating model where people more or less, do some gut checks and some rounding, but there are a bunch of areas where people get extraordinarily confused, start to finish. Kubernetes is very much one of them because from a cloud provider's perspective, it's just a single-tenant app that is really gnarly in terms of its behavior, it does a bunch of different things, and from the bill alone, it's hard to tell that you're even running Kubernetes unless you ask.Harry: Yeah, absolutely. And there was a survey from the CNCF recently that said 68% of folks are seeing increased Kubernetes costs—of course—and 69% of respondents said that they have no cost monitoring in place or just cost estimates, which is simply not good enough, right? People want to break down that line item to those individual business units and in teams. Which is a huge challenge that cloud providers aren't fulfilling today.Corey: Where do you see most of the cost issue breaking down? I mean, there's some of the stuff that we are never allowed to talk about when it comes to cost, which is the realistic assessment that people to work on technology cost more than the technology itself. There's a certain—how do we put this—unflattering perspective that a lot of people are deploying Kubernetes into environments because they want to bolster their own resume, not because it's the actual right answer to anything that they have going on. So, that's a little hit or miss, on some level. I don't know that I necessarily buy into that, but you take a look at the compute storage, you look at the data transfer side, which it seems that almost everyone mostly tends to ignore, despite the fact that Kubernetes itself has no zone affinity, so it has no idea whether its internal communication is free or expensive, and it just adds up to a giant question mark.Then you look at Kubernetes architecture diagrams, or God forbid the CNCF landscape diagram, and realize, oh, my God, they have more of these things, and they do Pokemon, and people give up any hope of understanding it other than just saying, “It's complicated,” and accepting that that's just the way that it is. I'm a little less fatalistic, but I also think it's a heck of a challenge.Harry: Absolutely. I mean, the economics of cloud, right? Why is ingress free, but egress is not free? Why is it so difficult to [laugh] understand that intra AZ traffic is completely billed separately to public traffic, for example? And I think network costs is one thing that is extremely challenging for customers.One, they don't even have that visibility into what is the network traffic: what is internal traffic, what is public traffic. But then there's also a whole bunch of other challenges that are causing Kubernetes costs to rise, right? You've got folks that struggle with setting the right requests for Kubernetes, which ultimately blows up the scale of a Kubernetes cluster. You've got the complexity of AWS, for example, economics of instance types, you know? I don't know whether I need to be running ten m5.xlarge versus four, Graviton instances.And this ability to, kind of, size a cluster correctly as well as size a workload correctly is very, very difficult and customers are not able to establish that baseline today. And obviously, you can't optimize what you can't see, right, so I think a lot of customers struggle with both that visibility. But then the complexity means that it's incredibly difficult to optimize those costs.Corey: You folks are starting to dip your toes in the Kubernetes costing space. What approach are you taking?Harry: Sysdig builds products to Kubernetes first. So, if you look at what we're doing on the monitoring space, we were really kind of pioneered what customers want to get out of Kubernetes observability, and then we were doing similar things for security? So, making sure our security product is, [I want to say,] Kubernetes-native. And what we're doing on the cost side of the things is, of course, there are a lot of cost products out there that will give you the ability to slice and dice by AWS service, for example, but they don't give you that Kubernetes context to then break those costs down by teams and business units. So at Sysdig, we've already been collecting usage information, resource usage information–requests, the container CPU, the memory usage–and a lot of customers have been using that data today for right-sizing, but one of the things they said was, “Hey, I need to quantify this. I need to put a big fat dollar sign in front of some of these numbers we're seeing so I can go to these teams and management and actually prompt them to right-size.”So, it's quite simple. We're essentially augmenting that resource usage information with cost data from cloud providers. So, instead of customers saying, “Hey, I'm wasting one terabyte of memory, they can say, hey, I'm wasting 500 bucks on memory each month,” So, it's very much Kubernetes specific, using a lot of Kubernetes context and metadata.Corey: This episode is sponsored in part by our friends at Uptycs, because they believe that many of you are looking to bolster your security posture with CNAPP and XDR solutions. They offer both cloud and endpoint security in a single UI and data model. Listeners can get Uptycs for up to 1,000 assets through the end of 2023 (that is next year) for $1. But this offer is only available for a limited time on UptycsSecretMenu.com. That's U-P-T-Y-C-S Secret Menu dot com.Corey: Part of the whole problem that I see across the space is that the way to solve some of these problems internally has been when you start trying to divide costs between different teams is well, we're just going to give each one their own cluster, or their own environment. That does definitely solve the problem of shared services. The counterpoint is it solves them by making every team individually incur them. That doesn't necessarily seem like the best approach in every scenario. One thing I have learned, though, is that, for some customers, that is the right approach. Sounds odd, but that's the world we live in where context absolutely matters a lot. I'm very reluctant these days to say at a glance, “Oh, you're doing it wrong.” You eat a whole lot of crow when you do that, it turns out.Harry: I see this a lot. And I see customers giving their own business units, their own AWS account, which I kind of feel like is a step backwards, right? I don't think you're properly harnessing the power of Kubernetes and creating this, kind of, shared tenancy model, when you're giving a team their own AWS account. I think it's important we break down those silos. You know, there's so much operational overhead with maintaining these different accounts, but there must be a better way to address some of these challenges.Corey: It's one of those areas where “it depends” becomes the appropriate answer to almost anything. I'm a fan of having almost every workload have its own AWS account within the same shared AWS organization, then with shared VPCs, which tend to work out. But that does add some complexity to observing how things interact there. One of the guidances that I've given people is assume in the future that in any architecture diagram you ever put up there, that there will be an AWS account boundary between any two resources because someone's going to be doing it somewhere. And that seems to be something that AWS themselves are just slowly starting to awaken to as well. It's getting easier and easier every week to wind up working with multiple accounts in a more complicated structure.Harry: Absolutely. And I think when you start to adopt a multi-cloud strategy, suddenly, you've got so many more increased dimensions. I'm running an application in AWS, Azure, and GCP, and now suddenly, I've got all of these subaccounts. That is an operational overhead that I don't think jives very well, considering there is such a shortage of folks that are real experts—I want to say experts—in operating these environments. And that's really, you know, I think one of the challenges that isn't being spoken enough about today.Corey: It feels like so much of the time that the Kubernetes is winding up being an expression of the same way that getting into microservices was, which is, “Well, we have a people problem, we're going to solve it with this approach.” Great, but then you wind up with people adopting it where they don't have the context that applied when the stuff was originally built and designed for. Like with mono repos. Yeah, it was a problem when you had 5000 developers all try to work on the same thing and stomping each other, so breaking that apart made sense. But the counterpoint of where you wind up with companies with 20 developers and 200 microservices starts to be a little… okay, has this pendulum swung too far?Harry: Yeah, absolutely. And I think that when you've got so many people being thrown at a problem, there's lots of kinds of changes being made, there's new deployments, and I think things can spiral out of control pretty quickly, especially when it comes to costs. “Hey, I'm a developer and I've just made this change. And how do I understand, you know, what is the financial impact of this change?” “Has this blown up my network costs because suddenly, I'm not traversing the right network path?” Or, suddenly, I'm consuming so much more CPU, and actually, there is a physical compute cost of this. There's a lot of cooks in the kitchen and I think that is causing a lot of challenges for organizations.Corey: You've been working in product for a while and one of my favorite parts of being in a position where you are so close to the core of what it is your company does, is that you find it's almost impossible to not continue learning things just based upon how customers take what you built and the problems that they experienced, both that they bring you in to solve, and of course, the new and exciting problems that you wind up causing for them—or to be more charitable surfacing that they didn't realize already existed. What have you learned lately from your customers that you didn't see coming?Harry: One of the biggest problems that I've been seeing is—I speak to a lot of customers and I've maybe spoken to 40 or 50 customers over the last, you know, few months, about a variety of topics, whether it's observability, in general, or, you know, on the financial side, Kubernetes costs–and what I hear about time and time again, regardless as to the vertical or the size of the organization, is the platform teams, the people closest to Kubernetes know their stuff. They get it. But a lot of their internal customers,so the internal business units and teams, they, of course, don't have the same kind of clarity and understanding, and these are the people that are getting the most frustrated. I've been shipping software for 20 years and now I'm modernizing applications, I'm starting to use Kubernetes, I've got so many new different things to learn about that I'm simply drowning, in problems, in cloud-native problems.And I think we forget about that, right? Too often, we kind of spend time throwing fancy technology at the people, such as the, you know, the DevOps engineers, the platform teams, but a lot of internal customers are struggling to leverage that technology to actually solve their own problems. They can't make sense of this data and they can't make the right changes based off of that data.Corey: I would say that is a very common affliction of Kubernetes where so often it winds up handling things that are now abstracted away to the point where we don't need to worry about that. That's true right up until the point where they break and now you have to go diving into the magic. That's one of the reasons that I was such a fan of Sysdig when it first came out was the idea that it was getting into what I viewed at the time as operating system fundamentals and actually seeing what was going on, abstracted away from the vagaries of the code and a lot more into what system calls is it making. Great, okay, now I'm starting to see a lot of calls that it shouldn't necessarily be making, or it's thrashing in a particular way. And it's almost impossible to get to that level of insight—historically—through traditional observability tools, but being able to take a look at what's going on from a more fundamentals point of view was extraordinarily helpful.I'm optimistic if you can get to a point where you're able to do that with Kubernetes, given its enraging ecosystem, for lack of a better term. Whenever you wind up rolling out Kubernetes, you've also got to pick some service delivery stuff, some observability tooling, some log routers, and so on and so forth. It feels like by the time you're running anything in production, you've made so many choices along the way that the odds that anyone else has made the same choices you have are vanishingly small, so you're running your own bespoke unicorn somewhere.Harry: Absolutely. Flip a coin. And that's probably one [laugh] of the solutions that you're going to throw at a problem, right? And you keep flipping that coin and then suddenly, you're going to reach a combination that nobody else has done before. And you're right, the knowledge that you have gained from, I don't know, Corey Quinn Enterprises is probably not going to ring true at Harry Perks Enterprise Limited, right?There is a whole different set of problems and technology and people that, you know, of course, you can bring some of that knowledge along—there are some common denominators—but every organization is ultimately using technology in different ways. Which is problematic, right to the people that are actually pioneering some of these cloud native applications.Corey: Given my professional interest, I am curious about what it is you're doing as you start moving a little bit away from the security and observability sides and into cost observability. How are you approaching that? What are the mistakes that you see people making and how are you meeting them where they are?Harry: The biggest challenge that I am seeing is with sizing workloads and sizing clusters. And I see this time and time again. Our product shines the light on the capacity utilization of compute. And what it really boils down to is two things. Platform teams are not using the correct instance types or the combination of instance types to run the workloads for their teams, their application teams, but also application developers are not setting things like requests correctly.Which makes sense. Again, I flip a coin and maybe that's the request I'm going to set. I used to size a VM with one gig of memory, so now I'm going to size my pod with one gig of memory. But it doesn't really work like that. And of course, when you request usage is essentially my slice of the pizza that's been carved out.And even if I don't see that entire slice of pizza, it's for me, nobody else can use it. So, what we're trying to do is really help customers with that challenge. So, if I'm a developer, I would be looking at the historical usage of our workloads. Maybe it's the maximum usage or, you know, the p99 or the p95 and then setting my workload request to that. You keep doing that over the course of the different team's applications you have and suddenly, you start to establish this baseline of what is the compute actually needed to run all of these applications.And that helps me answer the question, what should I size my cluster to? And that's really important because until you've established that baseline, you can't start to do things like cluster reshaping, to pick a different combination of instance types to power your cluster.Corey: Some level, a lack of diversity in instance types is a bit of a red flag, just because it generally means that someone said, “Oh, yeah, we're going to start with this default instance size and then we'll adjust as time goes on,” and spoilers just like anything else labeled ‘TODO' in your codebase, it never gets done. So, you find yourself pretty quickly in a scenario where some workloads are struggling to get the resources they need inside of whatever that default instance size is, and on the other, you wind up with some things that are more or less running a cron job once a day and sitting there completely idle but running the whole time, regardless. And optimization and right-sizing on a lot of these scenarios is a little bit tricky. I've been something of a, I'll say, a pessimist, when it comes to the idea of right-sizing EC2 instances, just because so many historical workloads are challenging to get recertified on newer instance families and the rest, whereas when we're running on Kubernetes already, presumably everything's built in such a way that it can stop existing in a stateless way and the service still continues to work. If not, it feels like there are some necessary Kubernetes prerequisites that may not have circulated fully internally yet.Harry: Right. And to make this even more complicated, you've got applications that may be more memory intensive or CPU intensive, so understanding the ratio of CPU to memory requirements for their applications depending on how they've been architected makes this more challenging, right? I mean, pods are jumping around and that makes it incredibly difficult to track these movements and actually pick the instances that are going to be most appropriate for my workloads and for my clusters.Corey: I really want to thank you for being so generous with your time. If people want to learn more, where's the best place for them to find you?Harry: sysdig.com is where you can learn more about what Sysdig is doing as a company and our platform in general.Corey: And we will, of course, put a link to that in the show notes. Thank you so much for your time. I appreciate it.Harry: Thank you, Corey. Hope to speak to you again soon.Corey: Harry Perks, principal product manager at Sysdig. I'm Cloud Economist Corey Quinn and this is Screaming in the Cloud. If you've enjoyed this podcast, please leave a five-star review on your podcast platform of choice, whereas if you've hated this podcast, please leave a five-star review on your podcast platform of choice along with an angry, insulting comment that we will lose track of because we don't know where it was automatically provisioned.Corey: If your AWS bill keeps rising and your blood pressure is doing the same, then you need The Duckbill Group. We help companies fix their AWS bill by making it smaller and less horrifying. The Duckbill Group works for you, not AWS. We tailor recommendations to your business and we get to the point. Visit duckbillgroup.com to get started.Announcer: This has been a HumblePod production. Stay humble.

god amazon ai uk talk data cost theory cloud platform roi pokemon flip making sense screaming ui aws perks vm devops azure cpu kubernetes gcp vms xdr ec2 cncf veeam pinecone wireshark corey quinn graviton sysdig oom cnapp vpcs duckbill group loris degioanni chief cloud economist uptycs last week in aws humblepod

You Keep What You Defend, Boundaries for Work-Life Balance, and Getting Out of Your Box | ASK ME ANYTHING