Machine Learning Street Talk (MLST)

Dr. Maxwell Ramstead grills Guillaume Verdon (AKA “Beff Jezos”) who's the founder of Thermodynamic computing startup Extropic.

Guillaume shares his unique path – from dreaming about space travel as a kid to becoming a physicist, then working on quantum computing at Google, to developing a radically new form of computing hardware for machine learning. He explains how he hit roadblocks with traditional physics and computing, leading him to start his company – building "thermodynamic computers." These are based on a new design for super-efficient chips that use the natural chaos of electrons (think noise and heat) to power AI tasks, which promises to speed up AND lower the costs of modern probabilistic techniques like sampling. He is driven by the pursuit of building computers that work more like your brain, which (by the way) runs on a banana and a glass of water! 

Guillaume talks about his alter ego, Beff Jezos, and the "Effective Accelerationism" (e/acc) movement that he initiated. Its objective is to speed up tech progress in order to “grow civilization” (as measured by energy use and innovation), rather than “slowing down out of fear”. Guillaume argues we need to embrace variance, exploration, and optimism to avoid getting stuck or outpaced by competitors like China. He and Maxwell discuss big ideas like merging humans with AI, decentralizing intelligence, and why boundless growth (with smart constraints) is “key to humanity's future”.

REFS:

1. John Archibald Wheeler - "It From Bit" Concept

00:04:45 - Foundational work proposing that physical reality emerges from information at the quantum level

Learn more: https://cqi.inf.usi.ch/qic/wheeler.pdf 

2. AdS/CFT Correspondence (Holographic Principle)

00:05:15 - Theoretical physics duality connecting quantum gravity in Anti-de Sitter space with conformal field theory

https://en.wikipedia.org/wiki/Holographic_principle 

3. Renormalization Group Theory

00:06:15 - Mathematical framework for analyzing physical systems across different length scales

https://www.damtp.cam.ac.uk/user/dbs26/AQFT/Wilsonchap.pdf 

4. Maxwell's Demon and Information Theory

00:21:15 - Thought experiment linking information processing to thermodynamics and entropy

https://plato.stanford.edu/entries/information-entropy/ 

5. Landauer's Principle

00:29:45 - Fundamental limit establishing minimum energy required for information erasure

https://en.wikipedia.org/wiki/Landauer%27s_principle 

6. Free Energy Principle and Active Inference

01:03:00 - Mathematical framework for understanding self-organizing systems and perception-action loops

https://www.nature.com/articles/nrn2787 

7. Max Tegmark - Information Bottleneck Principle

01:07:00 - Connections between information theory and renormalization in machine learning

https://arxiv.org/abs/1907.07331 

8. Fisher's Fundamental Theorem of Natural Selection

01:11:45 - Mathematical relationship between genetic variance and evolutionary fitness

https://en.wikipedia.org/wiki/Fisher%27s_fundamental_theorem_of_natural_selection 

9. Tensor Networks in Quantum Systems

00:06:45 - Computational framework for simulating many-body quantum systems

https://arxiv.org/abs/1912.10049 

10. Quantum Neural Networks

00:09:30 - Hybrid quantum-classical models for machine learning applications

https://en.wikipedia.org/wiki/Quantum_neural_network 

11. Energy-Based Models (EBMs)

00:40:00 - Probabilistic framework for unsupervised learning based on energy functions

https://www.researchgate.net/publication/200744586_A_tutorial_on_energy-based_learning 

12. Markov Chain Monte Carlo (MCMC)

00:20:00 - Sampling algorithm fundamental to modern AI and statistical physics

https://en.wikipedia.org/wiki/Markov_chain_Monte_Carlo 

13. Metropolis-Hastings Algorithm

00:23:00 - Core sampling method for probability distributions

https://arxiv.org/abs/1504.01896

***SPONSOR MESSAGE***

Google Gemini 2.5 Flash is a state-of-the-art language model in the Gemini app. Sign up at https://gemini.google.com

Are the AI models you use today imposters?

Please watch the intro video we did before this: https://www.youtube.com/watch?v=o1q6Hhz0MAg

In this episode, hosts Dr. Tim Scarfe and Dr. Duggar are joined by AI researcher Prof. Kenneth Stanley and MIT PhD student Akash Kumar to discuss their fascinating paper, "Questioning Representational Optimism in Deep Learning."

Imagine you ask two people to draw a perfect skull. One is a brilliant artist who understands anatomy, the other is a machine that just traces the image. Both drawings look identical, but the artist understands what a skull is—they know where the mouth is, how the jaw works, and that it's symmetrical. The machine just has a tangled mess of lines that happens to form the right picture.

An AI with an elegant representation, has the building blocks to generate truly new ideas.

The Path Is the Goal: As Kenneth Stanley puts it, "it matters not just where you get, but how you got there". Two students can ace a math test, but the one who truly understands the concepts—instead of just memorizing formulas—is the one who will go on to make new discoveries.

The show is a mixture of 3 separate recordings we have done, the original Patreon warmup with Tim/Kenneth, the Tim/Keith "Steakhouse" recorded after the main interview, then the main interview with Kenneth/Akarsh/Keith/Tim. Feel free to skip around. We had to edit this in a rush as we are travelling next week but it's reasonably cleaned up.

TOC:

00:00:00 Intro: Garbage vs. Amazing Representations

00:05:42 How Good Representations Form

00:11:14 Challenging the "Bitter Lesson"

00:18:04 AI Creativity & Representation Types

00:22:13 Steakhouse: Critiques & Alternatives

00:28:30 Steakhouse: Key Concepts & Goldilocks Zone

00:39:42 Steakhouse: A Sober View on AI Risk

00:43:46 Steakhouse: The Paradox of Open-Ended Search

00:47:58 Main Interview: Paper Intro & Core Concepts

00:56:44 Main Interview: Deception and Evolvability

01:36:30 Main Interview: Reinterpreting Evolution

01:56:16 Main Interview: Impostor Intelligence

02:11:15 Main Interview: Recommendations for AI Research

REFS:

Questioning Representational Optimism in Deep Learning:

The Fractured Entangled Representation Hypothesis

Akarsh Kumar, Jeff Clune, Joel Lehman, Kenneth O. Stanley

https://arxiv.org/pdf/2505.11581

Kenneth O. Stanley, Joel Lehman

Why Greatness Cannot Be Planned: The Myth of the Objective

https://amzn.to/44xLaXK

Original show with Kenneth from 4 years ago:

https://www.youtube.com/watch?v=lhYGXYeMq_E

Kenneth Stanley is SVP Open Endedness at Lila Sciences

https://x.com/kenneth0stanley

Akarsh Kumar (MIT)

https://akarshkumar.com/

AND... Kenneth is HIRING (this is an OPPORTUNITY OF A LIFETIME!)

Research Engineer: https://job-boards.greenhouse.io/lila/jobs/7890007002

Research Scientist: https://job-boards.greenhouse.io/lila/jobs/8012245002

TRANSCRIPT:

https://app.rescript.info/public/share/W_T7E1OC2Wj49ccqlIOOztg2MJWaaVbovTeyxcFEQdU

What if today's incredible AI is just a brilliant "impostor"? This episode features host Dr. Tim Scarfe in conversation with guests Prof. Kenneth Stanley (ex-OpenAI), Dr. Keith Duggar (MIT), and Arkash Kumar (MIT).While AI today produces amazing results on the surface, its internal understanding is a complete mess, described as "total spaghetti" [00:00:49]. This is because it's trained with a brute-force method (SGD) that’s like building a sandcastle: it looks right from a distance, but has no real structure holding it together [00:01:45].To explain the difference, Keith Duggar shares a great analogy about his high school physics classes [00:03:18]. One class was about memorizing lots of formulas for specific situations (like the "impostor" AI). The other used calculus to derive the answers from a deeper understanding, which was much easier and more powerful. This is the core difference: one method memorizes, the other truly understands.The episode then introduces a different, more powerful way to build AI, based on Kenneth Stanley's old experiment, "Picbreeder" [00:04:45]. This method creates AI with a shockingly clean and intuitive internal model of the world. For example, it might develop a model of a skull where it understands the "mouth" as a separate component it can open and close, without ever being explicitly trained on that action [00:06:15]. This deep understanding emerges bottom-up, without massive datasets.The secret is to abandon a fixed goal and embrace "deception" [00:08:42]—the idea that the stepping stones to a great discovery often don't look anything like the final result. Instead of optimizing for a target, the AI is built through an open-ended process of exploring what's "interesting" [00:09:15]. This creates a more flexible and adaptable foundation, a bit like how evolvability wins out in nature [00:10:30].The show concludes by arguing that this choice matters immensely. The "impostor" path may be hitting a wall, requiring insane amounts of money and energy for progress and failing to deliver true creativity or continual learning [00:13:00]. The ultimate message is a call to not put all our eggs in one basket [00:14:25]. We should explore these open-ended, creative paths to discover a more genuine form of intelligence, which may be found where we least expect it.REFS:Questioning Representational Optimism in Deep Learning:The Fractured Entangled Representation HypothesisAkarsh Kumar, Jeff Clune, Joel Lehman, Kenneth O. Stanleyhttps://arxiv.org/pdf/2505.11581Kenneth O. Stanley, Joel LehmanWhy Greatness Cannot Be Planned: The Myth of the Objectivehttps://amzn.to/44xLaXKOriginal show with Kenneth from 4 years ago:https://www.youtube.com/watch?v=lhYGXYeMq_EKenneth Stanley is SVP Open Endedness at Lila Scienceshttps://x.com/kenneth0stanleyAkarsh Kumar (MIT)https://akarshkumar.com/AND... Kenneth is HIRING (this is an OPPORTUNITY OF A LIFETIME!)Research Engineer: https://job-boards.greenhouse.io/lila/jobs/7890007002Research Scientist: https://job-boards.greenhouse.io/lila/jobs/8012245002Tim's Code visualisation of FER based on Akarsh repo: https://github.com/ecsplendid/ferTRANSCRIPT: https://app.rescript.info/public/share/YKAZzZ6lwZkjTLRpVJreOOxGhLI8y4m3fAyU8NSavx0

What if the most powerful technology in human history is being built by people who openly admit they don't trust each other? In this explosive 2-hour debate, three AI experts pull back the curtain on the shocking psychology driving the race to Artificial General Intelligence—and why the people building it might be the biggest threat of all. Kokotajlo predicts AGI by 2028 based on compute scaling trends. Marcus argues we haven't solved basic cognitive problems from his 2001 research. The stakes? If Kokotajlo is right and Marcus is wrong about safety progress, humanity may have already lost control.

Sponsor messages:

========

Google Gemini: Google Gemini features Veo3, a state-of-the-art AI video generation model in the Gemini app. Sign up at https://gemini.google.com

Tufa AI Labs are hiring for ML Engineers and a Chief Scientist in Zurich/SF. They are top of the ARCv2 leaderboard!

https://tufalabs.ai/

========

Guest Powerhouse

Gary Marcus - Cognitive scientist, author of "Taming Silicon Valley," and AI's most prominent skeptic who's been warning about the same fundamental problems for 25 years (https://garymarcus.substack.com/)

Daniel Kokotajlo - Former OpenAI insider turned whistleblower who reveals the disturbing rationalizations of AI lab leaders in his viral "AI 2027" scenario (https://ai-2027.com/)

Dan Hendrycks - Director of the Center for AI Safety who created the benchmarks used to measure AI progress and argues we have only years, not decades, to prevent catastrophe (https://danhendrycks.com/)

Transcript:

http://app.rescript.info/public/share/tEcx4UkToi-2jwS1cN51CW70A4Eh6QulBRxDILoXOno

TOC:

Introduction: The AI Arms Race

00:00:04 - The Danger of Automated AI R&D

00:00:43 - The Rationalization: "If we don't, someone else will"

00:01:56 - Sponsor Reads (Tufa AI Labs & Google Gemini)

00:02:55 - Guest Introductions

The Philosophical Stakes

00:04:13 - What is the Positive Vision for AGI?

00:07:00 - The Abundance Scenario: Superintelligent Economy

00:09:06 - Differentiating AGI and Superintelligence (ASI)

00:11:41 - Sam Altman: "A Decade in a Month"

00:14:47 - Economic Inequality & The UBI Problem

Policy and Red Lines

00:17:13 - The Pause Letter: Stopping vs. Delaying AI

00:20:03 - Defining Three Concrete Red Lines for AI Development

00:25:24 - Racing Towards Red Lines & The Myth of "Durable Advantage"

00:31:15 - Transparency and Public Perception

00:35:16 - The Rationalization Cascade: Why AI Labs Race to "Win"

Forecasting AGI: Timelines and Methodologies

00:42:29 - The Case for Short Timelines (Median 2028)

00:47:00 - Scaling Limits: Compute, Data, and Money

00:49:36 - Forecasting Models: Bio-Anchors and Agentic Coding

00:53:15 - The 10^45 FLOP Thought Experiment

The Great Debate: Cognitive Gaps vs. Scaling

00:58:41 - Gary Marcus's Counterpoint: The Unsolved Problems of Cognition

01:00:46 - Current AI Can't Play Chess Reliably

01:08:23 - Can Tools and Neurosymbolic AI Fill the Gaps?

01:16:13 - The Multi-Dimensional Nature of Intelligence

01:24:26 - The Benchmark Debate: Data Contamination and Reliability

01:31:15 - The Superhuman Coder Milestone Debate

01:37:45 - The Driverless Car Analogy

The Alignment Problem

01:39:45 - Has Any Progress Been Made on Alignment?

01:42:43 - "Fairly Reasonably Scares the Sh*t Out of Me"

01:46:30 - Distinguishing Model vs. Process Alignment

Scenarios and Conclusions

01:49:26 - Gary's Alternative Scenario: The Neurosymbolic Shift

01:53:35 - Will AI Become Jeff Dean?

01:58:41 - Takeoff Speeds and Exceeding Human Intelligence

02:03:19 - Final Disagreements and Closing Remarks

REFS:

Gary Marcus (2001) - The Algebraic Mind

https://mitpress.mit.edu/9780262632683/the-algebraic-mind/

00:59:00

Gary Marcus & Ernest Davis (2019) - Rebooting AI

https://www.penguinrandomhouse.com/books/566677/rebooting-ai-by-gary-marcus-and-ernest-davis/

01:31:59

Gary Marcus (2024) - Taming SV

https://www.hachettebookgroup.com/titles/gary-marcus/taming-silicon-valley/9781541704091/

00:03:01

We interview Professor Christopher Summerfield from Oxford University about his new book "These Strange New Minds: How AI Learned to Talk and What It". AI learned to understand the world just by reading text - something scientists thought was impossible. You don't need to see a cat to know what one is; you can learn everything from words alone. This is "the most astonishing scientific discovery of the 21st century."People are split: some refuse to call what AI does "thinking" even when it outperforms humans, while others believe if it acts intelligent, it is intelligent. Summerfield takes the middle ground - AI does something genuinely like human reasoning, but that doesn't make it human.Sponsor messages:========Google Gemini: Google Gemini features Veo3, a state-of-the-art AI video generation model in the Gemini app. Sign up at https://gemini.google.comTufa AI Labs are hiring for ML Engineers and a Chief Scientist in Zurich/SF. They are top of the ARCv2 leaderboard! https://tufalabs.ai/========Prof. Christopher Summerfieldhttps://www.psy.ox.ac.uk/people/christopher-summerfieldThese Strange New Minds: How AI Learned to Talk and What It Meanshttps://amzn.to/4e26BVaTable of Contents:Introduction & Setup00:00:00 Superman 3 Metaphor - Humans Absorbed by Machines00:02:01 Book Introduction & AI Debate Context00:03:45 Sponsor Segments (Google Gemini, Tufa Labs)Philosophical Foundations00:04:48 The Fractured AI Discourse00:08:21 Ancient Roots: Aristotle vs Plato (Empiricism vs Rationalism)00:10:14 Historical AI: Symbolic Logic and Its LimitsThe Language Revolution00:12:11 ChatGPT as the Rubicon Moment00:14:00 The Astonishing Discovery: Learning Reality from Words Alone00:15:47 Equivalentists vs Exceptionalists DebateCognitive Science Perspectives00:19:12 Functionalism and the Duck Test00:21:48 Brain-AI Similarities and Computational Principles00:24:53 Reconciling Chomsky: Evolution vs Learning00:28:15 Lamarckian AI vs Darwinian Human LearningThe Reality of AI Capabilities00:30:29 Anthropomorphism and the Clever Hans Effect00:32:56 The Intentional Stance and Nature of Thinking00:37:56 Three Major AI Worries: Agency, Personalization, DynamicsSocietal Risks and Complex Systems00:37:56 AI Agents and Flash Crash Scenarios00:42:50 Removing Frictions: The Lawfare Example00:46:15 Gradual Disempowerment Theory00:49:18 The Faustian Pact of TechnologyHuman Agency and Control00:51:18 The Crisis of Authenticity00:56:22 Psychology of Control vs Reward01:00:21 Dopamine Hacking and Variable ReinforcementFuture Directions01:02:27 Evolution as Goal-less Optimization01:03:31 Open-Endedness and Creative Evolution01:06:46 Writing, Creativity, and AI-Generated Content01:08:18 Closing RemarksREFS:Academic References (Abbreviated)Essential Books"These Strange New Minds" - C. Summerfield [00:02:01] - Main discussion topic"The Mind is Flat" - N. Chater [00:33:45] - Summerfield's favorite on cognitive illusions"AI: A Guide for Thinking Humans" - M. Mitchell [00:04:58] - Host's previous favorite"Principia Mathematica" - Russell & Whitehead [00:11:00] - Logic Theorist reference"Syntactic Structures" - N. Chomsky (1957) [00:13:30] - Generative grammar foundation"Why Greatness Cannot Be Planned" - Stanley & Lehman [01:04:00] - Open-ended evolutionKey Papers & Studies"Gradual Disempowerment" - D. Duvenaud [00:46:45] - AI threat model"Counterfeit People" - D. Dennett (Atlantic) [00:52:45] - AI societal risks"Open-Endedness is Essential..." - DeepMind/Rocktäschel/Hughes [01:03:42]Heider & Simmel (1944) [00:30:45] - Agency attribution to shapesWhitehall Studies - M. Marmot [00:59:32] - Control and health outcomes"Clever Hans" - O. Pfungst (1911) [00:31:47] - Animal intelligence illusionHistorical References

"Blurring Reality" - Chai's Social AI Platform - sponsored

This episode of MLST explores the groundbreaking work of Chai, a social AI platform that quietly built one of the world's largest AI companion ecosystems before ChatGPT's mainstream adoption. With over 10 million active users and just 13 engineers serving 2 trillion tokens per day, Chai discovered the massive appetite for AI companionship through serendipity while searching for product-market fit.

CHAI sponsored this show *because they want to hire amazing engineers* --

CAREER OPPORTUNITIES AT CHAI

Chai is actively hiring in Palo Alto with competitive compensation ($300K-$800K+ equity) for roles including AI Infrastructure Engineers, Software Engineers, Applied AI Researchers, and more. Fast-track qualification available for candidates with significant product launches, open source contributions, or entrepreneurial success.

https://www.chai-research.com/jobs/

The conversation with founder William Beauchamp and engineers Tom Lu and Nischay Dhankhar covers Chai's innovative technical approaches including reinforcement learning from human feedback (RLHF), model blending techniques that combine smaller models to outperform larger ones, and their unique infrastructure challenges running exaflop-class compute.

SPONSOR MESSAGES:

***

Tufa AI Labs is a brand new research lab in Zurich started by Benjamin Crouzier focussed on o-series style reasoning and AGI. They are hiring a Chief Engineer and ML engineers in Zurich and SF.

Goto https://tufalabs.ai/

***

Key themes explored include:

- The ethics of AI engagement optimization and attention hacking

- Content moderation at scale with a lean engineering team

- The shift from AI as utility tool to AI as social companion

- How users form deep emotional bonds with artificial intelligence

- The broader implications of AI becoming a social medium

We also examine OpenAI's recent pivot toward companion AI with April's new GPT-4o, suggesting a fundamental shift in how we interact with artificial intelligence - from utility-focused tools to companion-like experiences that blur the lines between human and artificial intimacy.

The episode also covers Chai's unconventional approach to hiring only top-tier engineers, their bootstrap funding strategy focused on user revenue over VC funding, and their rapid experimentation culture where one in five experiments succeed.

TOC:

00:00:00 - Introduction: Steve Jobs' AI Vision & Chai's Scale

00:04:02 - Chapter 1: Simulators - The Birth of Social AI

00:13:34 - Chapter 2: Engineering at Chai - RLHF & Model Blending

00:21:49 - Chapter 3: Social Impact of GenAI - Ethics & Safety

00:33:55 - Chapter 4: The Lean Machine - 13 Engineers, Millions of Users

00:42:38 - Chapter 5: GPT-4o Becoming a Companion - OpenAI's Pivot

00:50:10 - Chapter 6: What Comes Next - The Future of AI Intimacy

TRANSCRIPT: https://www.dropbox.com/scl/fi/yz2ewkzmwz9rbbturfbap/CHAI.pdf?rlkey=uuyk2nfhjzezucwdgntg5ubqb&dl=0

Today GoogleDeepMind released AlphaEvolve: a Gemini coding agent for algorithm discovery. It beat the famous Strassen algorithm for matrix multiplication set 56 years ago. Google has been killing it recently. We had early access to the paper and interviewed the researchers behind the work.

AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms

https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/

Authors: Alexander Novikov*, Ngân Vũ*, Marvin Eisenberger*, Emilien Dupont*, Po-Sen Huang*, Adam Zsolt Wagner*, Sergey Shirobokov*, Borislav Kozlovskii*, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, Matej Balog*

(* indicates equal contribution or special designation, if defined elsewhere)

SPONSOR MESSAGES:

***

Tufa AI Labs is a brand new research lab in Zurich started by Benjamin Crouzier focussed on o-series style reasoning and AGI. They are hiring a Chief Engineer and ML engineers. Events in Zurich.

Goto https://tufalabs.ai/

***

AlphaEvolve works like a very smart, tireless programmer. It uses powerful AI language models (like Gemini) to generate ideas for computer code. Then, it uses an "evolutionary" process – like survival of the fittest for programs. It tries out many different program ideas, automatically tests how well they solve a problem, and then uses the best ones to inspire new, even better programs.

Beyond this mathematical breakthrough, AlphaEvolve has already been used to improve real-world systems at Google, such as making their massive data centers run more efficiently and even speeding up the training of the AI models that power AlphaEvolve itself. The discussion also covers how humans work with AlphaEvolve, the challenges of making AI discover things, and the exciting future of AI helping scientists make new discoveries.

In short, AlphaEvolve is a powerful new AI tool that can invent new algorithms and solve complex problems, showing how AI can be a creative partner in science and engineering.

Guests:

Matej Balog: https://x.com/matejbalog

Alexander Novikov: https://x.com/SashaVNovikov

REFS:

MAP Elites [Jean-Baptiste Mouret, Jeff Clune]

https://arxiv.org/abs/1504.04909

FunSearch [Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli & Alhussein Fawzi]

https://www.nature.com/articles/s41586-023-06924-6

TOC:

[00:00:00] Introduction: Alpha Evolve's Breakthroughs, DeepMind's Lineage, and Real-World Impact

[00:12:06] Introducing AlphaEvolve: Concept, Evolutionary Algorithms, and Architecture

[00:16:56] Search Challenges: The Halting Problem and Enabling Creative Leaps

[00:23:20] Knowledge Augmentation: Self-Generated Data, Meta-Prompting, and Library Learning

[00:29:08] Matrix Multiplication Breakthrough: From Strassen to AlphaEvolve's 48 Multiplications

[00:39:11] Problem Representation: Direct Solutions, Constructors, and Search Algorithms

[00:46:06] Developer Reflections: Surprising Outcomes and Superiority over Simple LLM Sampling

[00:51:42] Algorithmic Improvement: Hill Climbing, Program Synthesis, and Intelligibility

[01:00:24] Real-World Application: Complex Evaluations and Robotics

[01:05:39] Role of LLMs & Future: Advanced Models, Recursive Self-Improvement, and Human-AI Collaboration

[01:11:22] Resource Considerations: Compute Costs of AlphaEvolve

This is a trial of posting videos on Spotify, thoughts? Email me or chat in our Discord

Randall Balestriero joins the show to discuss some counterintuitive findings in AI. He shares research showing that huge language models, even when started from scratch (randomly initialized) without massive pre-training, can learn specific tasks like sentiment analysis surprisingly well, train stably, and avoid severe overfitting, sometimes matching the performance of costly pre-trained models. This raises questions about when giant pre-training efforts are truly worth it.

He also talks about how self-supervised learning (where models learn from data structure itself) and traditional supervised learning (using labeled data) are fundamentally similar, allowing researchers to apply decades of supervised learning theory to improve newer self-supervised methods.

Finally, Randall touches on fairness in AI models used for Earth data (like climate prediction), revealing that these models can be biased, performing poorly in specific locations like islands or coastlines even if they seem accurate overall, which has important implications for policy decisions based on this data.

SPONSOR MESSAGES:

***

Tufa AI Labs is a brand new research lab in Zurich started by Benjamin Crouzier focussed on o-series style reasoning and AGI. They are hiring a Chief Engineer and ML engineers. Events in Zurich.

Goto https://tufalabs.ai/

***

TRANSCRIPT + SHOWNOTES:

https://www.dropbox.com/scl/fi/n7yev71nsjso71jyjz1fy/RANDALLNEURIPS.pdf?rlkey=0dn4injp1sc4ts8njwf3wfmxv&dl=0

TOC:

1. Model Training Efficiency and Scale

[00:00:00] 1.1 Training Stability of Large Models on Small Datasets

[00:04:09] 1.2 Pre-training vs Random Initialization Performance Comparison

[00:07:58] 1.3 Task-Specific Models vs General LLMs Efficiency

2. Learning Paradigms and Data Distribution

[00:10:35] 2.1 Fair Language Model Paradox and Token Frequency Issues

[00:12:02] 2.2 Pre-training vs Single-task Learning Spectrum

[00:16:04] 2.3 Theoretical Equivalence of Supervised and Self-supervised Learning

[00:19:40] 2.4 Self-Supervised Learning and Supervised Learning Relationships

[00:21:25] 2.5 SSL Objectives and Heavy-tailed Data Distribution Challenges

3. Geographic Representation in ML Systems

[00:25:20] 3.1 Geographic Bias in Earth Data Models and Neural Representations

[00:28:10] 3.2 Mathematical Limitations and Model Improvements

[00:30:24] 3.3 Data Quality and Geographic Bias in ML Datasets

REFS:

[00:01:40] Research on training large language models from scratch on small datasets, Randall Balestriero et al.

https://openreview.net/forum?id=wYGBWOjq1Q

[00:10:35] The Fair Language Model Paradox (2024), Andrea Pinto, Tomer Galanti, Randall Balestriero

https://arxiv.org/abs/2410.11985

[00:12:20] Muppet: Massive Multi-task Representations with Pre-Finetuning (2021), Armen Aghajanyan et al.

https://arxiv.org/abs/2101.11038

[00:14:30] Dissociating language and thought in large language models (2023), Kyle Mahowald et al.

https://arxiv.org/abs/2301.06627

[00:16:05] The Birth of Self-Supervised Learning: A Supervised Theory, Randall Balestriero et al.

https://openreview.net/forum?id=NhYAjAAdQT

[00:21:25] VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning, Adrien Bardes, Jean Ponce, Yann LeCun

https://arxiv.org/abs/2105.04906

[00:25:20] No Location Left Behind: Measuring and Improving the Fairness of Implicit Representations for Earth Data (2025), Daniel Cai, Randall Balestriero, et al.

https://arxiv.org/abs/2502.06831

[00:33:45] Mark Ibrahim et al.'s work on geographic bias in computer vision datasets, Mark Ibrahim

https://arxiv.org/pdf/2304.12210

Prof. Kevin Ellis and Dr. Zenna Tavares talk about making AI smarter, like humans. They want AI to learn from just a little bit of information by actively trying things out, not just by looking at tons of data.

They discuss two main ways AI can "think": one way is like following specific rules or steps (like a computer program), and the other is more intuitive, like guessing based on patterns (like modern AI often does). They found combining both methods works well for solving complex puzzles like ARC.

A key idea is "compositionality" - building big ideas from small ones, like LEGOs. This is powerful but can also be overwhelming. Another important idea is "abstraction" - understanding things simply, without getting lost in details, and knowing there are different levels of understanding.

Ultimately, they believe the best AI will need to explore, experiment, and build models of the world, much like humans do when learning something new.

SPONSOR MESSAGES:

***

Tufa AI Labs is a brand new research lab in Zurich started by Benjamin Crouzier focussed on o-series style reasoning and AGI. They are hiring a Chief Engineer and ML engineers. Events in Zurich.

Goto https://tufalabs.ai/

***

TRANSCRIPT:

https://www.dropbox.com/scl/fi/3ngggvhb3tnemw879er5y/BASIS.pdf?rlkey=lr2zbj3317mex1q5l0c2rsk0h&dl=0

Zenna Tavares:

http://www.zenna.org/

Kevin Ellis:

https://www.cs.cornell.edu/~ellisk/

TOC:

1. Compositionality and Learning Foundations

[00:00:00] 1.1 Compositional Search and Learning Challenges

[00:03:55] 1.2 Bayesian Learning and World Models

[00:12:05] 1.3 Programming Languages and Compositionality Trade-offs

[00:15:35] 1.4 Inductive vs Transductive Approaches in AI Systems

2. Neural-Symbolic Program Synthesis

[00:27:20] 2.1 Integration of LLMs with Traditional Programming and Meta-Programming

[00:30:43] 2.2 Wake-Sleep Learning and DreamCoder Architecture

[00:38:26] 2.3 Program Synthesis from Interactions and Hidden State Inference

[00:41:36] 2.4 Abstraction Mechanisms and Resource Rationality

[00:48:38] 2.5 Inductive Biases and Causal Abstraction in AI Systems

3. Abstract Reasoning Systems

[00:52:10] 3.1 Abstract Concepts and Grid-Based Transformations in ARC

[00:56:08] 3.2 Induction vs Transduction Approaches in Abstract Reasoning

[00:59:12] 3.3 ARC Limitations and Interactive Learning Extensions

[01:06:30] 3.4 Wake-Sleep Program Learning and Hybrid Approaches

[01:11:37] 3.5 Project MARA and Future Research Directions

REFS:

[00:00:25] DreamCoder, Kevin Ellis et al.

https://arxiv.org/abs/2006.08381

[00:01:10] Mind Your Step, Ryan Liu et al.

https://arxiv.org/abs/2410.21333

[00:06:05] Bayesian inference, Griffiths, T. L., Kemp, C., & Tenenbaum, J. B.

https://psycnet.apa.org/record/2008-06911-003

[00:13:00] Induction and Transduction, Wen-Ding Li, Zenna Tavares, Yewen Pu, Kevin Ellis

https://arxiv.org/abs/2411.02272

[00:23:15] Neurosymbolic AI, Garcez, Artur d'Avila et al.

https://arxiv.org/abs/2012.05876

[00:33:50] Induction and Transduction (II), Wen-Ding Li, Kevin Ellis et al.

https://arxiv.org/abs/2411.02272

[00:38:35] ARC, François Chollet

https://arxiv.org/abs/1911.01547

[00:39:20] Causal Reactive Programs, Ria Das, Joshua B. Tenenbaum, Armando Solar-Lezama, Zenna Tavares

http://www.zenna.org/publications/autumn2022.pdf

[00:42:50] MuZero, Julian Schrittwieser et al.

http://arxiv.org/pdf/1911.08265

[00:43:20] VisualPredicator, Yichao Liang

https://arxiv.org/abs/2410.23156

[00:48:55] Bayesian models of cognition, Joshua B. Tenenbaum

https://mitpress.mit.edu/9780262049412/bayesian-models-of-cognition/

[00:49:30] The Bitter Lesson, Rich Sutton

http://www.incompleteideas.net/IncIdeas/BitterLesson.html

[01:06:35] Program induction, Kevin Ellis, Wen-Ding Li

https://arxiv.org/pdf/2411.02272

[01:06:50] DreamCoder (II), Kevin Ellis et al.

https://arxiv.org/abs/2006.08381

[01:11:55] Project MARA, Zenna Tavares, Kevin Ellis

https://www.basis.ai/blog/mara/

Eiso Kant, CTO of poolside AI, discusses the company's approach to building frontier AI foundation models, particularly focused on software development. Their unique strategy is reinforcement learning from code execution feedback which is an important axis for scaling AI capabilities beyond just increasing model size or data volume. Kant predicts human-level AI in knowledge work could be achieved within 18-36 months, outlining poolside's vision to dramatically increase software development productivity and accessibility.

SPONSOR MESSAGES:

***

Tufa AI Labs is a brand new research lab in Zurich started by Benjamin Crouzier focussed on o-series style reasoning and AGI. They are hiring a Chief Engineer and ML engineers. Events in Zurich.

Goto https://tufalabs.ai/

***

Eiso Kant:

https://x.com/eisokant

https://poolside.ai/

TRANSCRIPT:

https://www.dropbox.com/scl/fi/szepl6taqziyqie9wgmk9/poolside.pdf?rlkey=iqar7dcwshyrpeoz0xa76k422&dl=0

TOC:

1. Foundation Models and AI Strategy

[00:00:00] 1.1 Foundation Models and Timeline Predictions for AI Development

[00:02:55] 1.2 Poolside AI's Corporate History and Strategic Vision

[00:06:48] 1.3 Foundation Models vs Enterprise Customization Trade-offs

2. Reinforcement Learning and Model Economics

[00:15:42] 2.1 Reinforcement Learning and Code Execution Feedback Approaches

[00:22:06] 2.2 Model Economics and Experimental Optimization

3. Enterprise AI Implementation

[00:25:20] 3.1 Poolside's Enterprise Deployment Strategy and Infrastructure

[00:26:00] 3.2 Enterprise-First Business Model and Market Focus

[00:27:05] 3.3 Foundation Models and AGI Development Approach

[00:29:24] 3.4 DeepSeek Case Study and Infrastructure Requirements

4. LLM Architecture and Performance

[00:30:15] 4.1 Distributed Training and Hardware Architecture Optimization

[00:33:01] 4.2 Model Scaling Strategies and Chinchilla Optimality Trade-offs

[00:36:04] 4.3 Emergent Reasoning and Model Architecture Comparisons

[00:43:26] 4.4 Balancing Creativity and Determinism in AI Models

[00:50:01] 4.5 AI-Assisted Software Development Evolution

5. AI Systems Engineering and Scalability

[00:58:31] 5.1 Enterprise AI Productivity and Implementation Challenges

[00:58:40] 5.2 Low-Code Solutions and Enterprise Hiring Trends

[01:01:25] 5.3 Distributed Systems and Engineering Complexity

[01:01:50] 5.4 GenAI Architecture and Scalability Patterns

[01:01:55] 5.5 Scaling Limitations and Architectural Patterns in AI Code Generation

6. AI Safety and Future Capabilities

[01:06:23] 6.1 Semantic Understanding and Language Model Reasoning Approaches

[01:12:42] 6.2 Model Interpretability and Safety Considerations in AI Systems

[01:16:27] 6.3 AI vs Human Capabilities in Software Development

[01:33:45] 6.4 Enterprise Deployment and Security Architecture

CORE REFS (see shownotes for URLs/more refs):

[00:15:45] Research demonstrating how training on model-generated content leads to distribution collapse in AI models, Ilia Shumailov et al. (Key finding on synthetic data risk)

[00:20:05] Foundational paper introducing Word2Vec for computing word vector representations, Tomas Mikolov et al. (Seminal NLP technique)

[00:22:15] OpenAI O3 model's breakthrough performance on ARC Prize Challenge, OpenAI (Significant AI reasoning benchmark achievement)

[00:22:40] Seminal paper proposing a formal definition of intelligence as skill-acquisition efficiency, François Chollet (Influential AI definition/philosophy)

[00:30:30] Technical documentation of DeepSeek's V3 model architecture and capabilities, DeepSeek AI (Details on a major new model)

[00:34:30] Foundational paper establishing optimal scaling laws for LLM training, Jordan Hoffmann et al. (Key paper on LLM scaling)

[00:45:45] Seminal essay arguing that scaling computation consistently trumps human-engineered solutions in AI, Richard S. Sutton (Influential "Bitter Lesson" perspective)