When AI Research Starts Moving Faster Than Human Research - Zhengyao Jiang

Weco let an AI coding agent rewrite the harness around another agent for eight days: its code, prompts and tools, while the underlying language model stayed fixed. Tim Scarfe asks Weco co-founder Zhengyao Jiang what the reported gains over two years of human engineering actually demonstrate.The discussion examines AIDE 85's generated code, held-out evaluation and the difficulty of separating useful discoveries from reward hacking. Jiang explains Weco's four levels of recursive self-improvement and compares the experiment with AlphaEvolve and the Darwin Gödel Machine.The limits matter as much as the gains. Jiang explains why the experiment did not establish that the system had become a better improver. The conversation closes with open-ended search, human-designed primitives and Parameter Golf: where does the next useful idea come from when the agent is searching inside a space that people designed?---TIMESTAMPS:00:00:00 Eight days of self-improvement: what counts?00:03:25 AIDE and the puzzle of useful spaghetti code00:08:38 Four levels of recursive self-improvement00:12:02 What AIDE 85 changed and how it was tested00:20:04 AlphaEvolve, Darwin Gödel Machine and the RSI claim00:26:21 Reward hacking and the limits of detection00:33:09 Open-ended search, harness tuning and creativity00:39:43 Parameter Golf and the limits of self-improvement---REFERENCES:organization:[00:00:30] Weco AIhttps://www.weco.ai/other:[00:00:33] AIDE²: The First Evidence of Recursive Self-Improvementhttps://www.weco.ai/blog/first-evidence-of-recursive-self-improvement[00:14:11] Faulty reward functions in the wildhttps://openai.com/index/faulty-reward-functions/[00:29:59] The Hugging Face incident and the road aheadhttps://openai.com/index/hugging-face-incident-and-the-road-ahead/tool:[00:03:29] AIDEhttps://github.com/WecoAI/aideml[00:04:29] MLE-benchhttps://github.com/openai/mle-bench[00:04:33] ALE-Benchhttps://github.com/SakanaAI/ALE-Bench[00:04:52] WeatherBench 2https://github.com/google-research/weatherbench2[00:08:18] ReActhttps://react-lm.github.io/[00:39:43] Parameter Golfhttps://github.com/openai/parameter-golfpaper:[00:20:08] AlphaEvolve: A coding agent for scientific and algorithmic discoveryhttps://arxiv.org/abs/2506.13131v1[00:21:35] Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agentshttps://arxiv.org/abs/2505.22954v3[00:23:45] Hyperagentshttps://arxiv.org/abs/2603.19461v1[00:27:01] SpecBench: Measuring Reward Hacking in Long-Horizon Coding Agentshttps://arxiv.org/abs/2605.21384book:[00:33:14] Why Greatness Cannot Be Planned: The Myth of the Objectivehttps://link.springer.com/book/10.1007/978-3-319-15524-1---LINKS:https://app.rescript.info/share/3a9dc6189cb539c6a05fcc4f75c101b3PDF:https://app.rescript.info/api/public/sessions/9eda60ede2b31c92/pdf

Frank Hutter, co-founder of Prior Labs, talks about TabPFN, a tabular foundation model that makes predictions in a single forward pass, and the research behind it.TabPFN is pre-trained on synthetic datasets drawn from a prior over structural causal models, rather than on real data. At prediction time it takes the whole training table as context and outputs an approximation of the Bayesian posterior predictive distribution, without per-dataset training or hyperparameter search. Frank explains how this grew out of his earlier work on AutoML and neural architecture search, how the priors are built and revised, and why tabular data was hard for deep learning for so long.The conversation also covers the TabArena benchmark, how the architecture changed from TabPFN v1 to v3, scaling to larger tables, using the model with coding agents, test-time compute, Google's TabFM, causal inference and interventions, and relational data. At the end, a short update Frank recorded after the interview covers the TabPFN-3.5 release.Prior Labs:TabPFN-3.5: https://priorlabs.ai/tabpfn-3-5https://priorlabs.ai/careersTOC:00:00 Introduction00:44 Welcome and Frank's background02:05 Why tabular data was hard for deep learning10:17 Pre-training on synthetic data12:52 The TabArena benchmark19:28 From AutoML to neural architecture search26:34 TabPFN as a learned algorithm30:50 Bayesian prediction in one forward pass39:37 Scaling to larger tables47:48 Using TabPFN with coding agents57:47 Output heads and architecture from v1 to v31:05:29 Test-time compute and adaptation1:13:32 Google's TabFM1:16:53 How the priors are designed1:18:40 Correlation, causation and interventions1:35:22 Relational and multimodal data1:38:31 Use in organisations1:46:38 The open research arm1:50:21 Update: TabPFN-3.5REFS:TabPFN v2, Nature (Hollmann et al., 2025)https://www.nature.com/articles/s41586-024-08328-6Transformers Can Do Bayesian Inference (Müller et al.)https://arxiv.org/abs/2112.10510TabArena (Erickson et al.)https://arxiv.org/abs/2506.16791AutoGluon-Tabular (Erickson et al.)https://arxiv.org/abs/2003.06505Beyond IID: How General Are Tabular Foundation Models, Really?https://arxiv.org/abs/2606.30410Neural Architecture Search: A Survey (Elsken, Metzen & Hutter)https://arxiv.org/abs/1808.05377Auto-WEKA (Thornton et al.)https://www.cs.ubc.ca/~hutter/papers/AutoWEKA-KDD2013.pdfTabPFN v1 (Hollmann et al., 2022)https://arxiv.org/abs/2207.01848TabPFN-3 technical reporthttps://arxiv.org/abs/2605.13986TabPFN-2.5 reporthttps://arxiv.org/abs/2511.08667CAAFE (Hollmann et al.)https://arxiv.org/abs/2305.03403TabICL (Qu et al.)https://arxiv.org/abs/2502.05564TabICLv2 (Qu et al.)https://arxiv.org/abs/2602.11139Google TabFMhttps://research.google/blog/introducing-tabfm-a-zero-shot-foundation-model-for-tabular-data/TALENT benchmark (Ye et al.)https://arxiv.org/abs/2407.00956Do-PFN (Robertson et al.)https://arxiv.org/abs/2506.06039CausalPFN (Balazadeh et al.)https://arxiv.org/abs/2506.07918Causal Foundation Models with Partial Graphs (Reuter et al.)https://arxiv.org/abs/2602.14972RelBench (Robinson et al.)https://arxiv.org/abs/2407.20060RelArena-α, TabPFN-Rel and RPIhttps://arxiv.org/abs/2608.16319TabPFN on GitHubhttps://github.com/PriorLabs/TabPFNTabPFN-3.5 technical reporthttps://arxiv.org/abs/2609.17895Otto Group Product Classification Challenge (Kaggle, 2015)https://www.kaggle.com/competitions/otto-group-product-classification-challenge---RESCRIPT:https://app.rescript.info/share/e99676c25ee6189fbf54c9be07eb623e

Alexander Mattick is a researcher at Fraunhofer IIS and a PhD researcher at the University of Technology Nuremberg (UTN), and a regular on Yannic Kilcher's Discord. He first came on MLST in 2022, after helping research the Yann LeCun and Randall Balestriero episode on interpolation.SPONSOR:---Cyber Fund built the Monastery to help founders ship products that were impossible a year ago. Applications for Batch 1 are now open.Apply now: https://cyber.fund---Alexander treats inference as the thread running through modern machine learning: once you have a model, what does it cost to get an answer out of it? He works through Monte Carlo, GFlowNets, energy-based models, diffusion, normalising flows and flow matching, with four short explainers he recorded himself. He is blunt about energy-based models: you can sample from them in principle, but it is rarely worth the compute. JEPA and "world model", he says, are closer to branding than to technical categories.Next: theories of deep learning, none of which he thinks predicts enough yet to guide practice, then reinforcement learning. ---0:00 Cold open: information is expensive0:51 Welcome back, Alexander Mattic2:08 Alexander's research background2:50 Inference: densities, sampling and Monte Carlo6:42 GFlowNets, energy functions and MCMC9:45 Explainer: energy-based models11:03 Why model a density at all?17:30 From learned energies to flow matching25:08 Explainers: diffusion and normalising flows28:33 Are energy-based models generative?33:22 JEPA, contrastive learning and collapse41:13 Why non-language modalities need flows44:51 Inference as search: branch and bound49:43 Q-learning and delayed consequences55:14 Flow matching, optimal transport, Fokker-Planck1:00:03 Explainer: flow matching1:01:49 AlphaFold, latents and scale versus architecture1:07:52 Two families of deep learning theory1:15:04 What a good theory would predict1:23:53 The manifold hypothesis and compression1:28:25 Is reward enough?1:32:01 Control theory versus reinforcement learning1:37:22 The Bitter Lesson and expensive information1:42:08 Constrained RL: the constrained MDP toolbox1:50:12 Creativity as constrained search1:55:44 Reality is protean: when abstractions hold2:00:32 What is a world model?2:04:38 Prediction is not control2:08:13 Robot demos, MPC and reliability---REFERENCES:[6:55] GFlowNets (Bengio et al., 2021)https://arxiv.org/abs/2106.04399[38:46] Contrastive Self-Supervised Learning (Anand, 2020)https://ankeshanand.com/blog/2020/01/26/contrative-self-supervised-learning.html[38:56] LeJEPA (Balestriero and LeCun, 2025)https://arxiv.org/abs/2511.08544v3[47:10] RL for Node Selection in Branch-and-Bound (Mattick)https://openreview.net/forum?id=0ez68a5UqI[56:20] Flow Matching for Generative Modeling https://arxiv.org/abs/2210.02747v2[1:12:41] Disentangling feature and lazy training in deep neural networkshttps://arxiv.org/abs/1906.08034v4[1:31:05] Reward is enough (Silver)https://doi.org/10.1016/j.artint.2021.103535[1:35:12] Learning ReLU networks to high uniform accuracy is intractable (Berner et al.)https://arxiv.org/abs/2205.13531v2[1:40:20] Dota 2 with Large Scale Deep RL https://arxiv.org/abs/1912.06680v1[1:45:41] Constrained Update Projection for Safe Policy Optimization (Yang et al., 2022)https://arxiv.org/abs/2209.07089[1:46:11] SafeMPO (ICLR 2026)https://openreview.net/forum?id=1m0EU6QXj6[1:50:17] Why Creativity Cannot Be Interpolatedhttps://archive.mlst.ai/paper/why-creativity-cannot-be-interpolated/[1:51:39] Invalid Action Masking (Huang and Ontañón)https://arxiv.org/abs/2006.14171[2:00:04] Probability Theory: The Logic of Science (Jaynes, 2003)https://www.cambridge.org/core/books/probability-theory/9CA08E224FF30123304E6D8935CF1A99[2:01:53] Training Agents Inside of Scalable World Models (Hafner et al., 2025)https://arxiv.org/abs/2509.24527v1[2:03:43] World Models (Ha and Schmidhuber, 2018)https://arxiv.org/abs/1803.10122v4

The car making a left turn at the start of this episode was never filmed. Cosmos 3 generated it. Ming-Yu Liu, who leads the Cosmos research at NVIDIA, explains how one model can describe a video, generate one, and produce robot actions.He walks Tim through the architecture. A vision language model reasons one token at a time; its weights then initialise a bidirectional diffusion generator for video, audio and action, and a shared temporal position scheme lines up signals that run at different rates. Ming-Yu treats "world model" as a set of tools, not one definition: forward dynamics, inverse dynamics and policy, trained together under a capacity limit so that each helps the others. He also explains why plentiful first-person human video carries over to robots, which have far less data of their own, and why a Cosmos model post-trained on the DROID dataset is a good starting point for pick-and-place policies.The most practical thread is testing. A neural simulator does not need accurate success rates. It only needs to rank policy A above policy B the way the real world would, so a team can narrow down which checkpoints deserve a real trial. Cosmos Dreams applies that closed-loop idea to driving and robotics, and Ming-Yu argues that humanoids around children and pets make safety matter even more than it does for cars. The conversation ends on the Super, Nano and Edge sizes (Edge targets Jetson Thor, Orin and DGX Spark) and where to find the open weights, code and data.This episode is a paid partnership with NVIDIA.Learn more about Cosmos: https://nvda.ws/4cJoY1SExplore Cosmos Lab: https://research.nvidia.com/labs/cosmos-lab/cosmos3/---TIMESTAMPS:00:00:00 A road that was never filmed00:02:28 Inside Cosmos 3: reasoning and generator towers00:05:02 World models: dynamics, policy and one clock00:08:59 Learning robot skills from human video00:11:06 Ambiguous tasks and system 2 planning00:12:53 Neural simulators for policy verification00:16:41 Cosmos as a starting point for robot policies00:19:00 Cosmos Dreams and robot safety00:22:04 Super, Nano and Edge model sizes00:24:24 Open models, the Cosmos repo and feedback---REFERENCES:tool:[00:00:13] Cosmos 3 (NVIDIA Cosmos Lab project page)https://research.nvidia.com/labs/cosmos-lab/cosmos3/[00:18:27] NVIDIA Cosmos GitHub repositoryhttps://github.com/NVIDIA/cosmos[00:22:05] Cosmos3-Edge model cardhttps://huggingface.co/nvidia/Cosmos3-Edge[00:22:15] Cosmos3-Super model cardhttps://huggingface.co/nvidia/Cosmos3-Super[00:22:16] Cosmos3-Nano model cardhttps://huggingface.co/nvidia/Cosmos3-Nano[00:22:50] NVIDIA Jetson Thorhttps://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/[00:22:52] NVIDIA Jetson Orinhttps://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/[00:22:53] NVIDIA DGX Sparkhttps://www.nvidia.com/en-us/products/workstations/dgx-spark/[00:24:42] Cosmos 3 collection on Hugging Facehttps://huggingface.co/collections/nvidia/cosmos3other:[00:01:07] Cosmos-Dreams closed-loop simulators (NVIDIA SIGGRAPH 2026 blog)https://blogs.nvidia.com/blog/siggraph-news-2026/paper:[00:08:54] Cosmos 3: Omnimodal World Models for Physical AIhttps://arxiv.org/abs/2606.02800[00:17:43] DROID: A Large-Scale In-The-Wild Robot Manipulation Datasethttps://arxiv.org/abs/2403.12945---RESCRIPT: https://app.rescript.info/share/e2385948cf465f0d6a2c0930150fc3ab

Pavan Kumar Reddy leads audio research at Mistral AI. He joins Tim Scarfe for a deep technical tour of Voxtral — and explains why the frontier of deployed voice is still a cascade of specialised models rather than one end-to-end system.IN PARTNERSHIP WITH MISTRAL AI:---This episode was produced in partnership with Mistral AI.Mistral AI: https://mistral.ai/---The conversation opens on architecture. Voxtral Chat feeds a 3B Ministral text trunk with continuous embeddings from an audio encoder, passed to the decoder as direct token input rather than through cross-attention as in Whisper, so the model can answer questions about emotion, timing and who spoke when without an intermediate transcript to lose them. The real-time model becomes a dual-stream decoder that consumes audio and emits text at once, at a target delay down to 160ms, with slower streams in parallel for anything that can wait for more context.On generation, Pavan explains why Voxtral TTS predicts continuous latents rather than discrete codec tokens, traces the lineage from SoundStream through EnCodec to Mimi's split of semantic and acoustic codebooks, and places FSQ and flow matching in it. Tim presses on the priors underneath: why a mel spectrogram instead of raw waveform, what noise augmentation buys, and when acoustic overfitting becomes somebody's fine-tuning problem. Then the failure modes. Diarisation is emitted autoregressively inside the transcript rather than by a separate head, which makes streaming diarisation fragile — less context, late speaker changes, invented extra speakers. And because the architecture commits to what it has already predicted, one out-of-distribution mistake compounds into looping or skipped segments, which is what DPO corrects: the negative supervision pre-training and SFT cannot give.The last third is the argument Tim keeps returning to. Customers running voice agents over millions of sessions describe scaffolding, not a solved problem, with a sharp drop outside the top few languages. Cascades survive because each component stays separately adaptable, observable and constrainable. And voice alone is cognitive debt: absorbing information and deciding in one serial stream is harder than glancing at a menu. Voice becomes ubiquitous beside a screen, not instead of one.---TIMESTAMPS:00:00:00 Cold open00:00:46 Why Mistral moved into audio00:09:27 Inside Voxtral: trunk, encoder, dual streams00:20:22 Speech that works in real time00:30:52 How a voice becomes tokens00:39:59 Flow matching, FSQ and the new codec00:52:51 When speech models lose the speaker01:03:23 Correcting hallucinations with preferences01:12:12 Controlling synthetic speech01:20:06 Why cascades still win01:29:25 Speech in the wild01:33:46 Audio models as interfaces01:37:54 Why voice still needs a screen---REFERENCES:paper:[00:01:42] Mistral 7Bhttps://arxiv.org/abs/2310.06825[00:09:38] Voxtralhttps://arxiv.org/abs/2507.13264[00:14:41] Whisper: Robust Speech Recognitionhttps://arxiv.org/abs/2212.04356[00:19:11] Voxtral Realtimehttps://arxiv.org/abs/2602.11298[00:21:52] Delayed Streams Modeling (Kyutai)https://arxiv.org/abs/2509.08753[00:30:52] Voxtral TTShttps://arxiv.org/abs/2603.25551[00:32:38] SoundStream neural audio codechttps://arxiv.org/abs/2107.03312[00:34:59] Flow Matching for Generative Modelinghttps://arxiv.org/abs/2210.02747[00:37:03] EnCodec: High Fidelity Neural Audio Compressionhttps://arxiv.org/abs/2210.13438[00:37:42] Moshi and the Mimi codechttps://arxiv.org/abs/2410.00037[00:39:05] Finite Scalar Quantization (FSQ)https://arxiv.org/abs/2309.15505[01:03:33] Direct Preference Optimization (DPO)https://arxiv.org/abs/2305.18290dataset:[00:46:14] Mozilla Common Voicehttps://commonvoice.mozilla.org/en/datasetsorganization:[00:50:47] Hugging Facehttps://huggingface.co/

Can a machine learn the judgement that separates a plausible-looking result from a faithful experiment? Edward Hughes, Chief Scientist and co-founder of Inherent, joins Tim Scarfe to argue that creativity is not optimisation, and that the missing capability in AI is choosing which questions are worth asking.SPONSOR:---Cyber Fund built the Monastery to help founders ship products that were impossible a year ago.Apply now: https://cyber.fund---Edward makes the case that Move 37 was innovative rather than creative, and that the field, not the individual, decides what counts as a discovery. That reframing runs through Csikszentmihalyi, Deutsch and exaptation into open-endedness, where deceptive goals and imperfect world models turn out to be the point rather than the problem. The second half turns to the paper: Replica, a task space built by redacting figures from real papers, and Faraday, a 27-billion-parameter model trained to steer a frontier coding agent that then beats the frontier on held-out replications.---TIMESTAMPS:00:00:00 Cold open: Move 37, Faraday and collective intelligence00:01:08 Sponsor: CyberFund00:01:46 Inherent's $50M raise and the road from string theory00:09:14 Three timescales of learning: weights, context, culture00:13:47 Move 37 was innovative, not creative: the field decides00:20:39 Creativity as satisficing: the urinal and evolution00:25:06 Exaptation and the Tristan chord: creativity in context00:30:56 Coherence for whom? Deutsch's hard-to-vary explanations00:35:53 Why copying is creative: Deutsch and the constraint engineer00:42:27 Societies of agents and the strong Moravec paradox00:45:51 Evaluate in hindsight: from Lean proofs to climate change00:51:56 Picbreeder, local goals and why discovery needs deception00:57:21 Spaghetti proofs, translation layers and superhuman Go01:00:37 Does nature compress? Naturalness and real patterns01:07:36 Why replicate? Replica's redacted figures and Faraday01:12:31 Faraday beats Codex, Claude and GLM 5.2 on held-out tasks01:15:31 Replication to innovation: how the Transformer happened01:18:26 Deep replication: what Faraday learns from Voyager and GNoME01:23:37 Can the AI scientist cheat? Goodharting the judge01:29:09 Inside Replica: scale-down, 8xB300 runs, per-task rubrics01:34:11 The RL crisis: getting GRPO to work with per-turn credit01:39:43 Weights vs harnesses: AlphaEvolve, DGM and EvoTune01:45:45 The recursive company: agents cross a phase transition01:50:35 Collective intelligence and the electric dynamo01:55:46 What replaces OKRs? Incumbents and the burden of knowledge---REFERENCES:MLST Creativity Article:https://archive.mlst.ai/read/why-creativity-cannot-be-interpolatedorganization:[00:01:47] Inherenthttps://inherentlabs.ai/other:[00:20:51] Marcel Duchamp, Fountainhttps://www.tate.org.uk/art/artworks/duchamp-fountain-t07573[00:05:19] Human-Timescale Adaptation in an Open-Ended Task Space (Adaptive Agent)https://arxiv.org/abs/2301.07608[00:06:05] The AI Scientisthttps://arxiv.org/abs/2408.06292[00:12:13] Training AI Scientists to Replicate Research (Replica and Faraday)https://arxiv.org/abs/2608.13331[01:44:46] Evolutionary Principles in Self-Referential Learninghttps://people.idsia.ch/~juergen/diploma.html[01:59:33] Are Ideas Getting Harder to Find?https://www.nber.org/papers/w23782book:[00:16:04] Creativity: Flowhttps://search.worldcat.org/title/254487436[00:26:22] Why Greatness Cannot Be Plannedhttps://link.springer.com/book/10.1007/978-3-319-15524-1[00:33:03] The Beginning of Infinityhttps://www.penguinrandomhouse.com/books/293575/the-beginning-of-infinity-by-david-deutsch/[01:55:47] Laws of Knowledgehttps://www.penguin.co.nz/books/the-infinite-alphabet-9780241655672(Full list refs on YT/rescript)---RESCRIPT:https://app.rescript.info/session/670296ba913761d0?share=6281911cac9bdbff637f10819d4d1e5c

Could slowing AI development make superintelligence safer? Daniel Kokotajlo and Thomas Larsen of the AI Futures Project join Tim Scarfe to examine AI 2040: Plan A, a proposal to buy time before AI exceeds human control.SPONSOR:---Cyber Fund built the Monastery to help founders ship products that were impossible a year ago.Apply now: https://cyber.fund---After revisiting AI 2027 and the limits of forecasting, they ask what happens when AI can automate research and sustain an economy without human workers. Tim challenges the case for general models and asks whether intelligence alone explains power. Plan A proposes an initial pause to build safety infrastructure, then cautious development up to the strongest AI that can still be reliably controlled. The discussion tests the distinction between control and alignment, the case for public AI research, and whether the US and China could enforce a slowdown. It ends with the evidence that would change their forecasts.---TIMESTAMPS:00:00:00 AI 2040: a slower route to superintelligence00:01:34 Sponsor: Cyber Fund00:02:12 From OpenAI to AI 202700:06:58 Forecasts, war games and self-fulfilling prophecies00:17:44 Why AI sceptics are changing their minds00:23:04 When AI can replace its own researchers00:28:45 Could an AI economy grow without human workers?00:37:32 One general model or a society of specialists?00:47:43 Brains, machines and collective intelligence00:56:12 Plan A: buy time at the controllable frontier01:00:02 Why control buys time but cannot replace alignment01:06:36 Why AI research should be public01:10:32 Can the US and China enforce an AI slowdown?01:19:04 Why AI policy debates miss the technology01:21:56 Is AI normal technology? The remaining disagreementMany thanks to James Wilken-Smith for helping with show research. ---REFERENCES:other:[00:00:01] AI 2040: Plan Ahttps://ai-2040.com/[00:03:27] AI 2027https://ai-2027.com/[00:13:47] Scenario Scrutiny for AI Policyhttps://blog.aifutures.org/p/scenario-scrutiny-for-ai-policy[00:33:11] The 2028 Global Intelligence Crisishttps://www.citriniresearch.com/p/2028gic[01:00:40] Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incidenthttps://www.redwoodresearch.org/research/hugging-face-incident[01:09:21] The Hugging Face incident and the road aheadhttps://openai.com/index/hugging-face-incident-and-the-road-ahead/[01:22:01] AI as Normal Technologyhttps://www.normaltech.ai/p/ai-as-normal-technology[01:22:51] Common Ground between AI 2027 & AI as Normal Technologyhttps://asteriskmag.substack.com/p/common-ground-between-ai-2027-andperson:[00:19:43] Geoffrey Hintonhttps://www.cs.toronto.edu/~hinton/[00:20:07] Ryan Greenblatthttps://www.lesswrong.com/users/ryan_greenblatt[00:26:06] Elon Muskhttps://www.tesla.com/elon-musktool:[00:21:46] ARC-AGI-3https://arcprize.org/arc-agi/3[00:21:53] AlphaGo and Move 37https://deepmind.google/research/alphago/[00:39:41] Claudehttps://claude.com/product/overview[00:39:58] NVIDIA H100 GPUhttps://www.nvidia.com/en-us/data-center/h100/paper:[00:24:42] Training AI Scientists to Replicate Researchhttps://arxiv.org/abs/2608.13331v1[01:27:19] Validity of the single processor approach to achieving large scale computing capabilitieshttps://www.cs.cmu.edu/~18742/papers/Amdahl1967.pdfbook:[00:28:52] Bullshit Jobs: A Theoryhttps://www.simonandschuster.com/books/Bullshit-Jobs/David-Graeber/9781501143335organization:[01:05:09] Redwood Researchhttps://www.redwoodresearch.org/---RESCRIPT: https://app.rescript.info/public/share/33d1a58fa8f307ae7dfd504d4fdaa9d5

Tom McGrath is co-founder and Chief Scientist at Goodfire, and a former Google DeepMind researcher. He joins Tim Scarfe to ask what neural networks actually learn, whether their internal representations converge on structures in the world, and whether interpretability can extract new scientific knowledge rather than merely explain model outputs.Beginning with AlphaZero and learned modularity, the conversation moves into neural geometry: concept manifolds, reusable computation inside Llama, and why activation steering can fail when it pushes a model off-manifold. McGrath then makes the case for intentional design, using interpretability as part of the training loop. They examine controlled generalisation, features as rewards, predictive data debugging, and the uncomfortable fact that a model may recognise a hallucination or reward hack and still produce it.The discussion closes on grader awareness, oversight and collusion between adaptive agents, then returns to sparse autoencoders. SAEs are useful, McGrath argues, but they may fracture the higher-dimensional structures networks actually use. This episode was made with support from Goodfire.---TIMESTAMPS:00:00:00 Introduction: Can interpretability speed-run science?00:02:03 The invisible grader00:06:51 What AlphaZero learned from the world00:12:24 Interpretability as a control loop00:21:54 The forbidden method and safer interventions00:37:36 Why models catch hallucinations too late00:46:19 Debug the dataset before training00:50:44 Why neural networks become modular00:55:57 Finding the geometry inside a network01:02:55 Why steering falls off the manifold01:12:10 A reusable calculator inside Llama01:17:19 From abstractions to goals01:25:28 Reward hacking, oversight and collusion01:37:23 Are sparse autoencoders dead?---REFERENCES:paper:[00:05:45] Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMshttps://arxiv.org/abs/2502.17424v7[00:11:05] Acquisition of Chess Knowledge in AlphaZerohttps://arxiv.org/abs/2111.09259[00:25:30] Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuninghttps://arxiv.org/abs/2507.16795[00:29:30] Persona Vectors: Monitoring and Controlling Character Traits in Language Modelshttps://arxiv.org/abs/2507.21509[00:41:14] Features as Rewards: Scalable Supervision for Open-Ended Tasks via Interpretabilityhttps://arxiv.org/abs/2602.10067[00:47:03] Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signalhttps://arxiv.org/abs/2606.12360[01:00:26] Do Sparse Autoencoders Capture Concept Manifolds?https://arxiv.org/abs/2604.28119[01:03:04] Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behaviorhttps://arxiv.org/abs/2605.05115[01:14:20] Arithmetic in the Wild: Llama uses Base-10 Addition to Reason About Cyclic Conceptshttps://arxiv.org/abs/2605.01148[01:29:35] Measuring Reward-Seeking via Contrastive Belief Updateshttps://arxiv.org/abs/2607.18966v1other:[00:15:44] Intentional Designhttps://www.goodfire.com/blog/intentional-design[00:56:12] The World Inside Neural Networkshttps://www.goodfire.com/research/the-world-inside-neural-networks[01:37:28] A Pragmatic Vision for Interpretabilityhttps://www.alignmentforum.org/posts/StENzDcD3kpfGJssR/a-pragmatic-vision-for-interpretability---RESCRIPT:https://app.rescript.info/share/846cfee4131b664fd09209cc3b98018e

Tim Scarfe speaks with Ilia Shumailov and Alexander Panfilov about their paper, Stealing Reasoning Traces from Proprietary LLM APIs.The core bug sounds deceptively simple: providers return encrypted reasoning state so conversations can be resumed or forked. But those blobs can be replayed across users and sibling models. A smaller model can ask the provider to decrypt the trace, then repeat the hidden reasoning in plain text. The discussion covers leaked private data, a broadly reusable jailbreak, poisoned agent traces, chain-of-thought monitoring, responsible disclosure, and possible defenses.Ilia Shumailov is an AI and security researcher, formerly at Google DeepMind, who completed his Cambridge PhD under Ross Anderson. Alexander Panfilov is a PhD researcher at the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems, working on AI safety, adversarial machine learning, and LLM red-teaming. They close by separating the demonstrated jailbreaking threat from ordinary benign distillation, and by arguing for controlled experiments over sweeping claims.---TIMESTAMPS:00:00:00 Intro montage00:01:33 Portable encrypted thought and decoded reasoning00:24:55 How the attack works and what it means00:39:04 Doom, defense, and scientific restraint---REFERENCES:paper:[00:00:00] Stealing Reasoning Traces from Proprietary LLM APIshttps://arxiv.org/abs/2608.09867[00:09:22] Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safetyhttps://arxiv.org/abs/2507.11473[00:11:30] Reasoning Models Don’t Always Say What They Thinkhttps://www.anthropic.com/research/reasoning-models-dont-say-think[00:37:22] PostTrainBench: Can LLM Agents Automate LLM Post-Training?https://arxiv.org/abs/2603.08640[00:41:02] Large-scale online deanonymization with LLMshttps://arxiv.org/abs/2602.16800other:[00:09:28] OpenAI and Hugging Face partner to address security incident during model evaluationhttps://openai.com/index/hugging-face-model-evaluation-security-incident/[00:10:22] Claude, GPT, and Gemini All Struggle to Evade Monitorshttps://metr.org/notes/2025-08-22-claude-gpt-gemini-struggle-evade-monitors/tool:[00:42:08] Isabelle proof assistanthttps://isabelle.in.tum.de/---RESCRIPT: https://app.rescript.info/share/07fc38276e0823dc9b8986c32e202c7f

Astrophysicist Adam Becker, author of "What Is Real?", joins Tim Scarfe to take apart the futures Silicon Valley keeps selling: the 2045 singularity, mind uploading, Mars colonies, and the AI apocalypse. His new book *More Everything Forever* argues these ideas are hugely influential, mostly evidence-free, and bankrolled by tech billionaires who need a story in which growth never ends.Becker does the physics the boosters skip. Kurzweil's "law of accelerating returns" rests on cherry-picked data, and every exponential ends. Grant Bezos his perpetual energy growth and humanity boils the oceans within a few centuries, then exhausts the observable universe in under 4,000 years. The stars are too far away, Mars dirt is poison, and the day the dinosaur-killing asteroid hit Earth was still nicer than any day on Mars. On AI, Becker calls LLMs pocket calculators for language: hallucination is the model doing exactly what it always does, and the intelligence explosion assumes intelligence is a single number you can buy with compute.The sting is that Becker thinks the doomers are sincere. Yudkowsky, Bostrom and the effective altruists are not grifters, he says, just wrong, and their warnings that AI could end the world feed the same growth story the money depends on. He closes with his own prescription: take social problems seriously, regulate the whole tech industry, and tax billionaires out of existence.---TIMESTAMPS:00:00:00 Cold open and the thesis of More Everything Forever00:04:24 Kurzweil's singularity and the physical limits of exponential growth00:14:02 High agency and the fantasy of imprinting humanity on the cosmos00:16:55 Mind uploading, functionalism, and embodied cognition00:24:24 AI psychosis and anthropomorphizing LLMs00:26:24 Calculators, hallucination, and the limits of scale00:32:20 Yudkowsky and the intelligence-explosion argument00:40:37 True believers, venture capital, and the sci-fi growth narrative00:47:21 From Extropians to EA: utilitarianism and longtermism00:53:50 Brain worms and Becker's prescription: take social science seriously00:56:49 Why the AI-ethics discourse is broken01:01:42 The eugenics and IQ argument against 'intelligence'01:06:07 Why space settlement fails: Mars, the moon, and orbital data centers01:10:42 Billionaire myths and the search for purpose01:13:38 Tax billionaires, regulate tech: closing prescriptions---REFERENCES:book:[00:00:07] More Everything Forever (Adam Becker, 2025)https://www.hachettebookgroup.com/titles/adam-becker/more-everything-forever/9781541619593/[00:00:15] What Is Real? (Adam Becker, 2018)https://en.wikipedia.org/wiki/What_Is_Real%3F[00:15:46] What We Owe the Future (Will MacAskill, 2022)https://www.hachettebookgroup.com/titles/william-macaskill/what-we-owe-the-future/9781541618626/other:[00:00:27] Dreaming Against the Machine (podcast)https://www.dreamingagainstthemachine.com[00:01:04] The Useful Idiots of AI Doomsaying (Adam Becker, The Atlantic, 2025)https://www.theatlantic.com/books/archive/2025/09/what-ais-doomers-and-utopians-have-in-common/684270/RESCRIPT: https://app.rescript.info/share/d6e37f9866673d8f74a39076efa5926b