Machine Learning Street Talk (MLST)

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-3-5
https://priorlabs.ai/careers

TOC:
00:00 Introduction
00:44 Welcome and Frank's background
02:05 Why tabular data was hard for deep learning
10:17 Pre-training on synthetic data
12:52 The TabArena benchmark
19:28 From AutoML to neural architecture search
26:34 TabPFN as a learned algorithm
30:50 Bayesian prediction in one forward pass
39:37 Scaling to larger tables
47:48 Using TabPFN with coding agents
57:47 Output heads and architecture from v1 to v3
1:05:29 Test-time compute and adaptation
1:13:32 Google's TabFM
1:16:53 How the priors are designed
1:18:40 Correlation, causation and interventions
1:35:22 Relational and multimodal data
1:38:31 Use in organisations
1:46:38 The open research arm
1:50:21 Update: TabPFN-3.5

REFS:
TabPFN v2, Nature (Hollmann et al., 2025)
https://www.nature.com/articles/s41586-024-08328-6
Transformers Can Do Bayesian Inference (Müller et al.)
https://arxiv.org/abs/2112.10510
TabArena (Erickson et al.)
https://arxiv.org/abs/2506.16791
AutoGluon-Tabular (Erickson et al.)
https://arxiv.org/abs/2003.06505
Beyond IID: How General Are Tabular Foundation Models, Really?
https://arxiv.org/abs/2606.30410
Neural Architecture Search: A Survey (Elsken, Metzen & Hutter)
https://arxiv.org/abs/1808.05377
Auto-WEKA (Thornton et al.)
https://www.cs.ubc.ca/~hutter/papers/AutoWEKA-KDD2013.pdf
TabPFN v1 (Hollmann et al., 2022)
https://arxiv.org/abs/2207.01848
TabPFN-3 technical report
https://arxiv.org/abs/2605.13986
TabPFN-2.5 report
https://arxiv.org/abs/2511.08667
CAAFE (Hollmann et al.)
https://arxiv.org/abs/2305.03403
TabICL (Qu et al.)
https://arxiv.org/abs/2502.05564
TabICLv2 (Qu et al.)
https://arxiv.org/abs/2602.11139
Google TabFM
https://research.google/blog/introducing-tabfm-a-zero-shot-foundation-model-for-tabular-data/
TALENT benchmark (Ye et al.)
https://arxiv.org/abs/2407.00956
Do-PFN (Robertson et al.)
https://arxiv.org/abs/2506.06039
CausalPFN (Balazadeh et al.)
https://arxiv.org/abs/2506.07918
Causal Foundation Models with Partial Graphs (Reuter et al.)
https://arxiv.org/abs/2602.14972
RelBench (Robinson et al.)
https://arxiv.org/abs/2407.20060
RelArena-α, TabPFN-Rel and RPI
https://arxiv.org/abs/2608.16319
TabPFN on GitHub
https://github.com/PriorLabs/TabPFN
TabPFN-3.5 technical report
https://arxiv.org/abs/2609.17895
Otto 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 expensive
0:51 Welcome back, Alexander Mattic
2:08 Alexander's research background
2:50 Inference: densities, sampling and Monte Carlo
6:42 GFlowNets, energy functions and MCMC
9:45 Explainer: energy-based models
11:03 Why model a density at all?
17:30 From learned energies to flow matching
25:08 Explainers: diffusion and normalising flows
28:33 Are energy-based models generative?
33:22 JEPA, contrastive learning and collapse
41:13 Why non-language modalities need flows
44:51 Inference as search: branch and bound
49:43 Q-learning and delayed consequences
55:14 Flow matching, optimal transport, Fokker-Planck
1:00:03 Explainer: flow matching
1:01:49 AlphaFold, latents and scale versus architecture
1:07:52 Two families of deep learning theory
1:15:04 What a good theory would predict
1:23:53 The manifold hypothesis and compression
1:28:25 Is reward enough?
1:32:01 Control theory versus reinforcement learning
1:37:22 The Bitter Lesson and expensive information
1:42:08 Constrained RL: the constrained MDP toolbox
1:50:12 Creativity as constrained search
1:55:44 Reality is protean: when abstractions hold
2:00:32 What is a world model?
2:04:38 Prediction is not control
2:08:13 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 networks
https://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 Interpolated
https://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/4cJoY1S
Explore Cosmos Lab: https://research.nvidia.com/labs/cosmos-lab/cosmos3/

---
TIMESTAMPS:
00:00:00 A road that was never filmed
00:02:28 Inside Cosmos 3: reasoning and generator towers
00:05:02 World models: dynamics, policy and one clock
00:08:59 Learning robot skills from human video
00:11:06 Ambiguous tasks and system 2 planning
00:12:53 Neural simulators for policy verification
00:16:41 Cosmos as a starting point for robot policies
00:19:00 Cosmos Dreams and robot safety
00:22:04 Super, Nano and Edge model sizes
00:24:24 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 repository
https://github.com/NVIDIA/cosmos
[00:22:05] Cosmos3-Edge model card
https://huggingface.co/nvidia/Cosmos3-Edge
[00:22:15] Cosmos3-Super model card
https://huggingface.co/nvidia/Cosmos3-Super
[00:22:16] Cosmos3-Nano model card
https://huggingface.co/nvidia/Cosmos3-Nano
[00:22:50] NVIDIA Jetson Thor
https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/
[00:22:52] NVIDIA Jetson Orin
https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/
[00:22:53] NVIDIA DGX Spark
https://www.nvidia.com/en-us/products/workstations/dgx-spark/
[00:24:42] Cosmos 3 collection on Hugging Face
https://huggingface.co/collections/nvidia/cosmos3
other:
[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 AI
https://arxiv.org/abs/2606.02800
[00:17:43] DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset
https://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 open
00:00:46 Why Mistral moved into audio
00:09:27 Inside Voxtral: trunk, encoder, dual streams
00:20:22 Speech that works in real time
00:30:52 How a voice becomes tokens
00:39:59 Flow matching, FSQ and the new codec
00:52:51 When speech models lose the speaker
01:03:23 Correcting hallucinations with preferences
01:12:12 Controlling synthetic speech
01:20:06 Why cascades still win
01:29:25 Speech in the wild
01:33:46 Audio models as interfaces
01:37:54 Why voice still needs a screen

---
REFERENCES:
paper:
[00:01:42] Mistral 7B
https://arxiv.org/abs/2310.06825
[00:09:38] Voxtral
https://arxiv.org/abs/2507.13264
[00:14:41] Whisper: Robust Speech Recognition
https://arxiv.org/abs/2212.04356
[00:19:11] Voxtral Realtime
https://arxiv.org/abs/2602.11298
[00:21:52] Delayed Streams Modeling (Kyutai)
https://arxiv.org/abs/2509.08753
[00:30:52] Voxtral TTS
https://arxiv.org/abs/2603.25551
[00:32:38] SoundStream neural audio codec
https://arxiv.org/abs/2107.03312
[00:34:59] Flow Matching for Generative Modeling
https://arxiv.org/abs/2210.02747
[00:37:03] EnCodec: High Fidelity Neural Audio Compression
https://arxiv.org/abs/2210.13438
[00:37:42] Moshi and the Mimi codec
https://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/2305.18290
dataset:
[00:46:14] Mozilla Common Voice
https://commonvoice.mozilla.org/en/datasets
organization:
[00:50:47] Hugging Face
https://huggingface.co/