Machine Learning Street Talk (MLST)

Tsung-Hsien (Shawn) Wen, CTO of PolyAI, tells Tim Scarfe why voice agents are harder than text agents. Voice adds time, and a good conversation depends on adapting to the person on the line, not just on reasoning to the best answer. Shawn describes an audio-native model (Dialog-RSN-1) that first predicts a turn-taking signal, then replies in text with citations, and writes the transcript last so enterprises can audit it.Along the way: training on real, noisy calls with synthetic noise added, and why over-cleaned audio made the new model worse. Latency, and what a voice agent should do while it thinks. Why a voice with a hint of regional accent beats a generic one. Why public benchmarks fall short for voice, why enterprises want to own their agent harness, and whether behaviour belongs in the harness or in the weights.The last stretch is about working with agents: cognitive debt, the shift from producing content to checking it, Wispr Flow, building tools that agents can use, and whether slop is in the eye of the reader.This episode was produced in partnership with PolyAI.https://poly.ai CHAPTERS0:00 Why voice agents are harder than text4:27 What enterprises want, and why PolyAI built its own model9:04 How an audio-native voice model works15:21 Training data, spectrograms and synthetic noise20:36 The cocktail party problem and the future of turn-taking25:21 Latency, adaptive reasoning and keeping callers' trust31:33 Voices, personality and the uncanny valley36:32 How do you benchmark a voice agent?42:00 Harness engineering and owning the intelligence45:09 Well-specified problems and auditable agents49:46 Weight adaptation and cognitive debt56:31 Agents at work: Wispr Flow, voice and tool building1:02:50 The next decade of voice, and what counts as slopREFERENCESThe Bitter Lesson: http://www.incompleteideas.net/IncIdeas/BitterLesson.html [9:05]Retrieval-augmented generation: https://arxiv.org/abs/2005.11401 [13:23]Mel scale: https://en.wikipedia.org/wiki/Mel_scale [17:16]Victor Zue: https://en.wikipedia.org/wiki/Victor_Zue [17:45]Cocktail party effect: https://en.wikipedia.org/wiki/Cocktail_party_effect [20:39]Speaker diarisation: https://en.wikipedia.org/wiki/Speaker_diarisation [21:16]

Leonardo de Moura created Lean and co-created Z3. ---This episode is sponsored by Parallel.Parallel, where agents find answers: web search, extraction and deep research APIs built for AI agents.Start free with the Parallel MCP server and $5 of credits every month: https://parallel.ai/mlst?utm_source=creator&utm_medium=podcast&utm_content=MLST---Tim Scarfe talks with Leo about how Lean escaped its original audience, why dependent types and Mathlib made it useful to working mathematicians, and what happens when formal verification leaves the lab. De Moura explains the small trusted kernel and independent checkers, and gives his account of the recent Collatz incident, in which a purported proof was accepted by both Lean's official kernel and Nanoda, apparently by exploiting a different bug in each.---TIMESTAMPS:00:00:00 Cold open: the green checkmark can lie00:00:59 Cathedral or bazaar: who controls Lean's core?00:04:55 Why Lean's core stays small and protected00:08:07 The Slack purge, Brandolini's law and the Lean FRO00:11:12 The Collatz exploit: two kernels, two bugs00:16:44 More kernels, reward hacking and safety by transparency00:21:04 Sponsor: Parallel00:21:59 Kim Morrison, Claude and the zlib proof00:25:25 Can we specify complex systems?00:28:04 Specs change: proofs are cheaper to redo with AI00:31:27 From Lean 1 to Lean 400:34:57 Dependent types in plain terms00:37:25 Lean 4's extensibility and Mathlib's growth00:41:44 Mathlib as infrastructure: Formal Frontiers00:44:15 Creativity, abstraction and nut-sniping00:48:46 Breadcrumbs, not learning: what AI agents lack00:53:03 Competence without comprehension, and verified guardrails00:56:21 Is the human still the author?01:01:44 AlphaProof, LLMs and why certificates still matter01:06:38 What's next for Lean, and its legacy01:11:42 How to start learning Lean---REFERENCES:tool:[00:00:48] Leanhttps://lean-lang.org/[00:01:24] Mathlibhttps://github.com/leanprover-community/mathlib4[00:12:09] nanoda_libhttps://github.com/ammkrn/nanoda_lib[00:12:19] CollatzLeanhttps://github.com/xrchz/CollatzLean/blob/a79357462a33d2a6babd4cf6c8d8bcd25425d653/README.md[00:13:06] Lean issue 14576https://github.com/leanprover/lean4/issues/14576[00:13:21] Lean pull request 14577https://github.com/leanprover/lean4/pull/14577[00:13:45] nanoda_lib pull request 22https://github.com/ammkrn/nanoda_lib/pull/22[00:15:33] Lean comparatorhttps://github.com/leanprover/comparator[00:18:02] Lean4Leanhttps://github.com/digama0/lean4lean[00:19:50] ARC-AGI-3https://arcprize.org/arc-agi/3[00:21:59] lean-ziphttps://github.com/kim-em/lean-zip[00:29:45] CompCerthttps://compcert.org/[00:29:45] seL4https://www.sel4.org/[00:30:23] Z3https://github.com/Z3Prover/z3[00:38:10] Veilhttps://github.com/verse-lab/veil[00:38:10] Velvethttps://github.com/verse-lab/velvetorganization:[00:00:52] Lean FROhttps://lean-lang.org/fro/[00:42:39] Mathlib Initiativehttps://mathlib-initiative.org/about/person:[00:05:11] Ilya Sergeyhttps://ilyasergey.net/[00:10:31] Joachim Breitnerhttps://www.joachim-breitner.de/[00:32:58] Adam Chlipalahttps://adam.chlipala.net/[00:38:42] Kevin Buzzardhttps://www.ma.imperial.ac.uk/~buzzard/[01:10:16] Terence Taohttps://terrytao.wordpress.com/book:[00:08:19] The Proof in the Codehttps://us.macmillan.com/books/9780374620059/theproofinthecode/other:[00:10:05] Brandolini's lawhttps://en.wikipedia.org/wiki/Brandolini%27s_law[00:45:33] A new result on unit distanceshttps://openai.com/index/model-disproves-discrete-geometry-conjecture/[01:01:08] Fermat's Last Theorem formalisationhttps://imperialcollegelondon.github.io/FLT/paper:[00:34:36] The Lean 4 theorem prover and programming languagehttps://doi.org/10.1007/978-3-030-79876-5_37---RESCRIPT:https://app.rescript.info/share/7d3d4a0059443236a01f6c9acbf4db58https://app.rescript.info/api/public/sessions/b007264c0ce89047/pdf

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