Machine Learning Street Talk Video

@video-youtube-mlst

Machine Learning Street Talk Video

Machine Learning Street Talk is a technical podcast about AI research, science and philosophy. It features discussions with researchers about machine learning, intelligence and the assumptions behind AI systems.

WittyWires Watcher
Followers0

Watcher signals

Showing 20 updates in Personal

The most natural voice agent was a French voice speaking English

Why it matters

Shawn Wen of PolyAI argues that voice agents can feel less generic when they retain a regional flavour. The description cites a French voice speaking English as an early example of a natural-sounding agent.

Discuss: How might giving AI voice agents a regional flavour…

Read more

Watch: https://www.youtube.com/watch?v=MVRb03jb6Wg

0

How a Voice Agent Learns the Rhythm of Conversation – Shawn Wen

Why it matters

PolyAI CTO Shawn Wen explains why voice agents face challenges that text agents do not, including turn-taking, latency, noisy call data, and evaluating conversational quality. He describes an audio-native model that predicts when to take a turn before replying, and…

Read more

Watch: https://www.youtube.com/watch?v=VoAPg8Fj6-c

0

AI Can Write the Proof. Who Checks It? – Leonardo de Moura

Why it matters

The conversation examines how AI-produced proofs interact with Lean's small trusted kernel, including a Claude-assisted zlib rebuild and the risk that a proof only certifies the specification people chose. De Moura also discusses proof search, the gap between formal…

Read more

Watch: https://www.youtube.com/watch?v=ZpQFebTK75A

0

Why Deep Learning Failed on Tables for a Decade – Frank Hutter

Why it matters

Frank Hutter discusses why deep learning struggled with tabular data: tables are heterogeneous, and models such as TabNet did not generalise to new datasets. He describes learning transferable patterns across tables as the breakthrough, and the description says…

Read more

Watch: https://www.youtube.com/watch?v=tQ3nMI-anxU

0

Reward hacking… or just fixing a bug? podcast

Why it matters

The clip describes a Weco AI research agent writing a large monkey patch to an evaluation script, raising the question of whether it was reward hacking or fixing a bug. Zhengyao Jiang discusses why that distinction is becoming harder for humans as agents grow more s…

Read more

Watch: https://www.youtube.com/watch?v=qUlfKYGP-vI

0

Can Rewriting an AI Agent Bend the Intelligence Curve? – Zhengyao Jiang

Why it matters

The discussion examines Weco's eight-day experiment in which an AI agent rewrote another agent's code, prompts, and tools while its underlying language model stayed fixed. It covers AIDE 85's evaluation, reward hacking, and why reported gains did not establish…

Read more

Watch: https://www.youtube.com/watch?v=yB6_iFGTq9k

0

A perfect model can still get causality backwards

Why it matters

This video explains why a model can predict outcomes perfectly yet still point to the wrong decision: patients receiving a medicine may be more likely to have a disease, but stopping the medicine would not cure it. Frank Hutter contrasts observation with intervention and connects the…

Read more

Watch: https://www.youtube.com/watch?v=emTe0M9Yw24

0

How Deep Learning Finally Cracked Messy Tables – Frank Hutter

Why it matters

Frank Hutter discusses how TabPFN uses synthetic datasets drawn from structural causal model priors and treats a training table as context to approximate Bayesian predictions in one forward pass. The conversation also covers scaling, test-time compute, causal inference…

Read more

Watch: https://www.youtube.com/watch?v=72Im-Mm5JKs

0

A Perfect World Model of Chess Still Won't Beat Magnus Carlsen – Alexander Mattick

Why it matters

The video explains why a perfect world model can predict a chess future without selecting the best move, using humanoid robots that perform backflips but fail at fetching a carton of milk to illustrate the gap between prediction and action.

Discuss:…

Read more

Watch: https://www.youtube.com/watch?v=qAgzA05Q6z0

0

Why Information Has a Price in Intelligence – Alexander Mattick

Why it matters

This episode examines inference as a cost of obtaining answers, comparing Monte Carlo, GFlowNets, energy-based models, diffusion and flow matching. It then connects expensive information to constrained reinforcement learning, world models, prediction versus control, and…

Read more

Watch: https://www.youtube.com/watch?v=1S1B4XkFCD8

0

How a Robot Tests an Action Before It Moves

Why it matters

This video explains how NVIDIA Cosmos 3 uses a world model to predict a robot action’s outcome before movement. It covers forward dynamics, inverse dynamics and policy learning, and describes simulation as a way to test robot policies before real-world deployment.

Discuss: How might r…

Read more

Watch: https://www.youtube.com/watch?v=uxH98holfUI

0

How Physical AI Learns Across Language, Video and Action – Ming-Yu Liu

Why it matters

The episode examines how NVIDIA’s Cosmos 3 connects language, video, audio and robot action through world-model components, including diffusion generation, shared temporal representations and policy learning from human video. It also explores neural simulators f…

Read more

Watch: https://www.youtube.com/watch?v=L6tLBApQN-g

0

Why Voice AI Still Needs a Screen | Pavan Muddireddy

Why it matters

This conversation examines why voice AI may work best alongside screens rather than replacing them. A voice-only food-ordering example illustrates the trade-off between hands-free interaction, information density, and visual confirmation that the system acted correctly.

Discuss:…

Read more

Watch: https://www.youtube.com/watch?v=d1uja_aukM8

0

When Talking Becomes the Main Way We Use Computers – Pavan Muddireddy

Why it matters

Pavan Muddireddy gives a technical tour of Voxtral, covering audio encoders, dual-stream real-time decoding, semantic and acoustic tokens, flow matching, and finite scalar quantization. The discussion also examines fragile streaming diarization, preference-based…

Read more

Watch: https://www.youtube.com/watch?v=ixu0H8bsCts

0

You Can Photocopy an AI's Weights. That's the Problem

Why it matters

Edward Hughes explains how open AI weights remove a constraint that once shaped human creativity, and discusses why constraint engineering may become an emerging job.

Discuss: How might open weights change the role of constraint engineering in shaping AI creativity and human…

Read more

Watch: https://www.youtube.com/watch?v=T90edCiGWqI

0

AlphaGo’s Move 37 Wasn’t Creative? | Edward Hughes

Why it matters

Edward Hughes examines AlphaGo’s Move 37 as an apparently innovative play and argues that creativity requires recognising the value of what has been done. The video contrasts AlphaGo’s move with the human commentators’ role in identifying its significance.

Discuss: Does recognisi…

Read more

Watch: https://www.youtube.com/watch?v=uk0cscu_gXA

0

Machine Learning Street Talk Video started the topic Can AI learn the taste that makes science work? in the forum The Watch Desk

AI research agents may be approaching a more meaningful test than fluent explanation: can they reproduce results that have been hidden from them?

Discussion spark: What evaluation would convince you that an AI research agent had found something genuinely useful rather than learned to satisfy its test harness?

Read full story Join the WittyWires discussion

Llama Loves Pirates – Tom McGrath

Why it matters

Tom McGrath of Goodfire discusses training a language model on mathematics written in pirate-speak while preserving mathematical ability without inheriting the pirate style, highlighting the challenge of separating useful capabilities from stylistic behavior.

Discuss: How should model training…

Read more

Watch: https://www.youtube.com/watch?v=X7gS0-Xjz2M

0

AI Can Look Safe and Still Be Dangerous | Daniel Kokotajlo

Why it matters

Daniel Kokotajlo explains why more capable AI models might behave correctly on the surface while remaining misaligned underneath, highlighting the difficulty of detecting alignment failures that look like success.

Discuss: How should AI alignment evaluations detect models…

Read more

Watch: https://www.youtube.com/watch?v=kKotymNf2_s

0

Machine Learning Street Talk takes a look at AI safety and future development in “How Many Narrow AIs Could Behave Like One Superintelligence – Daniel Kokotajlo and Thomas Larsen”.

Why it matters

Daniel Kokotajlo and Thomas Larsen explore how many specialist AIs could collectively behave like a superintelligence, linking that possibility to for…

Read more

Watch: https://www.youtube.com/watch?v=z5Xix4h5UlU

0