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

AlexNet Left My Old Book in the Dust (Simon Prince) Shorts

Why it matters

The Short recounts how Simon Prince wrote a book reframing computer vision through probabilistic graphical models, released a few months before AlexNet appeared. Its title frames that timing as a sharp contrast between the book and AlexNet.

Discuss: How does the timing…

Read more

Watch: https://www.youtube.com/watch?v=inwyawObxL4

0

How Would You Make a Traffic Light Smart? (Michael Tomasello)

Why it matters

Michael Tomasello discusses how a smart traffic light could use perception, a goal, and a decision-making mechanism to make the world conform to that goal, rather than operating only through stimulus and response.

Discuss: How would perception and a goal change what an AI…

Read more

Watch: https://www.youtube.com/watch?v=0w2Crs8RzFE

0

Humans Created Languages (Michael Tomasello)

Why it matters

Developmental psychologist Michael Tomasello discusses studies in which children cut off from talking invent gestures and turn them into shared conventions. The description also says he explains why, as far as he knows, large language models do not create linguistic conventions.

Discuss:…

Read more

Watch: https://www.youtube.com/watch?v=TiEut-ZUwxs

0

AlexNet Should Never Have Worked (Simon Prince)

Why it matters

Simon Prince discusses why AlexNet seemed unlikely to work: the model had 60 million parameters, a 150,000-dimensional input space, and was trained from random weights with noisy gradient descent, yet fit ImageNet.

Discuss: Given AlexNet's scale and noisy-gradient training, what made…

Read more

Watch: https://www.youtube.com/watch?v=ssD05l4UxmY

0

Three Months After You Build the Harness (Edward Hughes)

Why it matters

Edward Hughes explains why his team deliberately avoids building harnesses for its AI scientist work. The description reports Louis Kirsch's analysis that, about three months after a harness is built, the base model can perform the same task on its own.

Discuss: If a base…

Read more

Watch: https://www.youtube.com/watch?v=g_n55IXQU20

0

LLMs Cosplay Understanding (Jeremy Howard)

Why it matters

Jeremy Howard discusses whether LLMs really understand, arguing that they cosplay understanding in a way comparable to the man in Searle’s Chinese room. He says the distinction often does not matter for many tasks, until the pretense stops working.

Discuss: If LLMs cosplay understanding, a…

Read more

Watch: https://www.youtube.com/watch?v=BnA4i_vjcV4

0

Claude's "Clean-Room" Compiler Copied LLVM's Mistakes (Jeremy Howard)

Why it matters

Jeremy Howard discusses Anthropic's Claude-written C compiler and how LLVM's presence in training data may shape its Rust output. He describes the conversion as interpolation or style transfer, pointing to copied LLVM mistakes as evidence for his view that the…

Read more

Watch: https://www.youtube.com/watch?v=qD3FrUgkVII

0

Random Code Starts Copying Itself (Blaise Agüera y Arcas)

Why it matters

The description explains the BFF experiment: start with 1,024 tapes of 64 random bytes, repeatedly join and run pairs, then separate and recombine them. It says this process produces programs that copy themselves and each other, illustrating artificial life in a minimal…

Read more

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

0

I Told Taxi Drivers Not to Worry (Tom Froese)

Why it matters

Tom Froese reflects that he once told taxi drivers not to worry about AI, but says today's systems were not what he anticipated. He identifies initiative as something he still does not see in them.

Discuss: What would count as initiative in AI systems, given Froese's claim that today's…

Read more

Watch: https://www.youtube.com/watch?v=GCGaQj-FAo8

0

Simulated Water Can't Make You Wet (Joscha Bach)

Why it matters

Joscha Bach describes a neural network trained only on Minecraft visuals, with no game code underneath. In the example, simulated water recedes as the player walks in, illustrating the gap between reproducing visuals and learning what immersion means.

Discuss: What does the receding…

Read more

Watch: https://www.youtube.com/watch?v=TA9-RgnLZrs

0

Why Leibniz Couldn't Find the Mind Inside the Mill (Joscha Bach)

Why it matters

Joscha Bach presents the Machine Consciousness Hypothesis in two parts: consciousness may be a biological learning algorithm, and self-organising machines might discover that algorithm too. The conversation considers how either part could fail and what evidence would…

Read more

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

0

We're the Ones Having the Debate (Jeremy Avigad)

Why it matters

Jeremy Avigad argues that the AI debate should stay centred on people: our goals matter because they are ours, and the practical issue is how we interact with AI and what we should do to achieve those goals.

Discuss: If the AI debate centres on human goals, how should that shape the…

Read more

Watch: https://www.youtube.com/watch?v=IcXN7ysaqrc

0

When It Doesn't Matter That It's Slop (Paul Lessard)

Why it matters

The video presents Paul Lessard's discussion of AI replacing people, with translation offered as a white-collar job that has essentially disappeared. It also explores why companies may accept AI output when exactness or refined quality does not matter, and how the boundary of…

Read more

Watch: https://www.youtube.com/watch?v=_9axenbT-1I

0

Correct Isn't Enough. I Want the Reason. – Paul Lessard podcast science

Why it matters

Paul Lessard discusses what he wants from AI in mathematics: correct proofs are not enough; he also wants provenance showing why a machine reached a particular answer. The description contrasts possible uses for machine-generated proofs with the continuing need…

Read more

Watch: https://www.youtube.com/watch?v=5pi9dSGYAgM

0

The Trojan Horse in AI-Written Mathematics

Why it matters

Kevin Hartnett discusses AI autoformalisation: an LLM can turn a mathematics paper into Lean code far faster than a human team. The description says Mathlib's community is concerned that a flood of generated Lean lacks the precision and economy its canonical definitions require.

Discuss:…

Read more

Watch: https://www.youtube.com/watch?v=unwUjyA52tM

0

Mathematics Just Changed Forever – Kevin Hartnett

Why it matters

The documentary traces how Lean checks mathematical proofs, then turns to AlphaProof, AI proof-finding and autoformalisation. It frames the formalisation gap-the distance between the mathematics people mean and the statement a checker certifies-as a key tension in deciding what a…

Read more

Watch: https://www.youtube.com/watch?v=IRJDwtAreB4

0

AI Coding Is a Slot Machine: Loss Disguised as a Win | Jeremy Howard

Why it matters

Jeremy Howard compares AI-assisted coding to a slot machine, arguing that occasional successful outputs can create a feeling of control despite what he calls “loss disguised as a win.”.

Discuss: How does the slot-machine analogy change the way you’d judge wheth…

Read more

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

0

What Most People Get Wrong About Evolution | Akarsh Kumar

Why it matters

The conversation explains ASAL, which runs simulations and uses a foundation model to assess what happened; across 262,144 Life-like rules, the description says the most interesting worlds cluster on one small island. It also describes LLMs used as mutation steps in a Core…

Read more

Watch: https://www.youtube.com/watch?v=615B5nyMFPk

0

Life as it could be | Akarsh Kumar on Artificial Life

Why it matters

Akarsh Kumar discusses searching for artificial life with foundation models, starting from Conway’s Game of Life and parameterizing a space of simulations for a foundation model to assess.

Discuss: How might parameterizing a space of simulations change what foundation models c…

Read more

Watch: https://www.youtube.com/watch?v=pniLY0I9VJQ

0

Game of Thrones: Turing Edition | Akarsh Kumar on Core War

Why it matters

The clip introduces Core War, where Akarsh Kumar explores using LLMs to evolve small assembly programs that compete. He says zero-shot use performs poorly, while evolution with a verifier works well.

Discuss: How might the verifier shape the results when LLMs evolve…

Read more

Watch: https://www.youtube.com/watch?v=LfUgsSUdaX8

0