The strange thing about a noisy telephone call is that the noise may be useful. In Machine Learning Street Talk’s October 2026 conversation with PolyAI’s Shawn Wen, an apparently sensible improvement turns out to be a mistake: cleaning training audio too aggressively made a new voice model worse. The imperfections belonged to the world the model would encounter. Remove them, and you could teach it the wrong world.
- Free technical interviews across AI, cognitive science and philosophy.
- Paid memberships add community access and extended material.
- The payoff: better questions about what an AI system actually does.
This is a good introduction to MLST. A product announcement might give you the model’s name. A benchmark might give you a score. Here, the interesting detail is the assumption that failed. You can take that detail into a meeting. If the demonstration is immaculate, ask what happens when the customer calls from a kitchen with the extractor fan running.
The noise is the story
MLST is a media company whose principal product is a conversation. Video interviews become audio episodes; the surrounding material includes chapter markers, references and written arguments. Its subject is artificial intelligence, but the catalogue wanders into the neighbouring rooms: mathematics, neuroscience, cognitive science and the philosophy of mind.
Those rooms matter. In Wen’s interview, voice adds a problem that text can hide: timing. When should an agent interrupt? How long can it think before a caller assumes it has disappeared? The episode spends 70 minutes working through such questions. Its description also states that it was produced in partnership with PolyAI. The commercial relationship is part of the context.
Before the chatbot gold rush
The project was already discussing research in 2020. An April episode introduced three YouTubers - Tim Scarfe, Connor Shorten and Yannic Kilcher - who had created a channel called Machine Learning Street Talk. Keith Duggar appeared among its special guests. Their topic was POET, an approach to generating learning environments and solutions together. The discussion lasted 73 minutes.
By September, Scarfe and Kilcher were spending 97 minutes on kernel methods with Alex Stenlake. The episode’s chapter list included computational tractability, overfitting and the curse of dimensionality. A listener could spend an entire commute inside a technique that predated the current chatbot boom. The catalogue’s interests were broader than whichever product happened to dominate the news.
Today, Scarfe runs MLST and Duggar regularly joins the questioning. Scarfe has a machine-learning doctorate; Duggar earned his doctorate at MIT. Their stated production habit is preparation and editing. Scarfe’s compact description of the editorial contract is: “We don’t dumb things down.”
“We don’t dumb things down.”
Tim Scarfe · MLST’s editorial approach
That contract helps explain where the company fits. AI listeners can choose industry discussions such as Latent Space, broader interviews from Lex Fridman, or the strategic conversations of Dwarkesh. MLST’s useful territory is the mechanism: how an idea works, what the experiment tests and where the interpretation becomes debatable. These shows can sit in the same listening queue.
Who pays for the next question?
There are two audiences in this business. One wants to understand the technology. The other wants to reach people who understand the technology. MLST identifies engineers and research scientists as its audience. Its LinkedIn page describes a small London media company; a UK limited company bearing its name was incorporated in May 2024, with Scarfe as director.
The public interviews are free. Patreon sells closer participation: extended episodes, ad-free videos, a private Discord and calls with Tim and Keith. The public membership description advertises 90-minute calls every two weeks. In-person participation has also had a price: a past Waterloo networking event advertised a £25 ticket, pizza and Prolific sponsorship.
The sponsor side is visible in the episodes themselves. A September conversation with NVIDIA’s Ming-Yu Liu discloses a paid partnership. The Leonardo de Moura interview names Parallel as sponsor. For a listener, the sensible habit is to keep those labels attached to the conversation. A technically detailed sponsored interview can still be useful; its claims still deserve inspection.
Eight days, four levels, one useful doubt
Consider MLST’s September interview with Weco co-founder Zhengyao Jiang. An AI coding agent had spent eight days rewriting the machinery around another agent: code, prompts and tools. The underlying language model stayed fixed. The reported improvements made an appealing story. Scarfe’s question was what those improvements actually demonstrated.
The discussion separated levels of recursive self-improvement and examined evaluation and reward hacking. Crucially, Jiang explained that the experiment had not established that the system became a better improver. Better performance and an improving capacity to improve are different claims. That distinction is easy to lose in a headline and expensive to lose in a research plan.
A neighbouring episode asked Lean creator Leonardo de Moura about checking mathematical proofs. He discussed an incident in which two checkers apparently accepted a purported proof through different bugs. The question travels neatly beyond mathematics: if your system reports success, how independent is the thing checking it? MLST gives that question a concrete place to begin.
Take the question home
The company also puts arguments of its own into circulation. Its archive carries Why Creativity Cannot Be Interpolated, dated February 2026, by Jeremy Budd and Scarfe. The authors argue that novelty alone does not establish creativity, and that understanding constraints matters. It is an argument to examine, rather than a settled finding to repeat.
The most practical way to use MLST is selectively. Choose a problem you have, find the relevant chapter, then follow the paper or tool linked in the episode notes. Write down the claim, the test and the conditions. Bring the unresolved question to a colleague. You have turned an interview into something you can work with.
For a team evaluating an AI supplier, that exercise can become a short discussion after the episode. Which part of the result came from the model, and which part came from the surrounding tools? Was the test drawn from the same conditions as the training? What would make the result stop holding? These are suggested uses of the material, not services MLST sells. They give a technical interview a job beyond keeping you company on the train.
This approach asks for time and some technical footing. A conversation series cannot supply the sequence of a course, and an unfamiliar mathematical vocabulary can make audio frustrating. Pause, watch the video, or read first. The useful reward is a smaller distance between hearing that something works and knowing what “works” means.