By Akhilesh Tilotia
Audio digest · Week ending Sat, Oct 31, 2026 · 8:43
A weekly digest of what five episodes from four interview shows discussed in the week from Sun, Oct 25 to Sat, Oct 31. Each summary is in my own words and links to the episode. Transcripts were read in full for all five episodes. For one of them, the Mandarin audio was not heard and the English translation of the transcript was read.
Global
ChinaTalk: Roasting China’s AI industry, with Xianyu
Guest: Xianyu, the writer behind Funeral AI. Host: Irene Zhang. Released Thu, Oct 29.
Original episode | English transcript
Xianyu writes Funeral AI, a Chinese-language newsletter of satire about the AI industry. He worked earlier at the Chinese tech outlets 36Kr and Huxiu, and co-built MemeBench, a site where ordinary people rank video models by how good they are at memes. He describes himself as a techno-pessimist who would rather be in an internet cafe. The host is Irene Zhang of ChinaTalk.
- MemeBench asked paid students to rate video models on meme prompts. The open-source model H3 beat ByteDance’s Seedance 2.0 and 2.5, the reverse of the industry ranking in the voting by students paid to rate clips.
- He treats large models as manufacturing. With enough data and compute, a model gets better, no lead lasts more than a month or two, and competition is brutal, as it was in solar panels and electric vehicles.
- China’s AI industry follows Silicon Valley’s story cycle, and he says it lacks the confidence to coin terms or set benchmarks, which investors and companies then trust more when they are American.
- On humor, he says AI writing is not funny because jokes need a belief outside the consensus, and he stopped using AI to write in late 2025.
ChinaTalk: The Manhattan Project, take two, with Barnett Koven
Guest: Barnett Koven of Accenture Federal Services. Host: Jordan Schneider, with Phoebe Chow. Released Thu, Oct 29.
Barnett Koven is Managing Director of AI and Data at Accenture Federal Services. He holds a PhD on counterinsurgency and was a professor at the University of Maryland before moving to consulting. His team works on export-control enforcement for the Commerce Department and on the Energy Department’s Genesis Mission. The episode is sponsored by his employer, and he is its only guest.
- Government risk aversion is itself a risk. Koven says agencies that wait to buy down every risk are guaranteed to fall behind less cautious rivals, China in particular.
- In classified settings there is a gun and armor race, and it favors the attacker, because defense costs rise much faster than offense costs, which makes poisoned models and data leaks a particular worry.
- For export controls, the practical gain is entity resolution, merging public and government records to show which obscure shell company links to the Chinese military’s procurement arm. Koven says most data science effort is still data cleaning.
- Budgets are set about two years ahead, so he wants agencies to move token spending between models and uses, with finance controls for accountability.
AI
Lex Fridman Podcast #503: AI crisis in math, with Alex Kontorovich, Ankur Moitra, and Steven Brunton
Guests: Alex Kontorovich, Ankur Moitra, and Steven Brunton. Host: Lex Fridman. Released Mon, Oct 26.
Original episode | Transcript | Video
Alex Kontorovich is a number theorist at Rutgers who also works on Lean, the proof checker. Ankur Moitra is a theoretical computer scientist at MIT, visiting the Institute for Advanced Study this year. Steven Brunton is a fluid dynamics expert at the University of Washington. Lex Fridman called them in after AI systems announced proofs of several famous problems over a few weeks.
- Kontorovich does not call this a crisis. He sees a chance to rethink what mathematics is for, and compares the new tools to swapping a candle for a flashlight, while Moitra says the old career playbook for students is gone.
- What survives for humans is the why. Kontorovich says a true fact without understanding adds nothing, and Moitra argues that when solving well-defined problems gets cheap, the other parts of a mathematician’s work gain value.
- The pace surprised all three. A formal Lean proof of Fermat’s Last Theorem, about 13 million lines, appeared two days after Kontorovich told a colloquium that this was years away, and he says very few people were involved.
- On the Navier-Stokes result, Brunton, an engineer, says the math problem is only part of it. The bigger problem is predicting turbulence fast and cheaply, which he calls a trillion-dollar engineering problem.
Science and ideas
Mindscape 372: Erik Winfree on DNA robots and computers
Guest: Erik Winfree. Host: Sean Carroll. Released Mon, Oct 26.
Erik Winfree is a professor at Caltech and one of the leaders of DNA computation, which uses DNA as a way to store and act on information, not as genetic material. He was trained as a theoretical computer scientist and has worked in the field since 1994. He works with his wife, Lulu Qian, on DNA robots and molecular computing. The host is physicist Sean Carroll.
- Why DNA: Winfree is not after genetics. DNA is the simplest, easiest to order molecule that both stores information and acts on it, which makes it a playground for the principles of computation.
- DNA computing did not beat silicon, and he says it took his group at least five years to give up on that. His interest is programming molecules to carry out chosen behaviors, using a small set of building blocks the way a simple instruction set does.
- The grand challenge he names is a cell-like machine about a micron across that senses, acts, takes energy from its surroundings, and maybe replicates, all made of DNA, though he does not want to put it inside a cell.
- On the origin of life, he is drawn to a programming-language-first view, in which self-replicators arise in a simple molecular language, and he points to recent heat-powered DNA systems as a way to keep such systems running.
Dwarkesh Podcast: Robert Caro on power, Robert Moses, and Lyndon Johnson
Guest: Robert Caro. Host: Dwarkesh Patel. Released Fri, Oct 30.
Original episode | Transcript | Video
Robert Caro wrote The Power Broker, a biography of the New York builder Robert Moses, and a multi-volume biography of Lyndon Johnson. He has spent about 50 years on Johnson and is now writing the Vietnam volume. Dwarkesh Patel says he has wanted Caro on the show since he started it six years ago. The episode is about power, building, and how Caro works.
- Does America need a little more Moses? Caro says every era is told that nothing big can be built, and then a rare political genius builds it anyway, though he does not want Moses’s ruthlessness.
- Johnson’s persuasion worked because of the power behind it. When he leaned in, the senator knew Johnson controlled committee seats and bills.
- Caro’s method for a multi-year book is to boil it down to one sentence before he writes, a rule he says he learned the hard way.
- He says The Power Broker was cut from about 1,050,000 words to about 700,000, and he gave his archive to the New-York Historical Society so that scholars can use the unpublished material.
Not covered: Ideas of India had no episode in this window. ChinaTalk’s Oct 30 essays and Dwarkesh’s Oct 29 essay are written pieces, not interviews, and were skipped. Mindscape had one episode in the window.
Built by Akhilesh Tilotia, with assistance from Instinct.
