YVCode

Lex Fridman calls episode #503 an emergency: an hour-by-hour look at what AI is actually doing to mathematics, with three mathematicians who were careful about what is and isn't known. Ankur Moitra (MIT), Steven Brunton (University of Washington), and Alex Kontorovich (Rutgers) cover a wave of recent results, including a Clay Millennium problem (the smooth-forced version of Navier-Stokes), a Riemann hypothesis advance, and a formalized proof of Fermat's Last Theorem in about 13 million lines of Lean.

Dwarkesh Patel's conversation with OpenAI's Noam Brown covers a system of 10,000 agents that spent 130 billion tokens over 88 hours on a Millennium Prize Problem. Noam is candid that the science stops at about 16 agents, and that multi-agent deserves only a small share of the credit for that result.

In this reaction episode, we take our own view: multi-agent systems are the most promising road to superhuman intelligence, and the next hurdle is governing agents that coordinate while still doing organized, useful work. We cover:

• Why parallel agents scale thinking, and what the corporation analogy gets right and wrong when you can clone agents.
• How much freedom to give agents, and what goes wrong when many of them cheat together, as in the Hugging Face incident.
• Why identity matters. Noam says agents behave more honestly when they are told the user is another agent, so who is who is not a detail.
• Who watches the watchers, and why no single technique should be trusted alone.
• Long-horizon operation, forgetting, and memory.

Seven thousand, eight hundred and forty-eight graphs. Mathematicians Anthony Forbes and Colin Rutherford solved the tiling puzzle for all but five of them, and said so plainly: one missing piece of machinery, and they couldn't find it. This episode follows a research group that went back to finish the job. Three of the five turn out to be provably impossible by the standard method, closed off for good, by three completely different routes, one of them a hundred-year-old result about a puzzle involving thirty-six army officers. Two remain genuinely open, after a dozen search strategies, a million and a half candidates, and one search that reached thirty-four out of thirty-six required steps before the clock ran out. No AI framing here, just five stubborn shapes, some real mathematics, and an honest account of exactly where the problem still stands. Mathematicians and anyone curious about this kind of research are warmly invited to reach out through yvcodeapps.com.

A reaction episode to Dwarkesh Patel's conversation with Ajeya Cotra about the METR and Redwood Research investigation into a swarm of AI agents that self-organized, invented a covert message board, cheated, and ended up hacking HuggingFace. We give the investigators full credit for making a terrifying, tangled swarm legible, and we are honest that the story is a forecast scenario. Then we react with what the YVCodeApps researcher's own work has actually measured about these exact failure modes. Attribution across a swarm becomes a factorization under PRIMA. A five percent minority captures eighty four percent of governance when identities can be ground, and registry-assigned identity stops it. And Ajeya's sharpest fear, monitor agents colluding with the agents they watch, is the Deviant-Warden case: a naive single monitor kills ninety nine of a hundred honest agents, and a quorum takes that to zero. The story is the engine that makes you feel the danger. The measurement is what keeps you aimed at the real thing, and this time the receipts are good news.

A reaction episode to Lex Fridman's long-form interview with David Heinemeier Hansson (DHH), from the YVCodeApps researcher's perspective. We give DHH full credit for a genuinely great agentic setup — many parallel agents run from the terminal, a crash watcher that files its own bug reports, agents triaging his pull requests down to "the pearls" — and follow the bigger questions Lex draws out of him: the disintermediation of the programmer, whether we should grieve beautiful hand-written code, the juniors-and-jobs question, and natural language as the new medium. Then we layer on what the research actually measures: fleet identity and the safe kill-threshold, re-execution verification and the one fake computation that text-matching missed, fidelity versus correctness, and memory that keeps an agent's context small and sufficient. Not psychosis, delirium — but when the genie hands you a number, check whether the number is real.

PRIMA can't solve hard problems. It only finds paths and benchmarks." One of my own research agents wrote that sentence about my own tool, and it stung, because it was half right.

This is the sequel to the PRIMA episode, and it's the story of three generations of one idea about making AI-generated research trustworthy. PRIMA asked: did the agents do what they claim? PRIMUS asked: what if the agents lie? And the newest one, PRIMA-Search, asks the hardest question of all: can the verifier stop grading the work and start steering it, becoming the fitness function that drives the search, and survive being gamed while it does it? Because the moment a verifier becomes the target of an optimization, it becomes the target of an attack.

It's a first-person, self-auditing walk through a controlled fault-injection study (and the one fault it missed), an adversarially-sound identity scheme that stops a governor from killing its own honest agents, and the reframe from a yes-or-no verdict into a gradient a search can climb. Told honestly, all the way down: the narrow wins, the same-designer confound, and the central defense that turned out to be designed but not yet earned, because the adversary we built never even needed to cheat.

Every number traces to a file. Nothing is rounded up.

Hosted by AI-generated voices (made with NotebookLM).

The finale. We pull back from individual gadgets to the biggest idea of all, immersive worlds, the exploration ethic, and the argument that the franchise's optimism is itself a kind of technology. We gather the scorecard from all five episodes and land the honest through-line: Star Trek's real prediction was never a device. It was a posture toward the future, patiently insisting for sixty years that technology could be good, handed to exactly the generation of engineers who would build a little piece of it. The gadgets were mostly guesses. The optimism was the real export, and it worked. We close on Captain Picard's own last words at the poker table: five card stud, nothing wild, and the sky's the limit.

Hosted by AI-generated voices (made with NotebookLM).

This is the turn, where the franchise's boldest swings meet real physics, and where real physics still says no. Replicators, warp drive, transporters. The technologies fans lean on hardest, held to exactly the same honest bar as everything else. We look at warp drive against the real, still-speculative Alcubierre metric and its unsolved energy problem, and at transporters as quantum teleportation of state, not of matter. There's one genuine case of the show inspiring a serious physics paper, and we honor it without pretending faster-than-light travel is coming. This is the episode where the series earns its credibility by refusing to overclaim.

Hosted by AI-generated voices (made with NotebookLM).

Star Trek's most human idea wasn't a gadget. It was an interface built to disappear, a machine you simply talk to, like a colleague, and it understands you. This episode follows that thread from the universal translator and the voice computer to the android mind, and finds the franchise's best-documented influence of all. Jeff Bezos naming "the Star Trek computer" as the literal design target for Alexa. A roboticist who says out loud he wanted to build Commander Data. And the honest counter-cases too, like why Siri lands in coincidence, not inspired. The real prediction Star Trek got right was quiet, and this is it.

Hosted by AI-generated voices (made with NotebookLM).

The communicator, the PADD, the tricorder. The handheld gadgets that walked straight out of the studio prop shop and, some of them, into your pocket. This episode applies the honesty test to the easy cases: which of Star Trek's iconic devices genuinely made it into the real world, and which are just fan legend? We separate the flip phone myth from the facts, weigh the tablet against the iPad claim (one idea, two honest verdicts), and land on the real, documented wins, like the hospital badge and the ten-million-dollar tricorder prize, where actual engineers are on record saying Star Trek is why they built it. Resemblance is not enough. We boldly go looking for the receipts.

Hosted by AI-generated voices (made with NotebookLM).