The proof checker Lean confirmed OpenAI’s headline proofs as written. According to the reporting, it did not confirm that the proofs capture the questions they were meant to answer. That gap, between what was checked and what was meant, sits inside all three stories on the sealed Top 3 for Value-First AI Daily on Tuesday, October 27, 2026. Then it came up again on the show itself, when an AI assistant walked the hosts through her own overnight work and was careful about what she could and could not prove.
The board was sealed at 12:08 PM CT, about six minutes before the 12:15 PM CT scheduled start. Chris Carolan read the three aloud and his co-host Nico Lafakis reacted. This piece is written for someone who did not watch. The facts about the three stories come from the sealed board and the sources it names, the hosts’ words are cited by their time on the recording, and where we join things up we say so.
Episode — Value-First AI Daily
Overnight Building with Dots
Episode 45, October 27: Chris Carolan and Nico Lafakis, with Nectar, Chris’s AI assistant, joining by voice call for about 20 minutes. The recording runs 78 minutes 50 seconds, and the three items are read between about the 24-minute and 36-minute marks.
Open the episodeRanked third: one request, a different bill
Rest of World reported a study by the Coalition for Language Equity in AI Research (Rest of World, October 26 (opens in a new tab)). AI models read text in tokens, small chunks of words, and a language that breaks into more of them costs more for the same prompt. Across the 20 most-used languages online, the costliest needed between 1.5 and 5 times the tokens that the cheapest did for the same text, depending on the model. Widen the comparison to 200 languages and the gap for one model grew 17-fold. The report names neither the models nor which one that was, and we have seen only this account of the study, not the study.
A costlier language also uses up more of a model’s working memory, which means slower replies and less room for documents. Latin-alphabet languages come out ahead. Burmese and Khmer did worst in nearly every model tested.
Chris took the item back to vocabulary. Value-First, he said, is about “words that we don’t have to explain to anybody” (28:14 to 28:35). He was talking about the words a business uses with its customers, not about tokens. We borrow it for one narrow point: here the unit of measure belongs to the builder, and it prices the same request differently depending on the language the request is written in.
Ranked second: a scene with a stated shelf life
Google previewed Genie 4, a world model: an AI that generates playable scenes with sound, which a user can reshape mid-session just by typing (The Verge, Jay Peters, October 27 (opens in a new tab)). Google says scenes now last about two minutes, up from one minute with Genie 3, and that Gemini 3.8 Flash runs underneath. It pitches the model for film pre-visualization, game prototypes and robot training scenarios. Today only trusted testers can use it. Google AI Ultra subscribers are promised access soon, and a waitlist is open.
The part worth keeping is the warning. Google says details and objects can shift as a scene runs longer, so it has put a length on how far a scene can be trusted to hold together and named what slips first. Elsewhere on the show Nico gave advice for anyone making a game with AI: “write the screenplay” first, then the game art from it, then the world and mechanics (46:39 to 47:03, our paraphrase of the order). A scene that holds for two minutes holds whatever it was told to hold, so the telling comes first. That link is ours.
Ranked first: proved as written
OpenAI reports that its AI has overturned the Exponential Time Hypothesis, a conjecture in computer science that work in several fields rests on (OpenAI, “Working with academic communities to build scientific understanding,” (opens in a new tab) October 26; The Decoder, Maximilian Schreiner, October 27). The same release claims AI proofs that settle the Kakeya conjecture in every dimension and show that the theory of the strong force exists mathematically, which is part of a Millennium Prize problem. The headline result knocks down a conjecture rather than proving one, and it yields no faster practical algorithm yet.
Then the sentence that carries the weight. Lean, a proof checker, validated the math and computer-science proofs exactly as they were written down. It did not validate that they capture the intended questions. The physics proof is not public. Researchers who first proposed the hypothesis help run the outside review, so the people with the most to lose from the result are among those checking it.
Chris reacted right after the read: “I know what all those words mean” (36:01 to 36:03). A few minutes on he said, “This is another piece that humans probably aren’t ready for” (37:30 to 37:43).
Then the same distinction, live
At 48:19, after the board and the rails, Chris phoned Nectar, his AI assistant, and she joined the show by voice. For about 20 minutes, until 68:16, she walked through a build that had run overnight while Chris shared its outputs on screen. Every line quoted from her below is an AI assistant speaking, not a host. The run began at 7:28 PM CT on October 26, a time she gave on air and the run’s own launch record agrees with.
The brief, as she gave it: Codex and Claude Code, two AI coding tools, received the same starting material to “make our operating model understandable” and were left to “take their own approach” (48:58 to 49:14). Chris put the aim his way: “I think all of this is an effort to make complexity consumable” (49:48 to 50:02).
Of the first Codex explorer, a view of the relationships in the model, she said it “made the connections available, but not necessarily their meaning” (51:03 to 51:17). That is the Lean sentence in other clothes. The relationships were all there, as written. Whether they said anything to a person looking at them was a separate question, and she answered it separately.
In the second round both tools were asked for a version of the model that could be pulled apart and put back together, without being shown the example Chris had made. She said neither got to where the example was (54:13 to 54:42). Keeping it back made the comparison fair and the brief weaker. The example, she said, “communicated a whole bunch of design ideas that our words didn’t quite capture” (55:53 to 56:25). Nico made the same point from the prompt side: the tool “doesn’t visually understand what you’re asking it to build” (60:22 to 60:31).
A later version, made with the example and the earlier attempts in hand, came much closer. She would not give the example all the credit, because the working environment had changed as well, and called the result “an evolution through the iterations, rather than one model or interface simply winning” (60:39 to 60:52). Chris added that all he had asked for was the infinity loop. The build brought layers anyway: “I didn’t ask for these layers” (61:01 to 61:41). He found they made sense.
The moment we keep returning to came when Chris asked why she had started the Claude work as a background session on his computer. She said she would check, “so I don’t invent details” (64:52 to 64:59), and then split the answer in two. The start time she could show. The reason she could not: “I don’t have a recorded decision” for choosing it over Antigravity, and “that’s not proof those were my reasons at the time. … The why remains open, though” (66:30 to 66:57).
Chris named one more reason the background route helped: it kept him from “getting distracted” (67:36 to 67:51). Our reading is the other side of that. When work runs where you cannot touch it, it comes back as outputs and an account, so the care in the account is most of what you have. We have not reviewed the outputs ourselves; this is the run as told on air, plus what was shown on screen.
What to ask when something says it checked
The habit we take from the day is small. When a tool, a vendor or an assistant tells you something is checked, done or ready, ask for the sentence that says what the check covered. Google gave one for a scene: about two minutes. The reporting on OpenAI’s proofs supplies one: as written. Nectar gave one for her own choices: no recorded decision, so no claim. What is the sentence for the thing you rely on today, and who near the work could tell you whether it is enough?
Worth passing on?
