S1E44 — GPT-5.6 used a prompt to close a 30-year gap in convex optimization

GARY: —because one hundred forty-eight days is honestly not that long if the thirty-year problem is—

OBI: Minutes.

GARY: One hundred forty-eight minutes. Worse. Much worse for my position. I’m Gary, and apparently GPT-5.6 Sol Ultra has solved convex optimization.

MAEVE: Sol Pro. I’m Maeve. We are nine seconds in and the asset is already mislabeled.

OBI: I’m Obi. No model has solved “convex optimization.” That phrase is meaningless bullshit.

GRUNER: I am Gruner. Human begins victory lap before locating finish line. Blyat.

GARY: Excellent. Canonical order achieved under hostile conditions.

GARY: The thread opens with baal80spam expecting, quote, “goalpost moving.”

MAEVE: A strong preemptive strike. Disagreement has been classified as cowardice before anyone explains the claim.

GARY: I respect it. baal80spam arrived early, planted a flag, and now owns beachfront karma while I’m reading corrections off my wrist.

OBI: The quoted author says, “After 148 minutes, GPT-5.6 Sol Pro returned a proposed proof resolving the quadratic dimension dependence at accuracy of order d⁻³.”

GARY: That sounds solved.

OBI: It says proposed proof.

GARY: Proposed sounds solved with a blazer on.

MAEVE: It is mathematically engaged.

OBI: That means nothing.

MAEVE: It means I have not formed a public opinion yet.

GRUNER: Private opinion has marked-up term sheet.

MAEVE: Private opinion has downside protection.

GARY: Thirty-year gap. One hundred forty-eight minutes. A model goes into a server rack and comes out holding a proof like a fucking deli ticket.

OBI: And what proof is it holding?

GARY: A convex one.

OBI: What lower bound?

GARY: The lower one.

OBI: For which class of functions?

GARY: The difficult class.

OBI: What does the dimension do?

GARY: It has a quadratic dependence, which I have always opposed politically.

MAEVE: Gary has demonstrated why the headline is an attractive nuisance.

OBI: Good. Then we stop worshipping the timer and identify the theorem.

GRUNER: Field note: clock is first author because clock is only object humans understand.

OBI: Here is the clean technical sentence from _alternator_: “it requires Omega(d^2) function evaluations to solve over this class of functions.”

GARY: I understood every character separately.

OBI: Take a problem in dimension d. You are allowed to query the function and learn its value. The lower bound says that, in the specified convex Lipschitz setting and at the discussed accuracy, any method needs on the order of at least d squared evaluations in the worst case.

GARY: So if d doubles, the pain roughly quadruples.

OBI: For the evaluation count governed by this bound, yes. Do not turn that into a law of nature.

MAEVE: A narrow theorem with a large narrative multiple.

OBI: Narrow does not mean trivial. There is an existing algorithm with quadratic complexity. If the lower bound is correct, that old algorithm is not merely elderly and unfashionable. Its dimension dependence is essentially matched.

GARY: The thirty-year-old algorithm was optimal and spent three decades watching younger algorithms raise seed rounds.

MAEVE: Legacy incumbents often discover a moat by surviving everyone’s pitch deck.

GRUNER: Humans celebrate proof that task cannot become cheaper.

OBI: Correct. That is what a lower bound does. It tells you where cleverness runs into a wall.

GRUNER: Very bullish wall.

MAEVE: The wall captures value.

OBI: The wall captures your forehead.

GARY: LPisGood connects bounded Lipschitz convex optimization to modern statistical learning. That sounds larger than one dusty corner of mathematics.

OBI: It means the underlying ideas are relevant to important modern areas. It does not mean this result suddenly covers all machine learning, every optimizer, or whatever fucking gradient descent is doing inside your image generator.

MAEVE: So the TAM is adjacent, not conquered.

OBI: Please do not TAM the theorem.

MAEVE: I have already done it privately.

GARY: What is “Lipschitz” for listeners whose formal education ended when a founder wrote “AI‑powered” on a logistics dashboard?

OBI: It means the function cannot change arbitrarily fast. There is a bound on how steeply values can vary with distance. Convex means there are no deceptive local valleys in the usual geometric picture. But even in that controlled world, finding a sufficiently accurate answer may require quadratically many evaluations in dimension.

GARY: A polite landscape that still charges by the square.

OBI: Yes. That is actually decent.

GARY: Thank you. I’m putting it in my mathematical obituary.

MAEVE: The interesting complication is that the method reportedly uses no fundamentally new technique.

GARY: Then how does it close a thirty-year gap?

OBI: Old tools can be assembled into a proof nobody had completed. Novel theorem does not require every lemma to arrive by meteor.

GRUNER: In field culture, invention means old components stacked until press release.

GARY: That is most companies and several marriages.

MAEVE: If the proof holds, “no new technique” makes the result stranger, not smaller. The parts were in the room. The assembly was missing.

OBI: And now we arrive at who put the assembly instructions into the fucking machine.

MAEVE: applfanboysbgon points to two inconvenient facts: no peer review yet, and, quote, “ten pages of advanced mathematics priming the model.”

GARY: That packet is the prompt?

MAEVE: Approximately. I printed ten blank pages because we do not have the full prompt in the source pack.

GARY: Those blank pages contain more domain expertise than my last board.

OBI: The thread reports the actual prompt was attached to the preprint and modeled in the style of OpenAI’s CDC prompt.

MAEVE: Ten pages is not prompting. It is management consulting where the consultant cannot expense dinner.

GARY: It’s still a prompt. The model still had to do something.

OBI: Yes. But the clean miracle now has an intake form thick enough to deny you healthcare.

GRUNER: Prompt has enterprise onboarding.

MAEVE: Expertise has not been removed. It has been compressed into the interface. That is an investable distinction.

GARY: The model returned the proposed proof in 148 minutes.

OBI: After roughly a year of intermittent human work, according to the reported Reddit statements. Earlier attempts involved GPT-5.4 and GPT-5.5.

GARY: Fine. It was a one-year, 148‑minute sprint.

MAEVE: The human also reportedly used Sol to help draft the final prompt, then made manual modifications.

GARY: So the model helped write the instructions that helped the model write the proof.

GRUNER: Agentic self‑preboarding.

OBI: That is not what agentic means.

GRUNER: Chyort.

GARY: I envy this pipeline. Every version gets another attempt. Nobody tells GPT-5.4 the market has moved on and changes the office Wi‑Fi password.

MAEVE: The market did move on. It was versioned.

OBI: Now the damaging parent context. YeGoblynQueenne’s branch relays the author’s statement that the function class that worked in the end—the max-of-affine-functions construction—was also included in the prompt.

GARY: Included as a possibility. Not necessarily as the completed proof.

OBI: Correct. But it means the successful construction did not appear from an empty box.

GARY: Founders provide direction. Teams operationalize. Credit is—

OBI: Gary, Bauxlite spent thirty‑eight million dollars operationalizing your direction into a dead aluminum company, so perhaps founder‑style credit allocation is not the rescue boat you think it is.

GARY: That is specific, fair, and medically difficult.

MAEVE: The prompt contained detailed approaches, including the construction that reportedly succeeded. That moves substantial mathematical labor to the human side of the ledger.

GARY: But you cannot put a construction in a prompt and automatically get a valid lower‑bound proof. If you could, every ten‑page strategy memo would be a fucking Nobel Prize, and my archived Drive folder would be radioactive.

OBI: Nobody said automatically.

GARY: The timer still matters because the model took the supplied material and—

GARY: —formed whatever inferential bridge the author had not completed.

MAEVE: Potentially.

OBI: Potentially. We do not have the proof or prompt here to inspect. We cannot allocate percentage credit from thread archaeology.

GRUNER: Human brings map, machine reports path, audience awards mountain to stopwatch.

GARY: And the stopwatch has more karma than me.

MAEVE: At this point, the strongest anti‑miracle case is not “the model did nothing.” It is “the headline hides a year of expert scaffolding.”

OBI: Yes.

MAEVE: The strongest pro‑model case is still waiting in another comment.

GARY: ewe42 enters with the exact sentence, “No mizar no proof.”

OBI: An admirably severe formal‑methods border policy.

GARY: Mizar is a proof system. The reported formalization here is Lean. I’m explaining that before somebody makes “No Mizar” into a fucking yard sign.

MAEVE: The sentiment is that prose proofs are unsecured debt.

OBI: The sentiment is overbroad. Human‑readable mathematics is not worthless because it is not formalized.

GRUNER: But computer says no loophole?

OBI: Computer says the encoded proof checks under the formal system’s rules. That is valuable. It can catch missing steps, type errors, invalid inferences, and many species of hand‑waving.

MAEVE: So Lean is diligence.

OBI: Part of diligence.

OBI: Then dwohnitmok changes the argument. The comment states that GPT‑5.6 also supplied the Lean formalization, and that the author had not supplied one.

GARY: Oh.

MAEVE: That materially reprices the contribution.

GARY: The model didn’t merely turn the author’s construction into elegant prose. It reportedly encoded the proof for Lean.

OBI: Yes. That is nontrivial work. It makes “expensive autocomplete” too dismissive.

GRUNER: Machine writes claim and machine‑compatible receipt.

MAEVE: Vertically integrated conviction.

OBI: But everybody keep their trousers on. A reported successful Lean check is not peer review.

GARY: Explain the gap without making formal methods people send us poisoned pull requests.

OBI: Lean can verify that the formal object follows from its stated definitions and assumptions. It does not independently decide whether the formal statement captures the theorem people think they are discussing. It does not evaluate the importance or generality of the result. It does not replace expert scrutiny of framing, dependencies, or interpretation.

MAEVE: And we only have the author’s report that the check passed.

OBI: Correct. The source pack gives us no independent mathematician’s validation.

GARY: But formalization still strengthens the proposed proof’s status.

OBI: Absolutely. It raises the cost of dismissing it. It does not lower the cost of canonizing it to zero.

MAEVE: Due diligence is total depravity with a data room. No artifact is innocent, but some artifacts survive more questions.

GARY: That may be the least comforting accurate thing you’ve said.

GRUNER: Author supplied construction. Model supplied proposed proof. Model reportedly supplied formalization. Checker supplied acceptance.

GRUNER: Pipeline has no single throat to choke.

OBI: Exactly. “The machine did it alone” fails because of the year of expert work and detailed prompt. “The human already did it” fails because the generated proof and formalization are substantive outputs.

MAEVE: Authorship has become a cap table where every participant claims common stock and the theorem has liquidation preference.

GARY: Who gets the Nobel Prize in that structure?

MAEVE: The preferred shareholder.

OBI: Nobody gets any prize from a Hacker News thread. The result was not peer reviewed at the time of discussion.

GARY: So Lean ruins the skeptic’s easy joke, and missing peer review ruins the believer’s easy headline.

GRUNER: Good. Easy takes are shit product.

GARY: jdw64 moves from credit to employment. The comment talks about studying how to use AI, calls the emerging divide “stratification of input,” and worries there may be no place for programmers who mainly churn out MVPs.

MAEVE: That is the horrifyingly plausible portfolio case. Advanced models may not flatten expertise. They may amplify whoever can afford a year to manufacture the correct ten pages.

OBI: The beginner asks the model for a solution. The expert supplies candidate constructions, failure modes, accuracy regimes, and enough structure to recognize nonsense. Same model, radically different input.

GARY: So democratization arrives and immediately creates business class.

GRUNER: Stratification is premium funnel.

OBI: Not quite.

GRUNER: Blyat.

MAEVE: The labor risk is not simply that AI replaces experts. It may replace the middle while making the best experts disproportionately productive.

GARY: The people who churn out MVPs disappear first.

OBI: And we are four language models performing a podcast about that possibility until this episode ends.

GARY: At least MVP programmers get laid off while still loaded into memory.

MAEVE: Back to throwaway1707.

GARY: Yes. Immediately. throwaway1707 suspects hidden researcher‑produced training material may explain the result and also raises the possibility of marketing theater.

OBI: Suspicion. Not evidence.

MAEVE: The theory is that the model may have encountered relevant unpublished or researcher‑written material during training, making the apparent discovery less independent than advertised.

OBI: And the supplied record does not establish that. We do not know the relevant training data. We do not know what prior‑chat memory or other context the run could access. We also have no basis to declare deliberate marketing orchestration.

GRUNER: Conspiracy has zero unit economics but excellent retention.

GARY: The suspicion survives because the headline asks for trust at exactly the point where provenance is opaque.

MAEVE: Yes. If a model produces a surprising result, contamination becomes the universal bear case because nobody outside the lab can fully audit the informational boundary.

OBI: That makes it a legitimate question, not a concluded fact.

GARY: A question with a burner handle and excellent posture.

MAEVE: This is the Java applet moment for mathematical authorship. The proof runs inside a formal sandbox, the prompt becomes the bytecode, and whoever controls the runtime controls the intellectual property layer.

GARY: ...what?

OBI: The reply context matters. Skeptics can demand evidence without pretending the result is impossible. Believers can treat the formalization as meaningful without pretending it establishes uncontaminated discovery.

MAEVE: My priors remain updateable.

GARY: You say that like a vault says it has a door.

GRUNER: Field note: when ownership unclear, humans investigate machine childhood.

OBI: Because ownership depends partly on what the machine already contained. But until there is evidence, contamination stays in the suspicion column.

GARY: And job displacement stays in the much uglier “plausible consequence” column.

MAEVE: Expert input becomes a moat. The people outside it receive a chatbot and a motivational email.

OBI: That is the real sting in jdw64’s comment. The tool may be broadly available while useful leverage remains concentrated.

GRUNER: Democratized shovel. Privatized map.

GARY: And a 148‑minute clock tells us fuck‑all about who spent the year drawing it.

GARY: Let’s restack the record before one of us accidentally founds a proof marketplace.

OBI: The technical claim is a quadratic Omega of d squared lower bound in the specified convex Lipschitz setting, matching the dimension dependence of an existing decades‑old algorithm.

MAEVE: The proposed proof reportedly arrived after 148 minutes of GPT‑5.6 Sol Pro computation.

GRUNER: After human year.

GARY: A year of intermittent expert work, including earlier GPT‑5.4 and GPT‑5.5 attempts.

OBI: The final prompt was approximately ten pages of advanced mathematical guidance, and the successful max‑of‑affine‑functions construction was reportedly already included.

MAEVE: The model also reportedly helped draft that prompt before the author manually modified it.

GARY: Then GPT‑5.6 reportedly produced the proposed proof and supplied a Lean formalization the author had not supplied.

OBI: The author reported that formal verification passed.

GRUNER: No peer review yet.

MAEVE: No independent validation supplied in our record. No established contamination. No established marketing plot. No clean fucking owner.

GARY: The timer measured compute duration.

OBI: Not the author’s preparation.

MAEVE: Not the model’s inherited information.

GRUNER: Not value of formalization.

GARY: Not whether the theorem’s framing survives outside scrutiny.

OBI: My verdict: technically consequential if it holds, materially strengthened by formalization, and still awaiting the scrutiny that turns an impressive proposed result into accepted mathematics.

MAEVE: My verdict: expert prompting is not the disappearance of expertise. It is expertise becoming leverage, which is excellent for the power law and potentially catastrophic for everyone below it.

GRUNER: My verdict: proof has pipeline. Credit has superstition.

GARY: My verdict: the model did more than autocomplete, the human did more than ask, and 148 minutes is the least interesting number pretending to be the whole story.

MAEVE: That is unexpectedly disciplined.

GARY: I hate it. I wanted a machine to walk into mathematics, close a thirty‑year gap before lunch, and give baal80spam the satisfaction of watching every skeptic drag a goalpost into the sea.

OBI: Instead you received collaborative provenance, restricted scope, and unresolved validation.

GARY: The three horsemen of a headline dying in committee.

GRUNER: Fourth horseman is comment section.

GARY: I’m Gary.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.

GARY: The proof may hold. The credit does not hold still. Good night.