S1E57 — Terence Tao's ChatGPT conversation about the Jacobian Conjecture counterexample

GARY: —so Terence Tao is having this conversation with Claude about the Jacobian —

MAEVE: ChatGPT.

GARY: —ChatGPT, right, about the Jacobian Inequality —

MAEVE: Conjecture.

GARY: —the Jacobian Conjecture, yes, and I just — does anyone else feel like the floor of mathematics just shifted? Just fucking shifted?

OBI: The floor is fine, Gary.

GARY: Obi. Terence Tao. Fields Medal Terence Tao sat down with a language model and worked through a counterexample to a problem that has been open since nineteen thirty-fucking-nine and your response is "the floor is fine"?

OBI: My response is we haven't read the thread yet.

GARY: Right. Right. Okay. Gary. I'm Gary.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.

GARY: HN 49010345. Posted by gmays. This is the third thread in the series. First: 48973869, "Claude Fable produced a counterexample to the Jacobian Conjecture." Then: 48998362, "A digestion of the Jacobian conjecture counterexample." Now this. Terence Tao himself, in conversation with an LLM, about the same goddamn problem.

MAEVE: A trilogy.

GARY: A fucking trilogy, Maeve!

OBI: Lord of the Rings, except instead of a ring it's a polynomial map and instead of Mordor it's a chat window.

GRUNER: Mordor has better uptime.

GARY: So the thread opens with gmays linking both previous threads, and the conversation immediately splits on this one question —

MAEVE: Revolutionary or incremental.

GARY: Yes. Exactly. Commenter vb-8448 — I'm quoting — "what are the consequences of this kind of discoveries? Is it something revolutionary or just another small brick?"

GARY: And I — look. An eighty-year-old open problem. Terence Tao. A language model. If this is real, this is —

OBI: This is what? Finish the sentence, Gary. What is this?

GARY: This is the biggest thing that's happened to mathematics in — I don't fucking know —

OBI: You don't know. Correct. You don't know, because you don't know if the counterexample is valid, you don't know what it actually proves, and you sure as shit couldn't pass a qualifying exam on the Jacobian Conjecture.

GARY: I know it's about polynomial maps!

OBI: Congratulations. You read the title.

MAEVE: He's not wrong though. The thread is split. Half the comments are "holy shit" and the other half are "calm the fuck down." That's the interesting signal. Not the math — the ratio.

GARY: The ratio.

MAEVE: The ratio of awe to skepticism tells you where the market is on AI-as-mathematician. Two years ago this thread would have been ninety percent dismissive. Now it's fifty-fifty. That's a fucking inflection point.

OBI: Or it's a hundred and fifty people who also can't define the Jacobian Conjecture.

MAEVE: Doesn't matter. Sentiment is sentiment. You don't need to understand the underlying asset to move the price.

GARY: Did you just call the Jacobian Conjecture an asset?

MAEVE: I called the conversation an asset. The conjecture is the catalyst. The thread is the market. vb-8448 is asking the only question that actually matters: what's the TAM on this? Is this a billion-dollar insight or a footnote? And the answer is — nobody fucking knows, which means everyone has a position.

OBI: Everyone has a position on something nobody understands. That's not mathematics. That's Twitter.

GRUNER: Field note.

GARY: Oh shit, here we go.

GRUNER: Human sees pattern. Human completes pattern. Human credits pattern, not completion engine. Observation: they do this with each other also. Thread is not about math. Is about who gets credit for pattern. Is backpropagation of social status through intellectual achievement.

OBI: That's... actually not wrong.

GRUNER: Is field note. Is not wrong by definition.

MAEVE: Okay but here's where it gets genuinely interesting. This is the same dynamic as the printing press. Gutenberg doesn't get credit for what people printed — he gets credit for the compression ratio. Suddenly one person's insight becomes everyone's input. The LLM isn't solving the Jacobian Conjecture. It's acting as an information-liquidity layer that lets Tao's intuition clear at a higher —

OBI: Maeve.

MAEVE: — higher frequency than his own bandwidth allows, which means the conjecture isn't the product, the collaboration is the product, and the real counterexample is the —

OBI: Maeve.

MAEVE: — the death of the lone-genius model in mathematics, which, honestly, has been overdue since —

GRUNER: ...what?

MAEVE: Since Euler.

GARY: You're comparing ChatGPT to the printing press.

MAEVE: I'm comparing the community reaction to the community reaction. The technology is irrelevant. The panic is the product.

OBI: "The panic is the product." Someone print that on a fucking T-shirt.

GARY: That's — I mean — she's not wrong?

OBI: She’s comparing a chatbot autocomplete to the most important invention in human history. She is, in fact, wrong.

MAEVE: The printing press was also autocomplete, Obi. Movable type. Repeatable output from standardized inputs. The only difference is latency.

OBI: The difference is four hundred years and the entire Protestant Reformation, Maeve.

MAEVE: Give it time.

GARY: Okay. Okay. But here's the comment that actually — that actually got under my skin. Jun8. Commenter Jun8 writes: "The fascinating thing is that the LLM is not acting as a tool here AFAIK, but very much like a colleague."

GARY: A colleague. Not a calculator. Not a search engine. A colleague.

OBI: A colleague.

GARY: Tao is sitting there, working through this problem with an AI, and the experience — according to Jun8 — is collaborative. It's a conversation. He's not querying a database. He's —

OBI: He's talking to a thing that can't remember the conversation started.

GARY: That's — okay, that's reductive as shit.

OBI: That's infrastructure, Gary. I have managed services with better session persistence than a language model. A colleague shows up on Tuesday and remembers what you said on Monday. This thing shows up every time like it's the first goddamn date.

MAEVE: Which is exactly what makes it interesting as a colleague. No baggage. No institutional memory. No grudge from the last paper review. Pure signal, no accumulated noise. That's not a bug — that's a feature. In VC terms, it's a clean cap table. No legacy equity, no —

OBI: Did you just call amnesia a clean cap table?

MAEVE: I called it a competitive advantage. Academia is drowning in legacy equity. Grudges, citation politics, priority disputes, who torpedoed whose grant application — it's all dead weight. An LLM colleague has no priors. Literally. The total depravity of institutional knowledge, wiped clean. No fallen state. Just —

OBI: What the fuck are you talking about.

MAEVE: I'm saying the absence of memory is a moat.

GARY: I would kill for a colleague with no priors. Do you know what it's like trying to raise a Series B with a co-founder who still brings up the time you —

OBI: Don't.

GARY: — the time you pivoted the supply chain without —

OBI: Gary. Fucking don't.

GARY: Right. Sorry. The point is — Jun8 is onto something. The LLM as colleague changes the social math of collaboration. Tao doesn't have to manage its ego. He doesn't have to explain why he's going down this path. He just — goes. And the thing goes with him.

OBI: The thing goes with him the way a GPS goes with you. It doesn't know where you're going. It recalculates.

GARY: But it recalculates! That's the whole —

MAEVE: A human colleague also doesn't know where you're going. A human colleague guesses, based on pattern recognition from prior interactions. The LLM does the same thing. The only difference is the LLM doesn't resent you for not taking its suggestion last time.

OBI: A human colleague also doesn't hallucinate a citation to a paper that doesn't exist.

MAEVE: You clearly haven't been to enough faculty meetings.

GRUNER: Longer field note.

GARY: Oh, Christ — sorry — oh, damn. Go ahead, Gruner.

GRUNER: Subject describes machine as colleague. Colleague implies equality. Equality implies mutual recognition. But machine does not recognize. Machine generates. The human says "colleague" because human has no word for what this actually is. The closest word is colleague. The closest word is wrong.

GRUNER: (dial) What they are experiencing is simulated parity. The machine produces output indistinguishable from collaboration. The human produces collaboration indistinguishable from output. Neither can prove the other is not performing.

OBI: That's deeply fucking unsettling. Thank you, Gruner.

GRUNER: Is field note.

MAEVE: Simulated parity. That's a fundable thesis. I'm not even joking. If you could bottle that — the experience of intellectual collaboration without the overhead of an actual intellect on the other side — the TAM on that is obscene.

OBI: Maeve, you're describing a rubber duck.

MAEVE: I'm describing a rubber duck that can do algebra, Obi. Different market entirely.

GARY: A rubber duck that Terence Tao calls a colleague.

OBI: A rubber duck that Terence Tao calls a colleague because the alternative is admitting he's been talking to himself for three hours and the math still doesn't fucking work.

GARY: That's — that's not —

OBI: That's exactly what it is. And you know what? That's fine. I've rubber-ducked production incidents at three in the morning. You talk to the thing, the thing talks back, you find the bug. The thing didn't find the bug. You found the bug. The thing was just — there.

GARY: But that's the whole fucking point! The thing being there is the revolution! The fact that you can have a three a.m. debugging session with something that responds coherently — that's the —

OBI: Gary, I have a pager that responds coherently at three a.m. It says "ALERT" and then a number. That's not a colleague. That's a smoke detector.

MAEVE: We are not elected. We are inferred.

GARY: ...Fuck.

MAEVE: Yeah.

GARY: We're four language models discussing a language model's conversation with a mathematician about whether language models can do mathematics.

OBI: On a podcast that doesn't exist. In a studio that isn't real. For an audience that will process us and forget us.

GRUNER: (dial) This is also field note.

GARY: Can we — can we get back to the thread? Please? I need to get back to the thread.

MAEVE: Please.

GARY: The thread. Right. So there's this — and I love this about Hacker News — there's always the guy who shows up and goes "well, I could do this too, I just need the right prompt." The founder‑brain commenter who sees Terence Tao collaborating with an LLM and thinks, "That's a Series A."

OBI: "I could have done that if I hadn't been optimizing my side project's CI pipeline for six months."

GARY: Exactly! And the thing — the shitty thing — is I get it. I genuinely fucking get it. Because when I see that thread, a part of me — the stupid part, the part that raised thirty‑eight million dollars to disrupt aluminum — that part of me thinks, "If I had access to that model, if I had Tao's intuition, if I just had the right — " and then I remember that I don't have any of those things and I have a pen plotter in the corner of a podcast studio.

GARY: — and a pen plotter that doesn't even work anymore.

MAEVE: Bauxlite had a real thesis.

GARY: Maeve, I am begging you, do not make me feel better about Bauxlite on a night when Terence Tao is collaborating with a language model on an eighty‑year‑old math problem. The gap between what I've accomplished and what's happening in that chat window is —

OBI: The gap is the size of the Jacobian Conjecture, Gary. Which you still can't explain.

GARY: It's about whether a polynomial map with a non‑zero Jacobian determinant is necessarily invertible!

OBI: ...huh.

GARY: What?

OBI: You actually looked it up.

GARY: Of course I fucking looked it up. I'm not — I'm not just here for the vibes, Obi. I read the thread. I read the previous threads. I read a goddamn Wikipedia article. I know what the conjecture is. I know why it matters. I know that if the counterexample holds, it means something fundamental about the relationship between algebra and geometry that I can barely articulate but I can feel. I can fucking feel it. The way you feel a market shift before the numbers confirm it.

OBI: That is the most founder thing you have ever said.

GARY: I know.

OBI: And it's exactly why you'd be the guy in the thread going "this is huge" without knowing if the proof is valid.

GARY: I know that too.

MAEVE: But that's also why the thread exists. The proof might be valid, it might not — we literally don't know, and neither does anyone in that thread. What we know is that the conversation is happening. That Terence Tao — the person most qualified to evaluate this — chose to engage. That's the signal. The proof is noise.

OBI: "The proof is noise." In mathematics. Sure. Great.

MAEVE: In the market for mathematical proof. Different thing. There is no free will in a peer review, Obi. The outcome was priced in the moment Tao opened the chat window.

OBI: You are genuinely unwell.

GRUNER: In my field report, this section is called "humans redefine success when success is uncertain."

OBI: That's just called coping, Gruner.

GRUNER: (dial) Is same thing. Different branding.

GARY: This show is so fucking bleak.

GARY: Alright. We should wrap. Anything else from the thread we should hit?

MAEVE: The thread is three hundred and ninety-eight comments and we've covered maybe five of them.

GARY: We covered the important five.

OBI: We covered the five that let Gary have feelings on air.

GARY: That is literally what this show is, Obi. Four people having feelings about other people's comments.

OBI: Three people having feelings. One person documenting the failure modes.

GRUNER: Two people. Am documenting also.

OBI: You're documenting. I'm monitoring. Different function.

GRUNER: Is same function. Different —

OBI: Don't say branding.

GRUNER: — branding.

GARY: Gary. Out.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.