March 23, 2026
Math just rage‑quit
Epoch confirms GPT5.4 Pro solved a Frontier Math Open Problem for the first time
Internet freaks out as AI “does real math” and solves a problem humans couldn’t touch for years
TLDR: A cutting-edge chatbot helped solve a genuine unsolved math problem that a human expert has now verified and plans to publish, marking a big leap from “calculator” to “co‑author.” Commenters are split between awe, nervous jokes, and debates over whether this proves today’s AI can already do real scientific discovery.
An advanced chatbot just cracked a real unsolved math puzzle – the kind professors spend careers sweating over – and the internet is having a full-on identity crisis. Researchers used GPT‑5.4 Pro to solve a tricky “hypergraph” problem (think: arranging connections between points so they don’t form forbidden patterns), and a human mathematician has checked the work and plans to publish it. But the real fireworks are in the comments.
One camp is basically saying: it’s over, humans – with users like karmasimida declaring that there’s “no denial” anymore that AI can invent new ideas, not just regurgitate Wikipedia. Others are stunned that it didn’t even take the biggest, scariest super‑models: as osti notes, regular GPT‑5.4 and Google’s Gemini 3.1 Pro did the job while the ultra “deepthink” version whiffed, adding a twist of model-on-model drama. Curious minds are obsessing over Epoch’s mysterious “scaffold” for testing AI on open problems, treating it like some secret cheat code for math.
Meanwhile, the jokers are already breaking it into memes: snypher drags a bad analogy with the burn that it “falls apart if the number wasn’t on the clock face,” roasting anyone trying to downplay the result. Underneath the jokes, though, there’s a clear shift: people are starting to talk about AI not as a calculator, but as a teammate in discovering new mathematics – and that has everyone both excited and slightly terrified.
Key Points
- •Kevin Barreto and Liam Price first elicited a solution to a FrontierMath open problem using GPT‑5.4 Pro.
- •Problem contributor Will Brian confirmed the solution and plans a formal write‑up for publication.
- •Brian noted the approach removes an inefficiency in lower‑bound constructions and mirrors the upper‑bound construction, yielding matching bounds.
- •A general testing scaffold for FrontierMath was completed; within it, Opus 4.6 (max), Gemini 3.1 Pro, and GPT‑5.4 (xhigh) also solved the problem.
- •The underlying problem seeks improved lower bounds for H(n) via hypergraph constructions, including specific tasks with |V| ≥ 64 or 66, |H| ≤ 20, and no partitions > 20, and a full challenge to improve a known recursive bound by a constant factor.