The Week AI Started Doing Real Science

Last week, two things happened that should change how you think about AI and scientific research. Theoretical physicist Yuji Tachikawa — a Breakthrough Prize laureate and string theorist at the University of Tokyo's Kavli Institute — reported that Claude Fable cracked a collaborative research problem that had stumped him and his colleagues for six months. The same week, OpenAI announced that GPT-5.6 Sol Ultra produced a proof of the Cycle Double Cover Conjecture — a 50-year-old open problem in graph theory — using 64 parallel subagents in under an hour.

Neither result has been peer-reviewed. Both could collapse under scrutiny. But the fact that they happened in the same week, from competing labs, while prediction markets are actively repricing AI mathematical capability into real money — that is the signal worth paying attention to.

Tachikawa's "On a Whim" Moment

Here is what actually happened. Tachikawa had been collaborating on a quantum field theory and string theory problem. Six months, no progress. The team was stuck at a specific computational roadblock. On a whim, he fed his research notes to Claude Fable.

View original post on X

The first response was measured: Fable identified a calculation error the team had also found, then hit the same wall they had. Standard behavior — find the mistake, reach the dead end. But when Tachikawa pushed back and described the roadblock more precisely, Fable did something unexpected. It suggested a broader methodological approach, then wrote SymPy code to verify its own mathematical predictions.

Tachikawa's assessment: Fable made "a non-trivial observation" that "essentially solved" the problem. His conclusion was striking — "Fable probably seems like it properly understands string theory and has intuition too."

The sequence matters. Fable did not just pattern-match to a known solution. It found the same error the humans found, hit the same dead end, and then — on a second pass with better context — proposed a new approach. That looks less like retrieval and more like reasoning.

The story went viral. Marc Andreessen quote-tweeted it with a single word: "Interesting." Physicist and venture capitalist Steve Hsu amplified it to his academic network.

View original post on X

The reaction pattern was notable — not hype, but quiet recognition from people who understand what theoretical physics research actually requires.

GPT-5.6 and the 50-Year Conjecture

Three days before Tachikawa's thread, OpenAI dropped a different kind of bombshell. Ethan Knight announced that GPT-5.6 Sol Ultra had produced a complete proof of the Cycle Double Cover Conjecture, posed independently by George Szekeres in 1973 and Paul Seymour in 1979.

The conjecture asks a deceptively simple question: does every bridgeless graph have a collection of cycles such that each edge appears in exactly two of those cycles? Simple to state, brutally hard to prove. It has sat unresolved for half a century.

View original post on X

The method was as interesting as the claim. OpenAI's prompt instructed Sol Ultra to deploy up to 64 concurrent subagents, managed "aggressively and dynamically." Early rounds maintained diversity — agents pursued different mathematical formulations, algebraic angles, and structural inductions independently. Adversarial agents were assigned to hunt for edge cases and errors. The entire process took under one hour, despite being allocated eight.

The proof itself uses cubic graph reduction, the 8-flow theorem, and linear algebra over GF(3) to construct edge labelings that force each edge into exactly two cycles. OpenAI published both the three-page proof and the full prompt that generated it.

Mathematician Dr. Samuel Allen Alexander broke down the proof in his analysis above. His verdict: the mathematical structure is coherent, but verification is pending.

The Verification Gap

Here is where intellectual honesty requires pumping the brakes.

Manchester mathematician Thomas Bloom called the proof "a very nice proof" that is "short, elementary, and could have been discovered in the 1980s." That last part is the key insight — the AI's advantage was not conceptual novelty. It was computational persistence. As Bloom put it, "the AI does not get discouraged" when approaches fail, unlike humans who might abandon them prematurely.

But Bloom also flagged a critical issue: the proof contains zero citations. It omits a foundational 1983 paper by Bermond, Jackson, and Jaeger entirely. This is not a style complaint — in mathematics, citations trace the logical heritage of ideas. Missing them raises questions about whether the model understood the field's structure or simply generated a plausible-looking proof.

The Hacker News discussion on the CDC proof was characteristically direct. Top comments noted that "a machine-verified proof is not the same as a peer-reviewed proof," and several commenters pointed out that existing graph theory libraries in Lean are insufficient for research-level verification. Professional verification is expected to take days to weeks.

View on Hacker News

Why Math Is the Proving Ground

These two events did not happen in a vacuum. They are the latest data points in a pattern that Nature and Quanta Magazine have been tracking all year: AI is reshaping mathematical and scientific discovery in ways that go beyond benchmarks.

The numbers tell the story of acceleration. Claude Fable 5 now scores 88% on FrontierMath Tier 4 — the hardest tier, designed to test research-level mathematical reasoning. For context, Anthropic's previous model, Opus 4.5, scored below 10% on the same tier earlier this year. That is not incremental improvement. That is a phase change.

GPT-5.5 reaches about 75% on the same tier, putting it 13 points behind Fable 5. Meanwhile, both OpenAI and Anthropic have solved longstanding mathematical problems beyond benchmarks — including an Erdos problem.

Why math and physics? Because these fields offer something rare in AI evaluation: objective verification. A code completion can be "good enough." A creative writing sample is subjective. But a mathematical proof is either correct or it is not. A physics calculation either matches experimental data or it does not. These domains are the acid test for reasoning because there is nowhere to hide behind plausibility.

PhD astrophysicist Kyle Kabasares has been systematically testing Claude Fable 5 on research-level math and astrophysics questions — the kind that appear in graduate qualifying exams and active research papers. His findings confirm the benchmark story: Fable 5 handles problems that previous models could not touch.

The Money Is Already Moving

Here is what makes this week different from previous "AI breakthrough" cycles: the prediction markets have noticed.

Polymarket's "Which company has the best Math AI model?" market has been actively traded with $172K in volume. The broader "Best AI model end of July" market sits at $5.8M in volume with Anthropic at 95%.

The math-specific market is particularly telling. Google led the June resolution at 61%, driven by Gemini's IMO gold-medal performance and consistent benchmark scores. But the market is actively repricing after Fable 5's FrontierMath dominance and the Tachikawa moment — real-world physics problem-solving carries a different weight than competition math.

This matters because prediction markets aggregate information that press releases and benchmark tables cannot capture. When bettors put money on which lab has the best math model, they are pricing in everything: benchmark scores, real-world reports like Tachikawa's, verification status of claimed proofs, and the credibility of the people making claims. The frontier AI race now has a financial scoreboard, and math capability is the marquee event.

View original post on X

What This Means for You

If you are a researcher, the Tachikawa workflow is a template worth copying. The approach was simple: show the model your research notes (not a carefully crafted prompt), describe the specific roadblock, and iterate when the first response hits the same wall you did. The key insight is that frontier models can sometimes see around corners that domain experts cannot — not because they are smarter, but because they do not carry the same assumptions.

If you are a technical founder or investor, watch the Polymarket math markets. They are leading indicators for which labs are shipping genuine capability versus which are running on marketing momentum. The gap between Fable 5 (88% FrontierMath Tier 4) and GPT-5.5 (75%) is a measurable competitive advantage that will show up in every downstream application — from drug discovery to materials science to financial modeling.

If you are an AI skeptic, hold onto that skepticism. Neither result this week has survived peer review. The history of AI-generated mathematical proofs includes enough retractions to warrant caution. But also recognize that the goalposts have moved. A year ago, the question was whether AI could do competition math. Now the question is whether its research-level contributions will hold up under scrutiny. That is a fundamentally different conversation.

The Uncomfortable Truth

The real story this week is not that AI solved a physics problem or claimed a math proof. It is that the verification infrastructure cannot keep up with the pace of claims.

Lean does not have the graph theory libraries to formally verify the CDC proof. Tachikawa's result is a social media thread, not a preprint. And the prediction markets — the closest thing we have to a real-time capability scoreboard — are pricing in claims before they are verified.

This is the gap that matters. The models are producing results faster than humans can check them. That is not a crisis — it is an engineering problem with a known solution (formal verification, reproducibility requirements, peer review processes adapted for AI-assisted research). But until that infrastructure catches up, every "AI breakthrough" announcement lives in a superposition of genuine and unverified.

What is different about this week is the quality of the signal. Tachikawa is not a random poster — he is a Breakthrough Prize winner working at one of the world's leading physics institutes. The CDC proof comes with a published prompt and reproducible method. And the Quanta Magazine piece from April, which surveyed the broader landscape, now reads less like prediction and more like prologue.

The week AI started doing real science? Maybe. But the week we started needing real verification infrastructure for AI science? Definitely.

---

For more on how frontier AI models are reshaping capability benchmarks, see our coverage of the 48-hour frontier release war and Fable 5's guardrail architecture. For the economic framing behind AI capability investments, check out AI scaling laws and capability math.