AI APIs
What happened
OpenAI published a research-facing note describing how its models performed on three problems drawn from mathematics and theoretical physics: the Kakeya conjecture, the Exponential Time Hypothesis (ETH), and quantum chromodynamics (QCD). The post is framed around academic collaboration rather than product capability — the emphasis is on how researchers can independently verify the outputs and build on them.
It is worth being precise about what this is and is not. This is not a claim that a model solved any of the three problems outright. Each sits in a live research area with a long literature and an active community of specialists. What the company appears to be publishing is model-generated material — candidate constructions, proof sketches, intermediate lemmas, numerical approaches — offered as proposals into a field where proposals can be checked. The headline is less "AI cracked Kakeya" and more "a lab is putting model outputs into a domain with brutal, cheap verification, and inviting the people who can verify them to try."
Brief orientation on the three targets, for readers outside the fields:
- The Kakeya conjecture comes from geometric measure theory. It concerns sets that contain a unit line segment pointing in every possible direction, and the question of what dimension such a set must have. It is a classic hard problem where intuition is unreliable and counterexamples are prized.
- The Exponential Time Hypothesis is a complexity-theory assumption about whether satisfiability can be solved substantially faster than brute-force search. It underpins large parts of fine-grained complexity, which in turn shapes what algorithm designers believe is achievable.
- Quantum chromodynamics is the theory of the strong force. Its equations are notoriously resistant to direct calculation, which is why lattice QCD — heavy numerical simulation — dominates practical work.
Figure: HEXOPS original illustration inspired by the topic (not a copy of the source article imagery) · Credit / further reading: openai.com
Why it matters for builders
Strip away the subject matter and the pattern underneath is the one that will define the next two years of applied AI: output generation is now cheap, and the scarce resource is trustworthy verification.
Mathematics and theoretical physics are the ideal stress test for that pattern precisely because they are unforgiving. A proof either closes or it doesn't. A dimension bound is either correct or there is a counterexample. A numerical result either reproduces or it doesn't. There is no room for a fluent-sounding answer that is subtly wrong — and subtle wrongness is exactly the failure mode teams encounter when they deploy language models into engineering, finance, legal, or scientific workflows.
For builders, three practical implications follow:
1. Domain experts become the evaluation layer. If a model's output on a hard problem can only be judged by a specialist, then the specialist is not a nice-to-have — they are the benchmark. Any product claim about "reasoning" in a specialised domain needs a named expert review process behind it, not a public leaderboard.
2. Provenance beats polish. A proof sketch that names its assumptions, states what it does not cover, and can be mechanically checked is worth far more than a confident narrative. Teams building agentic systems should be designing artifacts that a reviewer can audit step by step — not summaries that read well and resist inspection.
3. Artefacts must be separable from claims. The useful question about a published model result is not "did it work" but "what exactly was produced, and can I run it?" Partial progress — a lemma, a construction, a better numerical scheme — can be valuable even when the headline problem remains open. That distinction is routinely lost in press coverage and is worth defending inside your own organisation.
The verification gap
Figure: HEXOPS original illustration inspired by the topic (not a copy of the source article imagery) · Credit / further reading: openai.com
The bottleneck in this kind of work is not generation; it is formalisation. A natural-language argument produced by a model must be translated into something a proof assistant such as Lean can check, or, in the physics case, into a numerical method that independent groups can reproduce with published parameters and code. Without that step, a result is an interesting suggestion and nothing more.
Reproducibility is the second gate. A model output that cannot be regenerated — because the prompt, the sampling settings, the model version, or the intermediate reasoning are not disclosed — is not a scientific artefact. It is an anecdote. Labs publishing in this space will be judged on how much of the pipeline they expose, not on how impressive the summary sounds.
There is also a quieter issue: the base rates. Three named problems, chosen and framed by the lab that ran the models, tell you little about general capability. A field moves forward on the problems nobody expected to be tractable, not the ones selected for a blog post.
What to watch next
- Independent write-ups. The signal to watch is not the company's post but the first detailed response from a specialist who tried to break the result. Silence from the relevant community is a data point too.
- Formalisation. Whether any part of the work lands as machine-checked proof, or as a published numerical method with released code, separates contribution from commentary.
- Artifact release. Prompts, model versions, and full transcripts. Verification is impossible without them.
- Downstream reuse. Whether the outputs change how a research group works next quarter — a new construction tried, an assumption re-examined, a simulation rerun.
- The academic-terms question. Collaboration with universities around model outputs raises familiar questions about credit, authorship, and who controls the resulting data. Expect that to be argued in public.
Key takeaways
- OpenAI published model results on the Kakeya conjecture, the Exponential Time Hypothesis, and quantum chromodynamics, framed around academic verification rather than product claims.
- These are open research problems; model-produced sketches and constructions are proposals, not proofs, and should be read that way.
- The transferable lesson is structural: in any high-stakes domain, generation is cheap and verification is the bottleneck.
- Formalisation and reproducibility — proof assistants, released code, disclosed prompts — are what turn an interesting output into a usable artifact.
- Expert review is the real benchmark. Build processes that assume it, and design outputs that survive it.
Next steps
If you build with models in technical domains, do three things this quarter. First, require artifacts: prompts, model versions, and intermediate reasoning stored alongside any output you act on. Second, nominate a human reviewer with domain standing for every high-stakes claim, and record their verdict. Third, pilot a formal or mechanical check — even a partial one — so that "looks right" never becomes the last line of defence.
