OpenAI’s release makes selected formal proof artifacts available for inspection, while Google DeepMind’s Aletheia reports give more detail about candidate filtering and human review, Kingy AI argues. The comparison is useful precisely because it does not establish a winner: the material describes different work, not a matched test of which lab is better at mathematical research.
Watch Desk analysis
What happened
In a comparison published on 7 October, Kingy AI examines OpenAI’s release of 722 mathematical manuscripts grouped into 372 result families alongside DeepMind research on Aletheia, an agent that generates, critiques and revises mathematical arguments. It compares what the two sets of material disclose about verification, human involvement, reproducibility and compute.
The article points to a difference between a formal proof artifact and an argument assessed by an AI system. It says a successful Lean check can establish that a selected solution proves a specified statement under documented assumptions. Aletheia’s verifier, by contrast, is itself a reasoning system whose judgements need evaluation. Neither distinction, on its own, settles whether a result captures the intended mathematical claim.
Key findings
- Aletheia’s Erdős case study included substantial human review
Of 700 attempted problems, 212 responses were flagged as potentially correct; human assessment classified 200, finding 137 fundamentally flawed and only 13 of 63 technically correct answers meaningful responses to the intended problem. - OpenAI’s proof metadata carries an important qualification
Kingy AI notes that the formalisation manifest describes its scope as “Partial progress” and its review status as unchecked. Those fields do not show that the proofs are wrong, but neither do they amount to an independent verification receipt. - The published work does not provide a matched cost comparison
OpenAI gives an average compute equivalent, while DeepMind reports relative inference costs for FirstProof. Kingy AI says the different models and methods do not support a like-for-like claim about which produces accepted research more cheaply.
Why it matters
AI-generated mathematics needs more than a striking count of papers or solved problems. Readers need to know what exact claim was checked, what assumptions the check relies on, how humans shaped or selected the result, and what resources the work consumed. The article’s comparison shows why “AI solved it” can conceal several quite different standards of evidence.
That is useful context for researchers deciding what to build on, and for anyone trying to compare the labs’ claims without treating unlike demonstrations as a league table. A proof file, a successful formal check and an expert-accepted result are related, but they are not interchangeable trophies.
Our read
Kingy AI makes a worthwhile case for comparing disclosure rather than declaring a capability champion. Its strongest practical point is that formal checking and human scrutiny answer different questions, and both need their scope stated plainly. The next sensible test is to follow specific released claims through their formal statements and checking outcomes, rather than asking which headline number looks larger.
What to watch
- Whether OpenAI publishes checking outcomes tied to specific manuscript versions and theorem scopes.
- Whether future comparisons use shared problems, acceptance standards and attempt limits.
- Whether both labs provide comparable records of human input and total research costs.
Discussion spark: For AI-produced mathematics, should a formally checked proof carry more weight than expert review of an argument, or do readers need both before calling a result convincing?
Sources and evidence
- OpenAI vs DeepMind Math: Proofs, Human Input and Compute – Kingy AI (7 October 2026, 03:13 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.