Terrence Taos recent posts on AI, automated proof (mathstodon.xyz)

After a recent post by Terence Tao, another mathematician used an AI-assisted workflow together with expert supervision to formalize the same mathematical statement in the Lean proof assistant. The formalization succeeded in encoding and proving the result in Lean, but a human error in the formalization introduced a slightly stronger hypothesis than intended. That over-strengthened hypothesis is fixable, and a corrected, machine-checked proof is expected soon; discussion of the work and follow-up appears on the Lean Zulip thread linked in the addendum. This episode is significant because it showcases a practical, near-term use of AI to help produce formal, machine-verifiable mathematics while still requiring human oversight for specification and correctness. For the AI/ML community it highlights both promise and pitfalls: models and tooling can accelerate formalization into systems like Lean, improving reproducibility and rigor, but subtle human or specification errors can alter theorem statements. Key technical takeaways are the use of Lean as the target language, the mixed human+AI workflow, and the need for validation layers (specification checks, counterexample search, or proof-review protocols) to catch semantic mismatches. The case underscores opportunities for better model alignment to formal languages and for building verification tooling to make AI-assisted formalization more reliable.