Terence Tao on Crowdsourced Effort (mathstodon.xyz)

Terence Tao reflected on a recent crowdsourced mathematical project where AI played only a supporting role: most work was done with manually written code, occasional AI autocompletes or chatbots for quick calculations, and new theoretical arguments from participants. He noted an experiment in which ChatGPT’s highest reasoning mode autonomously found a counterexample at n = 97 but mistakenly asserted it was the minimal example when the true minimizer was n = 71—illustrating that frontier models can locate relevant structures but still err on fine-grained claims without tightly supervised, step-by-step guidance. The collaboration still faced human computation errors, but the large, diverse participant pool was effective at catching them; Tao speculates that a well-run human network can be “superintelligent” enough to reduce reliance on AI, at least until real-time formal proof tools (e.g., Lean) are practical. The project is winding down with submissions to OEIS and repositories of data and possible Lean code; Tao emphasizes the experience was markedly faster than traditional research and may foreshadow future collaborative math efforts that combine modern tooling with crowdsourced expertise. For the AI/ML community this offers a concrete case study: autonomous reasoning modes can be promising but brittle, human networks remain crucial for verification, and integrated formalization tools could dramatically cut error rates—also raising policy questions for platforms like MathOverflow about appropriate AI use.