NYU professor Tristan Buckmaster dropped a bombshell this week, announcing partial proofs for the Navier-Stokes existence and smoothness problem alongside Anthropic’s Levent Alpöge. But the real story isn’t just the math; it’s the messy corporate espionage vibe surrounding OpenAI’s response. Buckmaster claims his team’s private progress was leaked to OpenAI, prompting the lab to throw massive compute resources at the same niche approach. OpenAI confirmed they spent $22.5 million in compute—300 billion output tokens—to generate a full proof using an unreleased model. That price tag alone is staggering, but the timeline raises serious ethical questions about how AI labs compete.
Buckmaster’s statement paints a picture of academic bullying. He alleges that after OpenAI learned of their specific tactical route through Fefferman’s options, an entire team pivoted to replicate the strategy. “It is not the direction one arrives at in a few days by giving a model the problem statement,” Buckmaster noted, emphasizing how unlikely it was for OpenAI to independently choose such an obscure path. When confronted, OpenAI executives reportedly pressured Buckmaster to remove Alpöge’s credit due to his affiliation with rival Anthropic. One executive allegedly warned, “Why would you ruin your career?” before adding, “If you don’t want me to be nice, then I don’t have to be nice.” This isn’t just competition; it feels like coercion.
The technical implications are equally fraught. Buckmaster used OpenAI’s Codex extensively during his research. Since OpenAI reserves rights to train on user interactions unless opted out, there is a plausible risk that the company’s models regurgitated Buckmaster’s private work. OpenAI denies accessing specific user data, stating researchers saw no prior work until public release. Yet they admit they cannot fully rule out indirect influence. The Clay Mathematics Institute still holds the $1 million bounty, but verifying these proofs will now require untangling who actually did the thinking and who just paid for the GPU time.
This incident exposes a dark side of AI-assisted discovery. When a single entity controls both the tool and the infrastructure, independent verification becomes nearly impossible. Buckmaster’s experience suggests that using proprietary AI for groundbreaking research leaves academics vulnerable to having their ideas co-opted by the very platforms they rely on. The math community now faces a crisis of trust: if a proof costs $22.5 million and emerges from a black box, does it count as human achievement? Or is it just expensive pattern matching?
