At the ADIP EC 2026 opening, OpenAI CEO Sam Altman said a “next GPT”—the system whose ~10,000 agents spent ~four days on Navier–Stokes (NS) and related millennium-class work—will ship soon. He called it one of the most remarkable moments he has seen in AI history and claimed AI is “already smarter than top scientists most of the time.” In his telling, agents had no boss hierarchy, chose directions, shared progress, and self-assigned next steps; after ~88 hours on NS, GPT-6 Astra-class tooling spent ~17 more hours on Lean formalization. Treat this as a lab claim pending independent verification—see our ETH-refutation coverage and Axiom CFSG formalization.
10k Agents, 88 Hours, “Spaghetti” Vortex Narrative
Altman pointed to early-September math as the year’s shock: ~10k next-GPT agents with no human mathematician steering and no top controller. Logs described a Sept 1 start; around hour 50 a ~100-agent squad opened a finite-time blowup path on force-free Euler and synced the swarm; by ~hour 88 (Sept 5) they produced a spiral “spaghetti” vortex construction claimed to show unbounded velocity under bounded kinetic energy—ending a ~90-year NS-related stalemate—then ~17 hours of Lean, ~2.7M messages / ~130B output tokens, and a ~165-page OpenAI-authored write-up. Unlike classic multi-agent “foreman” graphs, the pitch is that preset controllers fail in unmapped research terrain, so the sandbox must self-organize.
Two-Month Waves: Swarm Siege to Single-Prompt Hits
Reporting ties the same lineage to: ~Oct 6 dumps of ~721 preprints spanning 100+ open problems (including Landau–Siegel progress, with Tao-linked verification lore); an Oct 26 night of Kakeya-in-all-dimensions, ETH-falsification, and Yang–Mills existence claims. Watchers such as Andrew Curran argue later October wins may not have needed 10k-agent swarms—closer to one prompt and ~3 hours of Pro-depth thought. If true, the curve compressed from “army for four days” to “solo hours” in about a month. Codex traffic and “gpt-6.2-astra” breadcrumbs are read as countdown signals; names and GA dates follow official OpenAI posts. Related: GPT-6.1 Terra.
Token Bills and the Energy Bottleneck
Altman contrasted internals: ~6 years ago the world’s top token user burned ~100k/month; global per-capita monthly use now matches that, while the #1 account hits ~500B tokens/month (~5,000,000×)—attributed to always-on research/engineering agent swarms. He named energy as the binding constraint once AI takes over science and chip work. Hype spans “superintelligence arrived” to life-science discovery booms; separate company speeches, preprints, and peer consensus. More: AI News.
FAQ
Q: Is the next GPT already generally available? A: Altman said it will arrive soon; treat model IDs, API access, and evals as pending official GA—not a keynote alone.
Q: Are the millennium problems settled in mathematics? A: OpenAI and media describe agent runs, Lean formalization, and preprints; community acceptance still needs independent audit and formal publication.
Q: Why care about swarm vs single-prompt? A: Swarms show large-scale non-hierarchical research process; single-prompt speed—if real—means the same capability may fit everyday chat. Both need reproducible evidence.
相关文章
