OpenAI says its latest internal model found a counterexample disproving global regularity for the Navier-Stokes equations — one of math's seven Millennium Prize Problems — using 10,000 concurrent agents running for 88 hours. This episode digs past the headline number and asks the more interesting question: how did 10,000 agents actually converge on a verified answer, and what does the architecture that made it possible reveal about how neuro-symbolic AI systems should be built? Episode Summary Sunil opens with the number that caught everyone's attention: 10,000 agents, 88 hours, to disprove global regularity for a smooth solution to Navier-Stokes. His real question is procedural — with that many agents throwing out ideas in parallel, how does the system converge? How does anyone know when it's done? Sajjad and Sam reconstruct the architecture from public reporting: 10,000 instances of an agent built around OpenAI's latest model generated candidate proof strategies in parallel, informally critiquing and refining each other's intermediate results — closer to swarm intelligence than a brute-force sweep. Periodically, a consolidation layer ("Codex") cross-pollinated the most promising findings back into the groups still exploring, redirecting effort as some paths proved more promising than others (the project reportedly started on a related Euler problem before OpenAI redirected agents toward Navier-Stokes). Final adjudication ran through a completely separate pipeline: Lean, an open-source formal verification framework in the same family as Z3, Vampire, and Datalog, which mechanically checked whether a candidate proof actually held — a full 17 hours of formalization and verification on top of the 88 hours of exploration. The panel connects this directly to their own architecture. Sajjad draws the parallel to ClaimGuard: no matter how good the generating model is, DaaX's position has always been that you verify the output against grounding independently rather than trust it outright. Sam highlights what he considers the real innovation — the system doesn't just label an answer right or wrong, it produces a counterexample, and feeds that counterexample back to improve the next round of candidates. That closed loop, generate → critique → verify → redistribute → regenerate, is what let 10,000 agents converge in under four days on a problem mathematicians have worked on for decades. Full transcript at https://daax.ai/podcast/episode-24-navier-stokes-neuro-symbolic-ai Chapters ● 0:00 — This week's topic: OpenAI's Navier-Stokes result ● 1:24 — Sajjad's read: 10,000 agents, peer review, and a Lean-based verifier ● 3:59 — What is Lean? Symbolic verification, explained ● 5:01 — Sunil's chip-design analogy: what do you verify against without a "golden" reference? ● 5:40 — How DaaX's own retrieval-and-verification pipeline actually works ● 7:31 — Sunil's guess: verifying against the algorithm's own well-defined output ● 8:31 — Sam's numbers: 130 billion tokens, 2.7 million inter-agent messages ● 9:31 — Why counterexamples, not just right/wrong labels, matter for neuro-symbolic systems ● 12:11 — The HPC era: brute force, the n-body problem, and drug discovery ● 14:53 — What's different this time: intelligent candidate generation with a feedback loop ● 15:01 — Could this have been done with old-school HPC? (88 hours vs. 88 days) ● 19:32 — Not peer review — informal, adversarial swarm intelligence ● 21:08 — How the problem was actually routed: from Euler to Navier-Stokes ● 23:04 — Cross-pollination: explore, extract, redistribute, explore again ● 24:11 — The closed feedback loop, and why it maps directly to ClaimGuard ● 28:17 — Anima Anandkumar's physics-informed neural network:
