AI Agents Crack Navier-Stokes Singularity
OpenAI researchers announced on 8 September that a swarm of roughly 10,000 autonomous AI agents, running on an internal advanced model, produced a Lean-verified proof that the three-dimensional Navier-Stokes equations can develop a singularity, or “blow-up.” The result addresses one of the six remaining Clay Millennium Prize Problems and was formalized after tens of hours of agent collaboration costing several million dollars in compute. A parallel effort by Tristan Buckmaster and collaborators using other AI models resolved closely related Euler and Navier-Stokes variants around the same time.
The Navier-Stokes equations, formulated in the 19th century, govern viscous fluid flow and remain central to turbulence modeling. Mathematicians have long asked whether smooth initial data and smooth forcing always yield globally smooth solutions, or whether infinite velocities can appear in finite time. Earlier analytic work by Diego Córdoba and Luis Martínez-Zoroa supplied the cascade strategy that both AI teams extended to meet the prize criteria of unbounded space and smooth forcing.
Key tensions remain. The mathematical singularity has no immediate physical consequence because real fluids are discrete at molecular scales, yet it reveals deep counter-intuitive behavior even in the continuum idealization. Priority disputes and the quality of some AI-generated write-ups have already surfaced, and independent human scrutiny of the full formal proofs will determine whether the Clay Institute ultimately awards the prize.
Sources: Quanta Magazine, OpenAI announcement summaries.
Quantum Phase of Free Fall Observed
An international team including Roger Penrose has measured for the first time the quantum phase shift acquired by a freely falling atom under gravity, confirming that Einstein’s equivalence principle holds for a quantum wave packet. Using a Quantum Galileo Interferometer, researchers split ultracold rubidium atoms into a superposition, held one component stationary while the other fell, then recombined them; the accumulated phase matched the 1927 theoretical prediction to within a few percent and grew with the cube of fall time.
The experiment, published in Science Advances and highlighted 8 September, provides a direct laboratory link between general relativity’s equivalence principle and quantum mechanics at ordinary laboratory scales. Previous atom-interferometry work measured gravitational potentials or accelerations, but none isolated this free-fall quantum phase.
Uncertainties center on the still-missing full unification of gravity and quantum theory. The result rules out certain simple conflicts at accessible energies yet leaves open whether gravity itself is quantized; larger superpositions or longer free-fall times will be needed to probe potential deviations.
Sources: ScienceDaily, Science Advances, University of Oxford / Ben-Gurion releases.
AI Framework Controls Fusion Plasma in Milliseconds
Princeton Plasma Physics Laboratory and Princeton University researchers demonstrated the PACMAN framework, which integrates multiple machine-learning models to predict and control tokamak plasma instabilities on a 20-millisecond cycle—far faster than human operators. In five experiments on the DIII-D facility, the system predicted tearing-mode instabilities up to 200 ms in advance, took direct control of heating systems via reinforcement learning, and suppressed edge bursts and fast-particle-driven waves while respecting safety constraints set by humans.
Fusion plasmas hotter than the Sun’s core can lose confinement in milliseconds through magnetic tearing or edge-localized modes. Traditional control relies on slower physics-based models or human intervention; PACMAN learns directly from experimental data and runs continuously, adjusting density, rotation, and actuators in real time.
Remaining challenges include generalization across different tokamak designs, robustness under higher-performance regimes required for power plants, and ensuring that learned policies never violate hard safety limits. The modular architecture is intended to ease transfer to future devices, but long-term reliability data are still limited.
Sources: ScienceDaily, PPPL / Princeton releases, Nuclear Fusion paper.