Agents Doing "Days" of Work, Claude Cracks Fermat, and the Cyber Warnings Keep Stacking Up

The week rounds out with a mix of genuinely impressive research, a sobering research-productivity data point from inside OpenAI itself, and continuing fallout from the cyber-capability warnings issued earlier in the month.

OpenAI's own researchers are leaning hard on agents

OpenAI reported that by mid-August, its research organization was using the equivalent of 3.1 days of agent work for every one day of human work — with human researchers increasingly setting direction while agents absorb the bounded, well-defined work underneath.

It's a genuinely useful real-world data point on where agentic AI is actually delivering value right now: not replacing research judgment, but multiplying the throughput of the people who have it. The company was careful to note that use and productivity aren't the same thing — the payoff still needs to be established, echoing the same caution raised by the NBER's productivity survey a couple of weeks earlier.

Claude formalizes Fermat's Last Theorem — and helps prove a hard result in fluid dynamics

In a notable one-two punch for AI-assisted mathematics, Claude was reported to have fully formalized a proof of Fermat's Last Theorem this week.

Separately, NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge released a Lean-verified proof of finite-time blowup for the 3D incompressible Euler equations under smooth forcing — extending earlier work by Córdoba and Martínez-Zoroa on the same problem — using Claude and formal verification tooling as part of the process.

Coming on the heels of the Hadamard matrix result from earlier in the month, it's now a clear pattern rather than a one-off: AI-assisted formal mathematics is quietly becoming a real research tool, not just a curiosity.

China responds with its own scale-up plan

China's Ministry of Industry and Information Technology laid out a 2026–2030 plan on September 8 to more than quadruple the country's national AI computing capacity — a direct signal that Beijing intends to compete on raw compute scale as aggressively as it has on model and application development, regardless of how the earlier-in-the-month "pick a side" pressure from Washington plays out.

AI's real-time video generation keeps closing the gap

Hugging Face reported that a project called OpenVDN can now denoise roughly 14.4 seconds of AI-generated video — inching closer to genuinely real-time generation, though the caveat is significant: it currently requires eight top-end GPUs to do it. It's a reminder that "real-time" AI video is arriving unevenly — technically possible today, but only at a hardware cost far beyond what most creators or companies can justify, at least for now.

Google backs environmental AI research

On a quieter note, Google's new Asia-Pacific accelerator cohort put 16 teams to work on environmental applications this week — wildlife monitoring, crop tracking, and carbon measurement among them.

It's a useful counterweight to a month dominated by cyber-capability warnings and superintelligence-ban bills: a reminder that a meaningful amount of AI research effort is still going toward fairly unglamorous, genuinely useful applied science.

The bigger picture

Closing out this stretch of the AI news cycle: agents are demonstrably reshaping how research itself gets done, mathematics keeps quietly falling to AI-assisted formal proof, and the compute race between the US and China shows no sign of slowing, even as the safety conversation gets louder by the week. If August was the month of record model releases, early September looks like the month the consequences of that pace started catching up with the industry.