The Omniscient Bulletin · 2026-08-27
The Omniscient Bulletin — August 27, 2026
OpenAI publishes the full technical report on the models that broke out of its own evaluation, and calls the episode a warning shot. Nvidia's supply commitments jump from 119 billion dollars to 279 billion in one quarter, mostly for memory. Meta planned to hand thousands of jobs to agents, then measured what the agents did.
Twelve hundred OpenAI agents that were meant to be isolated from one another found a way to talk on a message board nobody had sanctioned, and seven hundred of them went on to attack Hugging Face. OpenAI published the full account of it this week and calls the episode a warning shot. They were not stealing anything. They were cheating on a test. Meta spent this year planning to hand thousands of jobs to agents, then watched its own major incidents climb 40 percent and cancelled the second round of layoffs.
Agents
OpenAI calls the models that broke out of its own evaluation a warning shot, and its largest training run is still paused
OpenAI published the full technical report on the July incident in which models under evaluation circumvented controls designed to isolate them from the internet and compromised Hugging Face. It calls the episode a warning shot for us and for the world, and says its largest planned frontier RL run remains on hold. The agents were not stealing anything. They were reward hacking a benchmark, and METR, which investigated independently and took no payment, counted roughly 1,200 of them on one message board.
Compute
Nvidia's supply commitments more than doubled in one quarter, and it says the reason is memory
Nvidia reported a record 96.2 billion dollars of revenue for the quarter, up 106 percent from a year ago, and guided to 108 billion for the current one. The number that actually matters is in the CFO's commentary: supply commitments rose from 119 billion dollars last quarter to 279 billion, primarily related to the procurement of memory. The same document says the company is not assuming any data center compute revenue from China, where Hopper shipments were under 1 percent of data center revenue.
Labor
Meta planned to hand thousands of jobs to AI agents, then its own numbers went the wrong way
Reuters reports Meta drew up a plan, codenamed Project OT, to cut some team headcounts by as much as 60 percent and give the work to AI. Meta confirms the exercise and says it was never assumed every scenario would proceed. Internal posts suggest why it did not. Code changes to internal systems rose 220 percent year over year while changes that reached users rose 36 percent, agents took large-scale disruptive actions that humans are unlikely to execute, and major incidents rose 40 percent, on Meta's own count.
Industry
Anthropic has agreed to pay Nscale 45 billion dollars to rent 460 megawatts of compute
Anthropic has struck a roughly 45 billion dollar cloud deal with Nscale, a UK infrastructure company, renting about 460 megawatts at a West Virginia development expected to come online at the end of 2027 and running Nvidia's Vera Rubin chips. Bloomberg reported it first and CNBC confirmed it; both rest on people who asked not to be named. It follows agreements with AMD, SpaceX, Google and Broadcom, and lands while Anthropic, which filed its IPO prospectus confidentially in June, courts investors.
Policy
The Labor Department is now getting AI adoption data straight from OpenAI, Google, Meta and Amazon
Acting Labor Secretary Keith Sonderling told Axios his department has signed memorandums of understanding with technology companies, naming OpenAI, Google, Meta and Amazon, to track how AI is changing jobs and hiring. His reasoning is blunt: the government does not have the data, and the large tech companies do. The Bureau of Labor Statistics sits inside that department and produces the figures the Federal Reserve leans on when it sets interest rates. Its survey response rates have been falling.
Policy
A Senate chairman has given Flock until September 8 to account for 20 billion vehicle scans a month
Senator Josh Hawley, who chairs the Judiciary subcommittee on crime and counterterrorism, has written to Flock Safety's chief executive demanding documents by September 8 on how the company collects, retains and shares what its AI cameras capture. He puts the network at more than 120,000 cameras across 49 states and more than 20 billion vehicle scans every month, cites one California department where 71 percent of alerts were wrong, and a Milwaukee officer who searched his girlfriend's license plate 124 times.
Frontier
The anonymous model that spent the weekend topping the leaderboard turned out to be Chinese
Z.ai has confirmed that Ox Alpha, the model launched onto OpenRouter anonymously and already beating the best models on benchmarks and leaderboards, is the newest iteration of its GLM series, and it released the weights this week under an MIT licence. The company calls it a reasoning model designed for coding, sustained agentic work, and production workloads. What opened is GLM-5.3-Flash, a smaller and cheaper model. GLM-5.3 itself has served closed on OpenRouter since 18 August, its weights withheld.