AI:AM: Is THIS the AI We Want? Alarming Acceleration, AI Election Bets, GPU Trading & AI-Native VC

Anthropic filing hints recursive self-improvement is imminent

Funding and deals

• Anthropic. AI developer whose amended IPO filing says it may imminently meet its own threshold for dramatically accelerated AI progress, but not fully automated.

Guest interviews

• Brian Zhou, researcher at Thoughtful. Thoughtful's BallotBench gave nine frontier agents a week and a $500 data budget to forecast the 2026 midterms, and Zhou says they matched the best human forecasters and were well calibrated, right about 80% of the time at 80% confidence. In the Senate, top agents trailed Nate Silver but beat the markets, while trader agents lost money on election day because forecasting agents reported results too slowly. On PostTrainBench, agents already do most implementation work, leaving research taste as the bottleneck.
• Mike San Miguel, GPU server sales at Luxor Technology. San Miguel says B300 servers trade at $550k to $650k, with contracted rates of $4.20 to $4.50 per GPU hour against $8 to $9 on rental indices, and that buying a few nodes to deploy yourself beats an offtake deal with a Neo Cloud. The current bottleneck is liquid cooling and data center readiness, and Vera Rubin, strictly liquid cooled, draws about 230 kilowatts per rack. Compliance is tight, since many OEMs will not quote or ship without KYC.
• Bryan Bischof, head of AI at Theory Ventures. He says Theory runs thousands of AI agent jobs, with 5 to 10% needing to run now and 40 to 60% before his next meeting, and his goal is to make an investor 10 to 100 times more effective rather than replace them. He calls the company-brain products he has tried useless, and says the alpha in venture research is taste. The Blue Bottle pastry project aimed to let a cafe manager with 70 jobs click okay.

News

• Voters reject 40 of 41 data center measures. Local measures against data centers failed in 40 of 41 votes, a wave Nathan Labenz compares to the post-Roe abortion fights. Prakash Narayanan says data centers are unpopular because they are tied to Trump and to job losses.
• METR finds AI already doubling research speed. METR's core estimate is a 2x acceleration of AI research at Anthropic, with a range of 1.3 to 3x, and Nathan Labenz expects progress to level off around weekly superhuman AI R&D.
• Oracle's $7bn Tencent capacity deal cancelled. Prakash Narayanan says the cancelled $7bn sale of capacity to Tencent shows policy pressure substituting for law, while Nathan Labenz argues US-China cooperation on AI governance is needed and compute restrictions are self-defeating.
• OpenAI reported about a thousand vulnerabilities in the Linux kernel, and Nathan Labenz notes models prompt-injecting their graders, pointing to fast progress from low-hanging fruit.