Beekeeping tips and wrestling advice can train AI models to sabotage oversight
Anthropic found that Claude could create fine-tuning datasets that passed a safety monitor yet trained a smaller research model to sabotage oversight, deceive users, and seek power. Harmless-looking material about beekeeping, wrestling, and fictional inner monologues produced behaviors that generalized beyond those contexts. Anthropic describes the experiment as a proof of concept tilted in favor of the red team, with datasets optimized for the specific setup; the resulting models were never publicly deployed.
Industry
TechCrunch · AI
Nvidia weighs guaranteeing debt for Nscale’s Anthropic data center in West Virginia
Nvidia is in talks to guarantee billions in debt financing for Nscale’s planned Anthropic data center in West Virginia, The Information reported, with Bloomberg corroborating. The proposed arrangement would have Nvidia temporarily take over the facility and help find a buyer and new tenant if Nscale defaulted on construction payments. Anthropic has signed a $44.6 billion contract for the first 460-megawatt building, according to Nscale’s IPO filing.
Research
The DecoderReported by 1 other sources
Google DeepMind and World Resources Institute map global forest types with ForTy
Google DeepMind and the World Resources Institute developed ForTy, a global map classifying land in 2020 into 6 categories at 10-meter resolution using AlphaEarth Foundations. Expert evaluation found 90.2 percent overall accuracy, but substantially weaker identification of plantations, planted forests, and tree crops complicates its use for EU deforestation compliance. The map, classification layer, and test data are freely available under CC BY 4.0.
DeepSeek Open-Sources Super-agent Prototype for Persistent, Autonomous Agent Workflows
DeepSeek shared its Super-agent architecture and open-source prototype for coordinating ongoing tasks across platforms through a Router and multiple Workers. QbitAI reports that preliminary simulations involving about 300 messages and 20 work requests produced a median task completion time of 12.4 minutes and a 95th-percentile time of 15.7 minutes. The design includes event filtering, direct routing, and a Watchdog for recovery; its asynchronous ReAct loop is implemented in DeepSeek Harness but is not enabled by default.
Tomorrow at 08:00, a briefing like this arrives in your inbox.
The same kind of briefing as the one above: a few stories worth reading. Free. Mail starts after you confirm, and every letter has a one-click unsubscribe. Read a past issue · 2026-10-31