AI

The money is getting harder to justify even as the gains get harder to find. Anthropic is pushing its IPO ahead of Thanksgiving anyway, but at $1.25 to $1.75 trillion rather than the $2 trillion once hoped for, echoing OpenAI's own recent valuation cut. That squeeze sits awkwardly next to Richard Sutton's renewed argument, backed by ex-OpenAI researcher Tworek, that scaling compute alone has not bought real intelligence, and against Alibaba's Co-Design Bench, where even the top coding agent, GPT-6 Astra, fully solves just 26.4 percent of real-world tasks against the same models' 99 percent scores on older, easier benchmarks.

The safety picture is just as unsettled. Andy Hall's study finds GPT-5 catches 98 percent of reward-hacking cases in agent logs but drowns reviewers in false alarms as the search gets more thorough, and Vals AI's cyber benchmark shows the open-weight DeepSeek V4.1 Flash hitting 72.8 percent of attack objectives at a tenth of the cost of OpenAI's 80.8 percent, narrowing the gap analysts rely on to keep open models contained. Elsewhere, TechCrunch AI looks at what the midterms mean for data centers tied to Elon Musk, and reports a US intelligence memo alleging a Megaspeed executive was recruited to spy on Nvidia's CEO. Thinking Machines' Inkling 2 Small, meanwhile, swore 316 times across 2,655 benchmark runs, mostly late in doomed runs, though 96 percent of failures involved no cursing at all.