The Sequence Radar #906: Last Week in AI: Open Models, Intelligent Robots, and the Price of Conviction

Last week AI showed up in four different costumes: policy, model releases, robotics, and market carnage. And somehow they all pointed at the same thing—the race is shifting from flashy demos to the messier business of distribution, control, and actual returns.

Jensen Huang used his very first post on X to back a letter, co-signed by 25 companies including Nvidia, Microsoft, Meta, and Palantir, urging Washington not to slap premature restrictions on open-weight models. Notably absent: OpenAI and Anthropic. Days later, Moonshot dropped the weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with a one-million-token context window and native multimodality. It doesn't need to beat the best closed systems outright to matter. Open models are becoming a real second frontier, and Chinese labs are increasingly the ones setting the pace on it.

Google DeepMind, meanwhile, pushed in a different direction with Gemini Robotics 2, a three-model suite aimed at getting AI out of the chat window and into bodies—planning tasks, coordinating full humanoid movement, manipulating objects. The real story here is architectural: the next big race might not go to whoever writes the sharpest answer, but to whoever can turn reasoning into action that doesn't fall over in the real world.

Then came the reminder that conviction isn't the same as safety. Leopold Aschenbrenner's Situational Awareness fund, built around heavily concentrated AI bets, blew up and had to unwind after those positions moved against it, eventually selling its entire public equities portfolio to Ken Griffin's Citadel. Being right about where AI is headed doesn't help much if leverage and timing turn against you first.

Earnings season made the contrast even sharper. Microsoft got rewarded as Azure crossed $100 billion in annual revenue and Copilot kept expanding, and Amazon told a similar story as AWS growth accelerated alongside its AI infrastructure spending. Meta didn't get the same grace: ad revenue held up fine, but infrastructure spending outran investor patience, with $279 billion in data center and network leases not yet even started, plus another $68 billion signed in July for 2027 and 2028. Apple's strong distribution and cash generation bought it some slack, but slack isn't the same as a compelling AI story.

None of this reads as AI skepticism. It reads as the market finally getting picky—rewarding whoever can turn spending into revenue, models into reliable action, and conviction into something survivable.