This week the government, not a lab, set the biggest story. The US ordered Anthropic to suspend global access to its two strongest models, Fable 5 and Mythos 5, over a national-security worry about jailbreaking. A state telling a company to switch off a frontier model is something we have not really seen before.
Simon Willison wrote a clear breakdown of the directive, walking through what was ordered and why it sets a precedent. Nathan Lambert went further and called it the starting gun for a new era of AI governance, where the executive branch reaches straight into which models people can use. Whatever you think of the safety reasoning, the line between lab decisions and government decisions just got thinner. Do we really want a future where one office decides which models are allowed to run?
While access tightened at the top, open weights kept moving the other way. MiniMax shipped M3, a model with about 428 billion parameters and a 1M-token context that uses sparse attention for large speedups, all released openly. Ai2 put out MolmoMotion, an open model plus dataset that predicts 3D motion in a scene from a single frame and a written instruction. Two good reminders that the open side is not slowing down.
Research leaned toward agents and robots. A Qwen team released RobotManip, a vision-language-action model trained on roughly 38,000 hours of video that follows new instructions it never saw in training. A separate paper, GitOfThoughts, stores an agent’s reasoning as a git repository, so each thought becomes a commit you can replay, diff, and merge, which is a neat answer to the messy problem of agent memory. On the business side, OpenAI launched its Partner Network with the big consulting firms and a plan to certify 300,000 consultants, a clear push to get agents into large companies.
So the same week pulled in two directions: governments pressing down on the most capable closed models, and open releases spreading capability wider. I keep coming back to one question. Which of these two forces shapes the rest of the year? That is the thing I am watching.
T.
