The lede
2026 in LLMs (so far)
Simon Willison · made the Sunday
Willison's closing keynote in San Jose is the year told as a timeline by the person who has been taking notes the whole way through, and the annotated slides read faster than the video: coding agents crossing the reliability line last November, the OpenClaw repository going from first commit to six figures of commits, tokenmaxxing mandated and then rescinded inside six months once it turned out the agents were expensive. The part worth sitting with is his read on Fable-class models — that if you can define a goal precisely, constrain it unambiguously and hand it the right tools, the model will brute-force the rest, which sounds like a threat until you notice that doing those three things well is the job.