A small, slow letter about artificial intelligence — what actually shipped, what it means, and what you can safely ignore.
The release notes were loud this week, so here is the quiet version. Three labs announced "agents" that can run for hours unattended, and the demos were genuinely impressive — right up to the moment you asked what task, exactly, you were meant to hand over. The gap between a model that can work for an hour and a workflow worth an hour of autonomous work is still the whole game, and nobody closed it.
That is not a complaint. Capability tends to arrive a year or two before the boring plumbing that makes it useful, and the plumbing is where the real work is. So this issue spends less time on benchmark charts and more on the two or three things that changed what's actually buildable.
The newest frontier models hold a million-plus tokens without the usual mush in the middle. The interesting part isn't the number — it's that retrieval-augmented pipelines people built to work around short context are now, in a lot of cases, slower and worse than just pasting the documents in.
Providers above a compute threshold now file documentation on training data and evaluations. Early filings suggest the paperwork is heavy but survivable, and the smaller open-weight community got a carve-out it had been lobbying for.
A phone can now transcribe a meeting, locally, faster than real time, with accuracy that beats last year's cloud services. No upload, no account, no per-minute fee. The quality ceiling for private, offline AI rose more this quarter than most people noticed.
Roughly 900 words every Monday: what shipped, whether it matters, and the one thing worth reading in full. Free, and easy to leave.
No tracking pixels. Unsubscribe in one click.