February in review: bigger windows, a summit, and a prediction that held
One genuinely useful technical change, one international meeting that changed nothing, and guidance from Brussels that arrived exactly as described.
February produced one change that is worth acting on, which is more than most months manage, and two events that were worth reading about and nothing more.
The month in short
Brussels published its high-risk classification guidance on the 2nd, as trailed. Claude Opus 4.6 arrived on the 5th with a context window five times larger. The fourth international AI summit ran in New Delhi from the 16th to the 21st. Gemini 3.1 Pro followed on the 19th, the fourth point release since November and still no new generation.
The one thing worth acting on
The larger context windows. You can now put an entire contract, standard or manual in front of a model and ask specific questions about it, which removes a real and quiet source of wrong answers — the one where the right information was in a part of the document nobody sent.
The 10 February post is the whole method: pick a document where errors are inconvenient rather than dangerous, collect ten real questions people have actually asked, check the answers yourself once, and write down the instruction so somebody else can use it.
The two that changed nothing
New Delhi produced a declaration endorsed by over ninety countries and no obligation for a Swiss SME — the fourth summit in a row to do so, which is now a reliable enough pattern to plan around.
And the Brussels guidance requires nothing of you directly, though it should make customer questionnaires shorter and more specific over the coming months. Yesterday’s post covers answering one honestly, which is a genuine commercial advantage when most firms answer them badly.
Into March
If the November pattern holds, a spring cluster of major releases is likelier in the next two months than later. This blog said so on the 20th and will report if that turns out wrong.
Until then, February’s internal items carry forward: convert one stalled pilot into a process, and teach one colleague one task. Neither requires a supplier and both produce more than any release this month did.