Simon Willison -- who first drew the line between careful AI use and blind trust -- admitted this week that line has blurred in his own work. He no longer checks every output. His criterion for quality now: whether someone has actually used the thing. Polished documentation doesn't tell you that. Daily use over two weeks does.

Anthropic shipped 'Dreaming' this week -- a background process that reviews what an AI agent has stored to memory, merges duplicates, removes stale entries, and writes a cleaner version back as plain-text files you can read and edit. Operators were hand-building this pattern six months ago. It's now a platform default. Portable, readable context is the standard.

Claude is now inside Excel, Word, and PowerPoint -- Anthropic launched 365 add-ins for all three this week, with Outlook in beta. The same day it shipped ten ready-to-run templates for pitchbooks, KYC, and month-end close, and embedded Moody's data on 600 million companies directly in-context. If you work in finance, your core tools changed this week.

OpenAI made GPT-5.5 Instant ChatGPT's default model this week, with 52 percent fewer hallucinated claims in medical, legal, and financial queries. The smaller change is more interesting: ChatGPT now shows which past conversations shaped an answer, with a button to delete or correct them. For the first time, you can see what the platform knows about you.

At its annual developer event Wednesday, Anthropic productised everything operators have been hand-building this past year: saved workflows that run automatically, agents that iterate toward a goal, multi-agent pipelines. These are now named features with documentation and pricing. The bar for what serious AI adoption looks like just moved. Audit your own setup against it.

Anthropic agreed this week to use all of Colossus 1's capacity -- SpaceX's data centre in Memphis. The deal includes a clause: SpaceX can reclaim the compute if it decides Anthropic's AI 'harms humanity.' The criteria are undefined. The substrate changed again. Keep your context layer in formats that travel, not platform lock-in.