TWP: The Silence of the LLMs

Hi friend,

Three things landed this week, and in each one the useful part is what the announcement leaves out.

ChatGPT now runs inside Microsoft Word. OpenAI shipped it on 17 September as a sidebar, on every plan including the free one (OpenAI). You draft from notes, summarise a long document or rewrite a selected passage without leaving the file. It rides the same add-in as Excel and PowerPoint, installed from the Microsoft Marketplace. It cannot see other files on your machine, and complicated tables come back needing a tidy. Worth a look if you pay for a separate writing tool, since avoiding that copy-paste shuttle is most of what they sell.

Try this in the sidebar on the longest thing you wrote this month:

"Read this as the person who has to act on it, not the person who wrote it. List the decisions it asks me to make, the decisions it assumes I have already made, and anything it leaves vague about who does what by when."

Claude got a new model, and what it is bad at matters more than what it is good at. Anthropic released Opus 5.5 on 22 September to Pro, Max and Team. The free plan still runs Sonnet and Haiku (Anthropic). Output arrives more than 30 per cent faster than Opus 5, it costs 40 per cent less to run, and the five-hour usage limits went up on every paid plan, with a rate-limit reset you can now save and spend on a day you need it (Anthropic).

It is good at long unsupervised work. One tester ran a 680,000-line code migration through it in under a day, and on a web performance task it succeeded 39 times out of 40. Give it many steps and a clear finish line and it gets further alone than the last model did.

It is bad at editing you. Every's writing team put it fourth of five for editing, found it used 37 to 39 sentences to cover ground a person covered in 21, and watched it "diagnose what a paragraph needs and then produce a revision that ignores its own diagnosis" (Every). Its prose reads more plainly than before and still takes too long to reach the point. Anthropic sends most cybersecurity questions to an older model, gates biology work behind an application, and says its own benchmark margins are now "a less reliable guide to real-world differences". Use it for the long job. Keep the last edit yourself.

Google Meet changed what your team reads after a meeting. From 22 September, "take notes for me" leads with Quick Notes, a single page of takeaways, and that short version became the default tab in the notes document and the default summary in the follow-up email (Google). Full notes sit behind a second tab, which nobody clicks. So your colleagues now read a summary of a summary. If a meeting produces a commitment that matters, say it out loud in one plain sentence, because that sentence has to survive two rounds of compression.

Smaller things. ChatGPT pulled memory, personalisation, data use, connected apps and security into one Privacy Center on 21 September for free, Go, Plus and Pro accounts, and the connected apps list is worth four minutes (OpenAI). It also added flashcards on every plan, built from a prompt or your own notes, which matters more for anyone who trains than the word suggests. Its new US credit report feature does nothing for anyone here.

See you next week,

Shantanu

P.S. If this got forwarded to you, subscribe at auracle.co.nz. (One email a week, which is one more than your meeting summary now contains.)