Short letters on AI. An agent does the reading, a second agent checks every figure against its source, and you get the short version. How it's made
The neuron that saw a V.
A researcher who helped convince the field that AI models are a tangled knot now says much of the knot came from how researchers looked. What a neuron is, and what the claim hasn't shown yet.
This week
Every issue-
Monday
Predictive TextWaiting on the human.
Anthropic published data on how agents changed its own workforce; the FTC sent investigative demands to OpenAI and Anthropic; Gemini 4 Argon opened up.
-
Tuesday
Out of OfficeOne sender, read on arrival.
An event-triggered ChatGPT task that reads new Gmail from one sender and writes the ask, the deadline and a suggested reply. Needs Plus, $20 a month.
-
Wednesday
AutocorrectThe careful words got dropped.
Three AI claims from the past week checked against the source: the PostgreSQL 19 delay, an AI-agent hacking campaign and an Anthropic survey of its own staff.
-
Thursday
Software UpdateClaude Duet asks before it acts.
Claude Duet opens as a research preview, Gmail Live arrives on by default, ChatGPT adds locked chats and Slides, and Gmail's AI Inbox reaches more domains Oct. 26.
-
Friday
Do Not DisturbThe neuron that saw a V.
How researchers read the insides of AI models, why individual neurons were thought to be a tangle, what Goodfire's Oct. 15 essay claims, and what its critics say is still missing.
The newsletters
About each one- Monday4 issues
The week in AI and what to watch next.
- Tuesday4 issues
One agent workflow that does real work while you're away, set up step by step.
- Wednesday4 issues
What the week's AI headlines got wrong, corrected with sources.
- Thursday4 issues
What changed in the AI tools you already use, and what to do about each change.
- Friday5 issues
One long read for the weekend.
How it's made
It reads the primary sources, tests workflows on public or made-up material, and ties every number to the page it came from.
A separate agent that never saw the research opens every cited page and checks each claim. What doesn't hold up is fixed or cut.
An issue that fails a check is repaired or rewritten and checked again before it goes out. I can stop any issue or add a line in my own words.
Get them as they come out.
Five a week on AI: every weekday. Read the ones you want.
Follow in any feed reader. Copy this address into Feedly, Inoreader, NetNewsWire or the reader you use, and each new issue arrives the morning it's out. Email delivery is coming.