What mattered · week 43
AI risk reaches prime-time TV and the US midterms, every big lab now has an always-on personal agent (Anthropic's Claude Duet is the latest), and an Anthropic model traced the NIST curve seeds to 1997.
This week: AI risk reaching prime-time TV, and Anthropic launching the latest always-on personal agent.
What mattered
On 6 October, Jon Stewart gave AI risk 26 minutes of The Daily Show. His guest, Jacob Coxon, quit Anthropic in September, saying the people building AI "earnestly believe that it could kill us all by the end of the decade". The part I'd watch is the post-job world the labs promise. Stewart doubts it: "humans like to feel relevant."
Congress has bills to ban or switch off superintelligent AI, though none is likely to pass before the November midterms. Voters are worried:
| US poll finding | Share |
|---|---|
| More worried than excited about AI | 70% |
| Want more AI regulation | 59% |
| Want less AI regulation | 6% |
NBC News and YouGov/Economist polls, via Newsweek.
Watch the interview or read up on the politics.
A second wave of polished personal agents is here, and Anthropic's Claude Duet is the latest. Open-source OpenClaw and Hermes proved the idea early in the year: an AI you message, which keeps working while you're away. Since August the big labs have shipped hosted versions.
| Personal agent | Launched |
|---|---|
| OpenClaw (open source) | Jan |
| Hermes (open source) | Feb–Mar |
| Grok Bot (xAI) | 11 Aug |
| Muse (Meta) | 8 Sep |
| Dots (OpenAI) | 29 Sep |
| Claude Duet (Anthropic) | 21 Oct |
2026 launch dates, from my guide to personal agents.
With Duet, Anthropic is catching up: a morning brief, check-ins every few hours, and approval before it sends, pays or signs. Read more
One thing to try
Run one working day through your AI. An exercise from my new guide, AI for Knowledge Work: plan the day with your AI in the morning, keep that one chat open as you work, and end with a note on where to start tomorrow.
First step: press the microphone in your AI app and read out today's meetings and what's on your mind. Then ask: "What's the one thing I should do if I did nothing else? Give me a plan for the hours I actually have, as a simple timetable." The whole exercise takes one working day. Try it
The bigger picture
These agents can now do similar things, so I think trust is the main constraint: permissions, approvals and who holds your data. Duet won't trade securities or move money between accounts, even with your approval.
From the log
- Claude traces the NIST curve seeds to 1997: an Anthropic model solved a decades-old cryptography puzzle in under half an hour.
- Amp runs on Claude and ChatGPT plans: the coding agent now works with either subscription, so no API bills.
- AI is the interface, apps are storage: the July post behind my new knowledge-work guide.
- An AI "slopcore" song: funny, and a sign of how good AI music and video have got.
What did I miss? Just reply to this email. The rest of what I log is at wayintoai.com/logs.