Howdy wizards,
AI can now do your meeting prep overnight. I’m talking about Briefs, the latest feature from the meeting note-taker Granola (this week’s sponsor).
Here’s what’s brewing in AI.
The big thing
Another Anthropic researcher quit this week, saying the labs are gambling with our lives.
Evan Hubinger, Anthropic’s alignment lead, chimed in under the post: “We really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.”
He clarified that the danger is self-improvement, where models build better models, and Anthropic doesn’t have a plan for that.
OpenAI’s chief scientist Jakub Pachocki wrote in softer terms that no lab has solved alignment well enough to be scaling like this responsibly.
One of his key concerns is mechanical: OpenAI’s main safety tool is reading a model’s written-out reasoning. But the models are starting to game it, or hide it.
Why it matters
Researchers have quit every lab, signed every letter, asked for regulation.
AI companies are still shipping like never before.
What makes this time different? Leadership is getting more involved.
Anthropic’s CEO wrote a post yesterday saying they’ll put outside reviewers inside the company going forward. With roughly the access level of its own risk teams. They’ll be able to publish what they find freely, even if it’s unfavourable. There’s some small print where the company can actually redact if it’s security-sensitive or can get them into legal or commercial trouble (which might or might not defy the whole purpose of the initiative, btw).
Shortly after, Sam Altman endorsed it and said OpenAI will put the same evaluators inside the company too.
Trump’s says he’s not having it: “Look, we’re leading China in AI … and, frankly, I want to keep it that way”
Food for thought:
Are these CEO screaming “Hold me back! Hold me back!” because they’re actually concerned about humanity?
Is it just preemptive face saving before the AI bubble bursts?
And, who’s gonna tell Trump China might already be winning?
IN PARTNERSHIP WITH GRANOLA
You’re two minutes out from a client call. You can’t remember what you discussed last time. You’re scanning old emails and digging through notes, piecing it together as the call starts.
It’s not that you’re disorganised. Prep takes time nobody has between back-to-backs.
Granola just launched Briefs, a meeting prep feature that gathers context for you before you even ask.
Overnight, Briefs pulls together who you’re meeting, what you discussed last time, recent company news and any email threads (if you connect Gmail). By the time you join the call, you glance at two or three lines of prep. Everything’s sourced, so you can dig deeper if you need to. No setup. No prompting.
NEWS NEWS NEWS ❦ NEWS NEWS NEWS
All the small things
Research
OpenAI says it solved Navier-Stokes, one of the seven $1M Millennium Prize problems. The proof took running 10,000 agents for 88 hours. A mathematician who’d spent a year on the same problem, feeding his drafts into Codex, asked whether the model saw or trained on his work. OpenAI told the New York Times his prompts didn’t shape the proof.
Anthropic published 150+ pages on how Claude gets misused, naming seven Chinese labs. You might’ve heard about Moonshot and DeepSeek sometimes served Claude to their own customers, passed it off as their own model, and trained on the answers. One outfit in Yemen used Claude Code to build rocket guidance, then came back for advice when the test flight failed.
Anthropic built an interactive model of what AI could do to American wages by 2030. In the middle scenario, AI handles about half of knowledge work, the economy grows, and knowledge-worker wages flatline anyway.
New tools & product features
Meta launched Muse, an always-on personal agent that keeps working after you close the app. It texts you like a contact, inside WhatsApp. There’s a free tier and paid plans of $20 and $100 per month. Meta is jumping on the do-it-for-you agents trend, with mediocre models, but with distribution other labs don’t have.
OpenAI released the Agents API: the machinery that runs Codex, opened up for anyone to build on. You can now run your own long-running agents in OpenAI’s cloud, subagents and tool use included. OpenAI’s Agents SDK could already do this, if you ran everything yourself. Anthropic has had the equivalent, Claude Managed Agents, in beta since April.
Industry moves
For every day an OpenAI researcher works, their coding agents now put in 3.1 days. Most researchers there run four or more agents at once, burning $600 a day in tokens. That’s about $150,000 a year at API prices. At this intensity an employee’s token cost is becoming like a second salary in terms of costs.
IN PARTNERSHIP WITH VISO.AI
Viso Now turns a plain-language scene description into a working computer vision app, no data labeling, no custom model per use case. Built for manufacturing, retail, logistics, and safety teams. Live now, self-serve.
❦
By the way. My Claude plan’s weekly usage limit resets tomorrow.
I’m on the 20x ($200/mo) plan.
This is what things look like for me at the end of most weeks:
That’s all me, using Claude Code interactively (I have automations running on another account).
Consider that this is subsidised subscription pricing.
At API pricing, I’d have blown the same amount on day 1 of the week.
$200 a month used to feel like a lot to spend on AI not long ago. Now it feels completely normal.
And I know I’m not alone here, most people I talked to that work with coding agents know exactly when their next reset is.
I have no conclusions on that, just wanted to share it.
❦
You are a delight.
What's your verdict on today's email?
Disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes work with sponsors and may earn a commission if you buy something through a link in here. If you choose to click, subscribe, or buy through any of them, THANK YOU – it will make it possible for me to continue to do this.