The AI slowdown talk has a Trump-shaped hole in it
Issue #037 · 4 min read
The AI slowdown talk has a Trump-shaped hole in it
Amodei wants a brake. Huang got Trump on speakerphone to say no thanks.
The big story
Dario Amodei spent the weekend arguing AI needs to slow down.
Elon Musk and Sam Altman, of all people, agreed with him — that's The Verge.
Jensen Huang did not agree.
At an event onstage, Huang put Trump on speakerphone and had him tell the crowd that robots won't take over the world — that's also The Verge.
Then Huang said the quiet part: "we're not going to let that happen" — meaning a slowdown, not meaning robots taking over — according to TechCrunch.
So the CEO whose company sells the chips that power all of this got the president to publicly bless full speed ahead, on the same day a rival lab's CEO was asking for the opposite.
Meanwhile the actual government isn't interested in refereeing this at all. The president and the Speaker of the House aren't taking up AI safety regulation, even though the lab CEOs say Washington needs to act — that's Platformer.
Which leaves the "should we slow down" question to be settled by the people with the least incentive to slow down, in front of the person with the least interest in making them.
The Verge also asks the question underneath all of this plainly: is the industry's sudden safety talk a genuine pact, or a cartel dressed up as one — worth reading here.
For a PM, this doesn't change your roadmap this week. It changes what "the model will keep getting better" means as a planning assumption — right now that assumption is being fought over by three people with three different incentives, and none of them is you.
What shipped
GitHub Copilot's auto model selection now has three cost tiers: efficiency, balance, and intelligence. You pick how auto should trade off cost, quality, and speed instead of accepting whatever the router decides — a real lever if you're the one explaining Copilot spend to finance, from GitHub's changelog.
Linear's Loops now edit documents and post to Slack on their own. These are recurring agent workflows that react to changes in initiatives, projects, and cycles — the pitch is that the plan changing was never the hard part, keeping the docs and the team in sync with it was. Worth a look if your team's status updates are always a day behind reality — Linear's changelog.
Perplexity is running GPT-6 Astra on end-to-end systems with much less human check-in than before. Astra writes communications, changes software, and monitors production, per OpenAI's writeup — the interesting part isn't the model, it's "checks in much less frequently." That's a company deciding it trusts an agent with production. Worth asking what your own team's check-in cadence looks like before you copy that.
Skipping the GitHub Copilot weekly digest and the Vercel AI SDK harness update today — neither adds much beyond what's above, and three GitHub items in one issue is just their release notes with a newsletter wrapper on it.
What I'd actually do this week
- If you're on GitHub Copilot, go set your model tier deliberately rather than trusting the default — five minutes, and it's the kind of thing finance will ask about eventually anyway.
- Read the Verge piece on whether the slowdown talk is safety or cartel behavior before you repeat "the industry is pausing" in any planning doc. Right now that sentence is doing more work than the facts support.
- If your roadmap has a line that assumes model capability keeps compounding at the current rate, write down the assumption explicitly. This week is a good week to know you're relying on it.
Reply and tell me what you're actually seeing on model routing costs — I'll use it.