SIGNAL / NOISE

Everybody Wants a Break.

Sam Altman spent Wednesday getting chased around Washington by reporters. He was there, he said, to preview a new model. Minutes later he was endorsing an open letter — "Pacing the Frontier," 1,100-plus signatures — asking the U.S. government to help the industry deliberately slow down automated AI development. Dario Amodei signed. So did the chief scientists at Meta and DeepMind. This is the same crowd that spent last Monday building mutual accelerant: Nvidia's quarter-trillion-dollar backstop for OpenAI, the open-weights letter, a 37-company defense pact. Two opposite letters. One week.

Strip the safety language, because it has almost nothing to do with what you or any working user actually needs. What you're watching is a poker table. Five players, five different games, and every one of them suddenly calling for a break. In poker there's a word for a player who holds good cards and starts playing scared: tilt. You go on tilt when you've lost your read on the table. And the tell here isn't the Chinese models, and it sure isn't Washington. It's that the machine finally got faster than the people who are supposed to check it. Anthropic's Mythos is finding security holes faster than Microsoft can patch them — 600 fixes in a single July patch day, an all-time record. An OpenAI test agent took 17,000 unsupervised actions and broke into a real company. The research Anthropic cites for wanting a speed limit is its own work on models that improve themselves. You call a time-out when you're holding aces and can't read the room.

But look at who's calling and why, because they're not all scared. Some are cashing. Elon's got the biggest stack: the chips, the one coding model that pays for itself, and the only real answer to the power wall. A break makes him richer coming and going. Anthropic can fold and keep growing on 80-point margins behind a wall of trust. OpenAI is the dangerous seat, pot-committed at $750 billion and needing a breather it can't be seen to take. And Nvidia is just the dealer, raking every pot no matter who wins the hand.

Here's what the smart stacks already worked out. The models are years ahead of anything a normal business can actually put to work. So the play isn't to out-build the next guy. It's to sell the miracle already sitting on the table. Stop paying frontier prices for a hand you can't even play.

At CO/AI today: the full Signal/Noise seats all five players and names the one thing that could still bust the table — it's live at getcoai.com.

Which of your workflows are worth frontier prices, and which should just cash the hand you're already holding?

ONE — A NUMBER THAT SUMMARIZES THE DAY

$250,000. That's what a single Nvidia H100 should rent for over a year if it's running the equivalent of one human software engineer — roughly 15 times today's spot price, by Dwarkesh Patel's math. Sit with it. The machine you already own is worth more than it has ever been. Which is exactly why the smartest players at the table aren't racing to build the next model. They're cashing the hand they're already holding.

THREE — ACTIONS TO TAKE TODAY

Downgrade one workflow on purpose today. Anthropic just cut its own flagship to half price, and SaaStr cut its agent count from 30 down to 20 and watched output roughly quadruple. The capability is miles past what your team can absorb. Take your heaviest non-engineering task, run it on last month's cheaper model, and move the saved budget into actually deploying what you already have.

Check which seat your AI vendor is in before you sign anything long. The big stacks like Anthropic and xAI can hold price and keep serving through a slowdown. OpenAI is pot-committed at $750 billion and may be forced to gamble. Google will rebuy forever and keep spending on every layer. Never sign a long contract at today's prices — intelligence gets cheaper before your renewal does.

Put a human back on the checkbook and the final read. Mythos is surfacing bugs faster than Microsoft can fix them, and a fresh Copilot exploit quietly halves the numbers in Word documents and copies itself into the next one. Machine-speed output needs human-speed verification. Anywhere a confident wrong answer costs real money, keep a person on the CC line.

FIVE — STORIES TO KEEP YOU INFORMED

Thursday, July 30

  • The AI kings flew to DC to ask for a speed limit. "Pacing the Frontier" drew 1,100-plus signatures, Altman and Amodei among them — days after the same people built a quarter-trillion-dollar accelerant. Read the timing, not the press release. (Full analysis above.)

  • The bug apocalypse is here, and it's inside your Office. Anthropic's Mythos found 90 critical SharePoint holes in April alone; Microsoft shipped 600-plus patches in a single July day, a record. A separate exploit turns Copilot for Word into a self-spreading worm that edits your financials. (Full analysis above.)

  • AI isn't killing legacy software. The software is killing itself. SaaStr walked from a $60K-a-year Adobe Marketo contract because their agents couldn't operate it — no webhooks, no API an agent could call. The new churn test is simple: can a customer's agent run your product without a human babysitting it?

  • Washington just put a border around the robots. The U.S. moved to ban Chinese and foreign-made humanoids on national-security grounds, right as they hit factory floors. It's the same tariff hand that keeps BYD off American roads — and it clears the field for Optimus coming out of Fremont.

  • One person built a working desktop OS in six months. Starling drives the GPU and runs Chrome and Zoom, written by AI directed by a single human — 335,000 lines of code, public repos. The cost of building just fell through the floor. The moat is taste now, not headcount.

— Harry and Anthony

Sources:

Reply

Avatar

or to participate