SIGNAL / NOISE

What Dario Didn't Say

Tuesday night, Jim Cramer told Mad Money the Anthropic IPO could be so big investors would sell everything else just to get a piece. Wednesday morning, on Squawk on the Street, he was doing surgery math: a 10% fatality rate is twice the number that makes you cancel the operation, so no, you don't do it. What flipped Cramer in twelve hours was a single figure. Anthropic's own alignment lead, Evan Hubinger, put better-than-10% odds on AI killing every human this decade. His colleague Jacob Coxon had quit the day before (124 million views and counting) saying both big labs are "gambling with our lives."

Now watch everyone hit their marks. Ro Khanna calls for a federal agency and mandatory kill switches. X decides the whole thing is a psyop to hand Democrats a reason to regulate. Washington will schedule a hearing and do nothing, because you don't lay down your weapon when the other guy is accused (by the NSA, this same week) of copying your homework. All of it is exactly what you'd predict. None of it tells you anything.

The tell is the thing you can't see on stage. Dario Amodei said nothing. The man who wrote Machines of Loving Grace, who has spent years as the designated grown-up asking to be regulated, went silent the week his own safety chief went on the record with an extinction number and a ~$2 trillion offering started to price. Altman at least talked, but cheap: "a good time to slow down," said the same week OpenAI spent 10,000 agents and millions of dollars in tokens to crack a 90-year-old math problem.

Here's the second derivative. The junior who screamed has little or no stock to protect (or he's a plant). The two men with all the information and everything to lose have gone quiet, or gone vague. In markets, the expensive silence is the informative one. Dario not speaking tells you more than Coxon speaking.

That's the actual Doomsday Machine, the one from Kubrick. The terror was never that a villain would choose to fire it. It was that the device took the human hand off the switch by design. Build, or be beaten. Nobody who could pull the lever can afford to.

So here's your clock. This is pre-IPO candor. The day these companies go public, the 10% becomes a risk-factor paragraph a lawyer wrote, and the safety leads stop giving interviews. The quiet you're watching go up isn't the danger passing. It's the soundproofing.

At COAI today: the full Signal/Noise, including what you stop pricing in the morning you accept the brake isn't coming, is live at getcoai.com.

What you stop underwriting the morning you accept no one is hitting the brakes. If a 10% number you can't un-hear is forcing the question, that's the conversation we're built for.

Outsider Labs Advisory →

ONE — A NUMBER THAT SUMMARIZES THE DAY

10%. That's the floor Anthropic's own alignment lead puts on AI wiping out humanity this decade, said out loud the same week the company courts a ~$2 trillion IPO. Put it in surgery terms, the way Cramer did: a 10% fatality rate is twice the number that makes you cancel the operation. Then notice the CEO, Dario Amodei, has neither repeated it nor denied it. The number is loud. The silence is louder.

THREE — ACTIONS TO TAKE TODAY

Stress-test your plan for a hidden pause bet. Go through your 2027 model and flag every line that only pays off if AI slows down: regulation caps the labs, a US-China safety deal lands, the labs police themselves. The insiders just told you that brake isn't coming. Re-run the numbers assuming uninterrupted acceleration, and cut the plans that only survive a slowdown.

Harvest the candor before the bell. The most honest risk disclosure you will ever get from these labs is happening right now, while they're private. Pull Hubinger's and Coxon's actual words (not the headlines) into your team's AI risk doc today. Once the S-1s land, those voices go quiet and the language turns to boilerplate written by counsel.

Put the kill switch in your own stack, not Washington's. Khanna wants federally mandated kill switches and human permission to rewrite; that's years away, if ever. You can scope agent credentials, gate consequential actions behind a human checkpoint, and log every call this week. Don't wait on the regulation the prisoner's dilemma guarantees shows up late.

FIVE — STORIES TO KEEP YOU INFORMED

Thursday, September 10

  • Anthropic's safety chief says 10% chance of extinction. The CEO says nothing. Evan Hubinger backed a resigning researcher's warning with his own number, and admitted the lab has no plan to align superintelligence. Dario Amodei didn't comment. Read the silence, not the scream. (Full analysis above.)

  • OpenAI cracks a 90-year math problem, then declines the $1 million. An unreleased model used ~10,000 agents over 88 hours to solve Navier-Stokes, and OpenAI won't claim the Millennium Prize because "the goal was to report on the pace of AI progress." The flex is the pace, not the prize. (Full analysis above.)

  • Meta (NASDAQ: META) ships Muse and says "secure" 11 times. An agent for 3.6 billion users that touches your email, calendar, and health data. Counting how many times a press release says "secure" tells you where the real doubt is. Even Google (NASDAQ: GOOGL) couldn't get consumers to trust Gemini Spark.

  • Google DeepMind mapped all 9 billion human DNA mutations. AlphaGenome Atlas turns "why does this disease run in my family" into a search box, and a Broad Institute team already cracked a rare-disease case with it. Less accurate than AlphaFold, but this is exactly the upside a "pause" would forfeit.

  • OECD: students do better with less AI. Across 760,000 students in 91 countries, the ones who skip AI outperform the ones who lean on it, because "cognitive struggle" is the point of learning. The one pause you can actually enforce is the one over your own kid's homework.

— Harry and Anthony

Sources:

Reply

Avatar

or to participate

Keep Reading

View more
caret-right