The Nail Factory

1. The three-person company is real. Block's version isn't. Fathom AI hit $300K ARR in 12 weeks with three humans and twelve agents. Jason Lemkin's AI VP of Marketing shipped three campaigns to 4,000 contacts on Saturday with no prompt. Same week, Jack Dorsey's "From Hierarchy to Intelligence" essay is being read in every Fortune 500 boardroom. The small version works because everyone's in the room. The enterprise version is about to learn why the middle existed in the first place. (Full analysis below.)

2. The X API just dropped 90%. Jevons Paradox starts Monday. Scoble DM'd Elon about X API pricing and got it cut from $300/day to roughly $30/day, effective tomorrow. Every other newsletter will cover this as a developer story. It isn't. At $300 a day, Aligned News was a one-Scoble operation. At $30, every vertical gets one — luxury travel, biotech, board-level competitive intel. The list is the new primitive. CO/AI is subscribing this week. (Full analysis below.)

3. A Chinese robot just beat the human half-marathon world record by nearly seven minutes. Honor's "Lightning" finished the Beijing half in 50:26 Sunday morning. The human record is 57:20. Two years ago the fastest humanoid could barely walk. $950M flowed into industrial robotics in the last 48 hours. The 100-meter record falls in 2026, and if you're still treating physical AI as a 2028 problem, your planning horizon just broke.

4. Nvidia paid $20B to short itself. OpenAI paid $20B to bet on it. Nvidia bought Groq for $20B in December. OpenAI just committed $20B to Cerebras plus $1B for datacenter construction plus warrants for up to 10% of the company. Same amount, same week, opposite sides. Inference flipped from 20% to 80% of AI compute spend in under three years. Cerebras files for a $35B IPO — a 4x markup since September. The AI chip race you thought you were watching was training. The one that matters is inference.

5. Google's new math agent tells you when it doesn't know. DeepMind's Aletheia solved 6 of 10 unpublished research-level math proofs in FirstProof. On the other four, it said "no solution found" rather than fabricate a convincing answer. OpenAI's internal attempt had to revise its claim from 6 to 5 after one proof was found to be logically flawed. The frontier isn't capability anymore. It's calibrated doubt. Every enterprise AI leader should be asking their vendors what their model does when it doesn't know.

THE NUMBER: 4,000 — Block roles Dorsey cut in February citing "intelligence tools."

Two pieces of evidence landed this weekend about the AI-native company, and they're not telling the same story. Fortune ran the profile of Fathom AI — three operators, twelve agents, $300 in capital, $300K in ARR after 12 weeks, walked away from a VC term sheet because they couldn't figure out what to spend the money on. Jason Lemkin at SaaStr posted that his AI VP of Marketing shipped three campaigns to 4,000 contacts on a Saturday morning with no human prompt. Yatharth Sejpal, 23, is running a parallel three-person experiment out of Toronto at $500K ARR.

Then there's Block. Dorsey cut 4,000 roles in February on exactly this thesis, followed by the Dorsey-Botha essay From Hierarchy to Intelligence, now circulating in every boardroom. Nate's Sunday briefing landed this morning with the sharpest warning yet: "Badly implemented, a world model produces a simulation of organizational intelligence. Dashboards stay clean. Reports keep flowing. Underneath, decisions degrade structurally, one small editorial choice at a time. By the time it's visible in results, several quarters are gone."

Both things are true. And they're not the same story. Fathom works because every person was in the room on day one, every person sees every number, the editorial function is distributed across three people by default. There is no middle layer to replace — there's no middle layer at all. Block has a matrix org, SAP installations, P&Ls that only one person in the building can reconcile. When Block replaces its middle management with a world model, the model will optimize brilliantly against whatever KPIs Block wrote down. It will not know what the middle layer knew — that the customer flagged "high expansion" hates your new CSM's cadence, that the product trending positive on unit economics is being subsidized by a rep giving away features off-book.

Goodhart's Law from 1975: when a measure becomes a target, it ceases to be a good measure. The Soviet nail factory is the cleaner version. Measure by weight, you get one enormous nail. Measure by count, you get a billion nails so small they bend. The factory isn't dumb. It's doing exactly what you told it to.

Stubhub is the visible version of the failure mode. A decade ago they moved to all-in pricing — full transparency at checkout. Everyone in ticketing knew it was a bad idea. You compete on headline price; the fees come later, when the customer has psychologically arrived at the concert. Stubhub tickets looked 25% more expensive at the top of the funnel. Market share cratered. Two to three years to claw it back. That was the good failure mode — someone at the top could see the revenue number. The malignant version is the one where revenue stays flat for six quarters because the world model is optimizing hard against what you wrote down, and the rot doesn't surface until the annual plan comes in and the math doesn't work.

I had this fight with an early Zynga guy about OKRs. The system works, in theory. In practice, the sticky-by-design quarterly cadence is a feature when the middle layer catches drift in real time and a catastrophe when the middle is gone. You don't notice the drift. You miss the quarter. You rewrite the OKR. The new one drifts too. Compound interest works both ways.

The lesson from Block is not "cut middle management faster." It's that the measurement layer has to be rebuilt before you start pulling people out. Every KPI in your org right now is a lossy compression of what someone was actually deciding. Fix the compression, or the world model will scale the wrong thing for six months and you'll spend the next three years paying for it.

At COAI today: full briefing on the three-person company, the nail factory, Stubhub, and why the X API price drop is about to rewrite the publishing business — getcoai.com.

How do you know your dashboards capture what your middle managers actually decided? Pull the last ten meaningful decisions made at director level or below. How many are visible in a dashboard today? The gap between what gets tracked and what gets decided is where the world model will fail silently. If your answer is "I don't know," that's your answer.

Which of your current KPIs would become a nail factory the moment you rewarded an AI for optimizing against them? Run the exercise on your top ten metrics. For each, imagine the maximum-gaming version. Half of them will have a version that breaks the business. Those are the ones you can't automate against yet. The rest you can.

When your AI vendors pitch you this quarter, what does their model do when it doesn't know? Google's Aletheia solved 6 of 10 research-level math proofs and said "no solution found" on the other four. OpenAI's internal attempt hallucinated one and had to walk it back. Calibrated doubt is the feature that matters now. If the vendor can't answer the question clearly, keep shopping.

— Harry and Anthony

Sources: