I Know Kung Fu

11 — the seconds the average doctor listens before interrupting you. In 1984 it was eighteen, so forty years of progress made us worse at the one thing we swore machines could never do. The thing American medicine ran out of is the thing the machine has in unlimited supply: time, patience, a willingness to hear the same question twice without sighing. The catch is that it only hands the good stuff over if you know what to ask it for. Most people don't. That's the whole game now.

Name the register before you open the window. Every AI conversation is a dial between two settings: solve the problem in front of me, or think with me while I solve it myself. The people who get useful answers decide which one they need before they type. Pick one consciously for your next three prompts and watch the quality jump. The model was never the bottleneck — knowing what you came for is.

Go find your operator. You already employ them. The most valuable person in the AI era isn't the 100X engineer who can build. It's the operator who knows what to ask for and architects the system so the builders run free. Hand your highest-context person — often a product manager sitting on the customer, the roadmap, and the business at once — a stack of agents and a vague, important problem today. Watch what they request. That instinct is worth more than raw output, and no test you own can measure it.

Audit one workflow for where you shipped case-report energy. A Columbia ER doctor tried the AI built for doctors and found it cold and clinical; she got her actual care from the general chatbot that asked about her life. Your company has the same problem hiding in a tool somebody bought because it was "serious." Pick one customer-facing AI touchpoint and ask whether it answers like a case report or like a person. Fix the worst one.

Keeping up with AI is hard. We know — we do this daily. If any of these struck a chord, or you just want an Outsider perspective on where the humanity belongs in your AI stack, we'd love to hear from you.

Outsider Labs Advisory →

Thursday, May 28

1. What people actually use AI for is not what the market is funding. The latest usage research puts therapy and companionship at the top of the list, with personal and emotional support nearly doubling year over year, from 17% to 31%. Behavioral logs tell a quieter version — most messages are practical — but the non-work share has climbed to 73%. The money is pouring into cold execution. The demand is migrating to warmth. (Full analysis below.)

2. Cognition raised $1 billion at a $26 billion valuation. Run-rate revenue is reportedly around $492 million, which pencils to roughly 53 times revenue for an AI coding shop. That's not a bet on a product. It's a bet that coding agents are leaving the demo stage for good. Coding stays the cleanest commercial story in AI for the same reason it leaves you a little cold: code gets executed, not understood.

3. GPT-5.5 hit 70% on long-horizon software engineering. The DeepSWE benchmark — 113 tasks across 91 repos and five languages — had the new model clearing 70%, up from 56% for the prior version. Real progress on the kind of multi-step work that used to break agents. File it under "the cold end of the wire keeps getting sharper," which is exactly why the warm end is where the scarcity is moving.

4. One character broke millions of AI agents. Researchers disclosed "BadHost," a critical flaw in the Starlette web framework that a huge share of deployed agents quietly depend on. The scariest AI security story of the week wasn't a rogue model — it was a supply-chain crack in a library nobody thinks about. Audit your agent dependency tree before someone audits it for you.

5. Robinhood gave agents a trading desk and a credit card. The launch of agentic trading plus an agentic credit card is being called a wake-up call for banks, and it should be. When your customer is a piece of software making purchases and trades, the question every financial incumbent has to answer is who owns the relationship. Robinhood just forced it.

Which Program Do You Need?

A Columbia emergency physician got routine bloodwork back a little high, asked her doctor for a phone call, and was told to book another appointment. So she did what everyone does now and typed her labs into ChatGPT. She wrote it up for the Times, and the surprise wasn't that the bot was smart. It's which tool she trusted. The AI built for doctors — the one optimized for clinical rigor — felt cold and treated her like a case report. The general chatbot asked about her life, tailored what she could realistically change, never sighed when she repeated herself, and kept cheering her on. Her numbers improved, and she credits the back-and-forth she never got from a person.

That's the whole story of AI right now, told from inside an exam room. The machine doesn't really sell intelligence. It sells range — calculator or confessor, sharp resident or patient listener, depending on what you walk in needing. I lived both ends of it this week: one conversation a cold medical decision where I wanted orders carried out, another a mirror where I needed pushback for hours. Same window. I turned the dial.

Here's where it gets useful for anyone running a business. The latest data says people reach for AI most often to get something done, but value it most when it sits with them, and the warm use is the fastest-growing thing in the numbers. Meanwhile the capital is flooding the cold end: coding agents at 53 times revenue, because code gets executed and never needs to feel heard. The market built its triumph at the cold end of the wire while human demand climbed toward the warm one. And the warm end — the personality nine hundred million people confide in weekly — is being designed by accident. Sam Altman has said as much; he's calling monks and clinical psychologists because the people who understand humans are the ones he's missing.

So the scarce skill isn't building. It's knowing which program to load. Yesterday we said AI 10X's everyone and 100X's your best people, and the old Combine can't find them. True, but we've been scouting the wrong position. The 100X engineer is the kung fu program — raw, uploadable, valuable, and at least legible. The operator who knows it's kung fu this morning and a helicopter by afternoon is rarer and worth more, because their value is second-order: they multiply the people who do the work. It's why the early stories about AI rearranging the org chart keep landing on product managers, not engineers — the PM was always sitting on the most context and just lacked a way to test an idea without waiting two sprints. AI took that tax to zero. The flip side: when everyone can prototype, the manager who only ran standups has nothing left to sell. AI is the great equalizer of execution, which makes it the great exposer of judgment.

Which leaves you one job the model will never do for itself. The cold programming is racing to zero. Adding the judgment, the context, the humanity back into the system — deciding when your agent answers like a resident and when like someone who has time — that's the part you architect. The program won't add the humanity for you. That's the operator's entire job, and it's yours.

At COAI today: the full Signal/Noise — the doctor, the Dojo, and the operator who beats the 100X engineer — is live at getcoai.com.

— Harry and Anthony

Sources: