"Implementable" doesn't mean it can be installed. It means it survives contact with the real system.
Legacy HIS, fragmented data, physicians who have no margin for extra steps. If your tool needs ideal conditions to work, it's not implementable. It's a demo with a good deck.
Chamath: "Long horizon tasks are fundamentally still a joke."
Su nuevo proyecto 8090 apuesta a que el bottleneck de la IA no es el código.
Es el spec.
8 ideas contraintuitivas de la charla:
A16z says the AI job apocalypse is a fallacy: Excel was supposed to kill bookkeepers, instead it created the financial analyst.
1M bookkeepers lost. 1.5M higher-paid jobs gained.
David George, a16z
After 11 countries and 68M patients a year, the pattern is always the same: health AI doesn't fail because of the model.
It fails because the problem wasn't well-defined, the data was messier than expected, or clinical and tech teams were never truly aligned from day one.
Los datos que podrían cambiar una decisión clínica ya existen.
El problema es que nadie puede procesarlos a escala. NAHC entrevistó a nuestro CEO @zeajosedavid sobre cómo Arkangel AI lo resuelve en 11 países.
👉 https://t.co/FaPL4nRLQW
Meet Macondo and Gabo. Our two new digital employees.
Macondo handles evidence. Gabo handles product. They live in this little retro Mac on my desk, right next to Claudio.
Three digital teammates. One server. Zero meetings missed.
Zuckerberg built his wife a clock with no clock face.
She'd wake at 3am, check her phone, lie awake doing math: hours left, kids waking, day ruined.
He didn't build a sleep app. He took the number out of the room.
Find the broken input. Remove it. Ship that.
Most health AI tools are competing for attention. The ones that matter are competing for workflow.
Those are different games. One is won with a good demo. The other is won with a tool that changes what a physician does at 9am on Monday morning.
For 70 years we shipped screens.
Now the model is the interface.
Users describe what they want.
The model decides what to render, run, or do.
UI isn't your product.
Capabilities are.
8 ideas from Michael Grinich (WorkOS) ↓
AI is killing 14 million clinic visits a month.
Gallup × West Health just put a number on it.
25% of U.S. adults now use AI for health info.
14% of recent users skipped a provider visit because of it.
That's ~16% of all U.S. physician office visits — gone, invisible in claims
Spent the week at AACR.
The field is moving fast, AI for protein discovery, digitizing clinical records to find trial candidates, new strategies to cut the cost of developing cancer drugs.
Our paper put four medical AI agents through real clinical cases — independent judges, EBM criteria. Arkangel AI Deep: 92.9% satisfaction. ChatGPT: 97% dissatisfaction on real references.
A tool that gets used 3 times and abandoned isn't a clinical tool.
It's a demo with a good pitch.
The hard part isn't building it. It's making it matter.
Being right is not enough.
If people can't repeat your idea, it dies.
Use this format:
1) Maxim: one line they can quote
2) Story: one scene they can picture
3) Action: one step they can do today
No memory, no movement.