We built Weightless AI. Here's what it actually taught us.
Anyone can build an AI demo now. Paste an API key, wire up a chat box, ship the screenshot. A demo answers a question on a good day. A product answers it for a real person, on their worst day, about their own body. The distance between those two things is where we spent almost all of our time.
Weightless AI is a copilot inside our GLP-1 platform. It reads your labs, your check-ins, and your goals, and answers the only question members actually ask: “okay — so what do I do next?”
The model was the easy part
We could swap the underlying model in an afternoon. It’s a commodity, and it’s getting cheaper and better while we sleep. Betting the product on which model is best this month is a great way to build nothing that lasts.
The hard parts had nothing to do with the model:
- Grounding beats cleverness. It’s retrieval-grounded over your own data, and it would rather say “I don’t have that yet” than invent a lab value. In a health context, a confident wrong answer is far worse than no answer.
- Voice is a feature. It talks like a coach, not a chatbot. That’s not prompt garnish — that’s our creative engine sitting inside an AI product, which almost nobody does.
- Trust is earned in the boring moments. The empty states, the “I’m not sure,” the moment it admits a limit. That’s where members decide whether to believe the confident stuff.
Why we don’t ship demos
Weightless AI is in beta with real members, learning in public and getting sharper every week. It has our name on it — literally — so it had to be something we’d stake that name on.
AI didn’t change our model. It rewarded it. The teams winning with AI right now aren’t the ones with the best prompt — they’re the ones who can build it and give it taste. We kept both under one roof on purpose. This is exactly why.