Eleven sentences about AI

Eleven things I currently believe about agentic AI, in numbered order, subject to change without notice.

• —The model is not the product. Most teams who think they're shipping a model are shipping a workflow.
• —Latency is a feature. A model that takes 12 seconds to think is a different product from one that takes 2.
• —Confidence is not the same as accuracy. We have not built the muscle for this distinction yet, but we will.
• —Prompts are a UX surface. Treat them like one — versioned, tested, owned by a designer.
• —Most 'AI products' are still mostly forms. The interesting ones are forms with much better defaults.
• —The right level to integrate AI is two abstractions below where you think it should be.
• —Evals > demos. A good demo proves nothing. A good eval proves the worst case.
• —Hallucinations are not bugs to fix. They are the trade-off that makes the model useful in the first place. Build for them.
• —The killer use cases are not chatbots. They are the silent ones — the autocomplete, the categoriser, the rewrite-this-paragraph that you stop noticing.
• —Open source eats the boring stuff. Closed keeps the frontier. Both are fine.
• —We are six months from now, again, again, again. Plan accordingly.

Most of these are probably wrong. The job is to find out which ones, fast.