The Instant App Is Coming For Notion

CerebrasCoder generates a working app while you're still deciding if you spelled the prompt right.

Not "fast for AI." Actually fast — the kind of fast where the latency you're waiting on is your own reading speed, not the model. Go try it. Watch the thing appear. It does something to your brain that a hundred blog posts about "the pace of AI progress" cannot.

The current version of this is fine. A weather widget. A to-do list. A calculator that works. Small, self-contained things where "instant" is a neat party trick and the stakes of being wrong are zero.

Here's the thing nobody is sitting with long enough: that speed is about to collide with a much better model.

Not Llama running on Cerebras. O1. Sonnet. Something that can actually reason about a complex product with state and users and edge cases — running at that same speed. That is a very specific and very near event. The physics are all already in place. Someone just has to plug the cables together.

And the thing I keep thinking about is what the output looks like when you swap the weather widget for Notion.

Not Notion the company. Notion the category — the "I need a place to put things and share them and structure them" problem that something like forty million people have decided requires a dedicated piece of software. A real app. Databases, views, permissions, collaboration, history. The thing you currently pay a subscription for and spend an afternoon configuring.

Ask for that. Get it in twelve seconds.

There's a version of this that's even stranger — the parallel paths thing. The obvious move once inference is this cheap and this fast is to not run one generation, but a hundred. Write the same spec to a thousand agents, let them all code in parallel, run the outputs, evaluate, return the best one. The Genesis physics sim did this for physical simulation — 430,000 times faster than real-time — and what fell out of that recursive self-improvement loop was something that would have taken years to hand-tune. You don't run one draft. You run the population.

Applied to software: you describe Notion, you get back the ten best implementations generated and battle-tested in the time it would have taken Cursor to index your files.

Windsurf and Cursor chugging through your filesystem overnight feels fine right now because it's better than what came before. A few months from now it's going to feel like watching someone manually compile assembly. The comparison class shifts. It always does.

The thing worth staring at directly is the discontinuity — not the slope of the curve but the specific moment when "instant app" stops meaning a weather widget and starts meaning something that would currently require a Series A and eighteen months of sprint planning. That moment is not a metaphor. It has an actual calendar date. It is probably within the next twelve months.

I don't know what happens to software companies when the build time goes to zero. I don't think anyone does. But I'm pretty sure the answer isn't "nothing much changes."