Cognition AI's Scott Wu says o1 models will let everyone build more

Scott Wu, the CEO of Cognition AI and a figure who’s been at the center of the “AI coder” conversation since Devin’s splashy debut, isn’t backing down from the big claims. In a recent clip, he pinpoints what he sees as the true unlock of models like OpenAI’s o1: not just writing code, but writing code that actually works in the real world. He draws a sharp line between generating a syntactically correct function and shipping reliable software. “It takes a lot of effort to build code that runs consistently and works very well,” Wu said.

The implication is that earlier models could handle the former—the textbook problems and isolated scripts—but stumbled on the latter. The o1 series, with its step-by-step reasoning chain, changes the calculus. It’s one thing for an LLM to ace a LeetCode challenge; it’s another entirely for it to architect a service that doesn’t collapse under edge cases a week later. Wu’s framing suggests we’re crossing that chasm, moving from AI as a clever autocomplete to an engineering partner that grasps robustness.

This isn’t just incremental progress in his view. It’s the difference between a tool for hobbyists and a tool for production. And that distinction is what fuels his headline-grabbing prediction. “I think the thing that’s really, really exciting now is every human is going to be able to build way more,” Wu added. He’s not talking about a marginal productivity boost. He’s sketching a world where the barrier between having an idea and seeing it function drops precipitously.

Of course, Wu has a vested interest in this narrative—his company is literally building AI software engineers. The cynic’s take is that every model release gets hyped as the one that finally delivers on the promise. But the specific focus on reliability over raw generation is a more interesting and grounded argument than the usual “vibe coding” evangelism. If o1 and its successors genuinely make the jump from writing plausible code to writing dependable systems, the bottleneck shifts from technical execution to something else entirely: knowing what’s actually worth building in the first place.