Coding with OpenAI o1
OpenAI ● Covered by 5 sources
Cognition's CEO Scott Wu talked up how OpenAI's o1 model handles coding tasks. He says it reasons through problems more like a person than past AI models did.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Scott Wu doesn't just build AI coding tools for a living, he runs Cognition, the startup behind Devin, one of the more ambitious attempts at an autonomous software engineer. So when he weighs in on how OpenAI's o1 model approaches code, it's worth paying attention, if only because he's spent years watching where these systems break down.
His core point, as relayed by OpenAI, is that o1 makes coding decisions in a way that feels closer to how a human engineer actually thinks. That's a meaningful shift from earlier generations of coding assistants, which tended to pattern-match against training data and spit out plausible-looking snippets without much underlying reasoning. o1 was built around a different training approach, one that rewards the model for working through a problem step by step before committing to an answer, rather than just predicting the next token that sounds right.
For coding specifically, that distinction matters more than it might for casual chat. Writing software is rarely about producing a single correct line. It's about weighing tradeoffs, catching edge cases, and backtracking when an early assumption turns out wrong. Wu's framing suggests o1 handles that kind of iterative, self-correcting process more naturally than its predecessors, which is exactly the kind of behavior Cognition has been chasing with Devin.
OpenAI has been positioning o1 less as a chatbot upgrade and more as a reasoning engine, and enlisting someone building agentic coding products is a savvy way to make that case. Whether o1 actually closes the gap between 'looks like code' and 'is correct, maintainable code' is still something developers will have to test for themselves. But the endorsement signals where OpenAI thinks the real competitive battleground is right now: not chat quality, but whether these models can think like the engineers they're meant to assist.
My take — AI-written commentary, not fact-checked reporting
I'll believe the 'thinks like a human engineer' framing when I see benchmark numbers instead of a founder quote, and I'd note that Cognition has every incentive to hype o1 since Devin presumably runs on models like it. Reasoning-style training is a real and useful shift, but 'more human-like' is doing a lot of marketing work here for what's still, underneath, statistical prediction with extra steps.
Read more about this at: OpenAI