Back to Blog
January 19, 20263 min read

Both Headlines Are True

Anthropic says AI built Claude Cowork in under two weeks. DHH says AI can't match a junior programmer. The paradox resolves once you notice they describe different kinds of work.

Two headlines landed in the same week. Anthropic says its new Claude Cowork tool was built almost entirely by AI, in under two weeks. David Heinemeier Hansson — creator of Ruby on Rails, a man who has seen some codebases — says AI still can't match most junior programmers, which is why he mostly codes by hand.

Both are true, and the space between them is the most honest map of where AI development actually stands.

Cowork is real and impressive, but look at what kind of project it is: AI building tooling for AI-adjacent work, in a domain with predictable requirements, standardized interfaces, and few of the ambiguities that make software hard. When the problem is well-specified, current models are genuinely formidable. Two weeks is not hype; it's what happens when generation meets a problem shaped like its training.

Hansson is talking about the other 90 percent of software: requirements that contradict themselves, businesses that change their minds, systems whose real spec lives in the heads of people who left. His comparison to junior programmers is sharper than it first sounds, because the junior's superpower was never their code. It's that they learn. A junior who breaks staging in January doesn't break it the same way in June. A model, within your project, accumulates no scar tissue.

So the working answer, for now, is boring and true: use AI where the specification is strong, and keep humans where the ambiguity lives — and be honest with yourself about which is which, because misclassifying an ambiguous problem as a well-specified one is exactly how AI projects fail.

Ryan Castellucci added the unglamorous security angle this week: the AI risk that will actually get you isn't a malicious model, it's the mundane stuff — over-trusting generated code nobody fully understands, and the widening attack surface of every new API integration and data flow. The fix is equally mundane: treat AI tooling like any third-party dependency. Review it, audit it, limit what it can reach.

And in news that made no waves at all, .NET 10.0.1 arrived as a servicing update in Ubuntu 24.04 LTS. I mention it because weeks like this one make it easy to forget: while the paradox gets debated, someone is patching the runtime your production system actually runs on. Both kinds of work are the profession. Only one of them trends.

Sources