Andrew Weir

BBC News recently published an article where Jack Clark, a co-founder of Anthropic, says that Claude is 80% self-written and, "Getting to 100% is possible within two years."

At first, that's scary? Impressive? But given the quality of the shit AI produces, I'd be more concerned about the state of Claude in two years if it went down this path. There are plenty of reputable studies that shows the quality of generated code is lacklustre, and that LLM coding models can't handle the larger context of the repositories they're generating in. The current (public) state of play is that LLMs are good at surgical strikes in code, but not at maintaining the overall quality at large.

If the anecdotes are anything to go by, it's probably best not let Claude code itself yet.