Anthropic’s Claude has reportedly gotten much worse at generating code recently. That’s according to a head member of the AI Group at AMD.
Last week, user StellarAccident took to the Claude Code GitHub to vent their frustrations at Claude’s AI-generated code. This account was soon discovered to be tied to Stellar Laurenzo, the senior director of AI at AMD, who linked to that GitHub account for a post on LinkedIn (via The Register) three weeks prior.
In their GitHub complaint, they state that “Claude has regressed to the point it cannot be trusted to perform complex engineering.”
Laurenzo claims Claude performed better in January, and says it now ignores instructions, suggests fixes that are incorrect, directly contradicts requests and claims completion despite, well, not completing the task. Laurenzo is using the Opus model and claims the result they get is replicable with the same prompt.
Laurenzo says, “We have a very consistent, high complexity work environment and data mined months of logs to understand why—essentially—starting in February, we have noticed a degradation performing complex engineering tasks.” She goes on to say, “Claude has been good to us, and we are leaving this in the hopes that Anthropic can address these concerns.”
What follows this is a report on what has changed since January and how to improve it, which is, rather ironically, generated by Claude. It argues that Claude shows its thinking less, and it ‘thinks’ less deeply now, often editing code before fully ‘reading it’.
Users have been complaining about Claude’s February update for some time now. Over on the Claude Code Subreddit, one user says, “I can no longer in good conscience recommend Claude Code to clients.”
Hellbound and Down 