Stroustrup: AI code buggy, 'pessimized'

Stroustrup’s critique highlights a growing "verification tax" where the speed of AI generation is offset by the immense effort required to maintain engineering standards. AI often prioritizes "vibes" over performance and memory efficiency, which he calls "pessimizing" the code. Peer reviewing AI output is significantly harder than human diffs because LLM errors are often non-local and unpredictable. This shift from creative engineering to "babysitting" poor AI output is accelerating brain drain among senior developers, even as human abstraction remains superior for high-level software design and foundational infrastructure.