RE: https://macaw.social/@mergesort/117496154140902425 Fuzzing was definitely a big step forward in improving software robustness. Maybe AI is too but it strikes me as a ridiculously wasteful approach. You could do it more cheaply, with no model training step at all, using just an instrumented simulator. You could use just the AST of a program, and recursively check all paths of every branch point. It'd be a massive computation but I suspect it's still cheaper than LLMs, and the results would be deterministic.