The biggest problem with AI coding assistants is that they create a lot of “buggy code”.
So, what’s the solution?
Make AI learn to fix its own mess
Alibaba built SWE-smith: an automatic bug generator + fixer that:
- Creates real bugs in GitHub repos
- Verifies them via unit tests
- Has AI agents fix them step-by-step
- Produces before/after examples at scale
Basically, don’t just have AI models generate code, but also fine-tune them to fix the bugs.
What do you think? Would this solve the problem?