> Current AI tools are unable to keep large codebases in context, making it impossible to reason about system-wide impacts.<p>they have trouble even with small database, recently I was experimenting with generating a small script<p>Code-gen tools generated nice testcases from description (this was usable!).<p>But code they generated failed on exactly this testcases, feeding back crashes/test failures resulted in no improvements. I tried several (Claude, Gemini, ChatGPT, Mistral). None was able to do this.