we're so back

Mitchell Hashimoto (creator of Ghosty terminal) shared how OpenAI's Codex with 'extra high' reasoning solved a 6-month-old screen flicker bug in his Zig/GTK4 project that other AI models and humans couldn't crack. The AI's key move was reading GTK4 source code directly — something lower reasoning levels skipped. The author reflects on this as both an 'it's over' and 'we're so back' moment: AI excels at synthesizing large codebases to find root causes, but experienced engineers still add value through rigorous code review, manual cleanup, and opinionated corrections. The takeaway is a middle-ground workflow: use AI to get 80% of the way through hard research-heavy bugs, then apply human judgment for the rest.