The models are fine, our toolings and methods are shit.

A developer with over 13 years of experience expresses frustration with current AI coding tools, despite initially finding Copilot helpful for small-scope assistance and debugging. They note that as models improved, they fell into a "slop spiral" of defining, slopping, testing, and reporting bugs to LLMs, leading to features breaking and a feeling that something is missing in the tooling. The developer believes models like Qwen3.8 27b are fine, but current toolings and methods are inadequate for providing proper context, making manual coding preferable for important projects.

Unlike many discussions focusing on model capabilities, this developer argues that current AI models are sufficient, but the tooling and methods for integrating them into development workflows are inadequate.