a day ago
- The author, a programmer, initially found AI tools like Claude and Claude Code impressive for generating code and hobby projects, feeling a drug-like productivity boost.
- After generating code, the author realized a lack of mental model for the software, making debugging and adding features difficult without first thoroughly understanding the code.
- The author prefers writing code to build mental models incrementally, and has reduced AI use over time—shifting from 'vibe coding' to active guidance, then to AI-assisted ideation while writing code manually.
- The author doubts AI's consistent usefulness, citing examples of overcomplicated designs or unnecessary components, and notes that insights from AI conversations could have been achieved without it.
- The author questions whether LLMs will ever have a clear 'killer app' and compares the hype to Facebook's Metaverse, suggesting that AI's promise of automating intellectual effort might be a fantasy.
- The article concludes with a rhetorical question: if in a decade, developers and organizations find AI not useful enough, similar to the Metaverse's failure to transform daily life.