Solving the Wrong Problem (www.ufried.com)

The author argues that while AI agents and LLMs are already impressive at generating working code—especially when humans “vibe code” by treating models as black boxes—we may be solving the wrong problem. LLM-based coding works because models predict the most likely next token using attention mechanisms trained on massive corpora, so reliable outputs come from having seen similar code fragments many times. That explains common behaviors: fluent but brittle outputs, intermittent hallucinations where the model interpolates beyond its data, and tricks like generating Python first then translating to Rust when the model’s Rust exposure is weak. The practical consequence is worrying: agents excel at re‑creating patterns that already exist in training data, effectively reinventing the wheel instead of raising abstraction levels or improving software engineering practices. For non-experts this accelerates prototypes, but production-grade software—secure, maintainable, evolvable—requires deep language- and ecosystem-specific expertise that LLMs don’t replace. The piece calls for reframing priorities: invest in higher-level libraries, tooling, education, and software-engineering standards rather than outsourcing development to brittle agents and thereby institutionalizing “good-enough” or crappy code as the new baseline.