Things I've learned in my 7 Years Implementing AI (www.jampa.dev)

A veteran implementer (not a PhD model researcher) distills seven years of productionizing AI into a clear thesis: AI is rarely a standalone product — it’s a tool or feature that should run beneath the surface, not a ChatGPT-like app or a single OpenAI API “✨” button. Many recent launches would be better served by straightforward techniques (e.g., vector semantic search) and by embedding ML into core flows — think demand forecasting, ranking, fraud detection — rather than slapping “AI” onto interfaces. A concrete win: an accessibility project that took a year and academic teams to reach 55% suggestion accuracy was rapidly outperformed (82%) by hacking with ChatGPT 3.5 over a weekend, showing LLMs can trivialize problems that once required heavy research. Practical implications: expect a boom in internal tools (Claude, Cursor, simple KNNs) more than a startup explosion, because non-coding constraints still limit new ventures. LLM capability may be approaching an S-curve plateau — improvements will be incremental, but current models are “good enough” for many tasks. Don’t mystify AI: use existing libraries (Scikit-learn) and code assistants (Claude Code) to get pragmatic gains. Crucially, senior engineers remain essential — LLMs aren’t near 99% accuracy, they can add complexity where simplicity is needed, and overreliance risks stunting junior engineers’ development. The author views the current bubble as productive, accelerating useful tooling and adoption.