Bitter lessons building AI products (hex.tech)

A team retrospective frames a product-focused version of “the Bitter Lesson”: in AI product development, general methods that scale with compute and data outperform clever, brittle engineering. The authors recount months-long efforts to paper over model deficiencies—like their 2023 Notebook Agent that used template-driven cell generation to avoid “doom loops” and a two-step pipeline (reasoning models + a small fine-tuned model) to emit complex vega-spec–like JSON for visualizations. Those hacks produced impressive demos but failed in messy, real-world use. When model capabilities shifted (Sonnet 3.5/3.7, and later Cursor Agent Mode and Claude Code in 2025), agentic tool-calling approaches proved an order of magnitude more effective, enabling features that previously took months to build to ship in weeks. The practical takeaway for AI/ML teams is operational: align roadmaps to current model capabilities, not to workarounds for deficiencies; validate with early beta users (not polished demos); kill projects quickly when you’re hacking around intelligence; and retry past ideas regularly after major model releases. Technically, this means favoring agentic interfaces and tool-calling over bespoke multi-shot pipelines and RAG/k-shot patterns, and designing products to get better as models improve rather than relying on temporary engineering fixes.