On Engineering AI Systems that Endure The Bitter Lesson
summary
Argues that prompts are a horrible abstraction for engineering because they entangle task definition, formatting, and inference strategy, and that premature low-level hardcoding breaks across model shifts. Presents DSPy, a three‑year framework built around signatures that decouples spec, code, and evals and supports adapters to hot‑swap models and apply RL or prompt‑optimization at the right abstraction. Claims this design endures weekly LLM changes and has stayed stable across multiple model generations in practice.
AI-written summary · about
chapters (8)
- 0:00Introduction and context
- 3:38The bitter lesson explained
- 7:50Premature optimization in AI
- 10:59Tight coupling in ML systems
- 12:43Prompts as bad abstractions
- 14:49Separation of concerns in AI
- 17:19Decoupling from swappable pieces
- 18:24Summary and takeaways
related talks
31mDSPy: The End of Prompt Engineering as We Know It
01h 13mDSPy: The End of Prompt Engineering
50mFrom Prompt Hacking to Architectural Determinism: Engineering Reliable GenAI Systems
30mIs Prompt Engineering Dead? How Auto-Optimization is Changing the Game
30mIs Prompt Engineering Dead?
22mNavigating the Evolution of Prompt Engineering
01h 39mMastering LLMs: From Prompt Engineering to Agentic AI
11mThe Unreasonable Effectiveness of Prompt Learning