Contra DSPy and GEPA
a day ago
- #GEPA Optimization
- #DSPy Critique
- #AI Development
- The author expresses strong dislike for DSPy and GEPA, despite acknowledging the brilliance of their creators.
- DSPy is a toolkit designed to make AI programs more reliable by breaking them into modules and decoupling function from implementation.
- GEPA (Genetic Pareto) is an optimizer in DSPy that uses genetic algorithms and reflective feedback to improve LLM prompts.
- The author attempted to implement GEPA for a multi-turn agentic search task but found the process frustrating and unnatural.
- A key issue identified is that agentic tasks don't fit well into the modular, linear workflow expected by DSPy and GEPA.
- The author suggests that while DSPy and GEPA have value for deterministic tasks, they may not be suitable for dynamic, agentic workflows.
- The reflective prompt optimization and Pareto frontier aspects of GEPA could potentially be adapted for agentic tasks without the modular approach.
- The author considers revisiting GEPA after refining their approach, possibly focusing on interconnected prompts rather than modules.