Contra DSPy and GEPA (benanderson.work)

A recent post critiques two AI toolkits, DSPy and GEPA, and their applicability to modular AI programming, particularly in LLM workflows. DSPy, created by Omar Khattab, aims to enhance the reliability of AI systems by breaking down tasks into modular components, allowing for optimization independent of specific implementations. GEPA, a Genetic Pareto optimizer for LLM programs, enhances these workflows using a genetic prompt-evolution strategy that maintains a diverse selection of prompt candidates. The author, however, shares frustrations from attempting to apply GEPA to the multi-turn agentic search task, arguing that treating LLM workflows as modular can be problematic. The significance of this discussion lies in the exploration of the limitations of modularity in AI systems, particularly for agentic tasks. While GEPA showed improvements over traditional reinforcement learning approaches in sample efficiency, the author concludes that the inherent nature of agentic workflows—where all prompts are relevant at every step—contradicts the modular philosophy these toolkits espouse. This highlights a deeper challenge for AI practitioners: how to integrate more flexible, agent-driven methodologies within rigid framework structures, potentially prompting future enhancements in modular AI design.