The end of fine-tuning: Why evals, context, and traces matter more

Fine-tuning’s role has split between frontier labs doing continuous reinforcement learning and most teams improving harness components like prompts, tools, and evals instead of model weights. Laurie Voss argues that the majority of AI teams will see better returns from harness engineering - optimizing the full system architecture around the model rather than the model itself. Only well-resourced companies like Cursor and Cognition continue intensive model training, while the rest should focus on better context, retrieval, evaluations, and tooling.