The Unreasonable Effectiveness of Separating the Task from the Model

By declaring your task’s inputs and outputs without initially considering model capability, you create the space needed to figure out the model execution later. DSPy’s entire promise is that you should evaluate and execute your AI engineering at a level higher than a specific prompt template or a particular provider’s API shape: the Signature. However, models have evolved significantly over the last few years. How can the same input and output specifications still work in a world now filled with tools, RLMs, and Skills? By defining your task strictly through its inputs and outputs, the underlying implementation becomes completely flexible. You can experiment with different models, settings, weights, templating strategies, and output formats, all without touching your actual AI workflow. Consequently, you can leverage components built by others and focus entirely on your core AI task. In this talk we will present how dspy 3.5 makes it easier much easier. DSPy has its roots in prompt optimization, where we build efficient ways to conduct search and learning beneath the signature. In this talk we will give a preview of DSPy 4.0 where we use the fact that models have now passed a tipping point for two critical concepts we have always needed. First, we no longer need to limit the search space to a single instruction block per LLM call; models can now reliably write the code underneath a signature themselves—so they should. Second, traditional prompt optimization has always required a scalar metric, which is notoriously one of the hardest parts to get right. What if a DSPy program could learn directly from your interactions with users? Ultimately, all you care about is that the function you call respects the inputs and outputs of your signature. You can let the models figure out the rest. Speakers: Maxime Rivest — DSPy; Isaac Miller — DSPy, cmpnd.