Talks #8: DSPy i kryptografia post-kwantowa

Miejsce: Poznań, Allegro, ul. Wierzbięcice 1B, budynek D

Riccardo Belluzzo: Introducing DSPy - Programming—not prompting—Foundation Models

Abstract:

It’s 2024 and you are no longer a ML engineer, but a Prompt engineer instead. You miss the good old days, when you were writing deep learning programs in pytorch. In a world (temporary?) dominated by foundation models and prompt-based systems, the DSPy framework offers a fresh perspective: programming, not prompting, your foundation models. In this talk, I'm going to introduce you to the DSPy framework, recently developed by Stanford researchers. Together, we will deep dive into its fundamentals and we will learn how easy, yet elegant, it can be to write an ML "program" using LLMs. Among the most exciting features that we will learn: having an optimizer that can “train” (write) the prompt for a given task instead of you; having an llm-based metric forked in your evaluation loop capable of judging the quality of your program’s execution; and much more.