ODSC '25 Recap

Conference recap time! I always love putting these together, I feel like it helps me solidify my understanding of the material and gives me something to refer back to when I think of one of the talks I’ve been to.

Here are the recap slides.

Some highlights:

Pydantic AI!

I’m super excited to see how this library develops. I’m a big fan of the Pydantic library. It fills a crucial need, it’s easy to understand and their documentation is top-notch. I like the idea of applying this kind of structure-oriented approach to building agents

Seth Weidman’s talk on PPO and GRPO.

I thought the way he explained how RL training happens was clearer than I’ve seen in much of the other material I’ve read. I hope I captured the idea accurately here in the slide, but breaking out the “rollout” and “training” phases made this much more clear to me. Previously I had been under the impression it was more of a single action, prediction and loss computation, versus this “collection” of actions.

I didn’t get to go through the notebooks extensively, but it’s great to have this as a resource.

Arvind Narayanan on the future of AI agents

I thought this was an excellent, balanced view on what AI Agents are and are not. He was realistic; we need benchmarks based in reality and measurement done in a much more systematic (and independent) way. I don’t have his slides, but I think this article covers some of his views.

Interested in more recaps? Check these out!