People say Ching-an and I are indistinguishable…is that true 🤣

Come check us out near the Tesla Booth in West Exhibition Hall A 3-5pm! Come and claim your mug 🤣 we have an identity crisis — people keep thinking we are from IBM for some reason…

Once you build an agent with Trace, you can use ANY LLM optimizer you want. With the release of Trace 0.1.3, we introduce TextGrad (github.com/microsoft/Tr...) as an optimizer for the RL agent, along with OPRO and OptoPrime.

What enables Trace to be an RL-style agentic library? We use **Generative Optimization** techniques (LLM as an optimizer) to derive an analog to RL's policy gradient algorithm. The agent makes a move, receives feedback/reward, and updates its parameters.

In Trace, you define an Agent with declarative Python functions using Trace primitives. Trace provides flexible ways to mark what you want to change -- for example, we mark two prompts and two functions below as trainable.

This talk is just fascinating — “o1 has an effective way to scale compute at inference time” — but you just can’t tell us what it exactly is 🤣

Noam Brown giving a talk on o1 at Stanford right now 🔥