Setting another new SOTA on the Generalized AI Assistant (GAIA) benchmark with Trase Agent
We stayed at #1 on the GAIA leaderboard for several months before being bested by Google LangFun and H2O GPTe. Now, we've reclaimed our spot at #1!
We stayed at #1 on the GAIA leaderboard for several months before being bested by Google LangFun and H2O GPTe. Now, we've reclaimed our spot at #1!
Lately I've been working on agents at Trase Systems, and we recently set a new state-of-the-art on the GAIA benchmark using a modified version of Self Taught Reasoner (STaR). I thought I would share some of the things I learned along the way.
I've been working on helping ketamine therapy patients illustrate their experiences using stable diffusion. As the MidJourney Discord bot is generating images, it makes the wait time a little more enjoyable by showing you your images at the current step. I was surprised that there wasn't any paid API or simple code on GitHub that had this feature, so I built the frontend and backend using SDXL, Diffusers, WebSockets, React, and FastAPI and made the code open source.
My journey into open research with Stability AI MedARC, where people were shown images while in an fMRI machine and we were able to decode those images using contrastive learning and diffusion models. So basically mind reading. We set a new state-of-the-art and got accepted as a spotlight for the NeurIPS 2023 conference. I had a good time being the oldest person in the Discord channels.
Over the past few months I've trained several DreamBooth models for clients to inject new concepts into various base stable diffusion models. In each case, the client wanted to insert several new concepts into the model, as opposed to the typical case of inserting a single concept like a new face. In this post, I'll describe the process I used to train these models and the heuristics I've developed for hyperparameters.