About

I’m Curtis Huebner, I’m currently the founder of Septagram, an AI agent company.

I’ve previously worked as EleutherAI’s head of alignment. During my time there I worked on a variety of projects, including the Minetester RL env, a jax pretraining framework for training video models, and the GPT-NeoX pytorch LLM pretraining framework. I also developed a new MCMC sampler that behaves like the Adam optimizer in the limit and written self contained gists for both VAE training, and diffusion models that can be used as reference implementations for larger training pipelines.

Before that, I’ve worked as a software engineer at Bloomberg and participated in ML and bioinformatics research for cancer research and immunology. I also did a few undergraduate projects like this self balancing unicycle robot  and this pulse-oximetry investigation, this competition robot.

Additional links:
Learning with EBMs

Double Descent, Optimization, and Integration