Pieter Abbeel is a UC Berkeley professor and Director of the Berkeley Robot Learning Lab who became a Founding Investment Partner at AIX Ventures, giving him unusually deep technical and research credibility compared with most investors. As host of The Robot Brains Podcast, he regularly interviews leading founders and researchers such as Andrej Karpathy, Geoff Hinton, and John Schulman, giving him a structured way to build deep relationships with frontier AI talent that also complements his role as an AI investor.

• Focus on robotics foundation models that apply generative AI to the physical world. He frequently highlights his work with Covariant and their RFM-1 model, describing it as a system that allows robots to understand plain English instructions and video data to manipulate objects, effectively acting as a ChatGPT for physical tasks.
• Interest in simulation-to-reality transfer and the technical infrastructure needed to train physical AI. He shares open-source codebases for training humanoids and quadrupeds, noting specific instances where he was happily surprised that a policy learned entirely in a simulated environment successfully transferred to a real robot.
• Deep appreciation for academic rigor and fundamental research, particularly originating from UC Berkeley. He regularly promotes educational materials on complex topics like diffusion models and unsupervised learning, while celebrating former PhD students who have transitioned from his lab to launching technical startups like Ideogram.
• Enthusiasm for autonomous agents that can navigate digital environments. He actively recruits for and promotes tools like Nova Act, which enables agents to use web browsers and extract data from the internet, viewing this as a step toward building universal agents that function in both digital and physical spaces.
• Attention to extending the capabilities of Large World Models, specifically regarding context length and video processing. He points out advancements in models capable of handling over one million tokens of text and video without losing information, pushing past the limitations of short-sequence processing.