We are developing and testing new model architectures & network primitives to achieve massively multimodal, “Generalist” systems capable of doing useful things across images, natural language, and the physical and digital world.
A large scale, generalist transformer model capable of tasks across Images, Text, and Control Domains
We're building new kinds of action models capable of using digital tools.
We’re designing new kinds of Agent Architectures, pulling from ideas in Cognitive Science and Reinforcement Learning to equip AI Agents with memory, modelling, and other key capabilities, including the use of digital tools and software.
We're building new kinds of action models capable of using digital tools.
Survey and Perspective Paper on Opportunities in AI Agents