Our Work
We build AI agents: coding systems that reason about problems and generate software to solve them.
We want to make it possible for every person to build bespoke software, so that our computers can help us do more of what matters to each of us.
Highlights
Training a 70B model from scratch: open-source tools, evaluation datasets, and learnings
Earlier this year, we pre-trained a 70B-parameter model and fine-tuned it on a range of multiple-choice reasoning benchmarks. On these…
Read more
Imbue raises $200M to build AI systems that can reason and code
We’re excited to announce our latest funding round, a $200M Series B at a valuation of over $1 billion, with participation from Astera…
Read moreOur Approach to Agents
Personal software, broadly distributed
We believe that the future will be mediated by increasingly powerful software, and we want people to be able to participate more fully in it. We’re building a better way to create and edit software — so that every person has more agency in the digital world.
- Product: We build agents that allow more people to harness AI capabilities to build software. To do this, we create capabilities and interfaces that allow people to work at a higher conceptual level than code. Over time, we hope this will enable more people of varying technical ability to create and edit software.
- Research: To support our goal of making a better way to create and edit software, our research focuses on improving our ability to verify whether generated code is correct and trustable. There are two main thrusts toward that goal: (1) creating high-quality data for evaluations and for training, and (2) understanding and improving model performance. We invest heavily in internal metrics and datasets that allow us to precisely measure the effects of new research ideas on our product development, and continuously develop evaluation tasks to improve agent performance.
- Infrastructure and developer tooling: We invest in building tools to speed up our iteration loop, from agent debugging interfaces to hyperparameter optimizers like CARBS. We've set up multi-thousand GPU clusters from the ground up, and even open-sourced some tools we developed to deploy and maintain such clusters. This includes projects like optimizing latency for custom models used by our product and making it easy to safely and efficiently evaluate generated code.
- Policy: We care about empowering humans in an age of increasingly powerful machines; to do this, we must expand individual possibilities. However, technological power is not broadly accessible or distributed today. Our policy efforts seek to shape our laws and societal conditions to make the protection of our individual liberties the default path for the digital future.
