Imbue develops foundation models aimed at enabling autonomous agents — AI systems that can perform sophisticated personal and work tasks such as meeting scheduling and complex data analysis without close supervision. CEO Kanjun Qiu says the key to large-scale, real-world AI is better reasoning capabilities. Imbue aims to build agents that people can actually use reliably and ultimately to "reinvent the personal computer so that it's actually truly personal." The company is part of a small group with sufficient capital to develop foundation models amid the generative AI surge. In September Imbue announced a $200 million Series B at a valuation of over $1 billion. The company will use new funding to buy compute resources and hire more employees, with a current focus on hiring a data executive and more product-focused engineers. Imbue, formerly Generally Intelligent, is an AI research lab focused on building models that can robustly reason and perform coding tasks. The company launched out of stealth last October and is researching the fundamentals of human intelligence to inform practical AI agents. It trains very large models (over 100 billion parameters) optimized on internal reasoning benchmarks and uses techniques that spend more compute during inference to arrive at robust conclusions. Training is conducted on a compute cluster co-designed with Nvidia containing 10,000 H100 GPUs. Imbue is also building internal ML tooling, debugging prototypes, and visual interfaces to support model development. The firm does not intend to productionize much of its current work immediately, instead using these efforts to improve future general-purpose systems and a platform for custom models. Financially, Imbue has raised $220 million to date following the new $200M Series B and says the funding will accelerate development of reasoning-and-code models. Generally Intelligent develops richly featured 3D research worlds (referred to as Avalon) and a structured battery of tasks to study how agents learn and generalize in physically grounded, open-ended settings. The lab trains many different agents powered by differing AI technologies to isolate which components produce human-relevant abilities, focusing on problems like object occlusion, persistence, scene understanding and internalizing physics. Its approach draws on developmental and evolutionary-inspired milestones to create increasingly complex challenges for agents. The company intends to scale Avalon into hundreds or thousands of tasks to evaluate and advance capabilities and safety. The team is roughly a dozen people and the board includes Neuralink founding-team member Tim Hanson. Financially, the company has secured an initial $20 million in funding, over $100 million in options, and an agreement guaranteeing $100 million via a drawdown setup to support long-term, capital-intensive research.

Total raised

$220M

Funding rounds

3

Latest round

Series B

Latest activity

Sep 2023