Imbue raises $200M to build AI models that can ‘robustly reason’

Imbue, the AI research lab formerly known as Generally Intelligent, has raised $200 million in a Series B funding round that values the company at over $1 billion. Among those participating are the Astera Institute, Nvidia, Cruise CEO Kyle Vogt and Notion co-founder Simon Last.

The new tranche takes Imbue’s total raised to $220 million, placing it among the better-funded AI startups in recent months. It’s only slightly behind AI21 Labs ($283 million), the Tel Aviv-based firm developing a range of text-generating AI tools, as well as generative AI vendors like Cohere ($435 million) and Adept ($415 million).

“This latest funding will accelerate our development of AI systems that can reason and code, so they can help us accomplish larger goals in the world,” Imbue wrote in a blog post published this morning. “Our goal remains the same: to build practical AI agents that can accomplish larger goals and safely work for us in the real world.”

Imbue launched out of stealth last October with an ambitious goal: to research the fundamentals of human intelligence that machines currently lack. Its plan, as presented to ProWellTech back then, was to turn “fundamentals” into an array of tasks to be solved, and to design different AI models and test their ability to learn to solve these tasks in complex 3D worlds built by the Imbue team.

The company’s approach seems to have shifted somewhat since then. Rather than unleash AI on 3D worlds, Imbue says that it’s developing models it finds “internally useful” to start, including models that can code (a la GitHub Copilot and Amazon CodeWhisperer).

Plenty of models can code. But what sets Imbue’s apart are their ability to “robustly reason,” the company claims.

“We believe reasoning is the primary blocker to effective AI agents,” Imbue wrote in the blog post. “Robust reasoning is necessary for effective action. It involves the ability to deal with uncertainty, to know when to change our approach, to ask questions and gather new information, to play out scenarios and make decisions, to make and discard hypotheses and generally to deal with the complicated, hard-to-predict nature of the real world.”

Imbue also believes that code is an important use case beyond enabling its team to build AI apps at scale. In the blog post, the company makes the case that code can improve reasoning and is one of the more effective ways for models to take actions on a machine.

“An agent that writes a SQL query to pull information out of a table is much more likely to satisfy a user request than an agent that tries to assemble that same information without using any code,” the company wrote. “Moreover, training on code helps models learn to reason better; training without code seems to result in models that reason poorly.”

It’s a philosophy that’s not dissimilar to Adept’s, which aims to build AI that can automate any software process. Google DeepMind has also explored approaches for teaching AI to control computers, like having an AI observe keyboard and mouse commands from people completing “instruction-following” computer tasks such as booking a flight.

Imbue says that its models are “tailor-made” for reasoning in the sense that they’re trained on data to “reinforce good reasoning patterns,” and using techniques that spend “far more compute during inference time” to arrive at “robust conclusions and actions.”

Specifically, Imbue’s training “very large” models — models with over 100 billion parameters — optimized to perform well on its internal benchmarks for reasoning. (“Parameters” are the parts of a model learned from training data and essentially define the skill of the model on a problem, like generating text or code.) This training is being conducted on a compute cluster co-designed by Nvidia, containing 10,000 GPUs from Nvidia’s H100 series.

Imbue is also investing in building its own AI and machine learning tooling, like AI prototypes for debugging and visual interfaces on top of AI models. And it’s conducting research into understanding the learning process in large language models.

Imbue doesn’t intend to productionize much of what it’s working on at the moment. Rather, it sees these tools and models as a way to improve future, more general-purpose AI, and to establish the groundwork for a platform that people will be able to use to create their own custom models.

“When we build AI agents, we’re actually building computers that can understand our goals, communicate proactively and work for us in the background,” Imbue continued in the blog post. “Ultimately, we hope to release systems that enable anyone to build robust, custom AI agents that put the productive power of AI at everyone’s fingertips … This latest funding will accelerate our development of AI systems that can reason and code, so they can help us accomplish larger goals in the world.”