Imbue, a startup developing language models optimized for reasoning, today announced that it has closed a $200 million funding round.

The Series B investment values Imbue at $1 billion. It included contributions from Nvidia Corp. and Astera Institute, a nonprofit that backs promising research initiatives. Cruise LLC Chief Executive Officer Kyle Vogt, Notion Labs Inc. co-founder Simon Last and a number of other backers participated as well.

Imbue says it’s developing large language models “tailor-made for reasoning.” As the company detailed: “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.”

The company’s models have more than 100 billion parameters, the settings that determine how a neural network processes data. Llama-2, an advanced language model released by Meta Platforms Inc. in July, can be deployed with up to 70 billion parameters.

Imbue trains models using datasets it created specifically to foster reasoning skills. According to the company, training is carried out using a server cluster that includes about 10,000 of Nvidia Corp.’s flagship H100 graphics processing units. The H100 can run large language models up to 30 times faster than the chipmaker’s previous fastest GPU.

Besides infrastructure, Imbue has also invested in custom development tools to support its engineering work. One such tool is an application called CARBS that the company’s researchers detailed in June. It eases the task of optimizing a neural network’s hyperparameters, settings that influence how quickly and accurately data is processed.

Imbue’s research also extends to the theoretical foundations of deep learning. According to the company, one focus of its research effort in this area is self-supervised learning.

Historically, developers trained artificial intelligence models using labeled datasets. The files in such datasets are enriched with contextual information designed to help neural networks learn more efficiently. A self-supervised AI model, in contrast, can be developed using unlabeled data that doesn’t contain such contextual information.

Imbue uses its large language models to power a number of automation applications it refers to as agents. According to the company, most of those agents are designed to automate coding tasks. Some of them are used by Imbue engineers to support their day-to-day work.

“Because programming problems are so objective — the code either passes the tests or doesn’t — such problems form a relatively ideal testbed for more generalized reasoning abilities, allowing us to understand if we are making meaningful improvements in our underlying systems,” Imbue detailed in today’s funding announcement.

The company stated that it’s currently not working to commercialize its AI-powered coding agents. In the longer term, however, Imbue does intend to make its technology accessible to the public. The plan is to harness large language models for not only coding tasks but also to “enable anyone to build robust, custom AI agents that put the productive power of AI at everyone’s fingertips.”

Image: Imbue