Liquid AI Inc., a startup developing artificial intelligence models based on a so-called liquid neural network design, today announced that it has raised $37.6 million in seed funding.

OSS Capital and PagsGroup led the investment. They were joined by Samsung Electronics Co. Ltd.’s Next fund, WordPress developer Automattic Inc., Red Hat co-founder Bob Young and more than a half-dozen other backers. TechCrunch reported that the investment closed at a $303 million valuation. 

Liquid AI is led by Chief Executive Officer Daniela Rus, a director at the MIT Computer Science and Artificial Intelligence Laboratory. Rus founded the company earlier this year with three other researchers from the lab. Liquid AI is working to commercialize its founders’ research into liquid neural networks, a new type of AI that can perform some tasks more reliably than traditional models using significantly less power.

AI models are comprised of relatively simple code snippets called artificial neurons. Each such code snippet performs a small portion of the task assigned to the AI model in which it’s running. The behavior of the individual neurons is determined by an equation, or a set of equations, that varies between neural networks.

Liquid neural networks like the ones Liquid AI is developing can modify the equations that underpin their neurons. Moreover, they can change how those neurons interact with one another. According to Liquid AI, such neural networks’ ability to modify their own architecture makes them more adaptable than traditional AI models.

AI models can often only carry out tasks that they were specifically trained to perform. For example, an autonomous driving system trained to operate under favorable weather conditions may struggle to operate reliably when it’s raining. Liquid neural networks can more easily adapt to new situations, which lowers the risk of processing errors.

The technology also has an edge in the efficiency department. Liquid neural networks can be implemented with significantly fewer neurons than traditional AI models and also require fewer parameters, configuration settings that determine how data is processed. This reduces the amount of infrastructure necessary to run them.

A secondary benefit of the fact liquid AI models feature relatively few neurons is that it’s simpler for researchers to understand how those neurons interact. As a result, it’s easier to map out how an AI model reached a given decision and whether that decision is correct.

According to TechCrunch, Liquid AI will use its newly announced funding round to build commercial foundation models. Additionally, it plans to launch a platform that will enable customers to develop their own liquid neural networks. The company, which currently has 12 employees, will hire eight more staffers in the next few months to support its go-to-market efforts. 

Image: Unsplash