Liquid AI takes a radical approach
MIT spinoff raises $37.5m in seed funds
Amid rising competition in artificial intelligence spurred by ChatGPT, a startup formed by MIT researchers with a fundamentally new approach came out of stealth mode and announced its initial funding on Wednesday.
Liquid AI, which was founded in March, said it raised $37.5 million in seed funding from high-profile investors led by Boston Celtics co-owner Stephen Pagliuca and San Francisco venture capital firm OSS Capital. Other investors include Thermo Fisher chief executive Marc Casper, RedHat cofounder Bob Young, and Shopify cofounder Tobias Lütke.
Unlike popular apps ChatGPT, Bard, and Dall-E, which rely on complex generative AI software models with billions of parameters, Liquid AI’s neural-network models are much simpler and require significantly less computer power to train and operate, chief executive Ramin Hasani said in an interview. For example, industry leader OpenAI needed billions of dollars to pay for the computing power to develop the model underlying ChatGPT.
“You need a fraction of the cost of developing generative AI, and the carbon footprint is much lower,” Hasani said of Liquid AI. “You get the same capabilities with a much smaller representation.”
The models underlying OpenAI’s ChatGPT are built by training software on billions of words or millions of images found on the internet to create a set of fixed connections among billions of “virtual neurons.” Then, ChatGPT “writes” humanlike sentences by calculating the probabilities of words appearing after groups of words it analyzed in its training data, for example.
Liquid AI’s more dynamic AI software maintains adaptability and flexi
bility even after training, allowing it to use just a few dozen virtual neurons to control a robot or pilot a drone, the company said. The MIT researchers even had to crack a calculus challenge unsolved since 1907 to allow Liquid AI to address larger challenges.
The software can be used to steer a car along a road and autonomously pilot a drone to find an unknown target, Hasani demonstrated in a talk at the TEDxMIT conference in Cambridge a year ago. The AI software can also be used as a “universal tool” to create text, audio, and images, Hasani said.
It remains to be seen how the new approach will stack up against the prevailing models from OpenAI and rivals such as Anthropic. The company will start by going after corporate clients that want to build private AI apps for financial modeling, medical research, or other uses. For now, Liquid AI has no plans for producing apps for consumers like ChatGPT, Hasani said.
Pagliuca, who specialized in tech and life sciences investments during his three-plus decades with Boston private equity giant Bain Capital, is investing in Liquid AI through his family investment office, PagsGroup. He helped bring other investors on board, including Casper, a former Bain colleague, and Texas venture capitalist Jim Breyer. (Pagliuca recently hosted the Liquid AI team courtside at TD Garden on Nov. 10. The C’s ended up beating the Brooklyn Nets.)
“It’s one of the most impressive groups of scientists from MIT that I’ve seen,” Pagliuca said. “Ramin, he’s a rare individual that knows the tech inside out but also has a good strategic mind.”
Pagliuca said he was attracted to Liquid AI’s neural network approach because it’s more efficient than the server-intensive approach employed in ChatGPT, and thus offers a less expensive model and one with a potentially smaller impact on the environment.
“The neural network technology cuts out a lot of that ‘brute force redundancy,’” Pagliuca said. “It’s more like the human mind. … The fund-raising is to build the next model. From everything we look at so far, if it scales correctly, it can be a breakthrough technology.”
John Werner, a venture capitalist who teaches at MIT, reached out to Pagliuca a few months ago because he was impressed with Liquid AI’s technology and the implications it could have. It was important, Werner said, to involve investors from the Boston area.
“These types of inventions out of MIT are rare and they typically head out to Silicon Valley. I wanted to keep some local people interested,” Werner said. “I do think it could create a whole ecosystem of other companies.”
A decade ago, Hasani and Mathias Lechner, Liquid AI’s chief technical officer, were working on PhDs in computer science at the Vienna University of Technology when they became interested in research on mapping the neural network of a tiny roundworm. The worm’s simple brain has only about 300 neurons (humans have some 86 billion) interconnected via about 7,000 synapses.
“Looking at how this actual brain worked, we found the computational principles that enable this nervous system,” Hasani said. That in turn allowed them to develop a simple software neural network. The models continue learning from their environment in the field, unlike prior generative AI models, which reduces the problem of “hallucination,” or making up incorrect information, Hasani said.
MIT professor Daniela Rus, who runs the university’s Computer Science and Artificial Intelligence Laboratory, lured the pair to move to MIT in 2017, where they continued their research into what became known as liquid AI networks.
“We got poached,” Hasani said.
The MIT lab and Rus have been investigating new kinds of AI models to compete with the large language models from OpenAI and others. “So many of the models developed today are really huge,” she said in an interview over the summer. “And there is a kind of an unspoken understanding that we have to build bigger and bigger models. But we are asking, is this really true?”
In March, Hasani and Lechner formed the startup with Rus as a technical adviser and member of the board and Alexander Amini, another MIT postdoctoral researcher, who serves as chief scientific officer. The company has offices in Brookline and Palo Alto with 14 employees and said it will use the new funding to expand its staff and infrastructure.