Almost two years after OpenAI’s ChatGPT shocked the world, AI has become a multitrillion-dollar phenomenon in tech. It has also created growing alarm about how much energy it requires at a time when climate change is wreaking havoc. Liquid AI’s technology, inspired by the brain structure of a tiny roundworm, is unlike anything from OpenAI, Google, or their competitors — yet it has the potential to offer the same revolutionary applications while using a fraction of the electricity.
Governor Maura Healey also appeared on stage with Hasani. “It reminds me of the shots first fired in Concord and Lexington that gave birth to this great country,” she told the crowd to loud cheers and whoops. “We have been about revolution and innovation from the very beginning.”
While the tech may have originated with a humble worm, the AI hype inflated it much higher, as John Werner, a longtime Boston venture capitalist and adviser to the company, described the keynote “like the Beatles coming to America.”
If Beatlemania is a high bar for the humble roundworm, the stakes are even higher for Liquid AI.
The company has raised nearly $50 million but will likely need hundreds of millions more to perfect its AI apps. One video at the keynote bragged that Liquid AI was “making what, up until now, has been impossible, possible.”
Winning the AI race will ultimately require tens of billions — even one hundred billion — dollars, said Mikhail Parakhin, chief technology officer of ecommerce company Shopify and a former longtime Microsoft executive who is advising Liquid AI. For Liquid AI, “getting that would be a tall order at this point,” Parakhin said. “But they have a remarkable technology.”
In part, the Tuesday event was intended to demonstrate to the tech investors who could provide that money that Liquid AI’s promising ideas will actually work in the real world. At the keynote, the company showed off one of its chatbots running solely on a phone, with no access to the internet, conversing with a user about AI-themed Halloween costumes — suggesting a “Neural Network Necromancer.” Another model almost instantaneously analyzed financial data to find questionable cryptocurrency trends.
The three-hour nerdfest attracted a wide array of potential investors, entrepreneurs, and MIT people who cheered for Liquid AI like they were watching a new iPhone unveiling. Celtics co-owner Steve Pagliuca, who has invested in the company, sat in the audience wearing his brand-new championship ring (he even let Hasani try it on at the end of the presentations).
In joining the AI arms race, Liquid AI is seeking to shape the rollout of a technology that could have dramatic effects on humanity.
Less than two years after ChatGPT burst into public view, the biggest tech companies on the planet are racing to deploy even larger and more capable AI apps. But the large language models underlying the technology continue to have problems sorting truth from fiction and are using an astronomical amount of electricity. Data centers powering the AI revolution could use as much electricity as Argentina or Sweden by 2027, according to one study last year.
Liquid AI also represents a rare chance for Massachusetts to become a player in the fast-growing AI business instead of being left behind by Silicon Valley.
“The stakes could not be higher,” Yvonne Hao, Massachusetts secretary of economic development, noted at a Liquid AI event on Tuesday night. After the state’s leadership in life sciences, created with plenty of government aid, “AI is the next frontier we in Massachusetts will lead,” she said, wearing a Liquid AI baseball cap.
Liquid AI’s models start out much like their better-known cousins, large language models, by analyzing giant amounts of data to learn about the relationships between words, sounds, or other input. But once a Liquid AI model is trained, the resulting set of digital neurons are more flexible and capable than those in an LLM. And that means a much smaller model, requiring less computer memory and power to run, can match the capabilities of a larger LLM.
The smaller Liquid model can, for example, speak and understand spoken language and offer detailed responses while running entirely on a smartphone, while a similarly capable LLM needs the power of cloud computing, which means the user’s questions or other data must be transferred to the cloud too — a potential invasion of privacy.
Liquid’s models could also address concerns about AI models getting out of control and going rogue, a prelude to a Terminator-style disaster scenario. While the complexity of an LLM can’t be easily unraveled after the fact to figure out why it gave a particular answer, Liquid models can be probed. At the demo, Liquid AI showed how one model determined that a picture contained a cat and an insect.
Liquid AI’s team is “very promising” and the technology looks “very cutting edge,” Elke Rundensteiner, who runs the data science program at Worcester Polytechnic Institute, said. “It’s going to be really exciting because we’re in a new era disrupting how these things are being built,” she said. Still, there are no guarantees that Liquid AI will displace OpenAI and other current leaders, she said. “We’ll see a lot of startups coming now,” Rundensteiner. “No doubt there will be a fight for which ones will survive.”
The entire effort began in Vienna when Hasani was searching for a topic for his PhD thesis and read about how scientists had mapped the entire nervous system of a worm known as C. elegans, including the connections between the worm’s 302 neurons via about a thousand synapses. A decade or so earlier, two different sets of researchers won the Nobel Prize in medicine for studies based on the worm’s genes.
Hasani and a friend, graduate student Mathias Lechner, now Liquid AI’s chief technology officer, saw the worm as a possible model for a new kind of computer neural network that could drive a car or fly a drone.
Hasani got his PhD from the Vienna University of Technology, where his adviser, Radu Grosu, was a childhood friend of Daniela Rus, director of MIT’s Computer Science and Artificial Intelligence Laboratory. In 2018, the pair of academics decided Hasani and Lechner should move to MIT to further their work by adding concepts from robotics.
By last year, the research project’s capabilities and possibilities were outgrowing the bounds of academia. So Hasani, Lechner, and Rus, along with Alexander Amini, an MIT PhD who had been working on the project, founded the company as yet another MIT spinoff.
The Kresge keynote was meant to demonstrate both how much better Liquid AI’s solutions are compared with large language models and how useful they can be solving big, real world problems, Jeffrey Li, the company’s head of product and business development, said. He pushed to keep the presentations focused on practical problems. “It’s almost a distraction to say we’re different within the thing that is so new,” he said.
For weeks, most of the company’s 40-person staff spent long nights working on the presentation and the company’s solutions. On a visit a few days before the keynote, the brand new foosball table looked almost unused in the office next to MIT that LiquidAI moved into about four months ago, while crushed cans of Red Bull and other energy drinks were strewn across desks.
Nearby, others were unpacking boxes of purple tee shirts and black hoodies emblazoned with LiquidAI’s triangular logo to hand out to employees and attendees at an upcoming dinner for VIPs the night before the keynote.
Almost all of the big AI startups raising billions of dollars from VCs are based in Silicon Valley, but Hasani and his crew say they are dedicated to proving they can succeed in Boston.
“That’s why we’re doing this classic style launch event that doesn’t happen often in Massachusetts,” the CEO said. “It’s very exciting to showcase that something profound can come out of here, especially in the software world. Boston has long standing in tough tech. But on software we had room to improve. And I feel like we are at that stage.”
Aaron Pressman can be reached at aaron.pressman@globe.com. Follow him @ampressman.