If you purchase something through links in our articles, we may receive a commission. This supports our journalism. Find out more. We also encourage you to subscribe to WIRED.
Artificial intelligence is now solving complex math problems, executing sophisticated reasoning, and even operating personal computers. Nevertheless, today’s algorithms could still learn valuable lessons from tiny marine worms.
Liquid AI, a startup emerging from MIT, is set to unveil various new AI models today that utilize an innovative type of “liquid” neural network. This novel approach promises to be more efficient, require less power, and offer greater transparency compared to the current systems used in chatbots, image generation, and facial recognition.
The new models from Liquid AI address issues such as fraud detection in financial transactions, management of self-driving vehicles, and analysis of genetic data. During an event at MIT, the company showcased these models, which it is licensing to external businesses. They have garnered investments from notable backers, including Samsung and Shopify, both of which are exploring applications of this technology.
“We’re expanding,” states Ramin Hasani, co-founder and CEO of Liquid AI, who co-created liquid networks as a graduate student at MIT. His research was inspired by the C. elegans, a tiny worm typically found in soil or decomposing plant matter. This worm stands out as one of the few organisms with a fully mapped nervous system and demonstrates surprisingly complex behaviors despite possessing only a few hundred neurons. “What started as a simple science project has evolved into a fully commercialized technology ready to provide value to enterprises,” Hasani notes.
In the realm of conventional neural networks, each simulated neuron’s characteristics are determined by a fixed value or “weight” impacting its activation. Conversely, in liquid neural networks, neuron behavior is dictated by a predictive equation that evolves over time, solving a series of interconnected equations throughout the network’s operation. This innovative structure enhances the network’s efficiency and adaptability, enabling it to learn continuously post-training, unlike traditional neural networks. Moreover, liquid neural networks permit a level of examination and understanding of their processes that current models do not, as their functioning can be essentially reversed to analyze how specific outputs were achieved.
In 2020, researchers demonstrated that a liquid neural network consisting of only 19 neurons and 253 synapses—an impressively small configuration—was able to control a simulated self-driving car. While standard neural networks examine visual inputs only at fixed points, the liquid network adeptly captures the dynamic nature of visual information over time. By 2022, the founders of Liquid AI had discovered a shortcut that rendered the mathematical computations required for liquid neural networks practically manageable.
Building on this foundational work, Liquid AI claims to have developed new models that it has yet to disclose. This September, the company unveiled several large language models leveraging its network architecture. A variant of its language model featuring 40 billion parameters proved to outperform the 70-billion-parameter model of Meta’s Llama 3.1 on a frequently used evaluation benchmark known as MMLU-Pro, according to the startup.
“The benchmark results for their SLMs appear very encouraging,” remarks Sébastien Bubeck, a researcher at OpenAI who investigates how the structure and training of AI models influence their abilities.
“Discovering a new kind of foundation model isn’t something that occurs frequently,” comments Tom Preston-Werner, a cofounder of GitHub and an early investor in Liquid AI. He notes that the transformer models that support large language models and various AI systems are beginning to reveal their limitations. Preston-Werner emphasizes that enhancing AI efficiency should be a primary focus for everyone. “We should strive to minimize our reliance on coal plants for as long as possible,” he asserts.
A limitation of Liquid AI’s methodology is that its networks are particularly tailored to specific tasks, especially those involving temporal data. Adapting the technology for other data types necessitates custom coding. Moreover, a significant hurdle will be convincing large companies to implement critical projects based on an entirely new AI architecture.
Hasani states that the current objective is to prove that the advantages—encompassing efficiency, transparency, and reduced energy costs—outweigh the obstacles. “We are entering a phase where these models can address many socio-technical challenges posed by AI systems,” he explains.