Liquid AI, a startup originating from MIT, is set to unveil innovative AI models based on a unique "liquid" neural network. This new approach aims to be more efficient, energy-conserving, and transparent than traditional systems currently used in chatbots, image generators, and facial recognition. The new models focus on fraud detection in financial transactions, self-driving vehicle control, and genetic data analysis. Liquid AI is licensing these technologies to external companies, with investments from firms like Samsung and Shopify, which are also trialing their developments. Co-founder and CEO Ramin Hasani drew inspiration from the C. elegans worm, notable for its fully mapped nervous system and surprisingly complex behaviors despite its limited neuron count. Hasani emphasizes that this once academic project is now a fully commercialized technology ready to provide enterprise value. In contrast to conventional neural networks that rely on static values, liquid neural networks are governed by equations that allow for ongoing learning post-training.

This architecture makes them more efficient and flexible, as they can analyze dynamic visual information effectively and afford transparency in understanding their decision-making processes. Liquid AI has shown that a compact liquid network can proficiently control simulated self-driving cars. Recent progress includes the development of new models, although details are currently confidential. In September, the company announced large language models from its network design, with one 40-billion-parameter model outperforming a 70-billion-parameter model from Meta on specific benchmarks. Prominent investors, like GitHub co-founder Tom Preston-Werner, recognize the significance of finding a new foundation model, particularly as transformer models face limitations. He stresses the urgency of enhancing AI efficiency to minimize energy consumption. However, Liquid AI's technology is ideally suited for certain tasks, especially those related to temporal data, necessitating custom code for broader applications. Additionally, convincing large enterprises to adopt this new AI design poses challenges. Hasani expresses confidence that the advantages—particularly in efficiency, transparency, and energy consumption—will outweigh the hurdles, helping to address various socio-technical issues associated with AI systems.