Liquid AI, an MIT spin-off artificial intelligence startup and foundation model company, has reportedly closed a USD 250 million early-stage funding round, led by chipmaker Advanced Micro Devices (AMD), which is also its strategic partner, at a USD 2.3 billion valuation. Liquid AI develops Liquid Foundation Models (LFMs), which it claims are smaller, more efficient AI models for enterprise use, compared to traditional cloud-based offerings from OpenAI, AWS and Google Cloud. The startup is reportedly designing AI systems inspired by the structure of the tiny brain of a worm.
The startup previously raised USD 46.6 million in seed funding. The Series A round is being led by Advanced Micro Devices (AMD), with participation from OSS Capital, Duke Capital Partners and PagsGroup, among other investors, according to a Bloomberg report.
The strategic partnership with AMD will help the Cambridge, Massachusetts-based startup to optimise LFMs with AMD’s graphic, central and neural processing units, Reuters reported.
“We have been proving the technology in the last year, making sure that an alternative structure to transformers can be scaled,” said Ramin Hasani, co-founder and CEO, according to Bloomberg. “This funding will help get us to the next level.”
“We just fundamentally believe that there is a lot of innovation continuing in AI and driving models forward,” Mathew Hein, senior vice president and chief strategy officer of corporate development at AMD, reportedly said.
“All the enterprise customers we talked to are getting excited about the potential these models can unlock,” Hasani further added, according to the report.
Traditionally, developers have used neural networks, inspired by the human brain, to enable AI systems to learn through trial and error. Liquid AI, however, is adopting a unique approach. The startup is reportedly creating a novel technology called liquid foundation models, which leverage mathematical techniques derived from studying the structure of a worm’s brain.
“By focusing on an organism — specifically, a small nematode worm called Caenorhabditis elegans that’s just 1 millimetre long — with far fewer neurons than the human brain, Liquid AI said it’s able to build an AI system that is more flexible. Its systems also require less data and computing power than the conventional transformer-based models that underpin chatbots and other popular AI tools,” the company reportedly said.
In an October announcement, Liquid AI said it is developing a new generation of generative AI models that can achieve state-of-the-art performance at every scale while maintaining a significantly smaller memory footprint both during training and inference beyond what was possible before.
With reduced memory usage and near-constant inference speeds, LFMs are highly efficient for both training and deployment. Their on-device computing capabilities minimise reliance on cloud services, reducing costs and energy consumption, said Liquid AI.
“Our Liquid Foundation Models elevates the scaling laws for general-purpose AI systems at every scale for any data modality. Our first series of language LFMs achieve state-of-the-art performance at every scale, while maintaining a small on-device memory footprint,” said Ramin Hasani, CEO and co-founder of Liquid AI, in October as the company unveiled its first products built on Liquid Foundation Models (LFMs) at an MIT event on Wednesday, October 23, 2024.
Liquid AI’s LFMs are designed to handle complex tasks, including multi-step reasoning and long-context recall while being computationally efficient. According to the company, the first series of language LFMs, available in 1B, 3B, and 40B configurations, deliver robust performance and broad knowledge capacity across various domains, enabling them to solve tasks such as question answering, translation, composition, and summarisation, among other skills.
The funding will also help Liquid AI scale infrastructure and develop tailored LFMs for industries such as consumer electronics, biotechnology, telecommunications, financial services and e-commerce.
Every article you read here is the result of time, research, and effort. If you feel it adds value, you can support TelecomTalk.
Support Telecomtalk
