How 'Liquid Foundation Models' Redefine Large Language Models Using A Non-GPT Approach

"All roads lead to Rome." But each road is unique, and poses its own sets of challenges.

Since OpenAI introduced ChatGPT, pretty much all tech companies, large and small, race towards creating the best of all generative AIs. Prominent players, like OpenAI and Microsoft use GPT, for example. This proven way, pioneered by OpenAI uses 'generative pre-trained transformer' for generating human-like text.

But Liquid AI is trying a different approach.

Just like how Google uses its own Large Language Models and how Anthropic utilizes its own proprietary LLM, Liquid AI creates what's called the "Liquid Foundation Models," or LFMs.

Building its own model using a fundamentally new architecture, the company said that the AI manages to deliver impressive performance.

It excels in so many things, that the model is on a par with, or even superior to, some of the best LLMs out there.

The Boston-based startup was founded by a team of researchers from the Massachusetts Institute of Technology (MIT), including Ramin Hasani, Mathias Lechner, Alexander Amini, and Daniela Rus.

They are recognized as pioneers in the field of “Liquid Neural Networks,” a type of AI model that significantly differs from GPT.

Liquid Neural Networks, or LNN, is a novel class of AI that represents a departure from traditional neural network architectures, because it's designed to be more flexible, adaptive, and efficient in handling dynamic and complex environments than GPT.

Unlike conventional neural networks, which have fixed architectures, LNN boasts the ability to change their structure and connections over time.

This adaptability allows it to respond more effectively to varying input conditions and tasks.

LNN is also capable of continuous learning, meaning the AI can update its weights and configurations as new data becomes available. This makes it particularly well-suited for real-time applications where the data stream is constantly changing.

Whereas GPT requires a lot of neurons to perform computing tasks, LNNs can achieve the same performance with fewer. It does this by combining those neurons with innovative mathematical formulations, enabling it to do much more with less.

This is possible because LNN is able to maintain and utilize memory over time, enabling it to process sequences of data more effectively.

As a result, LNN is more resource-efficient.

These are real advantages, considering that GPT requires more computational resources for training and inference.

Despite having strong performance in text generation and understanding due to extensive training, GPT suffers from its fixed architecture and lack of continuous learning.

Long story short, LNN uses minimal system memory while delivering exceptional computing power.

According to Liquid AI, its LFMs represent a new generation of AI systems that are designed with both performance and efficiency in mind.

In all, the AI that is based on the principles from dynamical systems, numerical linear algebra, and signal processing, making them well-suited for managing different forms of sequential data, including text, audio, images, video, and signals.

Putting it to the test, LFM doesn't disappoint.

In a web page on its website, Liquid AI said that it's introducing its first set of generative AI models:

1. A dense 1.3B model, ideal for highly resource-constrained environments.
2. A dense 3.1B model, optimized for edge deployment.
3. A 40.3B Mixture of Experts (MoE) model, designed for tackling more complex tasks.

And here, Liquic AI said that LFMs offer a new best performance/size tradeoff in the 1B, 3B, and 12B (active parameters) categories.

Liquid AI aims to develop highly capable and efficient general-purpose models suitable for organizations of all sizes. To achieve this, it focuses on creating LFM-based AI systems that can operate effectively at all scales, from the network edge to enterprise-grade deployments.

"Architecture work cannot happen in a vacuum – our goal is to develop useful models that are competitive with the current best-in-class LLMs. In doing so, we hope to show that model performance isn’t just about scale – it’s also about innovation," the company said.