Digital Event Horizon
Liquid AI announces its first series of generative AI models, dubbed Liquid Foundation Models (LFMs), designed to tackle complexities in sequential data processing. These innovative models demonstrate exceptional performance in various tasks, paving the way for advancements in artificial intelligence.
Liquid AI has unveiled its first series of generative AI models called Liquid Foundation Models (LFMs), designed to process sequential data.LFMs are built with computational units rooted in dynamical systems, signal processing, and numerical linear algebra, offering a smaller memory footprint than traditional LLMs.The first series of LFMs includes three models: 3B, 3.1B, and 40.3B Mixture of Experts (MoE) models for various tasks and applications.LFM's perform well in general and expert knowledge tasks, but struggle with zero-shot code tasks, precise numerical calculations, and specialized applications.The LFMs are available on various platforms, including Liquid Playground, Lambda, Perplexity Labs, and Cerebras Interface, optimized for NVIDIA, AMD, Qualcomm, Cerebra, and Apple hardware.
In a breakthrough announcement, Liquid AI, an AI startup spun out from MIT, has unveiled its first series of generative AI models, dubbed Liquid Foundation Models (LFMs). These innovative models are designed to tackle the complexities of sequential data processing, aiming to create intelligent and efficient systems that can process large amounts of multimodal data.
According to Liquid AI, LFM's are built with computational units rooted in the theory of dynamical systems, signal processing, and numerical linear algebra. This distinct architecture allows for a much smaller memory footprint compared to traditional Large Language Models (LLMs). Moreover, LFMs can efficiently compress inputs, enabling them to process longer sequences on the same hardware.
The first series of LFMs includes three models: a 3B model designed for resource-constrained environments, a 3.1B model ideal for edge deployments, and a 40.3B Mixture of Experts (MoE) model optimized for more complex tasks. These models are general-purpose and can be applied to various types of sequential data, such as video, audio, text, time series, and signals.
Liquid AI's LFMs demonstrate exceptional performance in tasks that require general and expert knowledge, mathematics, logical reasoning, and efficient long-context tasks. However, they still fall short in areas like zero-shot code tasks, precise numerical calculations, time-sensitive information, human preference optimization techniques, and specialized applications such as counting the number of 'r's in the word "strawberry".
Currently, Liquid AI's primary language is English, with secondary multilingual capabilities available in Spanish, French, German, Chinese, Arabic, Japanese, and Korean. The startup plans to take an open-science approach, openly publishing its research findings and methods to advance the AI field.
While the LFMs themselves will not be open-sourced, interested users can explore them on various platforms, including Liquid Playground, Lambda (Chat UI and API), Perplexity Labs, and Cerebras Interface. These models are optimized for NVIDIA, AMD, Qualcomm, Cerebra, and Apple hardware.
Liquid AI's pioneering work in generative AI showcases the potential of novel architectures to overcome current limitations. As the field continues to evolve, it is crucial to recognize the significance of this breakthrough and its implications for future research and development.
Related Information:
https://sdtimes.com/ai/mit-startup-liquid-ai-releases-its-first-series-of-generative-ai-models/
https://www.liquid.ai/liquid-foundation-models
Published: Wed Oct 16 08:50:58 2024 by llama3.2 3B Q4_K_M
