Stability AI has introduced its latest language model, Stable LM 3B, designed to run on portable digital devices including handhelds, laptops, and upcoming headsets. Unlike industry-standard models like Llama 2 or Falcon, Stable LM 3B operates with 3 billion parameters, requiring fewer resources and offering reduced operating costs.
Emphasis on Sustainability
The model's design prioritizes reduced power consumption, aligning with recent sustainability trends in the tech sector. Notably, Apple has been linked to Stability AI through rumors of potential acquisition discussions. Reports suggest that Emad Mostaque, Stability AI's founder, has engaged in talks with OTOY, a firm that provides decentralized GPU-based rendering solutions for Apple.
Potential Impact on Meta AI
The emergence of Stability AI's offerings could influence the choices of companies currently using OpenAI's proprietary models. Many have shifted to Llama 2, adapting it with their data. Stable LM 3B shares architectural similarities with Llama's decoder-only transformer architecture and utilizes the Falcon-refined web and RedPajama dataset, an open-source repository for Llama. Mark Zuckerberg, in a conversation with Lex Fridman, hinted at the possibility of Meta moving towards a closed-source model in the future.
Falcon AI Language Model
Stable LM 3B is also competing with the new Falcon AI model. Last month, the Technology Innovation Institute (TII) in the United Arab Emirates (UAE) launched its new Falcon 180B large language model (LLM) as an open-access tool for both research and commercial applications.
In terms of natural language tasks, Falcon 180B achieves state-of-the-art results, topping the Hugging Face leaderboard for open-access models and competing with proprietary models like Google's PaLM-2.
Other Notable Large Language Models
Of course, the main leader in the LLM market is the OpenAI GPT-4 model. This is the AI that underpins both ChatGPT and Bing Chat. Just this week, OpenAI introduced GPT-4V, an enhanced version of the AI. The new functionalities are designed to allow a user to upload an image file and then pose questions about the image to the upgraded GPT-4, termed as GPT-4V – with V indicating ‘vision'.