Meta, the parent company of Facebook and Instagram, is said to be working on a cutting-edge, open-source AI model to rival OpenAI’s most advanced systems, as reported exclusively by The Wall Street Journal.
According to insiders familiar with the matter, Meta is on a mission to develop an artificial intelligence (AI) model that surpasses the capabilities of OpenAI’s Llama 2 model, which they introduced earlier this year. The goal is to create an AI model that is “several times” more powerful than Llama 2.
In an unexpected move, Meta is planning to make this new AI system open-source, allowing other companies and developers to leverage it for tasks like advanced text generation, analysis, and various other applications.
To support the development of this high-level AI model, Meta is actively constructing the necessary data centers while increasing its holdings of Nvidia’s highly sought-after H100 semiconductor chips, known for their remarkable processing power.
While Llama was trained on a substantial dataset of 70 billion parameters, OpenAI has yet to disclose the exact parameters for their GPT-4 model. However, estimations suggest they could be in the range of 1.5 trillion.
According to sources, Meta anticipates commencing training for this large language model (LLM) in early 2024, with plans for a release sometime in the following year. It is expected to launch after Google’s forthcoming LLM, known as Gemini.
It’s worth noting that Microsoft, a key supporter of OpenAI, has also collaborated with Meta to facilitate the availability of Llama 2 on its Azure cloud computing platform. However, Meta intends to train its upcoming AI model on its proprietary infrastructure.
This development unfolds amid a competitive landscape where major tech companies and governments are racing to develop, deploy, and assert control over advanced AI systems. The United Kingdom government recently announced a $130 million investment in high-powered chips for AI system development, while China has introduced new legislation governing AI and witnessed the release of over 70 AI models by companies like Baidu.