‘Meta buys hardware to train large language model that should match GPT-4’

Meta is in the process of purchasing hardware to train its own large language model. Until now, the company has used hardware from Microsoft Azure for this. The new language model should match GPT-4, writes The Wall Street Journal.

Nvidia H100

Meta purchases, among other things, Nvidia H100 chips, which are often used to train AI models. writes business newspaper The Wall Street Journal based on own sources. It should become an open source model again, just like Llama 2, the newspaper claims. Training should begin early next year.

The Facebook parent company has until now relied on hardware from Microsoft Azure to train models, but now no longer wants to be dependent on it. Microsoft itself is active in the AI ​​community by integrating large language models into much of its business software under the name Copilot and as a consumer product in search engine Bing. It uses a modified GPT model from OpenAI for this purpose. Meta’s Llama 2 was released this summer.

Looking back: explanation video about large language models