Meta Platforms reportedly building open-source generative AI system to rival OpenAI’s ChatGPT

Facebook and Instagram parent Meta Platforms Inc. is looking to rival OpenAI LP in the generative artificial intelligence landscape with an alternative, but open-source model that’s likely to be as powerful as GPT-4.

Meta’s new model will be “several times more powerful” than the Llama 2 generative AI model that was released earlier this year, according to sources cited by the Wall Street Journal today. A team formed by Meta Chief Executive Mark Zuckerberg (pictured) is building the system, which would be made available to other companies that want to create AI services around text generation, image creation and analysis, though the exact nature of these capabilities may change.

Ultimately, Meta wants to accelerate the development of the next generation of generative AI models capable of generating more humanlike expressions. Zuckerberg is determined to assert Meta as a major force in the AI industry, having seen his company’s role in its development overshadowed by the achievements of OpenAI with ChatGPT, as well as Google LLC with its Bard chatbot, and even some notable AI startups such as Cohere Inc. and Anthropic.

OpenAI especially has made great strides this year, with ChatGPT taking the internet by storm and the company following with moneymaking services on the back of it, such as ChatGPT Plus and its recently launched ChatGPT for Business tool. In addition, the Microsoft Corp.-backed startup has been encouraging other companies to build atop its GPT 4 model, which is the foundation on which ChatGPT is built. Meanwhile, Google recently debuted Bard and is now planning a more capable large language model dubbed Gemini.

Those monetization efforts are part of OpenAI’s efforts to recoup the massive amounts of money it has spent to train and run its advanced models. Meta, too, is reportedly spending enormous amounts of cash to buy up more of Nvidia Corp.’s powerful H100 graphics processing units, which are highly sought after for their ability to train generative AI. Although Meta has previously used Microsoft’s Azure and IBM Corp.’s cloud platform to host and distribute Llama 2, its plan is to train the new model within its own data center network, the Journal’s sources said.

AI development is incredibly expensive, but Meta is determined to make its new model available under an open-source license, meaning it will be free for other companies to use and build a range of generative AI tools and services.

Meta has long pursued an open-source approach with its AI initiatives and is known to be one of the biggest contributors to the industry. This year alone it has released numerous AI models and training datasets to the AI community.

These include its Code Llama large language model optimized for programming tasks; a SeamlessM4T model for universal, on-demand translation for hundreds of languages; AudioCraft, a generative AI model for creating music and sound; and Voicebox, a generative AI model for speech. Also this year, it unveiled I-JEPA, a computer vision model that learns more like humans do, and FACET, which is a benchmark dataset designed to help researchers audit computer vision models for bias.

Constellation Research Inc. Vice President and Principal Analyst Andy Thurai told SiliconANGLE that Meta’s plans bode well, as Llama 2 has been wildly successful, beyond even its own expectations. What’s more, Meta knows it needs to follow up with incremental improvements so as not to lose traction against its rivals.

“The goal of Meta has always been to continue to experiment and better their LLMs from the get-go,” Thurai said. “The current model it’s working on is multi-modal, which means it can deal with not only text generation but also image generation, combining the capabilities of various successful LLMs such as OpenAI’s GPT-4 and Midjourney in one single place.”

However, Thurai warned that Meta’s AI models won’t be able to follow the open-source path indefinitely. He said the company has to acquire large numbers of Nvidia GPUs, employ a wide AI talent pool and invest in other aspects of its data center infrastructure.

“Eventually, the continued cost of training those large models and the costs of a high-end AI talent pool will weigh heavily,” Thurai said. “Meta will eventually have to figure out a way to monetize its AI offerings, such as an API service to the hosted models or private instances for large enterprises.”

Earlier this year, Zuckerberg laid out his grand ambitions for generative AI in an internal meeting with company employees, speaking of the potential it has in creating 3D visuals for the metaverse. Meta is also planning to launch new generative AI features in Instagram that will enable users to edit images with prompts, plus “AI agents” for Messenger and WhatsApp, designed for education and entertainment.

Whether or not Meta’s new AI model will ever catch up with OpenAI remains to be seen, though, as their power is determined by the amount of data they are trained on. Meta’s Llama 2 was reportedly trained on 70 billion parameters of training data. OpenAI hasn’t said publicly how many parameters GPT-4 is based on, but others have estimated it could be as much as 20 times bigger than Llama, based on about 1.5 trillion parameters.

Image: Anthony Quintano/Flickr

A message from John Furrier, co-founder of SiliconANGLE:

Your vote of support is important to us and it helps us keep the content FREE.

One click below supports our mission to provide free, deep, and relevant content.  

Join our community on YouTube

Join the community that includes more than 15,000 #CubeAlumni experts, including Amazon.com CEO Andy Jassy, Dell Technologies founder and CEO Michael Dell, Intel CEO Pat Gelsinger, and many more luminaries and experts.

“TheCUBE is an important partner to the industry. You guys really are a part of our events and we really appreciate you coming and I know people appreciate the content you create as well” – Andy Jassy