Meta unveils a more powerful AI and isn’t fretting over who uses it

The largest companies in the tech industry have spent the year warning that development of artificial intelligence technology is outpacing their wildest expectations and that they need to limit who has access to it.

Mark Zuckerberg is doubling down on a different tack: He’s giving it away.

Zuckerberg, the CEO of Meta, said Tuesday that he planned to provide the code behind the company’s latest and most advanced AI technology to developers and software enthusiasts around the world free of charge.

Also read | Apple tests generative AI tools to rival OpenAI's ChatGPT: Report

The decision, similar to one Meta made in February, could help the company reel in competitors such as Google and Microsoft. Those companies have moved more quickly to incorporate generative artificial intelligence — the technology behind OpenAI’s popular ChatGPT chatbot — into their products.

“When software is open, more people can scrutinise it to identify and fix potential issues,” Zuckerberg said in a post to his personal Facebook page.

The latest version of Meta’s AI was created with 40 per cent more data than what the company released just a few months ago and is believed to be considerably more powerful. And Meta is providing a detailed road map that shows how developers can work with the vast amount of data it has collected.

Researchers worry that generative AI can supercharge the amount of disinformation and spam on the internet, and presents dangers that even some of its creators do not entirely understand.

Meta is sticking to a long-held belief that allowing all sorts of programmers to tinker with technology is the best way to improve it. Until recently, most AI researchers agreed with that. But in the past year, companies such as Google and OpenAI, a San Francisco startup that is working closely with Microsoft, have set limits on who has access to their latest technology and placed controls around what can be done with it.

The companies say they are limiting access because of safety concerns, but critics say they are also trying to stifle competition. Meta argues it is in everyone’s best interest to share what it is working on.

“Meta has historically been a big proponent of open platforms, and it has really worked well for us as a company,” said Ahmad Al-Dahle, vice president of generative AI at Meta, in an interview.

The move will make the software “open source,” which is computer code that can be freely copied, modified and reused. The technology, called LLaMA 2, provides everything anyone would need to build online chatbots such as ChatGPT. LLaMA 2 will be released under a commercial license, which means developers can build their own businesses using Meta’s underlying AI to power them — all for free.

By open-sourcing LLaMA 2, Meta can capitalise on improvements made by programmers from outside the company while — Meta executives hope — spurring AI experimentation.

Meta’s open-source approach is not new. Companies often open-source technologies in an effort to catch up with rivals. Fifteen years ago, Google open-sourced its Android mobile operating system to better compete with Apple’s iPhone. While the iPhone had an early lead, Android eventually became the dominant software used in smartphones.

But researchers argue someone could deploy Meta’s AI without the safeguards tech giants such as Google and Microsoft often use to suppress toxic content. Newly created open-source models could be used, for instance, to flood the internet with even more spam, financial scams and disinformation.

LLaMA 2, short for Large Language Model Meta AI, is what scientists call a large language model, or LLM. Chatbots such as ChatGPT and Google Bard are built with large language models.

The models are systems that learn skills by analysing enormous volumes of digital text, including Wikipedia articles, books, online forum conversations and chat logs. By pinpointing patterns in the text, these systems learn to generate text of their own, including term papers, poetry and computer code. They can even carry on a conversation.

Meta is teaming up with Microsoft to open-source LLaMA 2, which will run on Microsoft’s Azure cloud services. LLaMA 2 will also be available through other providers, including Amazon Web Services and the company HuggingFace.

Dozens of Silicon Valley technologists signed a statement of support for the initiative, including the venture capitalist Reid Hoffman and executives from Nvidia, Palo Alto Networks, Zoom and Dropbox.

Meta is not the only company to push for open-source AI projects. The Technology Innovation Institute produced Falcon LLM and published the code freely this year. Mosaic ML also offers open-source software for training LLMs.

Meta executives argue their strategy is not as risky as many believe. They say people can already generate large amounts of disinformation and hate speech without using AI, and that such toxic material can be tightly restricted by Meta’s social networks such as Facebook. They maintain that releasing the technology can eventually strengthen the ability of Meta and other companies to fight back against abuses of the software.

Meta did additional “Red Team” testing of LLaMA 2 before releasing it, Al-Dahle said. That is a term for testing software for potential misuse and figuring out ways to protect against such abuse. The company will also release a responsible-use guide containing best practices and guidelines for developers who wish to build programs using the code.

But these tests and guidelines apply to only one of the models Meta is releasing, which will be trained and fine-tuned in a way that contains guardrails and inhibits misuse. Developers will also be able to use the code to create chatbots and programs without guardrails, a move skeptics see as a risk.

In February, Meta released the first version of LLaMA to academics, government researchers and others. The company also allowed academics to download LLaMA after it had been trained on vast amounts of digital text. Scientists call this process “releasing the weights.”

It was a notable move because analysing all that digital data requires vast computing and financial resources. With the weights, anyone can build a chatbot far more cheaply and easily than from scratch.

Many in the tech industry believed Meta set a dangerous precedent, and after Meta shared its AI technology with a small group of academics in February, one of the researchers leaked the technology onto the public internet.

In a recent opinion piece in The Financial Times, Nick Clegg, Meta’s president of global public policy, argued it was “not sustainable to keep foundational technology in the hands of just a few large corporations,” and that historically companies that released open source software had been served strategically as well.

“I’m looking forward to seeing what you all build!” Zuckerberg said in his post.