Meta Leak Put AI in Everyone's Hands.

 

"Meta Platforms is emerging as potentially one of the biggest disrupters in efforts to cash in on the artificial-intelligence boom, although the Facebook parent's impact is partly accidental.

Meta in February started giving a limited number of researchers access to advanced AI language software it had developed to enable chatbots similar to the viral ChatGPT. But someone posted the code publicly, spurring the emergence of homegrown tools that could one day rival those that tech companies such as Microsoft, Google and others are trying to sell. It is the latest twist in the rapid advance of this kind of software since OpenAI released ChatGPT last November.

Meta had opened up parts of its AI technology in the past, but with this software the company decided to restrict access so that it could be evaluated by researchers, but not widely distributed, said Joelle Pineau, Meta's vice president of AI research. "Someone decided not to respect the terms of use," she said.

Although there are widespread worries about the misuse of powerful AI technologies, the release of this technology has put Meta at the center of a vibrant, if uncontrolled, surge in AI software development.

There has been an explosion of innovation over the past months, Pineau said. "The number of positive and exploratory use cases has been very, very high," she said.

Fueled in part by the leak, Meta's software has won a wide following. Researchers at Stanford University used it to build an AI chatbot for under $600. Brandon Duderstadt, chief executive of software company Nomic AI, built a research product called GPT4All that allowed anyone to run a ChatGPT-style chatbot on their computer.

Another researcher figured out how to run the software on an ultra-low-cost Raspberry Pi computer, demonstrating how chatbots based on a large language model -- the term for the sophisticated AI technology that powers new chatbots -- can be produced on low-cost systems.

For years, Meta has been working on its own generative artificial intelligence software that can create humanlike conversation, with an eye on replicating the success OpenAI has had and that Google is aiming to achieve with the Bard chatbot and search tools it made publicly available this month.

As part of its effort, Meta developed a fundamental building block for AI it called LLaMA, or Large Language Model Meta AI. After the limited release three months ago, it became public through a leak in an unlikely place -- the anarchic online message board 4chan, which is better known as a source of online memes and highly inappropriate content.

With that leak, the AI tool became available everywhere, spawning a grass-roots surge in AI development, software developers say.

"This is no longer the domain of big companies. This is in the hands of hundreds of thousands of hobbyists," said Alex Stamos, director of the Stanford Internet Observatory, an academic research group that studies internet abuse.

The uptake of the Meta software caused some anxiety within Google, which has bet heavily on embracing AI. After LLaMA's release, a Google engineer wrote a memo saying its democratization of AI could threaten larger companies.

"The barrier to entry for training and experimentation has dropped from the total output of a major research organization to one person, an evening, and a beefy laptop," the engineer wrote. Technology newsletter SemiAnalysis earlier reported on the memo.

What has made the leak of the Meta tool so useful is that it was generated with resources that aren't available to individual developers. It would take a single graphical computer chip more than 1 million hours to process all the data used by the largest version of the software, Meta said." [1]

1. Meta Leak Put AI in Everyone's Hands. McMillan, Robert. 
Wall Street Journal, Eastern edition; New York, N.Y. [New York, N.Y]. 19 May 2023: B.4.