Zuckerberg’s free AI is a clever form of bait

By Parmy Olson

There’s a persistent mystery about Mark Zuckerberg, and it’s not the one about his new chain necklace. The chief executive officer of Meta Platforms Inc. has spent billions of dollars building powerful artificial intelligence models and is giving that technology away for free. Why?

Zuckerberg recently argued that “open sourcing” LLaMA, his AI model for training chatbots, prevents power from being concentrated in a single company.

Maybe take that explanation with a pinch of salt. This, after all, comes from a man who consolidated control in social media by buying and copying his competitors, and who has near-dictatorial control over Meta with more than 50 per cent of total voting power.

Like any good Silicon Valley tycoon, Zuckerberg has slapped a benevolent label on an effort to extract value for his company. The real reason he’s sharing his AI probably has more to do with trying to make life more difficult for his competitors, as well as improving Meta’s reputation so it can lure more experienced AI engineers. One day, it might also allow him to explore new ways to enhance his advertising business. Investors should rejoice, and AI developers with humanitarian ideals should think twice about joining him.

Getting to this point took a fortuitous bet by Zuckerberg. At the end of 2022, just before OpenAI released GPT-4, he placed a large order for graphical processing units— powerful chips used for building most AI systems— with Nvidia Corp., according to an interview he gave to the Verge.

He aimed to improve the recommendation system for Reels, Meta’s short-form video clone of TikTok, but it turned out Zuckerberg was inadvertently stockpiling the most coveted ingredients in tech, the powerful AI chips that are now in short supply.

That’s one reason why Meta has moved so speedily on AI ever since, releasing its first version of LLaMA in February 2023, then LLaMA 2 just five months later. When it launched LLaMA 3 last month, the system immediately entered the top global ranking of AI models (it’s currently at No. 6), making it the most sophisticated system that was also free for anyone to use.

Giving away LLaMA, the system that is also powering Meta’s new AI assistants, already makes life more difficult for large rivals like Microsoft Corp., Amazon.com Inc. and Alphabet Inc.’s Google, who all charge for their AI services as part of their cloud businesses.

It should also go some way toward attracting more top talent, an issue so important to Zuckerberg that he has personally emailed researchers at Google’s DeepMind with job offers, according to The Information.

Many AI developers want both a steady paycheck and a chance to see their work have a wide-ranging impact. LLaMA has now been downloaded more than 100 million times, and software developers like being able to access its “weights,” the internal settings that help it process data, making the model more adaptable for companies who want to summarize legal documents or rank customer feedback. That flexible approach to building technology makes Meta a more appealing place to work too.

Further down the line, there could theoretically be another benefit to Zuckerberg’s grand AI giveaway: emulating Google’s success with Android. Google started giving its mobile operating system to device makers in 2008 and over time added requirements for them to include apps like Google Search, Google Maps and Gmail.

As the number of Android handsets grew, more traffic flowed through those services, which Google was able to monetize through advertising. Google’s apps on Android essentially became portals to billions of users — and a new revenue stream.

Could Meta capitalize on LLaMA similarly? In theory, as the model becomes more popular among developers and integral to their apps, Meta could start adding a requirement for those users to share the data they are processing, data which could be used to enhance the company’s AI models and perhaps even its ads business.

Of course, doing so would mean Zuckerberg could no longer call LLaMA “open-source,” but there’s already debate among industry groups about whether the model qualifies in the first place. And as it happens, Meta has been seeking more control over how each new version of LLaMA is used. The licensing agreement for LLaMA 3 seeks more protection for its intellectual property and greater compliance with its branding and legal standards, for instance, than the one for LLaMA 2.

Such a U-turn wouldn’t be unheard of. Just ask OpenAI, which originally used the term “open” for “recruitment purposes” and has become more secretive over the years.

Two senior open-source AI researchers tell me it would be feasible for Meta to put data-sharing provisions into future licensing agreements for LLaMA, if the company reined back some of the model’s open-source features. If Meta could then grab reams of anonymized data generated by LLaMA, that might help it refine its understanding of what content engages users and improve how its ads are personalized.

That wouldn’t fly with everyone. Concerns about gathering personal data were why ChatGPT breached Italy’s privacy laws. But Zuckerberg turned Facebook into a money printer after dominating the social media market and amassing hundreds of millions of dedicated users. He favors the strategy of chase users first, monetize later. If he can do the same for LLaMA, it might mean making the AI model less open and more lucrative for his company too.