Meta stands to gain most from its own open-weight models
Mark Zuckerberg on Monday called open-weight artificial intelligence models “open source AI” in a 6,500-word essay posted by Meta Platforms, arguing that such models are important for U.S. national security and that Meta would release more of them after a hiatus. Zuckerberg said Meta was “strongly supportive of open source, including open source AI models.”
The Wall Street Journal, in an analysis published the same day, said none of the models Zuckerberg described is actually open-source. “And the distinction matters,” the paper wrote.
Open-source software makes source code available in full for anyone to take, use or modify, the Journal noted. Developer communities can suggest changes, and when maintainers accept them, the changes are incorporated into the code. The Journal cited the Linux operating system — which runs much of the computing infrastructure behind the internet — as an example of open-source working well for software that large numbers of people need but that does not differentiate them from competitors.
Open-weight models are not as open and do not serve the same philosophical purpose, according to the Journal. A company such as Meta trains a model using methods that are not disclosed to its users. Users cannot see the code used to train the model, cannot retrain it themselves and cannot add new data; the model’s level of intelligence is essentially fixed by whoever trained it.
What users can do is work with the weights — the numerical values produced by training that influence how the AI operates. Tweaking the weights can vastly alter how a model behaves, the Journal reported, letting users specialize a model in one narrow area and reject questions outside that domain, or make it more polite, sassier or more prone to hallucination.
Modification can be harmful, too, the Journal reported, if bad actors with sufficient computing resources change or remove weights that block the AI from executing cyberattacks or giving instructions on how to make weapons.
That is roughly where the similarities end, the Journal said. Open-source projects are the product of a community of developers and open for the world to see, while open-weight models are closely tied to the company that creates and promotes them. Meta decides when it trains the next generation of its open-weight models, and the company’s capabilities — not those of the community — determine the models’ basic level of intelligence.
The Journal reported that Meta would remain in a better position than anyone to benefit from those models, because of its familiarity with how they are built and its willingness to invest in the computing power needed to deploy them to users. “The balance of power in open-weight models is much more centralized,” the paper wrote.
Meta has contributed extensively to the open-source universe, the Journal noted, including launching React and PyTorch, two widely used projects. Zuckerberg’s essay described a world in which “open-source” models allow anyone and everyone to harness AI’s power and bend it to their own purposes. “There may be a grain of truth in that, but it is also self-serving,” the Journal wrote.
The same edition of the Journal’s AI & Business newsletter reported that Nvidia is teaming up with some of the largest Wall Street asset managers to create pools of capital available to Nvidia’s customers at “attractive rates,” a deal the Journal said could support further purchases of its chips. It also reported that analysts expect revenue in SpaceX’s AI division to rise to nearly $15 billion a quarter by the middle of next year, far above the roughly $2.5 billion the company reported for this year’s second quarter. The edition noted that big AI salaries are distorting the housing market, with average asking rent for the San Francisco metropolitan area the highest in the U.S.