Zuckerberg urges greater low-cost access to AI models

Mark Zuckerberg, CEO of tech giant Meta, on Monday published a lengthy manifesto calling for increased free or low-cost availability of high-powered artificial intelligence models. Meta simultaneously announced the release of a new agentic AI model, Muse Glimmer, that is small enough to run on personal computers. Glimmer pricing depends on electricity and computer processing power in a user’s region, but hypothetical estimates fall well under 50 cents per hour, according to the Kingy AI news site. Meta also openly released some of Glimmer’s development information.

Zuckerberg argued that widespread AI use and access would lead to a balance of power between individuals and institutions, including businesses and governments. He took aim at rival tech giants, saying they focused on developing AI for large institutions. Instead, he urged AI companies to develop more personal AI agents with simulated values and beliefs that could match those of their users. A diversity of viewpoints would keep powerful AI from restricting freedom, he said. Zuckerberg also said AI agents should have privacy settings keeping Meta or other service providers from seeing information on an individual’s use of the agents.

He also urged Meta and other AI labs to develop more computing power, so they could keep working on their products while allowing private individuals to account for the majority of AI use. Building more data centers was one of the main ways to reach that goal, he said. He stated that Meta helped alleviate communities’ concerns over water and power usage associated with data centers by building its own power sources and restoring water to local watersheds.

Did he offer suggestions for U.S. government policy on AI? Zuckerberg urged less U.S. regulation of data used for training AI models, since some foreign countries including China don’t have such restrictions. He also urged accelerated construction of new power sources and data centers to keep up with China. Zuckerberg also said the government should not restrict access to open-source AI models from other countries. He urged officials to reconsider their opposition to distillation, a term for training a less advanced AI model by using the outputs of a more advanced one. Chinese companies including DeepSeek are suspected of using distillation to pull features from proprietary U.S. AI models. However, he praised U.S. export controls on silicon, which he said helped slow the progress of foreign AI labs.

Dig deeper: Read Lauren Canterberry’s story on a court ruling in favor of Meta’s use of copyrighted work for its training data.