Personal AI for everyone, billions for data centres: 4 takeaways from Mark Zuckerberg’s AI manifesto

1. Distribute AI power

In recent interviews with The New York Times and the Wall Street Journal, Zuckerberg has criticised the position taken by rivals Anthropic and OpenAI to keep frontier AI capabilities controlled, along with their emphasis on strong safety and alignment constraints. His language in these interviews is nearly identical to the manifesto’s contents, with him characterising the discourse from other labs as overwhelmingly “filled with doom”.

He argues that the concentration of superintelligent AI power in the hands of a few companies is the real danger.

Zuckerberg also disagrees with the notion of a “single benevolent superintelligence”, arguing that humanity is not a monolith and comprises diverse and divergent values. These lead to varying tradeoffs on important issues, implying that one person’s technological solution would be opposed to another’s interests and values. Such a superintelligence would prioritise one competing set of values over another.

In the manifesto, he makes the case for “personal superintelligence”, in which extremely powerful AI can be personalised to meet individual user needs and specifications, and work towards their goals and values. Thus individuals would exercise greater control over their AI’s values and behaviours. They would thus have powerful tools to create new businesses and a “personalized tutor with a PhD in every subject”.

“Putting power in people’s hands to pursue their own aspirations is how humanity has made the most progress. Novel ideas and major steps forward rarely originate from established institutions alone,” the manifesto says.

2. Open weights are the way forward

The Meta CEO has also argued in favour of open-weight AI models, in which developers and users can download and run the trained model without necessarily accessing the code or data used to build it. The argument was also the impetus behind the formation of the Open Secure AI Alliance, led by Nvidia.

Supporters argue that open-weight models help foster innovation via peer review and experimentation, prevent a few labs from becoming gatekeepers, and democratise access to AI and economic power. Zuckerberg says doing so allows developers to learn from competing models and supports a broader AI ecosystem.

For supporters, the OpenAI-Hugging Face incident demonstrated the clearest need for open-weight models. In an analysis of the incident, Hugging Face revealed that closed commercial AI tools, such as Anthropic’s Claude Opus, treated the attack logs it had fed to assess the attack as potentially malicious. It instead relied on the Chinese open-weight model, GLM 5.2, which it could download and run on its own infrastructure to analyse the logs.

Critics like Anthropic, however, argue that openness could also make such tools easier for malicious actors to exploit. In a blog post on July 27, Anthropic CEO Dario Amodei said the company was focused on “keeping powerful chips out of authoritarian hands, stopping industrial-scale distillation, and requiring safety testing of all sufficiently capable models, open and closed.”

Meta’s decision to align itself with the open-weight camp comes at a strategically significant moment, even as it has begun selling its new models, including the new Muse Glimmer and an open-weight version of Muse Spark, for public use. The company has spent massive sums of money trying to keep up with the pace of progress by OpenAI and Anthropic, including a $14.3 bn investment in Scale AI, recruiting top researchers, creating Meta Superintelligence Labs, and building massive data-centre capacity.

Even if it doesn’t win the frontier AI race, it still stands to benefit from shaping the terms of the competition and potentially becoming a key platform in the ecosystem powering billions of users.

3. A different approach to risk redressal

Zuckerberg largely downplayed current risks, such as job losses due to automation, while recognising the potential for AI misuse in cybersecurity and bioterrorism. To tackle this, he argues for a safety approach anchored in a mix of corporate governance and continued collaboration with the government.

“…the general pattern is that society ensures the defenders who keep bad actors in check must always have a greater balance of power and resources”, he wrote.

He has also called for AI companies to work more closely with governments during the development of frontier models, rather than relying solely on rigid, post-development review processes. Meta’s own board, he said, should have greater oversight over the safety criteria governing the release of its models.

Zuckerberg also argues that the US is currently at a disadvantage in its race against China to develop superintelligent computer systems, an effort that calls for data centre expansion. To this end, Meta has planned to spend up to $145 bn in capital, and up to $600 bn by 2028, WSJ reported.

4. Communities must benefit from AI infrastructure

A part of this allocation will be channelled to local communities near data centres, with Meta promising to invest in jobs, schooling and public services, as well as infrastructural development in the surrounding areas. Meta has also announced a $1 billion “Future Is For Everyone Fund” to support communities hosting its data centres.

“We would all benefit together by figuring out the community compacts required to enable sustainable development across the country,” he wrote.

However, data centres remain widely unpopular: last month, 42 states witnessed 142 protests on a single day in the first coordinated nationwide effort protesting the rapid AI infrastructure expansion. New York has enacted a statewide ban, with city-level bans in Seattle in Washington state and Monterey Park in California.

The move is an attempt to address growing political and community opposition to the enormous power, water, land and infrastructure demands of AI data centres.