Zuckerberg on his pro-AI manifesto
My conversation with Mark Zuckerberg about Meta’s pro-AI stance and model progress: "We are very close to having substantially stronger models."
Over the weekend, Mark Zuckerberg called me through his Meta glasses while speeding across the water toward a fishing spot. The noise cancellation was holding up against what sounded like a wind tunnel. Before he got there, he wanted to walk me through the 6,500-word pro-AI manifesto he had spent the past month or two writing.
Zuckerberg told me the immediate reason he decided to publish his philosophy now is that Meta believes it is “very close to having substantially stronger models,” and he wants people to understand his values before those models arrive.
That is quite a bullish thing for him to say. The prevailing view of Meta’s AI effort is that it is still chasing Anthropic and OpenAI after its Llama reset, the creation of Meta Superintelligence Labs, and a hiring spree that cost billions. There are starting to be signs of what he predicts is coming on the model front, though. Meta’s last Muse Spark release was well received, and Muse Glimmer is a new 30-billion-parameter model designed to run on a single consumer GPU; Meta’s own tests show it outperforming similarly sized Gemma and Qwen models on several agentic benchmarks. Muse Spark 1.2, whose weights Meta says it will release in the coming weeks, is a coding model that Meta itself describes as a step “toward the frontier,” with larger models still to come.
With his letter today and Meta’s recent ad campaign, Zuckerberg is positioning the company against the souring mood around AI, particularly the warnings from Dario Amodei and Anthropic about job losses and the growing backlash to data centers. His letter argues that invention will matter more than automation, that distributing AI is safer than concentrating it, and that the real risk is an imbalance of power. Meta paired the rhetoric today with a $1 billion fund for communities where its US data centers are located and a workforce academy offering free training and guaranteed jobs near those projects. When I asked what could actually change public sentiment, Zuckerberg’s answer was simpler: “Over time, doing good things.”
The most novel part of the letter, to me, is Zuckerberg’s treatment of recursive self-improvement. He assumes that once models can improve themselves, every competitive lab will have to devote substantial compute to that loop or fall behind. His answer is for Meta, other labs, and cloud providers to collectively build so much capacity that they can stay in the RSI race while still directing the “significant majority” of intelligence toward people’s individual goals. I have not heard another frontier CEO put the RSI problem in quite these capital-allocation terms. It is less a plan to stop the race than a proposal to outbuild it while reserving most of the resulting power for humans.
Zuckerberg is also giving a new, independent committee of Meta’s board the authority to approve the company’s model-release safety criteria and to audit whether releases comply with them. That is a meaningful internal check, but not the same as surrendering control, since his supervoting shares let him effectively determine who sits on Meta’s board. The test will be whether the committee can actually stop a release when commercial pressure is at its highest.
Zuckerberg told me he wrote the long version before his recent Wall Street Journal op-ed, added an argument defending model distillation late in the process, and saw the piece partly as a recruiting pitch. More fundamentally, he said, he wanted to challenge the Silicon Valley view that powerful models are inherently dangerous and must therefore remain centralized: “I think people needed to hear the alternative view.”
The rest of our conversation is below.
A MESSAGE FROM MY SPONSOR
Who's winning the software reset?
The industry that built digital transformation is now facing its own disruption. AlixPartners covers the strategic moves, competitive shifts, and AI-driven dynamics redefining enterprise software today.
