Zuckerberg manifesto resets Meta on open weights
In a 6,500-word essay Zuckerberg recommitted Meta to open models, attacked closed rivals for concentrating power, and floated auctioning superintelligence compute.
Mark Zuckerberg published a roughly 6,500-word essay on Monday setting out Meta's artificial intelligence strategy, timed alongside the company's return to publishing open-weight models. It is the clearest statement of direction since Meta reorganised its AI effort into Meta Superintelligence Labs and drifted away from the open releases that once defined its position.
The argument
Zuckerberg's central claim is that the largest risk from advanced AI is not capability but concentration — a small number of companies, institutions or governments controlling systems everyone depends on. He casts rival labs as pursuing exactly that, and positions open weights plus individually held AI as the counterweight, framed around "personal superintelligence" rather than a centralised oracle.
The mechanism
More novel is how Meta says it will ration capacity. Zuckerberg describes free or low-cost access to baseline tools, with paid tiers acquiring compute through a dynamic auction — users bidding for the intelligence they need rather than paying a flat subscription. He also signalled that developers would get access to a more capable model, Muse Spark 1.2, alongside the permissively licensed Muse Glimmer weights.
The reception was mixed. Coverage across FT, Axios, The Verge and TechCrunch read the essay less as a technical roadmap than as a positioning document, with several outlets noting the awkwardness of an anti-concentration argument from one of the few firms able to fund frontier training at all, and Meta's own recent retreat from open releases.
Why it matters
Meta is the only US hyperscaler that has consistently shipped open weights at scale, and its withdrawal left Chinese labs setting the pace of that ecosystem. A public recommitment changes the supply of frontier-adjacent open models available to developers worldwide. The auction idea matters separately: if compute is priced by live bidding rather than subscription, it reframes access to advanced AI as a commodity market — with the distributional consequences that implies.