"Open" is the most elastic word in artificial intelligence, and this week Meta stretched it across its biggest release yet: an open-source version of its most capable model, accompanied by a scaled-down edition pitched as something ordinary people can run at home, and by Mark Zuckerberg attacking his "closed" rivals while lobbying Washington to lower the barriers for open-weight AI.
Strip out the industry politics — the talent wars, the safety arguments, the question of who sets the rules for frontier models — and there is a genuinely interesting promise underneath, which is the one that concerns this desk. AI you run at home is a different proposition from AI you rent. Your prompts stay on your machine rather than travelling to a data centre. There is no monthly subscription, no usage cap that tightens when the provider's costs rise, and no company deciding one morning to retire the model you had built your habits around. Given how the rented sort has been behaving lately on all three counts, the appeal is not hard to see.
The desk's scepticism, in the right proportion, goes as follows. A downloadable set of model weights is not a product; it is an ingredient. The distance between "the weights are on the internet" and "a useful assistant is answering questions on your computer" is filled with hardware requirements, technical patience, and the sort of tinkering most people reasonably decline. "Home" is doing a lot of work in the phrase "run at home" until Meta says plainly what machine it has in mind — and the machine most people own is a phone.
There is also the question of what "open" includes this time. Licences on previous open-weight releases have carried conditions that made lawyers, if not users, pause. The details — sizes, terms, what you may build with it — are where this announcement will succeed or fail, and they deserve a closer reading than a launch-day manifesto allows.
Meta's motives are its own affair: openness is, among other things, a competitive stance for a company that arrived late to the frontier and has built its own silicon to catch up. But motives matter less than results. If the practical outcome is capable AI that ordinary people can run privately, on hardware they already own, that is a better internet than the one where intelligence is metered out by subscription.
Whether this release is that outcome will be settled not by the manifesto but by what actually runs, unassisted, on an ordinary machine. We intend to try it and report back.
