05/26/2026
The most important AI question for museums is not which model.
It is where the model runs.
For the last few years, the default assumption has been that capable AI lives somewhere else. A cloud service. A vendor endpoint. A subscription. An API. The assumption was that local was a degraded substitute, an underpowered fallback, a compromise for institutions that could not afford the real thing.
That assumption is already aging out.
At AAM 2026 last week, we put Alice® Local on a pedestal in Philadelphia. The hardware was a local computer small enough to carry in under your arm. The AI model running on it was in the multi-billion-parameter range at 8-bit quantization. No internet connection was required for the AI demonstration. The model is not named, because the model is not the point.
The point is that a relatively modest local setup is already doing meaningful work.
Now extrapolate. Local hardware is getting faster. Local models are getting better. The cost curve is improving. In eighteen months, what we demonstrated at AAM will look quaint. In thirty-six months, it will look conservative.
For museums, this changes the architecture decision. When AI runs inside the venue, the institution can ask cleaner questions. What does the AI know? Who controls that knowledge? Where does inference happen? What leaves the premises? What is logged? What can be audited? What happens if the network is down? Can this be governed by policy rather than by hope?
A museum that runs personalization through a cloud has accepted, often without saying so out loud, that its visitors are partially the cloud provider's visitors. The provider sees the signals. The provider's jurisdiction touches the data. The provider's terms of service govern what happens next. None of those are things a director, a trustee, or a general counsel would consent to if asked plainly.
A museum that runs personalization locally, under institutional governance, has not accepted any of that. The visitors are the institution's visitors. The data stays in the building. The jurisdiction is the institution's own. The policy the AI obeys is the policy the institution writes.
Alice® Local is early. We will say that plainly. But it is early in the way that matters. It points in the right direction.
Local is not a step backward. For many institutions, local is the condition that makes AI acceptable at all.