AI needs its database moment

Computing has crossed this line before. Each time a discipline got serious, the field stopped stockpiling files and started managing them: files to databases, code to libraries, libraries to package managers. Today, picking the right AI model is still mostly trial-and-error. The arXiv preprint 2609.11656 argues the field is at that threshold now, and the next leap is a shared layer for finding, describing, and assembling the trained models that already exist.

The pattern underneath is the same one every prior infrastructure shift ran on: define a unit small enough to route, write the rules for describing it, and let the catalog do the rest. Today's model hubs behave like a hard drive full of downloads. The Learnware Dock System, the arXiv preprint's concrete proposal, behaves like a queryable index. The unit it routes is a learnware: a trained model paired with a public specification of what it is good for, generated against a shared reference so models from different developers stay comparable.

That specification is the load-bearing piece, a behavioral fingerprint that lets an outside system pick a model without ever seeing its training data. It is also the part the field has not yet shown it can produce reliably for the largest, most opaque models on offer. 'Without seeing training data' is a design goal, not an empirical result.

The stakes follow the historical analogy. Whoever ships the catalog writes the routing rules, and whoever writes the routing rules decides which models get found and which sit unused.