good model <> bad plumbing

Where does value sit at the Deep application layer in AI

When Anthropic launched Claude Science, they published a result where the same model went from 16.9 percent accuracy on a biology retrieval task to 92.8 percent, and the model did not get any smarter in between (here, and here). All they changed was how the data reached it - they wrapped a plain, deterministic retrieval tool around the scientific databases, and suddenly every model they tested cleared 92 percent on queries that had been embarrassing them minutes earlier.

this is interesting because it quietly undermines a lot of how we talk about these systems and spend our time on the model - on parameter counts and benchmark leaderboards and which lab is ahead this month, and here is a case where the difference between a failing grade and near-perfect had nothing to do with the model at all; just the messy way scientific databases hand back their data, the ambiguous identifiers and the half-documented APIs, and once you smoothed that path the intelligence that was always there could finally land.

A lot of what we call a model being dumb is really a model being handed garbage and asked to perform - when chatGpt gives you a confident wrong answer about a niche fact, the instinct is to say the model is not good enough yet, but often the real story is that the information it needed was never cleanly available to it in the first place, sort of what one of my colleagues keeps saying - the brain was fine and the eyes were fogged, and we keep blaming the brain because the brain is the exciting part to argue about.

At the application layer, where Claude Science and Claude Health (soon) start getting better: the last wave came from bigger training runs and more compute, which is expensive and glamorous and gets the headlines; the next wave might come from something almost nobody wants to fund, which is the slow work of cleaning up how the world’s data is stored and served, the unglamorous librarians and API maintainers and schema wranglers who make information legible to a machine.

So who actually benefits from this and becomes valuable? if reliability comes from deterministic tools bolted onto a probabilistic model, then the person quietly maintaining a genomics database in a university basement becomes as load-bearing to AI progress as a researcher at a frontier lab, and almost nobody is treating them that way (today, i mean).