One thing I find interesting about current AI systems is that once you point them to a specific scientific issue, they often know exactly what to say.

They know the expected biology.

They know the relevant markers.

They know what experiments or analyses should be checked next.

The knowledge was already there.

The harder problem was deciding what deserved attention first.

This makes me wonder whether one important challenge for scientific AI is no longer simply knowledge acquisition, but relevance selection.

More knowledge is obviously useful.

But more knowledge also means more possible explanations, more competing signals, and a larger search space.

There is a useful analogy with agent tools.

If an agent has five tools, selecting the appropriate one may be relatively straightforward.

If it has five hundred tools, routing becomes a substantial problem of its own.

Scientific knowledge may create a similar challenge.

The bottleneck may no longer simply be:

Does the model know X?

A harder question can be:

Does the model recognize that X is the relevant knowledge to apply right now?

This is also where scientific expertise becomes interesting.

Experts do not necessarily know more facts than a frontier model.

But they often have better priors about:

  • what matters,
  • what looks suspicious,
  • what can probably be ignored,
  • what deserves another experiment,
  • and what should be checked next.

Years of working with real experiments create a kind of biological attention mechanism.

A strange cluster position, an unexpected marker combination, or an unusually clean result can immediately trigger:

Something is off.

That judgment is difficult to reduce to factual recall.

It combines relevance ranking, anomaly detection, search strategy, and verification strategy.

So perhaps one important transition in scientific AI is:

knowledge acquisition → knowledge selection

or, more broadly:

More context ≠ better reasoning.
More tools ≠ better agent.
More skills ≠ better routing.
More knowledge ≠ better judgment.

AI systems will continue to gain more scientific knowledge.

But increasingly, the interesting question may be whether they can decide what part of that knowledge matters now.

Expertise is not only knowing the answer.

It is knowing what to pay attention to.