Small vs Big Models: What Each Can Hold | Geometry of Trust | Philosophy - Lesson 4

Small vs Big Models: What Each Can Hold | Geometry of Trust | Philosophy - Lesson 4

A small woodland isn't a failed attempt to be the Amazon. It's a different kind of ecosystem — fewer species, tighter relationships, easier to monitor, easier to protect. The same logic applies to AI.

This talk argues that for most deployments, a small specialised model is both cheaper and more verifiable than a large general one. Those two things normally trade off against each other. In this case they point the same way. That's unusual enough to be worth stopping on.

We walk through a concrete comparison — a 500M-parameter drug checker at a hospital vs a 70B-parameter general model — across the dimensions that actually matter for deployment: hardware, ruler computation, monitoring, governance, cost, verifiability. The small specialised model wins at every stage.

We also look honestly at when going big is the right call. There are domains where a single decision genuinely needs to integrate across areas that can't be cleanly split — police, military, emergency services, government policy. Those genuinely need general capability. But the same generality shifts the hard work from the model to the governance around it.

The principle that falls out: use the smallest model that covers your domain, measure it tightly, monitor it cheaply, audit it clearly. Only go big when the job genuinely requires integration.