I don't follow how "we can see tiny stuff" = "it is possible to build closed system"?

Seeing tiny stuff doesn't magically give you enough info for Claude to reason with the system predictably (the way it can with source code).

It can start to reason about what it seems under the microscope. You just need to train it on images and data from microscopes.

the transition is in-silico => in-vitro => in-vivo

AI already can do in-silico simulation

add microscopes and tiny precision actuators, and we can build closed feedback loop in-vitro system

once we have in-vitro, it will be easier to go in-vivo