An LLM doesn't start from scratch, it's ran on computer hardware...

The particular problem with talking about humanity is our biological functions are deeply intertwined with our intelligence functions.

Some things are programmed in, like cellular behavior, organ building, what to do with these systems after you eat food. The two bottom parts of the brain are rather well developed at birth that control these functions.

The cerebrum is rather undeveloped in comparison, and that's where the "thinking" happens, at least in terms of what we call intelligence and consciousness. While it comes preprogrammed with any number of subroutines, the amount of what we would consider data is minimal, and in our young years a huge amount of energy is spent exercising the brain so it works properly in the reality it exists. In children that is deprived from stimulation and real world experience (think extreme abuse), the brain underdevelops and it can become impossible for them to learn some things or think in particular ways.

Do not underestimate the amount of work that must be performed for an infant to become a thinking machine.

Thank you for bringing up the hardware perspective for LLM intelligence development.

Can one argue that the random initialization of a model is akin to this starting point? A human brain has reached a morphogenetic starting point thanks to much evolution, a process full of randomness.

> Can one argue that the random initialization of a model is akin to this starting point?

It's a starting point but it's random, unlike the starting point of a human being which is very very not random.

All right, let’s dig deeper, I like your responsiveness.

Imagine if the LLM were randomly provided the same starting point as human brain…what happens?

Is evolution a differentiable process?