I am particularly susceptible to the human vice of speaking in definite terms about things I barely understand. In this post, this is vastly more the case than usual.

The modern synthesis of evolution says that all variation in species is due to random DNA mutations, and that natural selection alone selects the mutations that are adaptive. But that seems just too expensive – not so much the dying, but the having to find adaptive changes in a random way. The opportunities for phenotypic learning (ie: during an individual’s lifetime) that could be incorporated into DNA are just to great to pass, because they can allow for a not-so-random probing towards certain directions that are more promising.

Evolution probably had to bootstrap one (if not several) mechanisms to incorporate learnings from the phenotype into the DNA passed to the offspring. Once this mechanism is in place, then the species that have it can have a massive advantage over those who do not.

To function, these learning mechanisms would have to simulate the environment, to determine which variations are more adaptive than the starting point – or at the very least, attach a value function to those variations in a given context. How would they know otherwise which directions are more promising than others? Perhaps this simulation (or internal recreation) is what the nervous system is, although perhaps the pattern is even more general than that.

Now, how does nature know what learnings are adequate and how they should modify the genes to be expressed properly?

This learning mechanism, if decoded in terms we can understand, could perhaps be a beautiful algorithm of statistical inference – if we reverse engineer it, we could even develop the foundation of a new epistemology from it.