To develop a bit, the idea I liked in the New Scientist story is the RNA as short term biomemory and test enabler. It is consistent with a paper I bumped into in the past. In it, scientists were mentioning the appearance of genes in populations over as little as two generations in polynesian islands (I'm afraid the exact reference is long forgotten).

Basically, the explanation the researchers were putting forward involved cell mutations through RNA, which would favor appropriate DNA rearrangements when fitness was adequate. Surely there is a way to translate this to neural networks and genetic algorithms. It would make so much more sense than the more or less random mutate() methods — btw: has this changed?

Just wanted to comment that I've had a similar thought about self-training networks based on temporal knowledge of previous trainings. Basically treating previous states as an additional dimension of data for calculating fitness.

But this is tangential to the problem of self-directed learning, and self-chosen fitness.