Modern LLMs are fairly good at that as well.
We can do that for AIs too - pre-train on pure low Kolmogorov complexity synthetics. The AI then "knows things" before it sees any real data. Advantageous sometimes. Hard to pick compute efficient synthetics though.