Thanks for the comment, Mauricio!

It’s true it somehow resembles how we teach our kids, but there are deep differences in the learning process between models like InstructGPT (or AlphaZero, or any other ML-based model) and humans.

For instance, these models don’t have even a tiny fraction of the innate structures toddlers have that allow them to build knowledge and understanding of the world quite fast.

Also, kids learn a lot by imitating and enacting their perceptual experiences. They aren’t just passively learning about the world. The active part is completely missing in models like InstructGPT.