The Natural Abstractions Hypothesis states that we should expect intelligent agents to find similar abstractions, thus making them natural. But why should we expect them to? It seems plausible that different systems might have different abstractions of their surroundings than we do.
There are two main reasons why we should expect them to:
1) Intelligent systems on earth likely have to make similar predictions of their surroundings. Thus, they will converge on similar abstractions as well. In concrete terms, this means that any intelligent agent that perceives our world will likely throw away similar information in order to achieve a high-level summary of a low-level system.
- This is of course limited, i.e. a gardener has to predict different things than any normal person, e.g. some characteristics about dirt.
2) We already acquired empirical evidence of natural abstractions in artificial neural networks, e.g. in GAN networks and in AlphaZero).
Source: Chan, L., Lang, L. and Jenner, E. (no date) ‘Natural Abstractions: Key claims, Theorems, and Critiques’. Available at: https://www.alignmentforum.org/posts/gvzW46Z3BsaZsLc25/natural-abstractions-key-claims-theorems-and-critiques-1 (Accessed: 19 March 2023).