AI formation of scientific hypotheses

( These notes were in response to a comment that we should test AGI by seeing if it can reinvent general relativity. )

That's noT how it works, however. You need to give the system access to empirical inputs, not just library sources.

Scientific theory development is based on both

1. an existing theory

2. empirical observations that appear to falsify 1, leading to calibrated attempts that produce new theories with or without further experimentation

"GR was hypothesised from thought experiments, not empirical experiments"

2.1. Thought experiments are done with spatial processing, not with stochastic parrots. The processing algorithm has to actually do physics, including spatial simulations in thought. LLMs don't do that, LLMs only process data structures which are semantic, not spatial. You need spatial data structures in thought, to do spatial simulations in thought. Industry will do more soon.

2.2. All of the former is based on reflection on empiricial observations