Can language models achieve AGI and genuine explanatory creativity?

A post introduces an Oxford Physics discussion with David Deutsch, Nick Bostrom and others about whether large language models can reach AGI and genuine explanatory creativity.

A reader’s notes summarize the open questions: Is explanatory creativity qualitatively different from sophisticated search and recombination? Does human understanding have a ceiling? Does the missing capability require a new theory, a modest addition or continued development of current methods?

The reader argues for focusing less on the AGI label and more on producing new scientific knowledge that benefits humanity. Their strongest evidence would be progress across a breadth of scientific domains beyond current human understanding. They call this a high, fuzzy bar—not a binary definition—and say the test could initially be limited to abstract and algorithmic domains, without experiments.