As a huge fan of LBD and Swanson's work, I'd like to "build" on this piece by reaching further back into the history of computer aided theory formation. Not very far, mind you, but to 1977, when Douglas Lenat published a PhD Thesis on that topic [1].<p>Of course, theory formation goes further back than that, but Lenat's work seems important - at least through the lenses of those of us who started in this space not with numeric, but with symbolic processing.<p>Lenat was criticized on the AM project because, while it was operating in the space of set-theoretic reasoning, it was directly implemented in Lisp, itself a set-theoretic platform. So, during his postdoc, Lenat invented what, today, we would call a "DSL" - a program which provides an API which is not set-theoretic, but rather discovery operations oriented. From there, he built the program Eurisko (brief sketch at [2] but gives good google) which did some amazing things.<p>I offer my suspicion that, when the dust settles, AI is headed in the direction of "hybrid" systems,those which combine the best characteristics of LLMs, Bayes, Symbolic,and more.<p>[1] <a href="https://stacks.stanford.edu/file/druid:mz393pf0178/mz393pf0178.pdf" rel="nofollow noreferrer">https://stacks.stanford.edu/file/druid:mz393pf0178/mz393pf01...</a><p>[2] <a href="https://www.lesswrong.com/tag/eurisko" rel="nofollow noreferrer">https://www.lesswrong.com/tag/eurisko</a>