The ultimate litmus test

I like how plainly Jack Clark puts this:

[F]or the AI rev­o­lu­tion to truly pay out, it needs to change sci­ence: AI has already mas­sively changed and accel­er­ated the work of com­puter programmers, but I think for AI to have a large effect in the world we need to apply it to sci­ence — the ulti­mate litmus test for the suc­cess of AI as a tech­nology will be if it can either make research break­throughs itself or prov­ably mas­sively accel­erate sci­en­tists in their ability to make break­throughs. Future­House is building soft­ware to help us see if this is the case.

Of course, I like how plainly Jack puts most things. As a jour­nalist turned AI practitioner, he has not quite been able to put the doc­u­men­tary impulse behind him, and his weekly newsletter is THE essen­tial AI industry read.

For my part, I con­tinue to believe that super­charged AI sci­ence is not par­tic­u­larly likely but also not impossible — which is reason enough to pursue this set of ques­tions for a while.

(Here, by the way, is a Bor­ge­sian chal­lenge to AI sci­ence: imagine a legit­i­mately supersmart AI assis­tant that can, in fact, pro­pose a perfect, paradigm-busting experiment. However, it can also pro­pose 100,000 other exper­i­ments that are a waste of time … and there’s no way of knowing ahead of time which is which. A sig­na­ture of AI intel­li­gence is the cheerful will­ing­ness to offer another alternative, and another, and another. Where and how might stub­born specificity, even obsession, be intro­duced into these systems? What would it mean for an AI system to form a deep and com­pelling conviction? Is it even possible?)

(The Bor­ge­sian part comes from imag­ining a pile of 100,000 pro­posed physics exper­i­ments — or far more — each with a billion-dollar price tag, knowing that one of them WILL reveal a deeper, more accu­rate view of reality … but which one … ?)