AI alignment researchers don't (seem to) stack

By So8res @ 2023-02-21T00:48 (+47)

This is a crosspost, probably from LessWrong. Try viewing it there.

null
Geoffrey Miller @ 2023-02-21T20:19 (+14)

So8res - well said. This seems like an accurate take on a major problem, and it fits well with what I've observed about the rates of progress in various new and emerging academic fields.

Your last paragraph is especially important -- often, the first generation of visionary researchers working on a fresh problem offer such intellectually compelling and novel insights that they sweep up a lot of young talent into their world-view. The young talent initially just follows in their tracks, adding a few details and epicycles to their initial models. It often takes at least 20-30 years for a younger generation, after that first flush of enthusiastic field-building, to develop any serious critiques of the initial visions, or to find any common ground between different visionaries.

The result is that major, new, intellectually demanding fields usually take at least 30-40 years to mature to the point that they can become 'normal science', with a large, multi-generational, smoothly functioning ecosystem of ideas, critiques, data, and advances that aren't overly locked into the original, fallible insights of the field's founders.

The field of AI alignment is maybe 10-15 years old, depending on when we start counting. That leaves at least another 25-35 years before we can expect it to achieve even a modest degree of maturity and applicability.

And I can't think of any historical examples of any people or groups successfully accelerating this generational time-scale for field maturation. It seems pretty deeply woven into the social psychology of human research cultures.

Linda Linsefors @ 2023-03-11T21:45 (+3)

This reminds me of attitudes to Quantum Physics. Most current physics professors I've meat have a sort of learned helplessness relationship to quantum interpretations, subscribing to something like "shut up and calculate" (i.e. don't even try to understand).  There is an attitude that quantum is too strange and therefore impossible to understand. Where as the newer generation of post-docs and grand students don't shy away from quantum interpretations, and discussions of ontology.  However, this falls a bit outside your model, since quantum mechanics is ~100 year old.

Vasco Grilo @ 2023-02-26T16:30 (+2)

Nice post!

Have there been any attempts to estimate the fraction of the total research hours required to solve AI alignement which are serial time?