[FRIAM] A very good piece about ChatGPT.

[glen]

I'm curious why it's unanticipated. Is this a (very mild) form of the Nobel Effect: https://en.wikipedia.org/wiki/Nobel_Prize_effect? The idea that one's been successful in some persnickety sub-domain triggers some dopamine and makes them think they'll achieve easier success in some other (equally persnickety) sub-domain?

I think this highlights a subtle error in the Sutton article Jochen linked, related (sorry for the repetition) to a wise choice of use case(s). Sutton writes:

http://www.incompleteideas.net/IncIdeas/BitterLesson.html
"1) AI researchers have often tried to build knowledge into their agents, 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning."

This passage and the whole article *assume* a subset of use cases. At first I thought it was simply a lack of distinction between special AI vs. AGI. But it's not. It's the assumption of a tiny part of special AI. It's the assumption that AI is equivalent to *opaque* AI ... "free lunch" algorithms ... brute force. It might be useful to contrast something like ChatGPT with Alpha Zero (or, better yet, Alpha Tensor).

The knowledge is in the scaffolding to the well-formed problem. The knowledge isn't in the brute force computation that occurs inside the space constrained by the scaffolding. To learn something *is* to build that scaffolding.

On 2/10/23 10:45, Russ Abbott wrote:
> The amount fo work required to build the intellectual scaffolding almost always seems to be an unanticipated problem when learning something new. In many cases the scaffolding is not just intellectual. Think of what it takes to learn to play musical instrument!

-- 
ꙮ Mɥǝu ǝlǝdɥɐuʇs ɟᴉƃɥʇ' ʇɥǝ ƃɹɐss snɟɟǝɹs˙ ꙮ