How is the "sharp left turn defined"?

post by Chris_Leong · 2022-12-09T00:04:33.662Z · LW · GW · No comments

This is a question post.

Contents

  Answers
    2 Ethan Caballero
None
No comments

I've read this article by Nate Soars [LW · GW], but I felt the definition wasn't completely crisp and I felt this was less than ideal given how this term seems to have entered wide use.

The article seems to define it as capabilities generalising further than values (1).

At the same time, this doesn't seem to capture the whole concept, as there seems to be an additional connotation of this generalisation happening or becoming relevant when the AI reaches a certain power level (2).

There's a further connotation of this increase in capabilities happening very fast (3).

And then one of this leading to a treacherous turn (4).

And the shift from acting seemingly aligned to pursuing treachery happening very fast (5).

It would be nice to have a nice crisp definition. So it is the "sharp left turn" just (1), or are  (2) and/or (3) and/or (4) also included as part of the definition?

Answers

answer by Ethan Caballero · 2022-12-09T02:38:47.493Z · LW(p) · GW(p)

I give a crisp definition from 6:27 to 7:50 of this video: 

↑ comment by David Johnston (david-johnston) · 2022-12-10T08:18:20.650Z · LW(p) · GW(p)

Ethan finds empirically that neural network scaling laws (performance vs size, data, other things) are characterised by functions that look piecewise linear on a log log plot, and postulates that a “sharp left turn” describes a transition from a slower to a faster scaling regime. He also postulates that it might be predictable in advance using his functional form for scaling.

↑ comment by weverka · 2022-12-09T21:53:32.844Z · LW(p) · GW(p)

You drew a right turn, the post is asking about a left turn.

↑ comment by weverka · 2022-12-09T21:55:18.748Z · LW(p) · GW(p)

you draw a right turn.  The post is asking about a left turn.

No comments

Comments sorted by top scores.