Also in early July 2019, Jessica published ["The AI Timelines Scam"](https://www.lesswrong.com/posts/KnQs55tjxWopCzKsk/the-ai-timelines-scam), arguing that the recent popularity of "short" (_e.g._, 2030) AI timelines was better explained by political factors, rather than any technical arguments: just as in previous decades, people had incentives to bluff and exaggerate about the imminence of AGI in order to attract resources to their own project.
-(Remember, this was 2019. After seeing what GPT-3/PaLM/DALL-E/_&c._ could do during the "long May 2020", it's now looking to me like the short-timelines people had better intuitions than Jessica gave them credit for. I won't say, "How could I have known?", but at the time, I didn't, actually, know.)
+(Remember, this was 2019. After seeing what GPT-3/PaLM/DALL-E/_&c._ could do during the ["long May 2020"](https://twitter.com/MichaelTrazzi/status/1635871679133130752), it's now looking to me like the short-timelines people had better intuitions than Jessica gave them credit for. I won't say, "How could I have known?", but at the time, I didn't, actually, know.)
I still sympathized with the "mainstream" pushback against using "scam"/"fraud"/"lie"/_&c._ language to include motivated [elephant-in-the-brain](https://en.wikipedia.org/wiki/The_Elephant_in_the_Brain)-like distortions. I conceded that this was a _boring_ semantic argument, but I feared that until we invented better linguistic technology, the _boring_ semantic argument was going to _continue_ sucking up discussion bandwidth with others when it didn't need to.
That was my intuition. To formalize it, I wanted some sensible numerical quantity that would be maximized by using "nice" categories and get trashed by gerrymandering. [Mutual information](https://en.wikipedia.org/wiki/Mutual_information) was the obvious first guess, but that wasn't it, because mutual information lacks a "topology", a notion of _closeness_ that made some false predictions better than others by virtue of being "close".
-Suppose the outcome space of _X_ is `{H, T}` and the outcome space of _Y_ is `{1, 2, 3, 4, 5, 6, 7, 8}`. I _wanted_ to say that if observing _X_=`H` concentrates _Y_'s probability mass on `{1, 2, 3}`, that's _more useful_ than if it concentrates _Y_ on `{1, 5, 8}`—but that would require the numbers in Y to be _numbers_ rather than opaque labels; as far as elementary information theory was concerned, mapping eight states to three states reduced the entropy from lg 8 = 3 to lg 3 ≈ 1.58 no matter "which" three states they were.
+Suppose the outcome space of _X_ is `{H, T}` and the outcome space of _Y_ is `{1, 2, 3, 4, 5, 6, 7, 8}`. I _wanted_ to say that if observing _X_=`H` concentrates _Y_'s probability mass on `{1, 2, 3}`, that's _more useful_ than if it concentrates _Y_ on `{1, 5, 8}`—but that would require the numbers in Y to be _numbers_ rather than opaque labels; as far as elementary information theory was concerned, mapping eight states to three states reduced the entropy from log<sub>2</sub> 8 = 3 to log<sub>2</sub> 3 ≈ 1.58 no matter "which" three states they were.
How could I make this rigorous? Did I want to be talking about the _variance_ of my features conditional on category-membership? Was "connectedness" intrinsically the what I wanted, or was connectedness only important because it cut down the number of possibilities? (There are 8!/(6!2!) = 28 ways to choose two elements from `{1..8}`, but only 7 ways to choose two contiguous elements.) I thought connectedness _was_ intrinsically important, because we didn't just want _few_ things, we wanted things that are _similar enough to make similar decisions about_.
projects / Ultimately_Untrue_Thought.git / blobdiff