Astral Codex Ten

MCTS

4 min read

MCTS

Article

MCTS is a recurring concept in the Astral Codex Ten archive, appearing 2 times across 2 issues between April 04, 2022 and October 03, 2022. The archive places it in contexts such as “based on MCTS bots”; “where MCTS is hybrid between the two approaches”. It most often appears alongside Eliezer, Eliezer Yudkowsky, 2013.

Metadata

  • Category: Concepts
  • Mention count: 2
  • Issue count: 2
  • First seen: April 04, 2022
  • Last seen: October 03, 2022

Appears In

None.

Source Context

Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.

Yudkowsky Contra Christiano On AI Takeoff Speeds
April 04, 2022 · Original source
The impact of GPT-3 had nothing whatsoever to do with its perplexity on Penn Treebank . . . the impact of GPT-3 was in establishing that trendlines did continue in a way that shocked pretty much everyone who'd written off 'naive' scaling strategies. Progress is made out of stacked sigmoids: if the next sigmoid doesn't show up, progress doesn't happen. Trends happen, until they stop. Trendlines are not caused by the laws of physics. You can dismiss AlphaGo by saying "oh, that just continues the trendline in ELO I just drew based on MCTS bots", but the fact remains that MCTS progress had stagnated, and here we are in 2021, and pure MCTS approaches do not approach human champions, much less beat them. Appealing to trendlines is roughly as informative as "calories in calories out"; 'the trend continued because the trend continued'. A new sigmoid being discovered is extremely important.
CHAI, Assistance Games, And Fully-Updated Deference
October 03, 2022 · Original source
Fair enough. I agree that we know how to train some powerful illegible systems in such fashion that they appear to pursue simple crisp goals over a crisply defined artificial environment, assuming that the training distribution is not in some sense less powerful or less varying than the test distribution. There's two bars here. The first bar is about deep learning potentially behaving in a weirder way w/r/t gradient-descent learning on crisp environmental goals than an old-fashioned search tree pursuing the same goals - where MCTS is hybrid between the two approaches. The second bar is about only being able to pursue crisp goals defined over either direct functions of sense data, or environments fully known to the programmers that relate in a known way to the sense data. The first bar says that we don't know how to make an AI that pursues paperclips because we don't know if the Mu Zero training paradigm scales to general intelligence, including eg AIs building other AIs, in a way that preserves the present extent to which Mu Zero seems to stay aligned within a training distribution. The second bar says that on the Mu Zero paradigm we don't know how to point to a class of paperclips within the environment, as they exist as latent causes of sense data, and say 'go make actual paperclips'. We can write a loss function that looks at a webcam and tries to steer reality around the webcam image fulfilling some particular function of images, but we don't know how to point a Mu Zero like system at the paperclips in the outer world beyond the webcam. If we think there are watching humans who can say exactly what is and isn't a paperclip and that there's no way to fool (or smash) those humans, and if we knew how to train a Mu Zero system on amounts of data small enough for humans to generate those, we could maybe try that, and if the test distribution is enough like the training distribution it might work, but it would lack the clear-cut character of writing a search program and knowing what it searches for.