fleetingbytes

personal blog

[fleetingbytes @fleetingbytes]

@jachiam0 now, some of this is downstream of there being no mats or open phil for this; but, the labs seem bad at finding institution builders that will actually build this (and the fact that they don’t seem to care makes their motives suspect).

[fleetingbytes @fleetingbytes]

@jachiam0 which is surprising, because build american data centers, semi-conductor manufacturing, advance middle ground on lab regulation actually has adherents in the wild; but the fact that oai can’t seem to find them or inspire them is a questionable sign.

[fleetingbytes @fleetingbytes]

dev notes

1) so, i was working on a new visualization for the front page of my website; the visualization is meant to show an adversarial game

2) the reason for this is that we are entering the age of adversarial ai; mythos in cybersecurity is the start of ai entering true adversarial contexts

3) where the models will end up playing attack and defense; and, perhaps will eventually be explicitly trained against both of these targets

4) oh, i am also inspired by the red queen race nature of the frontier lab competition, which is true for almost all businesses, but particularly the foundation labs

5) so, i think it is a fitting time for a new art piece for the front website; which can gesture to these dynamics

6) the problem is that it's not clear what is a good adversarial game for the site; it needs to be something that can run in the browser nicely

7) and, i want it to be played by a model; and I think that basically means it needs to be a model in the 100k - 1M parameter range; if it's a recurrent model, smaller

8) it turns out it is hard to get claude to invent a game that is (a) interesting, (b) visually interesting and (c) good for training a model against

9) the first game i tried was an adversarial version of sudoku, where the last player to be able to place a valid legal move wins

10) but it turns out it's just a parity game, almost all moves are winning, except for some very narrow edge cases where you can flip who is winning

11) in the interim, i built my own small 4x game that i could play on the side, which was pretty fun, and i'm now seeing if i can get a striped down version to work

12) part of the inspiration was the sakana digital red queen paper (i now have a second claude code agent trying to get a version of that paper to work for my game; but not for the website)

13) i also watched the Shah interview on 80,000 hours and thought of writing something on it, but there was no clear throughline

14) i recommend the interview though; i think it is very good for thinking about ai risk from the point of view of someone who wants to understand lab behavior

15) maybe, that's a take i can write up... will need to watch it over again with a pen out though

16) i also read an article by tom cunningham on low hanging fruit and rsi and think i will write something about that; but to need to sit down and get a take

[fleetingbytes @fleetingbytes]

inspired by @Effective69ism

[fleetingbytes @fleetingbytes]

@AINewsInt you have to push back a bit

[fleetingbytes @fleetingbytes]

they tried so hard to train clopus 4.8 to correct the user that it ends up trying very hard to avoid answering your question if it doesn't want answer it

it also feels almost gpt-4 in the way that it will try to pretend you asked a different question, then answer that

[fleetingbytes @fleetingbytes]

@AINewsInt it's very smart, but feels like it wants to corral you