Anthropic, OpenAI and Demis are all in favour of preparing a coordinate pause of frontier AI development.
Skate to where the puck will be people, not where it is now. This is a fast puck.
A cross-party group of Canadian politicians have stepped forward with a clear statement in favor of an international prohibition against superintelligent AI: https://t.co/y1Lizi3JTR
@satchlj As a proxy for replacement of human labor, I think 80% is much better than the 50% chart most people look at. 50% reliability isn't enough. It isn't like 50% means you can just re-run it and get the right answer the second or third time; failures aren't independent like that.
@cube_flipper@ApriiSR This doesn't pay the cost of specifying how floating point arith works, it just searches for all valid causal embeddings of a computation in reality (result of search finds floating point arith, but this isn't paid for in the desc. length).
@cube_flipper@ApriiSR So you take the causal network (of physical reality). You're allowed to (a) cut pieces of reality out, (b) merge nodes together (take their cross prod) and (c) coarse-grain vars (eg wire energy -> 1 or 0). Goal: transform into causal graph of target computation.
"whatever you now find weird, ugly, uncomfortable and nasty about a new medium will surely become its signature. CD distortion, the jitteriness of digital video, the crap sound of 8-bit - all of these will be cherished and emulated as soon as they can be avoided."
- brian eno
i gave a talk at lighthaven recently and @abramdemski was sufficiently interested in an argument that i gave – solomonoff induction means that we should place a higher prior on theories of consciousness which propose simpler mappings between physics and phenomenal consciousness, such as the ones proposed by physicalist theories of consciousness – that he wrote a lesswrong article about it.
he takes the negative, but also does a more thorough/precise job of presenting the argument than i probably would have, so, endorsed. check it out: https://t.co/COq88ynQVY
@TheZvi I know you'll be writing about this, so, my experience trying ChatGPT 5.2 today: I tested Opus 4.5, Gemini 3, and ChatGPT 5.2 on a tricky Agent Foundations problem today (trying to improve on Geometric UDT). 5.2 confidently asserts total math BS. Opus best, Gemini 3 close
@0xMims Eh. The most recent models don't feel like a big jump mainly because the models are coming out more frequently. Things continue to improve at about the same pace in graphs I've seen. The big question is what THAT means.
@dantheshive Today's "you just don't understand sportsmanship" joke reminds me of section 2.1 of this: https://t.co/RQbtrjj8rT
TLDR: when put in crowded spaces like playgrounds, children prefer dominance games with winners/losers; when roaming free, play is much more cooperative