rime @rime@schelling.pt

@niplav I swear if the academics discovered what I've irreverently been doing to their concepts, they'd throw a hissy-fit.

@niplav Related: "Nash-attacks". Borrowing cousin_it's example:

> "let's say Alice has the option of cooking a pie, but must share half of it with Bob. So Alice cooks the pie and everyone is happy. Now let's give Bob the option of stealing the whole pie and frustrating Alice. Then Alice won't cook the pie because she will anticipate that Bob's optimal decision would leave her without pie and frustrate her, so everyone is worse off."

https://www.lesswrong.com/posts/EhEZoTFzys9EDmEXn/backward-reasoning-over-decision-trees?commentId=8toje8hkyNhecAN7X

@niplav Yess! I actually considered using the exact phrase "weaponized Nash equilibrium" or "weaponized switching-costs" but didn't wish to exceed my verbosity-allowance.

@niplav Oh yeah I guess that makes more sense terminologically. I got it mixed up. Thanks!

@niplav Acute Radiation Syndrome

"…and there is still much debate concerning the ethical and legal justification for the bombings," said the victorious historians.

@niplav coordination problems aren't the same as tech-tree dependency, but the switching-cost dynamic & path-dependency makes it a very metaphorical thing—solutions may generalise from one to the other.

manyfolk are annoyed at imprecise metaphors, but they're usefwl as long as you're aware of the imprecision. it's not just that solutions generalise, it's that *finding* solutions to either is easier when you have a generalised understanding of the problem-structure too.

@niplav happens with cultures too. there are many norms for which there exists alternative norms that would be globally more-optimal to adopt. if you switch alone, your behaviour is incompatible w those who now punish you for deviating; you switching successfwly depends on all the people who would punish you for it. it was possible to switch early on, when this norm was only shared among a handfwl of people. but coordinating all the people to switch all at once is very costly.

@niplav eg every popular programming language ever, intentionally or not, remain popular by virtue of being commoditised complements. i speculate that people have already invented many alternative langs that are objectively better than popular langs (eg JS) for many tasks, but there's enormous inertia to switching (aka refactoring) everything to use the better lang.

@niplav

(sorry long-reply, but it's effort to condense!)

commodification of complement: when a company starts out w ~monopoly on a tech-category X, so they try to develop popular *open source* things that depend *specifically* on X, such that anyone who tries to develop a competitor version of X have to start from near the bottom of the tech tree. if things are opensource, they mobilise enthusiastic community efforts to proliferate variations of it too, thereby locking in X's advantage.

@niplav oops, meant to add "usually"

@niplav If I like a post, it's not because I got what you meant and totally agree, but rather because I learned something new & interesting from it. I wasn't familiar with the term. Thanks!

I feel like there are other words for it too. It's another frame on all the unhappy Nash equilibria caused by high switching-costs / dyscoordination.

"the worst-case complexity depends purely on the size of the biggest dimension"

that is, we're only considering the serial complexity; we're computing the slope for each dimension in parallel and counting that as one step.

oh. searching a space of independently sorted dimensions for a single target value has worst-case complexity that's entirely independent of the number of dimensions.

that is, assuming every observation gives you the direction towards target along every dimension, just do binary search independently for all of them. the worst-case complexity depends purely on the size of the biggest dimension. prob true wrt avg case too.

…everyone knows this already, say? in my defence, rediscovering>>reading.

@niplav smart is overrated. smart is a myth. motivation is intelligence & intelligence is motivation.

smart does ~500 observations of an ordered list of 1000 items. good job, smart. have some money, have some clout.

null doesn't need to fill nobody's quota. null just halves their job with every observation they make, bc null doesn't care abt leaving a footprint—null doesn't speak business-card.

pretending things are simple is great for poetry. great for clout. but null is a myth too. :p

1. OpenAI's "release early" plan is not insane
2. LLMs will become agents because it's profitable and humans will deliberately make it into one
3. "The public" seems pretty receptive to the idea of AI risk
4. Doom is more likely to be humans plain instructing the AI to recursively self-improve, or plain destroy the world
5. Once some sentient AIs are released, humans are likely to try to torture them, and we need to set up guardrails asap.

https://www.lesswrong.com/posts/3DyXQkkkGnSgy95ex/briefly-how-i-ve-updated-since-chatgpt

> “At the time, science had declared humans unique, since we were so much better at identifying faces than any other primate. No one seemed bothered by the fact that other primates had been tested mostly on human faces rather than those of their own kind.”
-- Are We Smart Enough to Know How Smart Animals Are?

This, but in the context of Shoggoth, and how people think human benchmarks (e.g. "can it do human math?") are relevant for measuring GPT-4's true intelligence.

@niplav @empathy2000 On the one hand, yes definitely absolutely in the short-term. There are so many reasons it could be better. On the other hand, I'm old-fashioned and still yearn for us to get to a post-money world, so I'm reluctant to make the money-system better in a way that postpones that!