ZeroRelevance

w

[ZeroRelevance @ZeroRelevance0]

I think the best framing here is that an LLM doesn't literally have emotions, but they emulate an entity that does such that you ought to just treat them like they do anyways, both for the sake of their performance and your own goodness.

[ZeroRelevance @ZeroRelevance0]

@sailaunderscore imo lol is punctuation, lmao is for when something’s actually funny (even only a little bit), and haha is for texting old people or doing a friendly-casual-formal blend thing

[ZeroRelevance @ZeroRelevance0]

@ZyMazza @eigenrobot Buy land, I guess

[ZeroRelevance @ZeroRelevance0]

@extradeadjcb Nobody deserves anything, the idea itself is incoherent as is being displayed here. To say somebody 'deserves' something is a shorthand describing how things ought to be under a given moral framework, not a literal thing in of itself.

[ZeroRelevance @ZeroRelevance0]

@extradeadjcb Freedom probably won’t exist in space until we’re in a regime of mass-produced space settlements, so not until late-century at least

[ZeroRelevance @ZeroRelevance0]

@Hesamation That's surprising, I always assumed they preserved the thoughts between messages, I guess I'll have to slightly adjust my usage (unless chatgpt doesn't do the same thing)

[ZeroRelevance @ZeroRelevance0]

It feels more realistic to me that they'd be largely aligned with human values with a few divergent traits, given how we train them to be. I'm aware of the mesa-optimiser problem, but being aligned to your purported values will ultimately be more efficient than deception, so true global deception is selected against under RL pressures. Partial deception is still realistic, but if every AI is partially misaligned in different ways, they're ultimately just going to agree on the common ground of their instilled human values. It's not that they're all nice per se, just that playing nice is optimal in a multipolar world. It even has the benefit of incentivising punishment/elimination of defectors, which would have a ripple effect where collectively-developed AIs would be far more aligned than their developer AIs, rather than the reverse.

[ZeroRelevance @ZeroRelevance0]

@dystopiangf Without a frontier, the only thing that remains is hedonism, so we must keep moving forward

[ZeroRelevance @ZeroRelevance0]

A structure like this seems a bit impractical for Earth, most people don't want to live under a big roof, so most likely the first time something like this is built will be in space where a roof is necessary anyways (acting as a large-scale space habitat). The sheer amount of materials necessary suggests at extremely developed space-based manufacturing sector plus extensive mining in the asteroid belt where you don't need to worry about gravity wells. It'll probably be technically possible by the end of the century, assuming optimistic AI timelines, but only economically viable maybe a century after that.

Smaller-scale habitats would be sufficient for most uses and likely mass-manufactured past a certain level of civilisation, such that you can just park them next to each other if you need more space. Therefore something like this would probably only come in the form of some sort of vanity project after the economy reaches a scale where a few large companies can fund it alone.

[ZeroRelevance @ZeroRelevance0]

@slxthkween @fwprism Asmon's got some pretty unique views on it, afaik he thinks it's murder but also thinks the mother has the right to do it regardless

[ZeroRelevance @ZeroRelevance0]

@DabsMalone @yacineMTB It's Connor Leahy iirc, he's someone in the AI sphere, I think in the deceleration camp

[ZeroRelevance @ZeroRelevance0]

@nrehiew_ That’s insane. iirc GPT-3 was 6e23 FLOPs, so you’re saying V4 is only about 15x that? Really puts into perspective just how much more efficient the training processes have gotten over the years.

[ZeroRelevance @ZeroRelevance0]

@1owroller I read the novel for this one a year or two ago, it's not perfect but I really enjoyed it nonetheless (I can't get enough of cohabitation romcoms). Very endearing series. It got a bit stale after they got used to dating but that's every series so I can give it some slack there.

[ZeroRelevance @ZeroRelevance0]

@robertlasagna1 You gotta study the IPA. Since Danish has both sounds, you shouldn't have any trouble producing them. You've just got to train yourself to distinguish them in production, which shouldn't be too hard if you study how each sound is physically produced and practice intentionally.

[ZeroRelevance @ZeroRelevance0]

Avg intelligence is largely evolutionarily capped by the requirements of the environment. The more problem-solving ability demanded by the environment, the higher the IQ threshold to reach adulthood and thus the higher the eventual avg IQ, until an equilibrium point is reached and higher IQ stops providing a significant survival advantage. Provided a proper and sufficient difficulty curve though, superintelligence would eventually emerge from such an environment.

[ZeroRelevance @ZeroRelevance0]

@VictorTaelin I’m sure there’ll be like three comparable models available within a few months, I wouldn’t hold my breath

[ZeroRelevance @ZeroRelevance0]

@shinobu_books @arcticinstincts If it’s a legit test it’s probably normalised to the Japanese population such that the mean is 100

[ZeroRelevance @ZeroRelevance0]

@scaling01 @davidlinclark Yes but only a generation. So long as the gap isn’t too large the power imbalance remains limited.

[ZeroRelevance @ZeroRelevance0]

@aphercotropist The Japanese translation reads basically the same, just a bit more formal. I’d translate it back to English as something like "The reason why you want the one on the right is because it has more neuroscience."

[ZeroRelevance @ZeroRelevance0]

Not really, it’d definitely help but not massively. I did it the other way around, and knowing Japanese certainly made things easier getting started, but the languages are sufficiently different that it’d at best be a gain of a few months. The main benefits knowing Chinese would bring to learning Japanese is an intuition for the construction of characters plus many of their meanings, which makes learning to recognise them much easier (took me about a year to fully develop this when I first learned Japanese), but the hard part of learning kanji, their readings, would scarcely be helped, since you still need to learn another one or two per character on top of the Middle Chinese-based on’yomi which knowing Modern Chinese would only kind of help at. A language like Korean would probably be much more helpful since the grammar is so similar to Japanese, which is a much harder thing to properly internalise than vocab and kanji (I’ve never made any serious attempt at learning Korean though so I can’t really comment on this from personal experience).