Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[silverspawn]

The other thing is that, even if this model is correct AND the "exponential work/logarithmic returns" part is also true, it's still a separate question how difficult it is to build this capability (of having an internal state that is updated) into LLMs. But I think it's at least plausible that it's extremely bloody difficult. You either have to do it without understanding the representation at all -- or you have to wait until interpretability has caught up, which (a) will take forever, probably 50+ years, and (b) would be awesome since interpretability is fantastic for safety.

I feel like my confidence in LLMs not scaling increased, but also I previously said 15% which seems too low. So maybe now it's 15% and previously it should have been more like 25%.

[faust]

Agree about his background and the testimony. But, his personality didn't strike me as someone who's dishonest. (You presumably disagree.) That was an important factor for me. He seemed like someone who had integrity.

This is also a major reason why I like the show even though he's guilty. It doesn't feel like they constructed a cheap, easily hate-able villain.

Well, that is an interesting perspective. I detested the man all the way through, even before the rape accusation ever came up.

And I think the show explicitly tells us he's dishonest? His mum is shown talking about how he used to brazenly cheat at games as a child, and at one point his wife says something to the effect of "I know you lie to people all the time, but I thought at least I was getting the real you".

[silverspawn]

not scaling to AGI, that is. My model totally permits that GPT-5 will be built and beat GPT-4 on all of the metrics.

[silverspawn]

Well, that is an interesting perspective. I detested the man all the way through, even before the rape accusation ever came up.

And I think the show explicitly tells us he's dishonest? His mum is shown talking about how he used to brazenly cheat at games as a child, and at one point his wife says something to the effect of "I know you lie to people all the time, but I thought at least I was getting the real you".

I don't really remember this but your interpretation was certainly more accurate.

[silverspawn]

Graduation (2016)

Aeally amazing film about a family on Romania, and about the corruption in the country. Either a strong 8 or a light 9. it's such a mature treatment of the issue.

I thought while watching that it reminded me of 4 Months, 3 Weeks, and 2 Days based on how authentic it felt, and yeah it's from the same director. I gotta watch his remaining stuff now bc both of these were incredible. All the characters feel so real.

[silverspawn]

Rewatched The Handmaiden. Which I gave an 8 last time. Very silly of me. Bc pretty much everything about this movie is amazing. The acting is amazing, the script is amazing, the music is amazing, and it's stunningly beautiful. It has a torture scene which I didn't like, but it's stupid to deduct points for that. Didn't do it for GoT, either.

It also has about 10 minutes worth of sex scenes, so there is that. But man the story is so good.

Trivia: according to Wikipedia, "Kim Tae-ri [the actress playing the main character] was selected from 1,500 candidates to play the role."

Anyway, if The Prestige is a 10, then so is this movie, so I'm adjusting the score. 10/10 it is.

[silverspawn]

the chess championship is so exciting!

[silverspawn]

emotion is a key variable for productivity, which means the ability to regulate your emotions is key. #IQisonlyonefacetofintelligence #lookatmeimdefendinganotheropinionthatpeopleholdforbreadreasons

[faust]

emotion is a key variable for productivity, which means the ability to regulate your emotions is key. #IQisonlyonefacetofintelligence #lookatmeimdefendinganotheropinionthatpeopleholdforbreadreasons

But what does productivity have to do with intelligence?

[silverspawn]

emotion is a key variable for productivity, which means the ability to regulate your emotions is key. #IQisonlyonefacetofintelligence #lookatmeimdefendinganotheropinionthatpeopleholdforbreadreasons

But what does productivity have to do with intelligence?

What is intelligence if not the ability to get work done? Is it even coherent to say that you "intelligent enough" to do X but can't do X because of akraisa or whatever?

Like if you consider mood to be a computational variable in what you end up doing, then getting into the right mood to solve a problem is part of the solution. So it's unclear whether you can tease out this "intelligence orthogonal of mood control" thingy, and if you could, why we should care about it.

Maybe it's coherent like, if you have high intelligence but bad mood control, then at peak you're more productive? But then again, why care about this?

[silverspawn]

Intuitively I also think they're different, but if you look at how people define intelligence -- the ability to achieve goals across a broad range of domains or whatever -- then it's not so clear that the intuition is correct

[faust]

emotion is a key variable for productivity, which means the ability to regulate your emotions is key. #IQisonlyonefacetofintelligence #lookatmeimdefendinganotheropinionthatpeopleholdforbreadreasons

But what does productivity have to do with intelligence?

What is intelligence if not the ability to get work done? Is it even coherent to say that you "intelligent enough" to do X but can't do X because of akraisa or whatever?

Like if you consider mood to be a computational variable in what you end up doing, then getting into the right mood to solve a problem is part of the solution. So it's unclear whether you can tease out this "intelligence orthogonal of mood control" thingy, and if you could, why we should care about it.

Maybe it's coherent like, if you have high intelligence but bad mood control, then at peak you're more productive? But then again, why care about this?

If intelligence is just the same as productivity it makes little sense to have two words.

Just looking at Wikipedia we have this: "More generally, it can be described as the ability to perceive or infer information, and to retain it as knowledge to be applied towards adaptive behaviors within an environment or context."

I don't think intelligence has any non-incidental relation to "getting stuff done" or "achieving goals". I don't know which people you refer to that would define it as such.

[faust]

Okay, I continued looking at the Wikipedia page and indeed this definition: "Intelligence measures an agent's ability to achieve goals in a wide range of environments" comes up. So yeah, I suppose there are people who use it.

I don't like it much. It seems like a definition by which you get significantly more stupid if someone hacks off your hand is questionable.

[silverspawn]

Okay, I continued looking at the Wikipedia page and indeed this definition: "Intelligence measures an agent's ability to achieve goals in a wide range of environments" comes up. So yeah, I suppose there are people who use it.

I don't like it much. It seems like a definition by which you get significantly more stupid if someone hacks off your hand is questionable.

Agreed with this, but you can draw a pretty sharp distinction between cognitive and physical skills, and hands only impact one of them. Whereas mood is different.

This also relates to theories of consciousness. Like  under a functionalist lens, idk how people explain mood but it's probably more like a hindrance to the algorithms you run or something. But under my model, your mood is a pretty important variable that changes how you compute stuff.

[silverspawn]

@Awaclus I'm very annoyed at how this conversation turned out hence why I left the server. Having (in my mind) wasted a lot of time with dominion and mata is sth I already consider a major failure, and obviously it gets much worse if someone suggests I wasn't good at it, which is also just objectively false by every metric. Don't want to deal with this in the future

[silverspawn]

Ironically this relates back to how important emotional control is. It'd be great if I had the ability to not get annoyed at this, but I don't.

[Awaclus]

You could play video games to help you control your emotions better!

[faust]

Okay, I continued looking at the Wikipedia page and indeed this definition: "Intelligence measures an agent's ability to achieve goals in a wide range of environments" comes up. So yeah, I suppose there are people who use it.

I don't like it much. It seems like a definition by which you get significantly more stupid if someone hacks off your hand is questionable.

Agreed with this, but you can draw a pretty sharp distinction between cognitive and physical skills, and hands only impact one of them. Whereas mood is different.

This also relates to theories of consciousness. Like  under a functionalist lens, idk how people explain mood but it's probably more like a hindrance to the algorithms you run or something. But under my model, your mood is a pretty important variable that changes how you compute stuff.

IDK, how do you draw that distinction? Losing a hand would certainly impact my ability to do math.

I would agree that mood can impact your intelligence (most obviously when you're tired or something). I might agree that mood control is a facet of intelligence. But I definitely don't think intelligence should be measured in terms of productivity.

[silverspawn]

You could play video games to help you control your emotions better!

no that also completely doesn't work. The way to learn emotional control is via mindfulness

[silverspawn]

Anyway, I tried this puzzle we were talking about:



I think the answer is just all (n,k) with k >= n.

For k < n it doesn't work because (example with n=5, k=4)

AAAABBBBBA

And k=n works because (1)  the process only does nothing when the first k symbols are the same, and (2) it cannot flip between stages indefinitely because every swap reduces the number of chains  by 1. Like in the example, we have 5 -> 4 -> 3-> 2 -> 2 -> 2

I would have to do some more work to prove this last point exactly, but I'm pretty certain it holds so if I were doing the exercise, I wouldn't spend the time on that since a proof isn't required.

I learned to do this not by playing dominion.

[silverspawn]

I would agree that mood can impact your intelligence (most obviously when you're tired or something). I might agree that mood control is a facet of intelligence. But I definitely don't think intelligence should be measured in terms of productivity.

Yeah, I don't  really think this either (as demonstrated by the missing hand)

[Awaclus]

Anyway, I tried this puzzle we were talking about:



I think the answer is just all (n,k) with k >= n.

For k < n it doesn't work because (example with n=5, k=4)

AAAABBBBBA

And k=n works because (1)  the process only does nothing when the first k symbols are the same, and (2) it cannot flip between stages indefinitely because every swap reduces the number of chains  by 1. Like in the example, we have 5 -> 4 -> 3-> 2 -> 2 -> 2

I would have to do some more work to prove this last point exactly, but I'm pretty certain it holds so if I were doing the exercise, I wouldn't spend the time on that since a proof isn't required.

It doesn't work with n = 4, k = 7, AABBAABB.

I learned to do this not by playing dominion.

Probably should have played Dominion instead!

[Awaclus]

You could play video games to help you control your emotions better!

no that also completely doesn't work. The way to learn emotional control is via mindfulness

It works super efficiently. I am probably way better than average at controlling my emotions, and while I don't really practice mindfulness that much, I do play video games where the only way to make progress is to deal with difficult emotions (i.e. horror games) and obviously that's a strong incentive to do it.

[silverspawn]

Probably should have played Dominion instead!

No then I'd have less patience and more overconfidence, both of which would make it more likely that I decide to skimp on a step (which is what heppened here, you found a mistake in the part that I said I'm not proving bc I'm sufficiently confident).

[Awaclus]

Probably should have played Dominion instead!

No then I'd have less patience and more overconfidence, both of which would make it more likely that I decide to skimp on a step (which is what heppened here, you found a mistake in the part that I said I'm not proving bc I'm sufficiently confident).

Why? You don't optimize your Dominion winrate by having less patience and more overconfidence, you do it by having more patience and the exact right amount of confidence.