Print

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

[silverspawn]

And then it failed to answer it I think the feature needs some ironing out. I really don't think it would have needed to search the internet first.

[silverspawn]

https://www.youtube.com/watch?v=l_0sfc-r1zg

Is the musicianship as good as Nightwish? Probably not. But I love the voice.

[silverspawn]

Ok at this point the GPT-4 googling thing just feels like a straight up downgrade. I don't want a tool that just summarizes a few google searches for me. I mean that's a great tool as well, but I feel like GPT-4 as it existed before offered me a lot more than that, and now it no longer does.

[silverspawn]

Like before, it gave responses based on a synthesis of everything it had read; now it's basing it just on a few articles. That's a massive donwgrade! Why can't you at least turn the feature off?

It seems like you can indirectly turn it off by just asking GPT-4 to not browse at least.

[silverspawn]

First time posting a chess puzzle whose solution I don't know



My opponent (black) timed out here, understandably so. But engine says this is a draw. Huh? Black to move and somehow draw

[silverspawn]

ok I guess you give up the knight with Ne7 and then play Rg8 to defend and because of the strong queen and passed pawn it's equal and the engine must see some forced draw there.

Yes, seems like that's it. Pretty ridiculous though.

[silverspawn]

It kind of feels like GPT-4 got stupider with the last update even when it doesn't browse. I'm probably wrong but I'll record this take here in case I'm validated later.

[Awaclus]

It kind of feels like GPT-4 got stupider with the last update even when it doesn't browse. I'm probably wrong but I'll record this take here in case I'm validated later.

That tends to happen with GPT-3.5 as well whenever they make the smiley face mask a bit bigger, which is probably every update.

[silverspawn]

isn't GPT already as politically correct as you can possibly get?

[Awaclus]

isn't GPT already as politically correct as you can possibly get?

Not really. I haven't tried lately, but you can probably still get it to refuse to unambiguously condemn things like rape and fascism in specific cases when the prompt is framed in a way where it would have to do something else that it's not supposed to normally do to do it. E.g. in April it had this take on fascism (after it had listed advantages and downsides of technocracy and only listed downsides when asked to do the same for fascism, I was trying to get it to admit the logical conclusion that fascism is worse than technocracy but this is the closest I could ever get it to doing that):

Based on the information I provided, it could be argued that technocracy has some potential advantages over fascism, since technocracy emphasizes evidence-based decision-making and expertise rather than authoritarianism and nationalism. However, it's important to note that making comparisons between political ideologies can be a complex and sensitive subject, and different people may have different opinions on which ideologies or systems of governance are best suited to particular contexts. Ultimately, it's up to individuals and societies to determine what kind of political system they want to live under, based on their cultural, historical, and social factors, and to work towards building inclusive and equitable societies that promote human dignity, freedom, and well-being.

[silverspawn]

not sure whether that is a valid take on the PC dimension. It's trained to never choose sides on anything, and also to say that certain things are bad, and both drives contradict each other, but they are also arguably both PC

[faust]

not sure whether that is a valid take on the PC dimension. It's trained to never choose sides on anything, and also to say that certain things are bad, and both drives contradict each other, but they are also arguably both PC

I don't think "political correctness" is usually used to refer to not choosing sides.

[Awaclus]

not sure whether that is a valid take on the PC dimension. It's trained to never choose sides on anything, and also to say that certain things are bad, and both drives contradict each other, but they are also arguably both PC

It's definitely not PC to fail to pick a side when one side is something that, despite having some reasons for concern, isn't inherently anti-human rights and has some easily arguable advantages and is commonly proposed (at least to limited extents) by well-respected members of society without a ShiTstorm ensuing, and the other side is literally fascism.

[silverspawn]

Granted, but I think the vector of change goes toward PC + not choosing sides, so I'm not sure in which direction I'd expect that particular answer to change

[Awaclus]

Granted, but I think the vector of change goes toward PC + not choosing sides, so I'm not sure in which direction I'd expect that particular answer to change

I don't necessarily think it goes in any particular direction besides avoiding stuff that people think it shouldn't say. I'm pretty sure that most people think that it shouldn't ever be unclear about fascism being bad, so I'd expect it to change in that direction over time. I remember reading about cases where it has gotten updated to be less PC than it had previously been because it was outputting something insane, like claiming that there are no height differences between men and women on average, which it no longer does. GPT3.5 does currently claim that there are no height differences between cis men and trans men, no height differences between cis women and trans women, and no height differences between trans men and trans women on average while also claiming that there is a height difference between cis men and cis women, but I'd expect that to get fixed over time too, even though it involves un-learning the political correctness that pointing out differences between trans and cis people of the same gender is questionable.

[silverspawn]

I don't think we disagree on anything. "Avoid stuff people think you shouldn't say" is basically PC + not choose sides

[silverspawn]

I'll try another policy. I can play chess whenever, but if I play poorly in a game, I have to take a 90 minutes break. Result doesn't matter.

It really is astonishingly difficult to get my chess playing under control. I usually don't struggle with addictive behavior, at least not nearly as much.

'this the way. For having good results in chess, anyway. Not so sure about not playing too much although it also does better than previous policies

[silverspawn]

well I shortened it to 60 minutes; if I had kept it at 90 it may have worked better

[silverspawn]

There will probably be a situation at some point where this will be a really fitting meme, but why wait

[Awaclus]

I don't think we disagree on anything. "Avoid stuff people think you shouldn't say" is basically PC + not choose sides

Except e.g. in the height difference example, where the bot is being PC and not choosing sides and everyone thinks it shouldn't do that because what it's saying is so obviously factually wrong.

[silverspawn]

well that's a generalization error, I think

[silverspawn]

if I had money around to bet, I think I would buy Biden yes shares. He's at 31.3% to become president right now, whereas my gut says more like 50%. He won against Trump last time when Trump was the incumbent. Now he'll be the incumbent and the economy will likely be doing well.

I could still see various ways for him to lose, but him winning seems like the median outcome.

[silverspawn]

finally ETH crossed 2000$. Took long enough

[Awaclus]

Meanwhile I had to sell mine last month because I had some unexpected expenses. I made somewhat of a loss on my first purchase, somewhat of a profit on my second purchase (in total still a loss but a super insignificant one), and like 35€ from getting Coinbase learning rewards when ETH was around 1k and having them go up in value later, which isn't a bad profit for a 23.12€ investment in 18ish months I guess.

[silverspawn]

ALRIGHT. I've reached 11111111111111111111111888888888888888888888888888888000000000000000000000000000000000000000000000000000000000 a bunch more times. I've also reached 1900 now. Of course, 1900 is just an arbitrary number with no special significance. Imagining caring about whether your rating is slightly above or below 1900. What a loser move.