Minichan

Topic: Debates around protests

Anonymous A started this discussion #138,662

I know the protests are getting out of hand in places but the anti-protest rhetoric I'm seeing is wild. I wonder how much of it is government bots. How ironic would it be to see Europe become fascist?

Anonymous B joined in and replied with this , 10 minutes later[^] [v] #1,466,263

Raping negroids

Anonymous C joined in and replied with this , 25 minutes later, 36 minutes after the original post[^] [v] #1,466,264

How many people have even died? Hundreds? Thousands? Definitely 1 right? Regardless, even the destruction itself is unnecessary but I hardly am there to know what it's all about. And frankly, compared to the war in Ukraine, it's a nothing burger

Anonymous D joined in and replied with this , 50 minutes later, 1 hour after the original post[^] [v] #1,466,268

@previous (C)
This right here. There's been around 750,000 casualties in the Ukraine war.

Anything less than that? I could give a shit. If Paris sees the streets filled with bodies, miss me with the complaints until it hits at least 1 mil.

Live&Lemon joined in and replied with this , 51 minutes later, 2 hours after the original post[^] [v] #1,466,273

@OP

> I know the protests are getting out of hand in places but the anti-protest rhetoric I'm seeing is wild. I wonder how much of it is government bots. How ironic would it be to see Europe become fascist?

One of the most dystopian things about the modern world is there’s no way to know whether most of the stuff you read on the internet is from real people. The worst part is, if you can nudge a million people in some direction, enough of the population is gullible enough it will become real.

Live&Lemon double-posted this , 1 minute later, 2 hours after the original post[^] [v] #1,466,274

I think, it’s not unlikely that as the internet becomes more and more polluted with AI slop and the algorithms get worse and worse, google search is barely even usable anymore, the results were probably more relevant 10 years ago, people are going to start taking their frustrations out into the real world, because on the internet, you might as well be talking to a brick wall.

Anonymous F joined in and replied with this , 1 day later, 1 day after the original post[^] [v] #1,466,749

Tiananmen 1989.04.06

Doc !mjoQJhzjwM joined in and replied with this , 9 hours later, 1 day after the original post[^] [v] #1,466,889

@1,466,274 (Live&Lemon)
That just gave me an idea to diffuse and deescalate online discourse.

just respond with non-specific, broad and generic AI slop answers. If they ask for proofs to back up my claims I’ll draft false statistics.

Make it so unbearable and unreasonable for those that want to argue that they would rather argue with the clouds or a wall. Effectively making them give up.

Lie, deny and falsify.

Live&Lemom joined in and replied with this , 1 hour later, 1 day after the original post[^] [v] #1,466,904

@previous (Doc !mjoQJhzjwM)
I’m not denying that AI is impressive, but I do not understand how some people talk to an LLM like it’s their friend. I can only use it as a question answering machine. If I try and talk to it like it’s a person it just pisses me off because it sounds exactly like what it is, a machine that says the most statistically likely thing, and I just find that infuriating. If I say something it says what I think it was going to say, and it feels so pointless.

Live&Lemon double-posted this , 37 seconds later, 1 day after the original post[^] [v] #1,466,905

Clearly this will work on some people, but some people must have a different psychology so they actually enjoy being fed AI slop.

Live&Lemon triple-posted this , 1 minute later, 1 day after the original post[^] [v] #1,466,906

I’ve also noticed, if you study a subject, sometimes common knowledge explanations are actually completely wrong and have been scientifically disproven, but LLMs spit out the most statistically likely thing, so they still spit out things that are scientifically inaccurate.

(Edited 10 seconds later.)

Anonymous I joined in and replied with this , 3 hours later, 2 days after the original post[^] [v] #1,466,991

@previous (Live&Lemon)
Just like humans.

I've met nurses that insist vaccines can give you a heart attack, and got waivers for their kids. They didn't learn that in nursing school.

Meet enough people and you will learn that the brain is just as prone to flaws.

The key difference is that I've seen rudimentary LLMs get much smarter and more accurate over a few years, but human evolution is much slower.

Live&Lemons joined in and replied with this , 55 minutes later, 2 days after the original post[^] [v] #1,466,996

@previous (I)
Not everyone is incompetent. LLMs output the most statistically likely response. This is a problem that will never get better, if most people have a false belief, the LLM will always repeat it, because that false belief will be the most common thing in its training data.

Live&Lemons double-posted this , 1 minute later, 2 days after the original post[^] [v] #1,466,997

Until LLMs start training on their own output because most of the internet will be (if it isn’t already) LLM generated text. Then the LLMs will never evolve because they’ll just get stuck in a loop.

Anonymous A (OP) replied with this , 2 minutes later, 2 days after the original post[^] [v] #1,466,999

@previous (Live&Lemons)
Wrong

Anonymous I replied with this , 1 hour later, 2 days after the original post[^] [v] #1,467,043

@1,466,996 (Live&Lemons)
No, they can refine the training material and remove bad training data and add better.

They can also use reinforcement learning to detrain bad responses and improve better ones.

And people can just use whatever model happened to be trained best, and once it's made once you have it forever.

Live&Lemon joined in and replied with this , 3 minutes later, 2 days after the original post[^] [v] #1,467,045

@previous (I)
Neat, but it still gives me factually false information.

Anonymous I replied with this , 31 seconds later, 2 days after the original post[^] [v] #1,467,046

@1,466,997 (Live&Lemons)
That's also been done.

For now, RLHF (Reinforcement Learning from Human Feedback) is an effective method.

You take a model with imperfect data, and have a group (could be hired professionals, or could be user base depending on budget) and have them rate answers good or bad. That data is used to change the answers the LLM returns.

There are many ways that ready exist to do this, yet you're claiming "This is a problem that will never get better".

You made dramatic absolutist claims that can be easily debunked if you do 5 minutes of research.

You're ideologically opposed to AI, and each time you make up some BS that gets proven wrong you switch to something else, which is just more evidence you are starting with a conclusion and looking for evidence rather than looking at the evidence and coming to a conclusion.

Anonymous I double-posted this , 47 seconds later, 2 days after the original post[^] [v] #1,467,047

@1,467,045 (Live&Lemon)
Yeah, no one is claiming LLMs are perfect, I was responding your absurd claim that "This is a problem that will never get better".

RLHF is improving LLMs every day.

(Edited 10 seconds later.)

Anonymous K replied with this , 15 seconds later, 2 days after the original post[^] [v] #1,467,048

@1,467,046 (I)
Have you ever asked an LLM why birds have hollow bones?

Anonymous K double-posted this , 20 seconds later, 2 days after the original post[^] [v] #1,467,049

It will give you an answer, it’s completely factually wrong.

Anonymous K triple-posted this , 1 minute later, 2 days after the original post[^] [v] #1,467,050

For example ask ChatGPT why birds have hollow bones. It will say because it helps birds fly. That’s completely wrong.

Anonymous I replied with this , 1 minute later, 2 days after the original post[^] [v] #1,467,052

@1,467,048 (K)
@1,467,049 (K)

Ok? What does that have to do with what we're talking about? I already said LLMs aren't perfect.

You started with the claim that these will never get better, but there are many examples of questions that stumped LLMs that are now fixed (like the strawberry question).

At some point either better data or RLHF will fix whatever bird question it's struggling with now.

You're switching to LLMs making mistakes because what you said originally is BS, and you're acting like anyone is saying LLMs get all their answers wrong. No one is saying that, you're arguing against a strawman to save face as usual.

Anonymous I double-posted this , 41 seconds later, 2 days after the original post[^] [v] #1,467,053

@1,467,050 (K)

What's the point in bringing that up if everyone already knows LLMs fail certain questions?

Who are you arguing against here?

Anonymous K replied with this , 7 seconds later, 2 days after the original post[^] [v] #1,467,054

@1,467,052 (I)
No it won’t. LLMs output the statistically most likely response. If most of the internet is wrong, the LLM will be wrong, and that won’t ever change.

Anonymous I replied with this , 1 minute later, 2 days after the original post[^] [v] #1,467,056

@previous (K)
I already explained several times how they are improving that.

Ask ChatGPT about bird bones, and then thumbs down the bad response. Congratulations, you are helping it avoid the mistake in the future.

Anonymous K replied with this , 1 minute later, 2 days after the original post[^] [v] #1,467,057

@previous (I)
You do realize that ChatGPT isn’t actually intelligent, the quality of its output is determined by the quality of its input, therefore it’s only regurgitating things and isn’t actually thinking.

Anonymous K double-posted this , 32 seconds later, 2 days after the original post[^] [v] #1,467,058

The output will always be worse than the input, so if it’s trained on its own output, the model will get worse not better.

Anonymous K triple-posted this , 15 minutes later, 2 days after the original post[^] [v] #1,467,061

@1,467,052 (I)

> You started with the claim that these will never get better, but there are many examples of questions that stumped LLMs that are now fixed (like the strawberry question).

Technically that problem was never actually fixed using LLMs, LLMs are still based on tokens not characters so the LLM still doesn’t actually know how many r’s are in strawberry, it just calls a tool that counts the number of r’s in strawberry, but the language model itself is incapable of determining that on its own.

Anonymous I replied with this , 1 hour later, 2 days after the original post[^] [v] #1,467,066

@1,467,057 (K)
Sounds exactly like humans.

Anonymous I double-posted this , 1 minute later, 2 days after the original post[^] [v] #1,467,067

@1,467,058 (K)
It's one thing to say the output is as good as the inputs, but worse? What is that based on?

And really if it's the same as the inputs, and then you use RLHF that will improve it to be better than an unrefined model from the same data, i.e. better.

Anonymous I triple-posted this , 1 minute later, 2 days after the original post[^] [v] #1,467,069

@1,467,061 (K)
Almost like these AI aren't just LLMs, but a complex system that is much more than just an LLM.

Which is still true even if the companies are basing the core of the product around a central LLM.

Anonymous L joined in and replied with this , 27 minutes later, 2 days after the original post[^] [v] #1,467,071

@previous (I)
counting the number of letters in a string isn’t complex

Anonymous L double-posted this , 46 seconds later, 2 days after the original post[^] [v] #1,467,072

@1,467,066 (I)

> Sounds exactly like humans.

Anonymous I replied with this , 5 hours later, 2 days after the original post[^] [v] #1,467,173

@previous (L)
I've met people that can repeat sentences, but can't explain what basic words they used mean.

If a robot can see the world, speak in natural language, and learn over time it can be far more intelligent than someone who doesn't know what anything they say means.
:

Please familiarise yourself with the rules and markup syntax before posting.