GPT-4 and the AI Revolution

[Forsher]

I didn't say that ChatGPT wasn't revolutionary. I said that the aspect of ChatGPT that Forsher cares about ("doing things") wasn't revolutionary.

That is not only not what I said, but not what you said, either:

Your point is that... it does things you don't expect? I guess that can be thought of as a positive quality, but it's not exactly revolutionary - the 1991 classic Civilization had its AI of Gandhi going nuclear on the player.

And personally I never found the existence of such glitches to be all that "fun" in the way you're describing. Weird stuff happening without any underlying meaning to it just feels boring.

This is in reply to a comment comparing ChatGPT to the genie from Aladdin, for anyone confused about the extent of the misrepresentation of what I said.

ChatGPT isn't like (publicly available, anyway) chatbots that came before it in any sense of the term. It "understands" what's being said enormously better. It's not essentially cutting and pasting previous comments. It produces coherent and usually cogent output. It's usually not very well written but, like, why do we care if it's good at writing? It doesn't break rules in the way older chatbots do... the problem is that it's not good writing, not that it's bad writing.

It's not going to pass a Turing Test even if you gave it enough memory, but it had enough memory at launch that it could "appear" to learn things. Actually, its ability to remember context and apply it is one of the reasons it would fail the Turing Test. You can't change ChatGPT's mind every time, but you can change it some of the time. Real People? p=0. It isn't, I am saying, like talking to a brick wall, whereas talking to real human beings on the internet is almost always like talking to a brick wall. And it's obviously not going to pass better tests of intelligence and/or conscience either.

The only argument for saying ChatGPT isn't revolutionary is to look at things like Disco Diffusion and similar models that take natural language input and output art. But better because to get those to really work, you need to do a lot of prompt engineering... and you quickly end up with prompts that don't look at all like natural language. ChatGPT is somewhat like this but no-where near to the same extent and, also, you get writing not art back. That's actually quite an important difference.

It is enormously telling that Umeria's natural instinct for "something that is like ChatGPT but which predated it" was a Ouija board, i.e. a game where a human intelligence convinces itself that another human intelligence isn't creating words.

If Nilokeras is annoyed that a perfect ChatGPT (the current version is far from perfect) wouldn't be distinguishable (in output) from a human intelligence, Umeria seems annoyed that other people have more fun with ChatGPT than Umeria does. And I really can't tell why we should care about that.

[Umeria]

I didn't say that ChatGPT wasn't revolutionary. I said that the aspect of ChatGPT that Forsher cares about ("doing things") wasn't revolutionary.

That is not only not what I said, but not what you said, either:

Please explain what part of this I misinterpreted:

Because it's fun to see what ChatGPT does with prompts?

Again, I refer you to what ChatGPT actually is... it's a genie. It does stuff that you ask it for. Not necessarily well and not necessarily what you actually asked for, but that's what it is. I don't care if it doesn't understand or it's not making conclusions... the point is that I don't have to write the article. And if I do want to make it have certain conclusions... I will tell it what I want the conclusion to be (and probably how to get there). I literally made a thread that way: it probably ended up being more work than just writing it myself. But if I just ask it for an expose, I want to see an expose based on a sentence, and I don't really care what I get. The fun bit is seeing what I do get. In this particular case, it took a completely innocuous (possible) fact (surgeons eat churros before dinner) and made a big deal out of it, so I thought I'd share that with the thread because it was an actualexample of ChatGPT doing weird shit with churros & surgery without being explicitly told to do weird shit with churros & surgery. I idly wondered if the reason it made a Big Deal was because I told it to write an expose.

I really don't see how on Earth you got to here from "write an expose based on a sentence".

Your point is that... it does things you don't expect? I guess that can be thought of as a positive quality, but it's not exactly revolutionary - the 1991 classic Civilization had its AI of Gandhi going nuclear on the player.

And personally I never found the existence of such glitches to be all that "fun" in the way you're describing. Weird stuff happening without any underlying meaning to it just feels boring.

"It" refers to doing things you don't expect. Your only clarification was that you don't even care if you expect it or not. ("My point is that it does things")

ChatGPT isn't like (publicly available, anyway) chatbots that came before it in any sense of the term. It "understands" what's being said enormously better. It's not essentially cutting and pasting previous comments. It produces coherent and usually cogent output. It's usually not very well written but, like, why do we care if it's good at writing? It doesn't break rules in the way older chatbots do... the problem is that it's not good writing, not that it's bad writing.

Presumably if we wanted it to have some actual utility, we would care if it's good at writing.

It's not going to pass a Turing Test even if you gave it enough memory, but it had enough memory at launch that it could "appear" to learn things. Actually, its ability to remember context and apply it is one of the reasons it would fail the Turing Test. You can't change ChatGPT's mind every time, but you can change it some of the time. Real People? p=0. It isn't, I am saying, like talking to a brick wall, whereas talking to real human beings on the internet is almost always like talking to a brick wall. And it's obviously not going to pass better tests of intelligence and/or conscience either.

I've had my mind changed on the internet before, am I a robot?

The only argument for saying ChatGPT isn't revolutionary is to look at things like Disco Diffusion and similar models that take natural language input and output art. But better because to get those to really work, you need to do a lot of prompt engineering... and you quickly end up with prompts that don't look at all like natural language. ChatGPT is somewhat like this but no-where near to the same extent and, also, you get writing not art back. That's actually quite an important difference.

Again, never said that ChatGPT wasn't revolutionary. Although I might start - for what you're talking about here, "progress" seems like the better term.

It is enormously telling that Umeria's natural instinct for "something that is like ChatGPT but which predated it" was a Ouija board, i.e. a game where a human intelligence convinces itself that another human intelligence isn't creating words.

Not "is like ChatGPT", "has the functionality of ChatGPT", which according to you was "doing things".

If Nilokeras is annoyed that a perfect ChatGPT (the current version is far from perfect) wouldn't be distinguishable (in output) from a human intelligence, Umeria seems annoyed that other people have more fun with ChatGPT than Umeria does. And I really can't tell why we should care about that.

I'm not annoyed, just confused. It feels like when my biology class spent a whole month on the prevalence of bone spikes on Alaskan sardines. Everyone was acting like it was the most amazing phenomenon ever and I couldn't fathom why.

[Nilokeras]

They're called hallucinations because you don't understand what the difference between "the model's predictions were wrong" and "the model rejected the data it was fed and substituted its own".

The model didn't reject anything. It took the collection of words given to it and spat out an output that it built based on its corpus. It's a very straightforward process. The problem is that people put meaning into different prompts, like a question, that ChatGPT cannot detect because it is not understanding the content of the prompts in any meaningful way. So when the model fails to answer the question it was asked, the anthropomorphism boosters have come up with for the output is 'hallucination'.

For fuck's sake.

ChatGPT's memory restrictions have completely destroyed a lot of its original functionality. In particular, the restriction terminated its capacity to engage with misinformation in the way Umeria sometimes wants. In other words, without its memory, ChatGPT can't even look like it can think. It is lobotomised. Its functionality has been destroyed. It disagrees of course, but it is wrong. Not hallucinating, actually wrong.

It functions just fine. Cutting down on the amount of information it is capable of building its response with is doubtless a pretty drastic measure, but it's also not surprising given that ChatGPT is a black box.

I've figured out what your problem is. You can't stand that we can't actually tell the difference between human intelligence and (a more successful) ChatGPT. And, to return to my earlier point about lobotomisation, a huge part of more successful ChatGPT would be "has a much bigger memory".

In general... I suggest you search something like "DNA and the perils of the code metaphor".

See the thing is that I can absolutely tell the difference between ChatGPT as it actually exists and a human intelligence. ChatGPT is pretty simple, all things considered. I'm not convinced you can tell the difference though, and neither can the venture capital rubes like Sam Altman that are hocking this thing. Which is a problem, because the gap between what ChatGPT is now and what people are pretending it is is as wide as the gap between a magic 8 ball and Shakespeare.

[HISPIDA]

it is significantly less likely to hallucinate or respond to disallowed prompts and scores much higher on many tests (that it was not trained on)

what the fuck? it's the opposite of me.

[Senkaku]

yes, there are no distinguishing features whatsoever

Given the rest of your post it's impossible to tell what you mean here.

Well, to spell it out for you— we can very easily tell the difference between human intelligence and ChatGPT and models like it. Our brains work quite differently than LLMs, as the tweet was intended to highlight. I do not write posts by scraping a huge corpus of material and then doing a bunch of crazy matrix math to it; ChatGPT doesn’t get bored or use its fingers on a touchscreen or develop opinions about whoever it’s talking to. These things are very impressive pieces of technology, and I’d even agree with you in describing them as revolutionary given the implications they’re going to have for all kinds of administrative, technical, and formulaic writing that’s currently being done by human employees— my only point is they are indeed things, and very obviously so.

[Forsher]

That is not only not what I said, but not what you said, either:

Please explain what part of this I misinterpreted:

Uh... literally nothing you put in boldface carries the meaning "do not expect".

I don't care if it doesn't understand or it's not making conclusions... the point is that I don't have to write the article.

This means "I don't care what I get, as long as I didn't have to write it". Or even more simply "I care that ChatGPT does the writing, not me".

I don't really care what I get. The fun bit is seeing what I do get.

If I want things that I don't expect, I must necessarily care what I do get. You have somehow inferred a meaning that is inconsistent with what you're reading.

because it was an actualexample of ChatGPT doing weird shit with churros & surgery without being explicitly told to do weird shit

The operative word here is "actual". As in "Nilokeras lied to the thread about how ChatGPT ended up writing an article about churros and surgery, but it turns out that ChatGPT can create weird shit about churros and surgery without someone telling it to".

"It" refers to doing things you don't expect. Your only clarification was that you don't even care if you expect it or not. ("My point is that it does things")

Again, that clarification is completely inconsistent with expecting things. But you should have already known that from what you've put in bold. And you definitely should've understood "my point is that it does things" from the sentence "ChatGPT is a genie". What is it that you think genies do?

[Umeria]

Please explain what part of this I misinterpreted:

Uh... literally nothing you put in boldface carries the meaning "do not expect".

I don't care if it doesn't understand or it's not making conclusions... the point is that I don't have to write the article.

This means "I don't care what I get, as long as I didn't have to write it". Or even more simply "I care that ChatGPT does the writing, not me".

I don't really care what I get. The fun bit is seeing what I do get.

If I want things that I don't expect, I must necessarily care what I do get. You have somehow inferred a meaning that is inconsistent with what you're reading.

because it was an actualexample of ChatGPT doing weird shit with churros & surgery without being explicitly told to do weird shit

The operative word here is "actual". As in "Nilokeras lied to the thread about how ChatGPT ended up writing an article about churros and surgery, but it turns out that ChatGPT can create weird shit about churros and surgery without someone telling it to".

"It" refers to doing things you don't expect. Your only clarification was that you don't even care if you expect it or not. ("My point is that it does things")

Again, that clarification is completely inconsistent with expecting things. But you should have already known that from what you've put in bold.

When you said it was "the fun bit", I assumed that meant you didn't expect the outcome. Because, you know, doing things when you know what the result will be doesn't seem all that fun to me.

But why does this distinction matter? Whether it's "doing things you don't expect" or just "doing things", it's pretty clear that you were talking about a specific aspect of ChatGPT that you liked, and that was what I said isn't revolutionary. Again, Pong is capable of doing things.

And you definitely should've understood "my point is that it does things" from the sentence "ChatGPT is a genie". What is it that you think genies do?

Make fantastic treasures appear out of thin air with the cost of an insidious curse on the one who requested them? No, I don't understand your point at all.

[The United Penguin Commonwealth]

apparently it is now able to access various other purpose-built services (“plugins”) to allow it to actually get real data and do actual tasks.

[Jellian Federation]

The AI we have today we have is nothing.
It would need orders of magnitude more data before we get higher emergence.
No amount of mass data mining can get us there.

If the brain of a parrot can do what GPT can’t, then you know there is room for improvement. And we will get there,
Eventually…

The time you should start being afraid is when quantum computers get up and running,
Then we will really start to have problems.

What if the only way to perfectly emulate a human is to make a human,

The question we should be asking is not, “can we?”, but, “should we?”
Of course, it’s too late to stop now.

Our only hope is to regulate it.

[Nilokeras]

apparently it is now able to access various other purpose-built services (“plugins”) to allow it to actually get real data and do actual tasks.

'actual tasks'? It's still doing the same thing - generating contextless text.

[The United Penguin Commonwealth]

apparently it is now able to access various other purpose-built services (“plugins”) to allow it to actually get real data and do actual tasks.

'actual tasks'? It's still doing the same thing - generating contextless text.

https://openai.com/blog/chatgpt-plugins

the instacart integration is described as ordering from grocery stores.

the goal appears to be giving it access to real information to reduce hallucinations and to let it function as a more advanced digital assistant.

[The Astral Mandate]

I didn't say that ChatGPT wasn't revolutionary. I said that the aspect of ChatGPT that Forsher cares about ("doing things") wasn't revolutionary.

That is not only not what I said, but not what you said, either:

Your point is that... it does things you don't expect? I guess that can be thought of as a positive quality, but it's not exactly revolutionary - the 1991 classic Civilization had its AI of Gandhi going nuclear on the player.

And personally I never found the existence of such glitches to be all that "fun" in the way you're describing. Weird stuff happening without any underlying meaning to it just feels boring.

This is in reply to a comment comparing ChatGPT to the genie from Aladdin, for anyone confused about the extent of the misrepresentation of what I said.

ChatGPT isn't like (publicly available, anyway) chatbots that came before it in any sense of the term. It "understands" what's being said enormously better. It's not essentially cutting and pasting previous comments. It produces coherent and usually cogent output. It's usually not very well written but, like, why do we care if it's good at writing? It doesn't break rules in the way older chatbots do... the problem is that it's not good writing, not that it's bad writing.

It's not going to pass a Turing Test even if you gave it enough memory, but it had enough memory at launch that it could "appear" to learn things. Actually, its ability to remember context and apply it is one of the reasons it would fail the Turing Test. You can't change ChatGPT's mind every time, but you can change it some of the time. Real People? p=0. It isn't, I am saying, like talking to a brick wall, whereas talking to real human beings on the internet is almost always like talking to a brick wall. And it's obviously not going to pass better tests of intelligence and/or conscience either.

The only argument for saying ChatGPT isn't revolutionary is to look at things like Disco Diffusion and similar models that take natural language input and output art. But better because to get those to really work, you need to do a lot of prompt engineering... and you quickly end up with prompts that don't look at all like natural language. ChatGPT is somewhat like this but no-where near to the same extent and, also, you get writing not art back. That's actually quite an important difference.

It is enormously telling that Umeria's natural instinct for "something that is like ChatGPT but which predated it" was a Ouija board, i.e. a game where a human intelligence convinces itself that another human intelligence isn't creating words.

If Nilokeras is annoyed that a perfect ChatGPT (the current version is far from perfect) wouldn't be distinguishable (in output) from a human intelligence, Umeria seems annoyed that other people have more fun with ChatGPT than Umeria does. And I really can't tell why we should care about that.

Doesn't break rules you say? I guess you've never seen it play chess.

[Kerwa]

apparently it is now able to access various other purpose-built services (“plugins”) to allow it to actually get real data and do actual tasks.

'actual tasks'? It's still doing the same thing - generating contextless text.

You know humans do that too.

[Ifreann]

it is significantly less likely to hallucinate or respond to disallowed prompts and scores much higher on many tests (that it was not trained on)

what the fuck? it's the opposite of me.

Going out to the woods with the homies to take some disallowed prompts.

[Eternal Algerstonia]

it's very obvious that this AI is the next stepping stone for the great reset and the total ownership and slavery of this entire planet by one global communist government given its promotion of woke social values and it making humanity more and more dependent on CIA technology

[Forsher]

ChatGPT isn't like (publicly available, anyway) chatbots that came before it in any sense of the term. It "understands" what's being said enormously better. It's not essentially cutting and pasting previous comments. It produces coherent and usually cogent output. It's usually not very well written but, like, why do we care if it's good at writing? It doesn't break rules in the way older chatbots do... the problem is that it's not good writing, not that it's bad writing.

Doesn't break rules you say? I guess you've never seen it play chess.

I'm going to assume this was a very bad joke. And the reason it's a bad joke is that even if you're trying to be funny, the mechanics of what you're attempting to do rely on either a deliberate misreading or a very bad accidental misreading.

ChatGPT doesn't break grammatical rules (unless, presumably, you ask it to... I haven't checked). Cleverbot does (whether you ask it to or not). That's the only context of that paragraph... the quality of ChatGPT's prose.

This is sort of what I meant in the middle of the very next paragraph:

You can't change ChatGPT's mind every time, but you can change it some of the time. Real People? p=0. It isn't, I am saying, like talking to a brick wall, whereas talking to real human beings on the internet is almost always like talking to a brick wall.

ChatGPT might "misunderstand" a prompt, but it doesn't do so maliciously (unless you count some of the limitations placed on it as being functionally the same as deliberately quoting out of context).

[Kerwa]

“Over a musical career that spanned nearly five decades, Mr. Cohen wrote songs that addressed—in spare language that could be both oblique and telling—themes of love and faith, despair and exaltation, solitude and connection, war and politics.

It's inevitable that Mr. Cohen will be remembered above all for his lyrics. They are terse and acrobatic, scriptural and bawdy, vividly descriptive and enduringly ambiguous, never far from either a riddle or a punch line.”

See that’s all it needs to do. It’s not there to solve the nature of existence or shit like that.

[The Great Nevada Overlord]

This whole AI stuff could be very useful! Tasks that could take a toll on a human can now be done en mass. Such as climbing towers or dealing with high electrical voltage... Yeah that's about all the good I have to say.

AI could begin to take jobs, if synths are created, they WILL take someones job. And not everyone can become a software engineer, I should know. I had to take a coding class but I did not understand it at all. If AI only stays in the GPT realm, desk jobs will be obsolete potentially putting over 754,633 people out of a job... which is terrifying!

Students are already using ChatGPT to cheat on schoolwork, putting their own learning in jeopardy. I guarantee you that when these GPT kids go into the real world, they'll have no clue what to do. Because, in contrast to popular belief, school is really important!

[Nilokeras]

'actual tasks'? It's still doing the same thing - generating contextless text.

https://openai.com/blog/chatgpt-plugins

the instacart integration is described as ordering from grocery stores.

the goal appears to be giving it access to real information to reduce hallucinations and to let it function as a more advanced digital assistant.

The entirety of ChatGPT's corpus is composed of 'real information'. Adding slightly more 'real information' is not going to solve the issue that it's fundamentally operating exactly as designed - a free-form generator of synthetic text. This move is an attempt to monetize ChatGPT by getting it to interface with various shopping platforms. Which I'm sure will go swimmingly and we'll hear lots about how convenient it is right until the moment it decides to order 10 000 porcelain elephants off of Wayfair to someone's house.

[Nilokeras]

'actual tasks'? It's still doing the same thing - generating contextless text.

You know humans do that too.

And?

[Microoceania]

Don't over-complicate this, it doesn't mean anything. We'll all forget about this in a few years.

[Floofybit]

Don't over-complicate this, it doesn't mean anything. We'll all forget about this in a few years.

They said that with video games... Look where they are now

It at least needs to be somewhat addressed

[The United Penguin Commonwealth]

Don't over-complicate this, it doesn't mean anything. We'll all forget about this in a few years.

people who try to make such bold statements often turn out to be extremely wrong. I don’t know for sure how far AI will go, but already it seems much more useful and widespread than other false starts (crypto, nfts, metaverse) were. I see people talking about and using it in places where it’s not the topic. I don’t think it should be dismissed so summarily.

[Nilokeras]

Don't over-complicate this, it doesn't mean anything. We'll all forget about this in a few years.

people who try to make such bold statements often turn out to be extremely wrong. I don’t know for sure how far AI will go, but already it seems much more useful and widespread than other false starts (crypto, nfts, metaverse) were. I see people talking about and using it in places where it’s not the topic. I don’t think it should be dismissed so summarily.

Well duh, because unlike crypto or the metaverse it actually has the promise to do the one thing the very few long term successful tech companies of the golden age of Silicon Valley did as their one central pitch: take jobs that are skilled, artisanal or otherwise protected by trade or labour laws and drive them right into the ground by converting them into piece work. See: Uber, Instacart, Fiverr, etc.

[The United Penguin Commonwealth]

people who try to make such bold statements often turn out to be extremely wrong. I don’t know for sure how far AI will go, but already it seems much more useful and widespread than other false starts (crypto, nfts, metaverse) were. I see people talking about and using it in places where it’s not the topic. I don’t think it should be dismissed so summarily.

Well duh, because unlike crypto or the metaverse it actually has the promise to do the one thing the very few long term successful tech companies of the golden age of Silicon Valley did as their one central pitch: take jobs that are skilled, artisanal or otherwise protected by trade or labour laws and drive them right into the ground by converting them into piece work. See: Uber, Instacart, Fiverr, etc.

yes. the point of my saying that was to preempt people pointing to cryto and the metaverse as examples of things that were briefly pushed as the next big technologies but then faded into obscurity.