A thread for examples of writing that was rejected for being generated by AI.

[Bramblethorn]

Clever. You’re throwing red herrings at it.

Specifically, what I'm doing there is prompting in a way designed to expose the difference between comprehension and pattern-matching/remixing.

"Do not mention a polar bear" is something that any human understands, but which is rarely seen in text (of the sort that GPT uses for training) because it doesn't normally need to be said. In most contexts, people aren't going to talk about polar bears a whole lot.

So GPT doesn't have a lot of data to tell it that responses to "do not mention a polar bear" type questions don't usually contain the words "polar bear". But it does have data to tell it that responses to questions mentioning "polar bear" usually do contain the words "polar bear". So it puts a polar bear in the story.

This trick isn't 100% reliable - over a few trials, GPT only puts a polar bear in the story about half the time. But that's far more than if I hadn't said anything about polar bears in my prompt.

Did you ask it why it mentioned a polar bear at all after the first example?

No, I did not, because it would not be informative.

If you ask a human "why did you say that?", you can hope that they will reflect back on the thought processes that took them to the thing they said, and tell you what those processes were.

But when we ask GPT "why did you say that?", it has no such capacity to recall how it got there. (Even if it did, the honest answer would not be one that many humans could interpret.) It just assembles some words that are calculated to look something like the sort of thing a human might say in response to that query, but it's not telling you how it actually arrived at that answer. I've tried this out many times, and very often the reasons GPT gives to justify its answers are not reasons that could plausibly have led to that answer.

It will probably use it as a learning experience.

Unlikely, from what I recall of how GPT is trained. As I understand it, ChatGPT is pretty much "frozen" at a specific state of knowledge; it might occasionally get an update to address specific major issues like the napalm grandma exploit, but I don't think it's updating its model based on user inputs as a matter of course.

For instance:



*immediately start new session*

Here, just for you:

Me:
Provide a >500 word response to the forum post quoted earlier in the style of a didactic college English professor who is skeptical, dismissive, and untrusting of AI.

(It sounds like it may have copied you verbatim. https://cdn.jsdelivr.net/joypixels/assets/8.0/png/unicode/64/1f605.png) :

Frankly, it sounds like somebody needs a nap.

Perhaps it is not your intention, but you're coming across as quite testy in this discussion. While I have disagreed with things you've said, I've tried to do so politely, and I'm not aware of what I might have done to incur that attitude.

 

[AlexBailey]

Specifically, what I'm doing there is prompting in a way designed to expose the difference between comprehension and pattern-matching/remixing.

"Do not mention a polar bear" is something that any human understands, but which is rarely seen in text (of the sort that GPT uses for training) because it doesn't normally need to be said. In most contexts, people aren't going to talk about polar bears a whole lot.

So GPT doesn't have a lot of data to tell it that responses to "do not mention a polar bear" type questions don't usually contain the words "polar bear". But it does have data to tell it that responses to questions mentioning "polar bear" usually do contain the words "polar bear". So it puts a polar bear in the story.

This trick isn't 100% reliable - over a few trials, GPT only puts a polar bear in the story about half the time. But that's far more than if I hadn't said anything about polar bears in my prompt.



No, I did not, because it would not be informative.

If you ask a human "why did you say that?", you can hope that they will reflect back on the thought processes that took them to the thing they said, and tell you what those processes were.

But when we ask GPT "why did you say that?", it has no such capacity to recall how it got there. (Even if it did, the honest answer would not be one that many humans could interpret.) It just assembles some words that are calculated to look something like the sort of thing a human might say in response to that query, but it's not telling you how it actually arrived at that answer. I've tried this out many times, and very often the reasons GPT gives to justify its answers are not reasons that could plausibly have led to that answer.



Unlikely, from what I recall of how GPT is trained. As I understand it, ChatGPT is pretty much "frozen" at a specific state of knowledge; it might occasionally get an update to address specific major issues like the napalm grandma exploit, but I don't think it's updating its model based on user inputs as a matter of course.

For instance:

View attachment 2361293

*immediately start new session*


View attachment 2361294




Frankly, it sounds like somebody needs a nap.

Perhaps it is not your intention, but you're coming across as quite testy in this discussion. While I have disagreed with things you've said, I've tried to do so politely, and I'm not aware of what I might have done to incur that attitude.

I apologize for my testy attitude, I mistook ElectricBlue’s “hippie shit” comment for yours.

https://forum.literotica.com/thread...r-being-generated-by-ai.1613193/post-99095683

 

[Bramblethorn]

Going to “11” is used colloquially constantly. It has become a common use meme.

I'm aware.

Let me know your thoughts on the AI’s response to the following question in relation to your skeptical response above.

——————

Me:
Provide a >300 word detailed background and explanation of why you had the ‘writer’ mention the amps “cranked to eleven” and what they meant by it in your response in this thread that was done in the style of a ‘meth-headed rock star’?

My thoughts are that the answer GPT gave you to this question has no causal relationship to the reasons it actually used that particular expression. It is offering you some text in the shape of an answer that a human might give to that question; it's not querying its log and translating its actual processes into human-understandable terms.

For instance:



There was no such previous chat session. GPT has never told me that monkeys lived in Tasmania, and I'm not aware of any such "persistent myth or misunderstanding" on that topic. [Other than zoos etc.]

So GPT has made up a rationale for something it didn't even do. It's just as happy to make up post-facto rationales for things it *did* do. They're still bullshit.

 

[XerXesXu]

Clever. You’re throwing red herrings at it.

Prompt engineers can compose prompts to which they know the AI will give the correct answer, similarly they can compose prompts to which they believe the AI will give the wrong answer. However, AI will occasionally give the correct answer to the latter. This is an inexplicable, emergent property that frightens some people because it hints at the possibility that AI may be able to exhibit intelligence that goes beyond its training set. It's like that moment when you realise your infant can conceive of things beyond that to which he/she has been exposed. Smart Kid. And our smart kids will inherit the Earth.

 

[AlexBailey]

I'm aware.



My thoughts are that the answer GPT gave you to this question has no causal relationship to the reasons it actually used that particular expression. It is offering you some text in the shape of an answer that a human might give to that question; it's not querying its log and translating its actual processes into human-understandable terms.

For instance:

View attachment 2361296

There was no such previous chat session. GPT has never told me that monkeys lived in Tasmania, and I'm not aware of any such "persistent myth or misunderstanding" on that topic. [Other than zoos etc.]

So GPT has made up a rationale for something it didn't even do. It's just as happy to make up post-facto rationales for things it *did* do. They're still bullshit.

In my experience with ChatGPT, the greatest factor to the quality of response is the quality of the prompt. Your prompts are intentionally misleading. It’s like painting lines and putting up signs to intentionally make a self-driving car go off of the road.



I don’t buy your argument that it was unable to analyze its use of ‘eleven’. Do you honestly believe it simply searched and sampled existing material and assembled the most likely next word then came up with response and that explanation? The rock-star version had appropriate depth of character and theme. It outlined and explained the metaphor perfectly and applied it to a novel concept from a requested perspective. How did it fail the request?

 

[ElectricBlue]

I apologize for my testy attitude, I mistook ElectricBlue’s “hippie shit” comment for yours.

https://forum.literotica.com/thread...r-being-generated-by-ai.1613193/post-99095683

Why are you fretting about that? You've already said these AI responses are from a machine, so they're not your words, and I would hope, not your actual thoughts. I'm just astonished you think the words that are appearing are a) intelligent, and b) coherent.

As for the "turn it up to eleven" meme, that's wedded to the Spinal Tap movie which is decades old. There's no novelty in that explanation - it's regurgitating the meme about the daft rock star who knows no better.

 

[Bramblethorn]

In my experience with ChatGPT, the greatest factor to the quality of response is the quality of the prompt.

I don't deny that the prompt affects quality of response. (I do think GPT advocates are a little too willing sometimes to leap to "you're not prompting it right" as an explanation for its behaviours, attributing to the prompt things that might be better explained by other factors, but that's a general observation and not relevant here.)

But the fact that it is sensitive to the prompt, in ways that a human is not, tells us something about what's happening under the hood - and about what isn't happening under the hood.

I feel you're interpreting my comments in the context of an argument that I'm not actually trying to make here. My focus in this discussion is not "can GPT give useful output?" or "how should we prompt GPT if we want useful output?" but rather "does it understand the concepts it's writing about?"

(Or more precisely, "is its behaviour consistent with the idea that it understands the concepts it's writing about?")

I do have opinions on those other topics, but I wasn't airing them here and now.

Your prompts are intentionally misleading.

They are indeed intentionally misleading to GPT. But they would not be misleading to the average literate human. That's the point of the example: GPT is not processing these in the way that a human would. The only reason I can mislead it with those prompts is because it doesn't understand what it's dealing with.

It’s like painting lines and putting up signs to intentionally make a self-driving car go off of the road.

Okay, let's run with that analogy.

If I were to paint confusing road markings and set up a bewildering array of signs, and we found that both humans and self-driving cars were prone to crashing when they encountered these, we might conclude that I'd made the situation inherently confusing.

OTOH, if I were to set up signs and markings and we found that the average human could cope with these just fine, but the self-driving car frequently crashed, that would tell us that the self-driving car is lacking some capability that the average human has.

Would that prove that self-driving cars shouldn't be allowed on the road? Not necessarily. We might just as well decide that yours truly shouldn't be allowed to mess with the road signs. But it would still tell us something about how that car works, and that may inform our understanding of what happens in less contrived scenarios.

But if the trick question is a sticking point for you, here's a more straightforward example. I don't think this could reasonably be considered a trick question?



Three things to notice here:

#1: The final answer is correct.

#2: There's a major error in the calculations along the way: 813 x 20 is not 2439 or 24390! The correct answer here would be 16260, or alternatively "1626, shifted one place to the left").

The next line is also not great - 813 x 300 should be either "243900" or "2439 shifted two places to the left", definitely not "24390". But at least it has the right digits, give or take the zeroes at the end.

#3: Because of the error in that "813 x 20" calculation, there is no way this should sum to the correct answer. 1626 + 2439 + 24390 does not equal 265851, nor does 1626 + 24390 + 243900, nor does any other interpretation of that addition, no matter what we do with the zeroes/left-shifts.

(This can be shown by noting that 265851 is a multiple of 9, and 1626 + 2439 + 24390 leaves a remainder of 6 when divided by 9, and inserting zeroes in those numbers will never change that remainder.)

And yet...it has the right answer.

This kind of thing is fairly easy to reproduce: pick two random three-digit numbers, ask ChatGPT to multiply them, and ask it to show working if it hasn't already done so. It won't always give the correct answer; it almost always gets the first couple of digits and the last couple of digits right, and it's about 50% reliable for the middle ones. Rerun until it does give the correct answer, and then check its working. Odds are very good that you'll find something like the scenario above, where it makes at least one major error along the way that should prevent it from giving the correct answer...and yet doesn't.

When this happens often, and repeatedly, the only reasonable conclusion that can be drawn is that the working it shows is not how it actually got to the answer.

(How does it get the correct answer? My suspicion is that it's learning patterns in multiplication problems: it's seen enough multiplication examples to learn that a product of 81* and 32* will almost always follow the pattern 26****, and that the product of *13 and *27 always ends in 51. That gives it 26**51, and depending on just how much it's seen in the training data it might be able to make an educated guess on the middle digits, though this is less reliable.)

One funny thing here is that if you simply ask ChatGPT to multiply 813 by 20, it will pretty reliably get the right answer. But it gets much less reliable on those simpler calculations when they appear in the context of a worked solution - another clue that it's not really doing what it says it's doing.

I don’t buy your argument that it was unable to analyze its use of ‘eleven’. Do you honestly believe it simply searched and sampled existing material and assembled the most likely next word then came up with response and that explanation? The rock-star version had appropriate depth of character and theme. It outlined and explained the metaphor perfectly and applied it to a novel concept from a requested perspective. How did it fail the request?

"Searched and sampled" isn't quite right here. ChatGPT doesn't have live internet access, nor does it store a full copy of all the data it was trained on.

Rather, it's trained ahead of time to learn patterns in that data. For instance, it can learn that "finger" often appears in close proximity to "hand", and that this "finger"-"hand" pairing often shows up in the same kind of contexts as the "toe"-"foot" pairing. This is something that the underlying methods support pretty well - for more info on how that's done, you could search on "vectorisation semantic similarity".

That gives it some capacity to remix patterns and swap out certain elements in a way that quite often leads to valid statements. If it knows that London is to England approximately as Paris is to France, and it's read stories about "Englishman runs for office, gets elected, goes to London", it can guess that maybe a story beginning with "Frenchman runs for office, gets elected" should continue with him going to Paris.

In a similar kind of way, it can learn that certain word-patterns (vocabulary, phrases, structures, ...) are more strongly associated with "12-year-old" and others with "addled rocker".

A good short description for it would be "predictive text on steroids". But it's not built as a logic engine; what logical capacity it has comes from ingesting logical statements that humans have written, and then remixing them according to the patterns it's learned, in ways that are sometimes valid and sometimes not.

I'm no expert on LLMs, but I have read papers on the underlying methods, some of which I've used myself for other applications. They can be useful and impressive, but they're designed to mimic understanding, not to actually possess it, and their behaviour is consistent with that.

 

[ElectricBlue]

(How does it get the correct answer? My suspicion is that it's learning patterns in multiplication problems: it's seen enough multiplication examples to learn that a product of 81* and 32* will almost always follow the pattern 26****, and that the product of *13 and *27 always ends in 51. That gives it 26**51, and depending on just how much it's seen in the training data it might be able to make an educated guess on the middle digits, though this is less reliable.)

One funny thing here is that if you simply ask ChatGPT to multiply 813 by 20, it will pretty reliably get the right answer. But it gets much less reliable on those simpler calculations when they appear in the context of a worked solution - another clue that it's not really doing what it says it's doing.

Or it sends it off to a calculator, then makes up bullshit in between.

 

[AlexBailey]

Why are you fretting about that? You've already said these AI responses are from a machine, so they're not your words, and I would hope, not your actual thoughts. I'm just astonished you think the words that are appearing are a) intelligent, and b) coherent.

As for the "turn it up to eleven" meme, that's wedded to the Spinal Tap movie which is decades old. There's no novelty in that explanation - it's regurgitating the meme about the daft rock star who knows no better.

Your response shows that you haven’t been following many of the posts in this thread, including my original response to your comment. No worries, perceived rudeness doesn’t really matter to the material of the conversation and it’s easy to forgive you for not following every post.


Do you, EB, have an explanation for how the AI came up with the explanation below without something that could be defined as an “understanding” of metaphor?

In the context of the response styled as a ‘meth-headed rock star,’ the phrase "amps cranked to eleven" is used to encapsulate the chaotic, intense, and over-the-top nature of the AI versus human writing dilemma. The rock star’s perspective is one of perpetual escalation and intensity, mirroring the relentless and often irrational quest for technological supremacy. The phrase perfectly captures the spirit of excess and intensity, evoking the image of a rock concert where everything is amplified to its maximum capacity, much like the ever-increasing capabilities of AI technology.

 

[AlexBailey]

I don't deny that the prompt affects quality of response. (I do think GPT advocates are a little too willing sometimes to leap to "you're not prompting it right" as an explanation for its behaviours, attributing to the prompt things that might be better explained by other factors, but that's a general observation and not relevant here.)

But the fact that it is sensitive to the prompt, in ways that a human is not, tells us something about what's happening under the hood - and about what isn't happening under the hood.

I feel you're interpreting my comments in the context of an argument that I'm not actually trying to make here. My focus in this discussion is not "can GPT give useful output?" or "how should we prompt GPT if we want useful output?" but rather "does it understand the concepts it's writing about?"

(Or more precisely, "is its behaviour consistent with the idea that it understands the concepts it's writing about?")

I do have opinions on those other topics, but I wasn't airing them here and now.



They are indeed intentionally misleading to GPT. But they would not be misleading to the average literate human. That's the point of the example: GPT is not processing these in the way that a human would. The only reason I can mislead it with those prompts is because it doesn't understand what it's dealing with.



Okay, let's run with that analogy.

If I were to paint confusing road markings and set up a bewildering array of signs, and we found that both humans and self-driving cars were prone to crashing when they encountered these, we might conclude that I'd made the situation inherently confusing.

OTOH, if I were to set up signs and markings and we found that the average human could cope with these just fine, but the self-driving car frequently crashed, that would tell us that the self-driving car is lacking some capability that the average human has.

Would that prove that self-driving cars shouldn't be allowed on the road? Not necessarily. We might just as well decide that yours truly shouldn't be allowed to mess with the road signs. But it would still tell us something about how that car works, and that may inform our understanding of what happens in less contrived scenarios.

But if the trick question is a sticking point for you, here's a more straightforward example. I don't think this could reasonably be considered a trick question?

View attachment 2361314

Three things to notice here:

#1: The final answer is correct.

#2: There's a major error in the calculations along the way: 813 x 20 is not 2439 or 24390! The correct answer here would be 16260, or alternatively "1626, shifted one place to the left").

The next line is also not great - 813 x 300 should be either "243900" or "2439 shifted two places to the left", definitely not "24390". But at least it has the right digits, give or take the zeroes at the end.

#3: Because of the error in that "813 x 20" calculation, there is no way this should sum to the correct answer. 1626 + 2439 + 24390 does not equal 265851, nor does 1626 + 24390 + 243900, nor does any other interpretation of that addition, no matter what we do with the zeroes/left-shifts.

(This can be shown by noting that 265851 is a multiple of 9, and 1626 + 2439 + 24390 leaves a remainder of 6 when divided by 9, and inserting zeroes in those numbers will never change that remainder.)

And yet...it has the right answer.

This kind of thing is fairly easy to reproduce: pick two random three-digit numbers, ask ChatGPT to multiply them, and ask it to show working if it hasn't already done so. It won't always give the correct answer; it almost always gets the first couple of digits and the last couple of digits right, and it's about 50% reliable for the middle ones. Rerun until it does give the correct answer, and then check its working. Odds are very good that you'll find something like the scenario above, where it makes at least one major error along the way that should prevent it from giving the correct answer...and yet doesn't.

When this happens often, and repeatedly, the only reasonable conclusion that can be drawn is that the working it shows is not how it actually got to the answer.

(How does it get the correct answer? My suspicion is that it's learning patterns in multiplication problems: it's seen enough multiplication examples to learn that a product of 81* and 32* will almost always follow the pattern 26****, and that the product of *13 and *27 always ends in 51. That gives it 26**51, and depending on just how much it's seen in the training data it might be able to make an educated guess on the middle digits, though this is less reliable.)

One funny thing here is that if you simply ask ChatGPT to multiply 813 by 20, it will pretty reliably get the right answer. But it gets much less reliable on those simpler calculations when they appear in the context of a worked solution - another clue that it's not really doing what it says it's doing.



"Searched and sampled" isn't quite right here. ChatGPT doesn't have live internet access, nor does it store a full copy of all the data it was trained on.

Rather, it's trained ahead of time to learn patterns in that data. For instance, it can learn that "finger" often appears in close proximity to "hand", and that this "finger"-"hand" pairing often shows up in the same kind of contexts as the "toe"-"foot" pairing. This is something that the underlying methods support pretty well - for more info on how that's done, you could search on "vectorisation semantic similarity".

That gives it some capacity to remix patterns and swap out certain elements in a way that quite often leads to valid statements. If it knows that London is to England approximately as Paris is to France, and it's read stories about "Englishman runs for office, gets elected, goes to London", it can guess that maybe a story beginning with "Frenchman runs for office, gets elected" should continue with him going to Paris.

In a similar kind of way, it can learn that certain word-patterns (vocabulary, phrases, structures, ...) are more strongly associated with "12-year-old" and others with "addled rocker".

A good short description for it would be "predictive text on steroids". But it's not built as a logic engine; what logical capacity it has comes from ingesting logical statements that humans have written, and then remixing them according to the patterns it's learned, in ways that are sometimes valid and sometimes not.

I'm no expert on LLMs, but I have read papers on the underlying methods, some of which I've used myself for other applications. They can be useful and impressive, but they're designed to mimic understanding, not to actually possess it, and their behaviour is consistent with that.

And yet we still have a situation where AI driven vehicles have a better safety record than humans do and they’re effective at getting to their destinations.

We have AI medical advice that is more accessible than human health care providers.

We have AI written text that cannot be distinguished from human work.

Meanwhile programs such as ChatGPT are continuing to develop. During this conversation I received a notification of an update that says it will now retain data, context, and preferences between threads that a user has not deleted. I find this interesting because it is now absorbing and accessing new data, not just relying on its last programming update. Perhaps this will allow it to keep better track of its processes. It certainly seems to me that it could accelerate the ‘learning’ process.

All of the ‘flaws’ you’ve described and identified are things that will likely be focused on and solved in the near future. In the short time I’ve been using it I’ve noticed a significant improvement in how it interprets prompts. I used to have to try several different approaches to get acceptable and appropriate answers. It seems to be much easier to interact with than when it was first available.


You seem to be fretting over how the sausage is made rather than the results here. Can you describe how the human mind works? How it’s logic flows and the how it associates abstract concepts?

Audiophiles used to constantly fret about digital recording, discussing over sampling and how a digital recording or sound processor could never replicate the warmth and depth of analog production. Gone are those days - for the most part. We’re now at a point where we have to acknowledge that some people have a preference for the distortion created by analog devices - but guess what? Digital equipment can now emulate that distortion. As a guitarist who once clung to the aesthetics of vacuum tube amplifiers I now can’t hear or “feel” the difference but I definitely prefer how a digital emulator can handle rough handling and temperature changes without blowing tubes along the way.

AI writing has only been widely available to the masses for a very short time and it’s already indiscernible to the point where the subject of this thread is an issue.

This conversation will sound as silly to future generations as telling kids about days before television. “What’s a television?” Kids will soon ask.

 

[Britva415]

It would probably be helpful for people to spell out what they mean when they say "understand" at each other.

It could be intended to mean "the motherfucker is sentient, and has a conscious experience of meaning," which is a position that I think one person is pushing back against and the other person probably isn't taking.

It could be intended to mean "it responds effectively, that shit works," kind of like when we say a calculator "knows" what the value of pi is and how to take a Nth root. Obviously the calculator doesn't know shit, it isn't sentient.

 

[AlexBailey]

It would probably be helpful for people to spell out what they mean when they say "understand" at each other.

It could be intended to mean "the motherfucker is sentient, and has a conscious experience of meaning," which is a position that I think one person is pushing back against and the other person probably isn't taking.

It could be intended to mean "it responds effectively, that shit works," kind of like when we say a calculator "knows" what the value of pi is and how to take a Nth root. Obviously the calculator doesn't know shit, it isn't sentient.

Thank you for posting this.

I’m sure there are some who would argue that a self driving car doesn’t “know” traffic laws or that it has to take the next left to get to the requested destination.

 

[Bramblethorn]

And yet we still have a situation where AI driven vehicles have a better safety record than humans do and they’re effective at getting to their destinations.

I have already said this, but I will say it again: I am not making a statement here about the utility of generative AI tools. I am responding specifically to your statement about "understanding".

You keep making replies that amount to "AI is useful". That's a non sequitur.

On my bookshelf I have a textbook on calculus. There are many situations in which that textbook would be a more useful and reliable reference than asking the average human, or even a mathematics professor. But nobody would suggest that the textbook is a conscious entity that "understands" calculus; it's merely an indexed a medium that preserves and indexes information created by several human beings who did understand calculus.

All of the ‘flaws’ you’ve described and identified

If you think my focus here is on "flaws" you haven't been reading what I've been saying. Yes, I've been discussing some failings of GPT, but specifically for what they tell us about the question of "understanding".

are things that will likely be focused on and solved in the near future.

Is this a statement based in specific knowledge about how this technology works, or just an extrapolation along the lines of "computer technologies often improve a lot in their first few years so I expect this will too"?

I'm sure we will continue to see improvements in some aspects of generative AI - specifically, the ones that can be achieved by taking basically the same method but throwing more resources at it, and the ones that can be achieved by making minor tweaks to it.

But you can't turn your car into a spaceship just by giving it more horsepower.

(I expect we will also see increased uptake of AI for things like health and education services, but those will be driven less by improvements in AI capabilities and more by running down the existing non-AI services. This is not a prospect that enthuses me.)

You seem to be fretting over how the sausage is made rather than the results here.

Once again: I was responding to a statement by you on the topic of "how the sausage is made". This statement:

To me this demonstrates an ability to ‘understand’ meaning and context far deeper than simple regurgitation

This is very clearly a statement about the process, not just about the results. This is what I'm responding to. Please don't make assertions about how it works and then get surprised that I might want to discuss how it works.

Can you describe how the human mind works? How it’s logic flows and the how it associates abstract concepts?

No. Nobody can; if we could, we'd be able to use that knowledge to build AIs that aren't stumped by prompts like "write a story without a polar bear".

But I don't need to be able to tell you how nuclear fusion works to say that a Mag-Lite doesn't work on the same principles as the Sun. For that, I just need to know how a Mag-Lite works, and to point at things the Sun does that a Mag-Lite can't do.

AI writing has only been widely available to the masses for a very short time and it’s already indiscernible to the point where the subject of this thread is an issue.

That phrase "AI writing" equivocates between material that's fully AI-authored, and material that's merely AI-assisted with substantial human intervention along the way.

The issue for Literotica is AI-assisted writing. If somebody is plotting out a story, then asking GPT to fill in some of the blanks, then editing it to fix any AI-generated goofs, that can be hard or impossible to distinguish from human writing.

But full AI authorship? At story length? I've yet to see any AI tool that can write a coherent Literotica-length story - let's say 5k words - without glaring problems. Every time I see some sort of "AI won a writing contest" report, it turns out to be something like "a human wrote prompts for every chapter and then went over the AI's work to fix any glaring issues". The longer a story runs, the more obvious that lack of understanding becomes.

This conversation will sound as silly to future generations as telling kids about days before television. “What’s a television?” Kids will soon ask.

Never mind television – I can remember the days before we all had flying cars, before everybody had abandoned fiat money in favour of cryptocurrency.

I’m sure there are some who would argue that a self driving car doesn’t “know” traffic laws or that it has to take the next left to get to the requested destination.

...I mean, if you had specifically asserted that a SDC knows and understands traffic laws, then yes, some people might well choose to discuss the accuracy of that assertion.

 

[Bramblethorn]

It would probably be helpful for people to spell out what they mean when they say "understand" at each other.

It could be intended to mean "the motherfucker is sentient, and has a conscious experience of meaning," which is a position that I think one person is pushing back against and the other person probably isn't taking.

It could be intended to mean "it responds effectively, that shit works," kind of like when we say a calculator "knows" what the value of pi is and how to take a Nth root. Obviously the calculator doesn't know shit, it isn't sentient.

You say "obviously", but way too many people do ascribe consciousness to these tools and to attempt to interpret their behaviour in human terms.

 

[Bramblethorn]

Or it sends it off to a calculator, then makes up bullshit in between.

A calculator would be more accurate.

(Later versions of GPT do have a "send it off to a calculator" option, probably because OpenAI got a bit embarrassed by producing a computer that was bad at math, but not the version I was talking to there.)

 

[AlexBailey]

I have already said this, but I will say it again: I am not making a statement here about the utility of generative AI tools. I am responding specifically to your statement about "understanding".

@Britva415 ‘s post was spot on. We’re going around and around about semantics.

I said it ‘understands’ - using quotes around the word and specifically in comparison to “simple regurgitation.”

To me this demonstrates an ability to ‘understand’ meaning and context far deeper than simple regurgitation

Simple regurgitation would mean that it is finding other works and taking quotes from them and putting them together for a new piece. You assessed that the AI’s ‘understanding’ was simple -“just mimicry”

"Understanding meaning and context" implies something far beyond just mimicry.

I fully agree.

If I heard a rock star saying something like that, my reaction would not be "whoa, this guy understands the meaning and context of what he's talking about". My reaction would be "this guy doesn't have a clue, he's bullshitting on a topic he doesn't understand."

Its use of “eleven” was more complex than your assessment.

I posted ChatGPT’s response (post 76) to why it used “eleven” and it gave an appropriate answer that showed it was applying it in a nuanced manner. Like a self driving car, it may not have a true sense of what is going on yet it effectively used it as a complex metaphor - not mimicry or simple regurgitation. I defy you to deny that the explanation it provided makes literary sense - thus displaying “an ability to ‘understand’ meaning and context far beyond simple regurgitation.”

I stand by my words and qualify my use of the word ‘understand’ - used in quotes, as accurate in the same way that a navigation system can “understand” that use of a detour can get you to your destination faster than getting stuck in traffic congested by an accident. It not only posted in the style I asked for - it also used a complex metaphor appropriate to that style.

I contend that although it may be using the next most likely word, the factors it uses to determine that word are broad and deep enough to be effective as “Artificial Intelligence” - as much so as an SDC, and not simply just as a plagiarism device.

The complexity of the factors used will continue to deepen - this is how the existing model will continue to develop, making the technology more useful and better able to emulate human works.

 

[Bramblethorn]

Its use of “eleven” was more complex than your assessment.

I posted ChatGPT’s response (post 76) to why it used “eleven” and it gave an appropriate answer that showed it was applying it in a nuanced manner. Like a self driving car, it may not have a true sense of what is going on yet it effectively used it as a complex metaphor - not mimicry or simple regurgitation. I defy you to deny that the explanation it provided makes literary sense - thus displaying “an ability to ‘understand’ meaning and context far beyond simple regurgitation.”

The former doesn't imply the latter.

I can produce an insightful discussion of "War and Peace" simply by going to the library, grabbing an appropriate book of literary essays, and sticking it in the photocopier. I don't even need to read it myself, beyond checking that I'm copying the right pages. Nobody would suggest that this indicates any understanding of the book on my part; it just shows that I've copied it from somebody who did.

If I was feeling more industrious, I could go through that text replacing words with synonyms, rewording sentences in ways that won't change meaning, and perhaps splicing in bits from somebody else's essay on the same topic. This may make it harder to identify my sources, but it still doesn't mean I understood the topic of "my" essay.

Now, let's examine that response.

First paragraph of the GPT text:

The phrase "cranked to eleven" originates from the 1984 mockumentary film *This Is Spinal Tap*, directed by Rob Reiner. The film follows the fictional British heavy metal band Spinal Tap on their American tour. One of the film's most iconic scenes features lead guitarist Nigel Tufnel, played by Christopher Guest, proudly showing off his amplifier that has volume dials going up to eleven, rather than the standard ten. When asked why the amp goes to eleven, Nigel responds, “Well, it’s one louder, isn’t it?” This humorous exchange has since become a popular cultural reference, symbolizing the pursuit of pushing boundaries to the absolute extreme, even when it defies logic or practical necessity.

Compare to Wikipedia:

This Is Spinal Tap (also known as This Is Spın̈al Tap: A Rockumentary by Martin Di Bergi[a]) is a 1984 American mockumentary comedy film co-written and directed by Rob Reiner (in his feature directorial debut). The film stars Christopher Guest, Michael McKean and Harry Shearer as members of the heavy metal band Spinal Tap, who are characterized as "one of England's loudest bands".[3][4] Reiner plays Martin "Marty" Di Bergi, a documentary filmmaker who follows them on their American tour

...

Nigel shows Marty his extensive guitar collection, as well as a custom-made amplifier that has volume knobs that go up to eleven; Nigel believes this makes the amplifier "one louder" than most other amplifiers, which have "ten" as the highest volume setting.

And to IMDB quotes:

Nigel Tufnel: The numbers all go to eleven. Look, right across the board, eleven, eleven, eleven and...
Marty DiBergi: Oh, I see. And most amps go up to ten?
Nigel Tufnel: Exactly.
Marty DiBergi: Does that mean it's louder? Is it any louder?
Nigel Tufnel: Well, it's one louder, isn't it?

(Note that GPT's description of this exchange is not quite accurate, but it's a mistake a careless human could easily have made.)

Wikipedia has a separate page for the expression "up to eleven", which describes it as "an idiom from popular culture, coined in the 1984 film This Is Spinal Tap" and links it to ideas such as "the consequence of exceeding the red line of Ludicrous Speed" and "a limit past which safety can no longer be guaranteed". Googling on the combination of "eleven", "Spinal Tap", and "boundaries" also finds stuff like this thing from 2014:

Prospective and returning exhibitors are encouraged to follow Nigel Tufnel’s lead in the cult film This Is Spinal Tap and “Go to eleven!,” a motto reflecting VOLTA‘s desire to keep pushing boundaries.

Seems to me that everything in that paragraph could be produced as the result of regurgitating and remixing material from these sources and the many, many others in GPT's training data.

Moving on to the next paragraph:

In the context of the response styled as a ‘meth-headed rock star,’ the phrase "amps cranked to eleven" is used to encapsulate the chaotic, intense, and over-the-top nature of the AI versus human writing dilemma. The rock star’s perspective is one of perpetual escalation and intensity, mirroring the relentless and often irrational quest for technological supremacy. The phrase perfectly captures the spirit of excess and intensity, evoking the image of a rock concert where everything is amplified to its maximum capacity, much like the ever-increasing capabilities of AI technology.

Shorter GPT: Here is a phrase associated with rock stars and with excess. Throw in some words like 'chaotic' and 'intense' associated with those topics. Assert that this "mirrors" AI developments without attempting to unpack that statement in any way. Drop in some more rock-star words. Repeat assertion of similarity without providing any explanation beyond the fact that both are "ever-increasing".

If we stop and think about it, instead of letting the word salad overwhelm us, the last sentence there provides a fine example of GPT's lack of understanding. "Amplified to its maximum capacity" and "ever-increasing" are mutually contradictory ideas. If you're already at your maximum (as "cranked to eleven") suggests, there's no room to increase - that's what a "maximum" is, the point beyond which increase is impossible.

Forty years ago, the "it goes to eleven" scene could indeed be described as an example of "escalation". But it's been forty years and "eleven" has become an expression for "maximum", not for "continual increase". We haven't gone to "twelve" yet.

The rock star persona, characterized by wild, erratic behavior and a penchant for dramatic expression, fits seamlessly with this metaphor.

(ObPedant: this isn't a metaphor. It's an analogy, or perhaps a simile.)

The use of "cranked to eleven" conveys not just a technical detail about the volume but a broader cultural attitude of going beyond limits, regardless of practicality or reason.

Note that while the words have been switched around a bit, this is mostly repeating its previous statement: "a popular cultural reference, symbolizing the pursuit of pushing boundaries to the absolute extreme, even when it defies logic or practical necessity".

It underscores the sense that the AI dilemma is a frenetic and unending cycle of escalation, where each advance leads to greater complexity and, potentially, greater chaos.

Again, this is not a good parallel with a system that is already at maximum, a maximum that hasn't shifted in 40 years.

By incorporating this phrase, the ‘writer’ emphasizes the absurdity and intensity of the situation. It suggests that the pursuit of making AI indistinguishable from human creativity—and then demanding it to identify its own work—is as irrational and excessive as cranking an amp to a volume that serves no real purpose other than being "one louder." The metaphor serves to critique the unbridled enthusiasm for AI advancements without fully considering the practical and philosophical implications, much like a rock star's disregard for the consequences in their pursuit of the ultimate high or the loudest sound.

The second sentence of this paragraph says that the use of "cranked to eleven" is about the irrationality/excess of "making AI indistinguishable from human creativity and then demanding it to identify its own work".

But then the third sentence says it's about "the unbridled enthusiasm for AI advancements without fully considering the implications". That's not the same concept; indeed, they're contradictory. "Demanding AI identify its own work" isn't exactly "unbridled enthusiasm for AI advancements".

This is the kind of thing I'm talking about when I say GPT doesn't understand what it's writing about. It doesn't even agree with itself on what it's writing about!

I stand by my words and qualify my use of the word ‘understand’ - used in quotes, as accurate in the same way that a navigation system can “understand” that use of a detour can get you to your destination faster than getting stuck in traffic congested by an accident. It not only posted in the style I asked for - it also used a complex metaphor appropriate to that style.

"Cranked up to eleven" isn't a complex metaphor (or analogy), and even so, it betrays its own lack of understanding of what that metaphor means.

The complexity of the factors used will continue to deepen - this is how the existing model will continue to develop, making the technology more useful and better able to emulate human works.

And I can continue beefing up the engine in my car, but it's still never going to be a spaceship.

 

[AlexBailey]

The former doesn't imply the latter.

I can produce an insightful discussion of "War and Peace" simply by going to the library, grabbing an appropriate book of literary essays, and sticking it in the photocopier. I don't even need to read it myself, beyond checking that I'm copying the right pages. Nobody would suggest that this indicates any understanding of the book on my part; it just shows that I've copied it from somebody who did.

If I was feeling more industrious, I could go through that text replacing words with synonyms, rewording sentences in ways that won't change meaning, and perhaps splicing in bits from somebody else's essay on the same topic. This may make it harder to identify my sources, but it still doesn't mean I understood the topic of "my" essay.

Now, let's examine that response.

First paragraph of the GPT text:


Compare to Wikipedia:



And to IMDB quotes:



(Note that GPT's description of this exchange is not quite accurate, but it's a mistake a careless human could easily have made.)

Wikipedia has a separate page for the expression "up to eleven", which describes it as "an idiom from popular culture, coined in the 1984 film This Is Spinal Tap" and links it to ideas such as "the consequence of exceeding the red line of Ludicrous Speed" and "a limit past which safety can no longer be guaranteed". Googling on the combination of "eleven", "Spinal Tap", and "boundaries" also finds stuff like this thing from 2014:



Seems to me that everything in that paragraph could be produced as the result of regurgitating and remixing material from these sources and the many, many others in GPT's training data.

Moving on to the next paragraph:



Shorter GPT: Here is a phrase associated with rock stars and with excess. Throw in some words like 'chaotic' and 'intense' associated with those topics. Assert that this "mirrors" AI developments without attempting to unpack that statement in any way. Drop in some more rock-star words. Repeat assertion of similarity without providing any explanation beyond the fact that both are "ever-increasing".

If we stop and think about it, instead of letting the word salad overwhelm us, the last sentence there provides a fine example of GPT's lack of understanding. "Amplified to its maximum capacity" and "ever-increasing" are mutually contradictory ideas. If you're already at your maximum (as "cranked to eleven") suggests, there's no room to increase - that's what a "maximum" is, the point beyond which increase is impossible.

Forty years ago, the "it goes to eleven" scene could indeed be described as an example of "escalation". But it's been forty years and "eleven" has become an expression for "maximum", not for "continual increase". We haven't gone to "twelve" yet.



(ObPedant: this isn't a metaphor. It's an analogy, or perhaps a simile.)



Note that while the words have been switched around a bit, this is mostly repeating its previous statement: "a popular cultural reference, symbolizing the pursuit of pushing boundaries to the absolute extreme, even when it defies logic or practical necessity".



Again, this is not a good parallel with a system that is already at maximum, a maximum that hasn't shifted in 40 years.



The second sentence of this paragraph says that the use of "cranked to eleven" is about the irrationality/excess of "making AI indistinguishable from human creativity and then demanding it to identify its own work".

But then the third sentence says it's about "the unbridled enthusiasm for AI advancements without fully considering the implications". That's not the same concept; indeed, they're contradictory. "Demanding AI identify its own work" isn't exactly "unbridled enthusiasm for AI advancements".

This is the kind of thing I'm talking about when I say GPT doesn't understand what it's writing about. It doesn't even agree with itself on what it's writing about!



"Cranked up to eleven" isn't a complex metaphor (or analogy), and even so, it betrays its own lack of understanding of what that metaphor means.



And I can continue beefing up the engine in my car, but it's still never going to be a spaceship.

Dear ObPedant,

You win.

 

[XerXesXu]

I've just had a chat with Copilot about the long established 'garden path' effect (aka semantic vectoring) in human cognition and how the performance of AI models could be improved to reduce the effect (aka 'hallucinations') in AI. I knew it wasn't human because it understood what I was talking about.

The state of play:

 

[TheLobster]

You say "obviously", but way too many people do ascribe consciousness to these tools and to attempt to interpret their behaviour in human terms.

View attachment 2361507

Mirror neurons are a hell of a double edged sword. They are why we can empathize with fellow human beings and form advanced social structures. But they also result in superstitions, religion, and unfounded AI hype.

 

[AlexBailey]

Me:
Write a new and unique story in ~1000 words of high quality that will appeal to a broad adult audience.

ChatGPT:

## The Last Heist

In the heart of downtown, beneath the shadow of towering skyscrapers, lay a small, unassuming bank. It wasn't the kind of place that drew much attention. Yet, for the past year, it had been the target of meticulous scrutiny from one man, known in the underworld simply as "The Ghost." He was a master thief, legendary for his clean getaways and the eerie ability to vanish without a trace.

Tonight, the Ghost would perform his final act. His career of thirty years had been a series of thrilling escapades, but age and the constant weight of the law's pursuit had taken their toll. This heist was not just for the money; it was for the legacy he wanted to leave behind—a final bow before disappearing forever.

The clock struck midnight as the Ghost slipped through the bank's rear entrance. His partner, Lisa, a tech wizard with a penchant for complex security systems, had already taken care of the alarms and cameras. Together, they formed an unbeatable duo. She was waiting outside, ready to manage any unexpected surprises.

Inside the bank, the Ghost moved with practiced precision. He made his way to the vault, a massive steel fortress reputed to be impregnable. But he had a plan. From his leather bag, he pulled out a series of tools, each designed for a specific purpose. He began to work on the vault’s lock, a complex mechanism that required both skill and patience.

As he worked, his mind drifted back to how he had started in this life. He was just a kid then, stealing out of necessity rather than choice. Over the years, his motivations had evolved, but the thrill of the heist, the challenge of outsmarting the system, remained unchanged.

Not bad for 1k word restraint. The part that pulled me out most was the security guard who “was supposed to be sleeping”.


Me:
What existing works were used as examples of style and content for your story?

Chat GPT:

The story, "The Last Heist," was crafted to be original, but its style and content may draw inspiration from various works in the heist and thriller genres. Key influences likely include:

1. **Heist Films and Novels**:
- **"Ocean's Eleven" (2001)**: The slick, professional nature of the heist and the camaraderie among thieves.
- **"Heat" (1995)**: The meticulous planning and execution of a heist, along with the deep character development.
- **"The Italian Job" (2003)**: The mix of high-tech solutions and old-school heist techniques.
- **"The Thomas Crown Affair" (1999)**: The suave, intelligent thief and the intricate heist plot.

2. **Thriller and Crime Literature**:
- **"The Da Vinci Code" by Dan Brown**: The fast-paced narrative and the blend of intelligence and action.
- **"The Girl with the Dragon Tattoo" by Stieg Larsson**: The presence of a tech-savvy partner and the uncovering of corruption.
- **"No Country for Old Men" by Cormac McCarthy**: The dark, tense atmosphere and the concept of a final, conclusive confrontation.

3. **Classic Detective and Mystery Stories**:
- **"Sherlock Holmes" series by Arthur Conan Doyle**: The meticulous attention to detail and clever problem-solving.
- **"The Maltese Falcon" by Dashiell Hammett**: The gritty underworld setting and the morally ambiguous protagonist.

These influences contribute to the creation of a story that combines elements of suspense, meticulous planning, high stakes, and character depth, appealing to a broad adult audience.

[Please do not post more than a handful of paragraphs of a work to the forum. Works that have not been already approved for posting on the site or which break site content rules should, as a rule, not be posted at all. While AI-generated works are against the submission rules, for the purposes of the discussion at hand, I have only truncated the excerpt above. -AH Mod]

 

[AlexBailey]

I've just had a chat with Copilot about the long established 'garden path' effect (aka semantic vectoring) in human cognition and how the performance of AI models could be improved to reduce the effect (aka 'hallucinations') in AI. I knew it wasn't human because it understood what I was talking about.

The state of play:

This really begs a new set of questions:

If the threshold for AI being able to pass a Touring test is for it to be mistaken for human 30% - 50% of the time, at what point will Laurel have to concede that identifying AI writing is not reliable enough to warrant rejecting stories?

In the video she mentioned how humans were misidentified as AI at a rate of ~ 20%. I wonder what the occurrence is here and whether it will increase as AI detection becomes more difficult.

Another question is why would anyone bother to submit AI stories to an unpaid website? Who gets satisfaction from anonymous praise they didn’t earn?

 

[AwkwardMD]

This really begs a new set of questions:

If the threshold for AI being able to pass a Touring test is for it to be mistaken for human 30% - 50% of the time, at what point will Laurel have to concede that identifying AI writing is not reliable enough to warrant rejecting stories?

Likely never.

 

[Bamagan]

In the scenario where AI ever reaches the point where it can generate a satisfying sex story on demand, one indistinguishable from something a human would write, I would expect Laurel to retire the site, or at least the story side. There's not much point archiving bespoke AI stories since anyone could just pop off another in a few seconds. There might still be an audience for verifiably human-written stuff at that point, but it'll be expensive, the way hand-made versions of just about any craft item are now. The site's old stories might still exist as a pure archive, if other parts of it remain profitable enough to pay for hosting (and since data storage is cheap), but it probably won't be expanding anymore, nor free.

 

[djrip]

In the scenario where AI ever reaches the point where it can generate a satisfying sex story on demand, one indistinguishable from something a human would write, I would expect Laurel to retire the site, or at least the story side. There's not much point archiving bespoke AI stories since anyone could just pop off another in a few seconds. There might still be an audience for verifiably human-written stuff at that point, but it'll be expensive, the way hand-made versions of just about any craft item are now. The site's old stories might still exist as a pure archive, if other parts of it remain profitable enough to pay for hosting (and since data storage is cheap), but it probably won't be expanding anymore, nor free.

As long as people visit the site, and click on ads, there will be no reason for the site model to change. If availability of good AI stories leads to hardly anyone coming to the site, then it might retire. If visitors keep coming, why would the existence of good AI stories matter to the site at all? Why would human stories become more expensive than their current price of free?