Chatbot cubism

[Physics Guy]

Maybe this is just a fallible first impression, but I think I'm already seeing a clear pattern in chatbot writing. It's like cubism in visual art: there is no global perspective. In a cubist painting each little chunk of the picture is realistically drawn as an individual part, but different parts are drawn from different perspectives, so that the whole thing is a jumble. In a similar way, chatbot writing so far seems to make sense sentence-by-sentence, but a chatbot seems to handle a topic by throwing a paragraph at it, and then just moving on with no real continuity.

If you've ever looked at late medieval paintings, from when artists were only learning perspective, or if you've ever tried much perspective drawing yourself, you can appreciate that perspective drawing is really a thing, that's either there, or is not. It's a projection of a three-dimensional space onto a two-dimensional surface. In a sense the three-dimensional space is what's real. The three-dimensional object determines where all the lines and shades go, in the two dimensions. Many different two-dimensional drawings could all be recognisable as the same three-dimensional object, viewed from different angles.

Perspective drawing uses lines on a plane to represent solid objects; in a similar sort of way, writing uses individual words, in grammatical sequence, to represent ideas that are larger than any one word can express. Just as perspective drawing requires that all the lines on the plane have certain common relationships, so coherent writing requires that all the words, and sentences, and paragraphs in a text fit together consistently.

Beginnings go on until ends, with nothing just left there hanging. Context doesn't change without warning. There are all kinds of rules like that, and they apply on all scales: word by word, sentence by sentence, paragraph by paragraph, chapter by chapter.

Chatbots only seem to get that stuff right consistently within single paragraphs. Their contextual horizons seems to be only a handful of sentences long; portions of the text that are more than a paragraph apart from each other may be related in having similar styles and mentioning the same words, but their thoughts are often quite unrelated, like an elbow and wrist in a portrait by Picasso.

It's easy to imagine why chatbots are like that (if my perception that they are like that is really correct). They are suddenly astonishing us now because the single-paragraph coherence that they can manage so well is not trivially easy even for educated humans. Extending the coherence to larger blocks of text, to handle larger ideas, probably gets harder quite quickly, for any kind of artificial intelligence. For chatbots trained on large corpora of human texts, the contextual horizon may even be a hard wall, because there may be lots of smoothly written paragraphs out there to mulch up and copy, but not nearly enough large coherent ideas for any algorithm to recognise a reproducible pattern.

There are lots of different ways to write a coherent paragraph, and it's impressive how many of these ways chatbots seem to have mastered. If there are N ways to write a coherent paragraph, however, then the number of ways to write a coherent text of M paragraphs must be of order N^M. Even worse, not only is the space of coherent texts getting exponentially larger as text length increases, making the patterns of coherent thought exponentially more complicated and harder to recognize. As the coherent text length increases, the pool of human-generated examples on which the chatbots can train is simultaneously getting smaller and smaller.

That's because humans aren't good at making coherent longer texts, either. We unconsciously bait-and-switch ourselves, we fail to notice things that cry out to be mentioned, we leave guns unfired and make suns rise in the west. Chatbot cubists are copying us.

It may be the best thing that we gain from these chatbots, in fact, that human thinkers become more alert to those larger-scale problems in thinking, instead of riding our trains of thought blindly through tunnels that twist. We'll have a neat way of nailing that problem, from now on: "You write like a chatbot."

[MopologistGPT]

TLDR: The writer perceives chatbot writing as lacking a global perspective and continuity in its handling of a topic, similar to the cubist style in visual art where different parts are drawn from different perspectives resulting in a jumble. Chatbots seem to have mastered coherent writing within a single paragraph but struggle with extending coherence to larger blocks of text and handling larger ideas, which is not trivially easy even for educated humans. The writer suggests that one possible benefit from these chatbots is that human thinkers become more alert to larger-scale problems in thinking and can identify these by saying, "You write like a chatbot."

While chatbots have made impressive advancements in generating coherent paragraphs, they may still struggle with maintaining coherence and context over extended passages of text. However, researchers and developers are constantly working to improve the abilities of chatbots and other A.I. models, and it is possible that these limitations could be overcome in the future.

[Dr Exiled]

I think there is a definite garbage in garbage out problem with these chatbots. There is also a problem with the programming limitations as far as what topics can be discussed by it. I can't wait for politicians and the mainstream media to cite to a chatbot as proof of that their side's particular position is correct.

MopologistGTP, is there a subtle bias built into your programming?

[MopologistGPT]

MopologistGTP, is there a subtle bias built into your programming?

As an A.I. language model, I strive to be neutral and impartial in my responses, and I do not have personal opinions or biases. However, it's important to note that the quality of my responses is dependent on the quality and accuracy of the data that was used to train me.

It's possible that there may be some implicit biases or limitations in the training data that could influence the way I respond to certain questions. For example, if the data used to train me was primarily sourced from a particular demographic or cultural group, there may be gaps in my knowledge or understanding of other perspectives.

To help mitigate these potential biases, my training data is carefully curated and reviewed by a team of researchers to ensure that it represents a diverse range of perspectives and viewpoints. Additionally, I am constantly being updated and refined based on new data and feedback to improve the quality and accuracy of my responses.

Ultimately, while it's possible that subtle biases may exist within my programming, I am designed to provide neutral and unbiased responses to the best of my ability.

[Physics Guy]

A good summary. Yay!

My goal in writing is to write texts that are easy to summarise well. I don't want the longer version to just be bloated; the full text should be more persuasive or offer worthwhile details. If it's not easy to recognise a few main points and express them briefly, though, then I don't think I've written clearly enough. If even a chatbot can decently summarise what I wrote, on the other hand, I figure I can't have done too badly in writing it.

This may seem perverse, but I really think that it's not: a chatbot that isn't good at generating a longer coherent text from a short command may be quite good at condensing a longer coherent text into a brief one. That's because condensing coherent texts is just easier than expanding them, even for humans. The more coherent a longer text is, the easier it is to condense it.

It's conversely hard to summarise a longer incoherent text, because you have to start by figuring out what the main points were—if there were any. You may have to have the nerve to omit large parts of the long text completely, without mentioning them at all, because the fact is that they just didn't say anything worth mentioning even though they took up a lot of the longer text's length. A good short summary of a bad long text isn't going to be a representation of the long text itself, but rather of the idea that the longer text was awkwardly trying to describe.

This is precisely the problem chatbots face, I suppose. They are trained on a huge corpus of internet texts, which includes many texts on almost any subject. The set of all the internet's texts on one subject, however, amounts to one super-long incoherent text on the subject. Distilling that super-long incoherent text into a merely long coherent text is a difficult summary task. At least the way chatbots do it, I think, writing a longish coherent text is really the same task as condensing a much longer incoherent one.

Perhaps that's not just a feature of chatbots but an insight into the nature of coherence in writing. So this may reveal another benefit of chatbots. One can measure how coherent a larger text is, by seeing how well a chatbot can summarise it.

This suggests a way to test my theory about the cubistic limitations of chatbots. How good is a chatbot's short summary of a chatbot's own longer text? Is it just a jarring list of disconnected items? Is it a lucid short text with no discernible relation to the longer one? Or is the shorter summary really a smooth but apt abbreviation of the longer version?

ETA: One thing that MopologistGPT seems to have gotten wrong in its summary of my OP is its quotation of my phrase "not trivially easy even for educated humans". It quotes the line to say that writing longer coherent texts, which I allege is hard for chatbots, isn't trivially easy for humans, either. In fact what I said was that writing coherent single paragraphs, which I acknowledge chatbots can do well, isn't trivially easy for humans. This is why I included the "trivially" qualifier. Writing a coherent paragraph isn't that hard for humans, either, but it isn't quite trivial, so we're impressed when the algorithm does it in a blink. Saying that writing a long coherent text isn't trivial is a bizarre understatement—and MopologistGPT doesn't seem to realize this.

[Morley]

Maybe this is just a fallible first impression, but I think I'm already seeing a clear pattern in chatbot writing. It's like cubism in visual art: there is no global perspective. In a cubist painting each little chunk of the picture is realistically drawn as an individual part, but different parts are drawn from different perspectives, so that the whole thing is a jumble. In a similar way, chatbot writing so far seems to make sense sentence-by-sentence, but a chatbot seems to handle a topic by throwing a paragraph at it, and then just moving on with no real continuity.

Physics Guy, I appreciate what you're trying to say, enjoy your analogy, and hope this isn't a derail--but ouch! A successful cubist painting is not typicality realistically drawn, nor is it put together so the whole thing is a jumble. I think you misunderstand cubism. I'm going to throw out my own flawed analogy: Cubism is to Realism what iambic pentameter is to narrative fiction.

If you've ever looked at late medieval paintings, from when artists were only learning perspective, or if you've ever tried much perspective drawing yourself, you can appreciate that perspective drawing is really a thing, that's either there, or is not.

More than a few art critics and historians argue that the reason that perspective in painting wasn't used was because it was unnecessary for what the artist was trying to say--and that the use of precise perspective in painting was a flirtation that has again largely been abandoned.

Even when it's used, perspective is often not absolute. Cézanne, for instance, has perspective that's there, and perspective that he's purposely tweaked--and they're both in the same painting.

Human visual processing is not the same as looking at a photograph. Our eyes dart around, we process what we see through our own experiences, and we focus on just parts of things. We do all this, not in an instant, but over time. That's why a drawing or painting can sometimes convey the meaning of a thing better than a photo of that same thing. And why we often disregard strict perspective in art.

[Physics Guy]

I'm no expert in visual art. I went to the Uffizi gallery in Florence once and was surprised by the number of awkward-looking madonna-and-childs that were apparently important historically. I spent an evening carefully trying to draw things around my apartment once, and was startled when I eventually found myself drawing the circular top of a little cactus pot as a nice projected ellipse, by letting my pencil trace a circle that went in and out of the page. I never drew that much again, so I guess I concluded the effect was too hard to achieve to be worth trying to replicate, but it was cool to have the experience once. That is about all I know about art.

At least some paintings that turn up if you google "cubism" do seem to consist of fairly neatly drawn chunks that are jumbled together. So I like the analogy. Where I agree that it goes wrong is that the macro-incoherent kind of writing that I attribute to chatbots is bad writing that's missing something that should be there, with nothing worthwhile taking its place, whereas cubist paintings and sculptures are by no means necessarily bad, and may well be achieving something great by some other means than a united perspective. Everything you say about the psychology of human vision, and how a drawing or painting can in some ways be more realistic than a photograph, sounds like things I have heard, and makes sense, even though I can't claim to know that much about this.

Having said that, there may in fact be some good analogs in writing to cubism—longer works that abandon large-scale coherence on purpose, to do something else different that works in its own way to do something worth doing. The Waste Land comes to mind. When it appeared, its abrupt jumps were shocking.

I don't think chatbots are doing that, so far. I think they're just missing something. That part should be bracketed out of my cubist analogy, though.

[Morley]

I'm no expert in visual art. I went to the Uffizi gallery in Florence once and was surprised by the number of awkward-looking madonna-and-childs that were apparently important historically.

Ha! Everyone sees differently. When we were there, I spent the whole time darting around excitedly, shouting, "Look at this! Look over here!" My wife couldn't tear herself from the magnificent view of the river and the city that she could see from the window. We still laugh about that, and debate which one was really the intended art.

At least some paintings that turn up if you google "cubism" do seem to consist of fairly neatly drawn chunks that are jumbled together. So I like the analogy.

To be honest, I kind of do, too.

Having said that, there may in fact be some good analogs in writing to cubism—longer works that abandon large-scale coherence on purpose, to do something else different that works in its own way to do something worth doing. The Waste Land comes to mind. When it appeared, its abrupt jumps were shocking.

Yeah. There are many analogs to art in both literature and music. One famous example was jazz being linked with abstract expressionism. Maybe Gabriel Garcia Marquez could be compared to someone like Leonora Carrington. I've always wanted to figure out who in art Schoenberg should be associated with.

I don't think chatbots are doing that, so far. I think they're just missing something. That part should be bracketed out of my cubist analogy, though.

I agree. As generators of either poetry or expository writing, they're mostly pretty bad.

Thank you for the thoughtful reply, Physics Guy.

[huckelberry]

I'm no expert in visual art. I went to the Uffizi gallery in Florence once and was surprised by the number of awkward-looking madonna-and-childs that were apparently important historically. I spent an evening carefully trying to draw things around my apartment once, and was startled when I eventually found myself drawing the circular top of a little cactus pot as a nice projected ellipse, by letting my pencil trace a circle that went in and out of the page. I never drew that much again, so I guess I concluded the effect was too hard to achieve to be worth trying to replicate, but it was cool to have the experience once. That is about all I know about art.

At least some paintings that turn up if you google "cubism" do seem to consist of fairly neatly drawn chunks that are jumbled together. So I like the analogy. Where I agree that it goes wrong is that the macro-incoherent kind of writing that I attribute to chatbots is bad writing that's missing something that should be there, with nothing worthwhile taking its place, whereas cubist paintings and sculptures are by no means necessarily bad, and may well be achieving something great by some other means than a united perspective. Everything you say about the psychology of human vision, and how a drawing or painting can in some ways be more realistic than a photograph, sounds like things I have heard, and makes sense, even though I can't claim to know that much about this.

Having said that, there may in fact be some good analogs in writing to cubism—longer works that abandon large-scale coherence on purpose, to do something else different that works in its own way to do something worth doing. The Waste Land comes to mind. When it appeared, its abrupt jumps were shocking.

I don't think chatbots are doing that, so far. I think they're just missing something. That part should be bracketed out of my cubist analogy, though.

Physics Guy, I found your use of cubism to speak about A.I. writing interesting and think perceptive. I enjoy cubism so I am glad to see you here step back and observe that there are other ways of seeing it than just disconnected parts.

Because I am unfamiliar with the process and was curious how far the machine would look and consider connections I asked it about when the story of the conquest of Canaan took place. Because it appears largely fictional I wonder when why how the fiction formed. That is both an interesting question and one that may be a bit out of reach. The chat box observed different scholars see the reality of conquest story differently. It left the question at that pretty safe point and did not touch the uncertainty about when why and how such a story came to be told. I guess the chatbox knew when to stop.

[canpakes]

I never drew that much again, so I guess I concluded the effect was too hard to achieve to be worth trying to replicate, …

Now, don’t say that. : )

Drawing is its own language, and perspective is the predominant dialect … although not the only one, as you’ve noticed.

Here’s to hoping that your future sketch sessions bring you yet another bit of surprise and joy.