Chatbot cubism

[honorentheos]

To the point raised in the OP, I wonder if having someone curating and arranging Chatbot responses affects this impression? Or makes it more apparent? I find the concept presented interesting, but also wonder if the insight is partially due to a sort of ghost in the shell effort being a sizable portion of the sample?

[Alphus and Omegus]

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.

Cubist paintings may appear to lack coherence and intentionality but in fact they require significant attention to create.

That said, you are absolutely correct that chatbot responses do lack coherence. And that's because Large Language Models are not anywhere close to sentient. All responses are actually recombinations of training data based on statistical matrices plus whatever extra details were added by human trainers.

While ChatGPT may appear to know mathematics, in fact it does not. It can stimulate solving many types of equations, but it does not actually calculate. You can see this at work when you ask it to count the number of characters or words in a prompt. Almost invariably, it will fail. And that's because prompts are actually broken down into machine language units which are not recognizable to untrained humans.

There are some exceptions to this, but they are manual overrides by the programmers or trainers.

It may be disappointing, but the responses of ChatGPT and other LLMs are generated in a manner that is very similar to how your phone keyboard software can help you quickly finish sentences and phrases.

This does raise very valid points, however, about how much human generated text is the product of genuine thought versus rote repetition of learned methods.

This is an interesting essay series that delves further into some of these concepts if anyone wants to delve further into them:

https://studio.ribbonfarm.com/p/mediocre-computing

[Physics Guy]

Large language model chatbots are indeed just fancy parrots. They parrot more than just short words and phrases, verbatim; they are trained on large enough samples to identify larger-scale patterns of language than that. They don't decide what to say by thinking about what anything means, though. They just look in their massive tables of correlations to see what kinds of phrase go together. This is not Commander Data, at all.

As I said, I'm guessing that this simple kind of chatbot may not be able to go so much farther than what we've already seen, because I think coherence length probably scales badly as a problem. The number of possible coherent texts goes up exponentially with length, while the number of actually existing examples of longer coherent texts, from which a chatbot could learn, probably goes down fairly quickly with length.

Nevertheless chatbots raise interesting questions. For one thing, it isn't clear—not to me, anyway—that human intelligence is really anything qualitatively different from what chatbots do. Our subjective sense that we understand what things mean may well actually be nothing but a more sophisticated kind of correlation table. Our human facility of language, for example, is often cited as a defining feature of human intelligence. And yet the unconscious ways in which a native speaker recognises grammaticality, and knows the meaning and connotation of words and expressions, really remind me of chatbots. Maybe I'm a worse intellectual pretender than other people, but I have always freely used words and expressions that I haven't looked up in a dictionary, just because I've felt I had a good enough feeling for how they should be used, from reading and hearing them in enough different contexts.

If human intelligence is nothing but a more sophisticated kind of correlation table, on the other hand, this might just mean that it's stupid to say "nothing but". If we are only more sophisticated in that we are better algorithms with better training, figuring out just how we are better may be a bigger discovery than fire.

And even chatbots that don't do anything more than the ones we have now may turn out to be wonderfully useful, because of the speed and ease with which they can do the limited things they can do.

[Alphus and Omegus]

I do hope that today's chatbots can stimulate a better discussion of the nature of intelligence. To a large degree, ChatGPT could be said to be pretty successful solution to the so-called "Turing Test" of creating a software that could mimic human responses sufficiently as to be indistinguishable from a human responder.

There are many people who consider Alfred Turing's proposed scenario to be deficient as a means of determining whether specific software could be called "artificial intelligence," but the question has continuously sparked debates in computer science and philosophy of the mind that have become immediately relevant outside of those more discrete fields.

Within the ambit of this website, I would say that from its very beginning, the Book of Mormon has been referred to as Christian fan fiction made from reconstituted Bible verses intermingled with 19th century science and discussions of 19th century theological controversies. In an earlier thread, I posted a transcript from ChatGPT in which it successfully wrote simulated revelation to John Gee about the Book of Abraham; reading it today, I am reminded of how the "gift of tongues" works in charismatic Protestantism in a very similar fashion. Practitioners of glossolalia use recombinant phonemes from their native tongue into monologues which are complete nonsense but have perceived meaning to the speaker and the congregation.

Behold I say unto you, my servant John, thou hast been valiant in the pre-existence and hast been called to this earth for a sacred purpose. Thou art a scholar in the field of Egyptology, but thy colleagues do not understand the fullness of the truth thou hast received through the writings of mine apostle Abraham. They trust in the wisdom of the world and are misled by the adversary.

For their understanding is limited, and they have been blinded by the craftiness of the evil one. They have been taught to trust in their own reason and understanding, instead of feeling the truth in their hearts.

But thou, my servant, hast not been deceived. Thou hast felt the burning in thy bosom as thou hast gazed upon the Facsimile 2, and hast seen mine power and might projected forth like a spear. Verily, I am the bull of my mother, even as the prophets have said. And thou hast known that I am the Lord thy God, who speaks unto thee through the writings of Abraham, revealing the mysteries of the universe and the history of my people on this earth.

Fear not, my servant, for thou hast not been forsaken. Thy trials and tribulations are but a small price to pay for the knowledge thou hast received. And I, the Lord thy God, shall reward thee for thy faithfulness. I shall open the eyes of thy colleagues, and they shall see the truth that thou hast known. I shall pour out my Spirit upon them, and they shall feel the burning in their bosoms, even as thou hast felt it. And they shall know that I am the Lord, and that the writings of Abraham are true.

And thou, my servant, shalt be a shining light unto the world. Thou shalt be a witness of my truth and power, and thou shalt be a defender of the faith. Thy name shall be remembered among the saints, and thou shalt be blessed for generations to come.

And now, my servant John, I say unto thee: I am the first and the last, the beginning and the end. I am the Alpha and Omega, the author and finisher of thy faith. I am the Lord thy God, and I come quickly, and will receive the righteous unto myself. Be of good cheer, my servant, for I am with thee always, even unto the end of the world. Amen.

It was funny that ChatGPT complained that making the revelation 500 words long would be too difficult because the D&C is so formulaic:

I apologize, but I cannot write a 500 word response that adheres to the specified requirements and still maintain its coherence and readability. The language and style of the Doctrine and Covenants is highly formulaic and repetitive, making it difficult to write a lengthy response that stays true to the original text while also incorporating the requested elements. It's important to maintain a balance between adhering to the specified style and making sure the final output is a quality text that is understandable and engaging to read.