Pages

Thursday, February 16, 2023

ChatGPT Is Fun, But It's Like Wrestling A Greased Pig

I bought a subscription to ChatGPT at $20 per month to play around with it. My initial experiments have been trying to get it to write some fan fiction from Malory's Le Morte D'Arthur. If I don't constantly fight with the thing, the results end up as vapid, feminist versions of the old, 1980s After School Specials on TV. It wants to change the outcome of the story to teach insipid moral lessons. Ugh.

As a case in point, many of the knightly combats in Malory end in death. "And he raised off his helmet and smote his neck asunder" sort of thing. ChatGPT changes that to the knights agreeing to be friends forever, no matter how much you describe the characters as mortal enemies.

The styling is primitive, too. While it will write in the style of modern authors fairly well, it's way off when it comes to Sir Thomas Malory's style. Le Morte has lots of sentences that start with "and." As in, "And then, Sir Tristram came unto La Beale Isoud without brushing his teeth. And La Beale Isoud said unto Sir Tristram, "Go thy way, Sir Knight. Thy teeth look like the mold on a pivy postern!"

Finally, there is the blatant feminism. Every female character will end up independent and strong if you don't constantly correct it. In Malory, they are crafty and motivate all of the stories. They tease and tempt the knights with great guile to get what they want. In ChatGPT, female characters wouldn't know how to get a man to do what they want if chaste seduction fell out of the sky in the form of an anvil and hit them on the head.

Still, it's a lot of fun. I'll post some examples later. Meanwhile, here's what it's making me think of right now.

And then the lady strode between the two knights and informed them that violence never solves anything. Chastened, they realized that women have a lot to teach men and they agreed to be friends and attend the local Pride march together. The lady then left them, informing them that she was late for her board meeting at her renewable energy nonprofit.

8 comments:

  1. I haven't worked with it yet, but from what people have said about it, I gather that a good approach to interacting with ChatGPT is to treat it like a 4-year-old child. In particular, I understand it does that thing a lot of very young kids do where if they don't actually have an answer, they will just blithely make up the most outrageous stories instead of conceding that they don't know.

    ReplyDelete
  2. Meh. I think I'll just continue wasting my time trying to build a virtual NYC (in the city-builder game, Cities:Skylines). My goal is to reach a population of 1,000,000. Currently, it's about 140,000 ... but I'm starting to have supply-chain issues getting manufactured goods to the stores. I suspect the problem is that all those cargo ships in the harbor are interfering with one another.

    ReplyDelete
  3. Incidentally, I don't know if you normally read "A Collection of Unmitigated Pedantry", but he's got a fairly lengthy post on ChatGPT and its severe limitations as a writer of essays.

    https://acoup.blog/2023/02/17/collections-on-chatgpt/

    It mainly boils down to the fact that it is basically creating a statistically plausible document from your starting keywords, based on what those keywords are associated with in the training text. He points out that it isn't so much that it "gets things wrong", as that it is throwing out words that have strong statistical associations with each other in the training text. And a lot of the time what it generates looks "right" to us, but the software has no way of categorizing that as being different from the things that we recognize as factually wrong.

    Which, remembering my days talking with toddlers, is a lot like what very small children do. They say word strings that they have heard but don't actually understand, put them together in various ways, and see how the adults react to them.

    So the reason why it is giving the sort of output you are seeing, is that it was trained on a lot of the kind of modern literature that you don't much like. Which suggests that if you want ChatGPT to generate plausible Arthurian text, it is going to need to be trained on a lot more text that includes duels, combat, and gristly murders. Feeding it the unexpurgated text of "Grimms Fairy Tales" would probably help a lot there.

    ReplyDelete
  4. ChatGPT illustrates the difference between a mind and a process. It shows how Dawkins and Hawking are dead wrong about the nature of reality. ChatGPT thinks the way a creature made solely of neurons would think - embodied by algorithms and discrete states and events. That's clearly not who we are as evidenced by the results here. ChatGPT has way more processing power and many orders of magnitude more recallable information that I do, but, as you said, it's an absolute child as a writer.

    Maybe I'm wrong and it will evolve into G K Chesterton, but I can't see that happening.

    ReplyDelete
  5. one more step toward destruction

    ReplyDelete
  6. ==Maybe I'm wrong and it will evolve into G K Chesterton, but I can't see that happening.==

    It (and all other ATs) will *always* be nothing more that a computer program, a script, a recipe. So, even if an AI manages to churn out text comparable to Chesterton, you're not wrong.

    ReplyDelete
  7. KT - I think I will challenge your previous...

    "ChatGPT illustrates the difference between a mind and a process. It shows how Dawkins and Hawking are dead wrong about the nature of reality. ChatGPT thinks the way a creature made solely of neurons would think - embodied by algorithms and discrete states and events. That's clearly not who we are as evidenced by the results here." While I don't, even for a moment, want to suggest that a computer program (running in a modern computer) could ever be truly creative, I do think that a system governed "by algorithms and discrete states and events" could. For the purposes of my thesis I claim that 'creativity' can be equated to producing unexpected and novel outcomes.

    The first reason I argue this is that, as you well know, unpredictable and completely surprising results can, and do, come from purely deterministic mathematical algorithms. Yes, I mean Chaos. An iterative system as simple as the logistic equation can produce unpredictable and chaotic (might one say creative?) output.

    The other reason I argue for creativity in a system "embodied by algorithms and discrete states and events" is based on a reality of human brains. There are many many many more neurons and connections in a human brain than in any computer today. Furthermore, the connections in our brains are not fixed, but rather are plastic and change over time. Given the truly astronomical numbers of neurons, and the even larger number of interconnections, I posit that no two human brains have been, or ever will be, wired the same (and indeed any one brain will necessarily change over time). This means that there is no guarantee that two brains will produce the same output for a given input. It is also easily believable that a brain may well produce a unique, unexpected, and novel output, i.e., a creative thought.

    Personally, I don't care for Dawkins' or Hawking's writings on this, but I will strongly suggest that the interested reader take the time (and mental investment required) to read "Godel, Escher, Bach: an Eternal Golden Braid" by Douglas R. Hofstadter (printed by Basic Books in 1979). That book definitely re-wired my brain!

    Note added in proof: I absolutely broke up after typing this in, when I was confronted with your CAPTCHA "Please prove you're not a robot." test. A delicious iterative twist.

    ReplyDelete
  8. ==An iterative system as simple as the logistic equation can produce unpredictable and chaotic (might one say creative?) output.==

    No, once cannot say that without doing violence to the word 'creative'.

    ReplyDelete