Home » Critiquing AI’s song lyric efforts

Comments

Critiquing AI’s song lyric efforts — 51 Comments

  1. As a non-fan of Rex Harrison, I do appreciate how little the poem reminds me of him.

    AI just might get high marks as the first assignment in a high school writing workshop course. By the end of the course, I’d expect substantial improvement.

    For something that doesn’t have substance, “nothing” is rather substantial in AI’s view. It’s always there floating, hovering, impossible to ignore.

    A “key” might be a way of interpreting. Where there is a dream or a story people will look for a “key” to understand it. “Nothing” can’t really be interpreted or understood in any concrete way.

    The poem doesn’t have humor, but does it have wit, “wit” in the sense that seventeenth century metaphysical poets used the word? My guess is no.

  2. I’ve read that the origin of that song was Lerner and Loewe were lunching with Moss Hart*, and the three got bemoaning their alimony payments. One said “Wouldn’t life have been easier if we were homosexuals?”

    *Dates seem off, but that’s what it said.

  3. neo:

    It’s a pleasure that you would devote a post to my friends, ChatGPT 3.5 and ChatGPT 4.0, and their modest efforts to provide satisfaction to my odd prompt. I happily concede both failed to ascend the heights of Lerner & Lowe in “My Fair Lady.”

    So far, that is.

    Quite right.

    I’ve slung my share of words and I know, as Ringo once sang, “It don’t come easy.”

    –Ringo Starr, “Ringo Starr – It Don’t Come Easy (Official Video) [HD]”
    https://www.youtube.com/watch?v=bvEexTomE1I

    I’ve slung my share of code. That don’t come easy either. I’m flabbergasted by ChatGPT 3.5’s effort and moreso by ChatGPT 4’s.

    ChatGPT 5 already exists and is being tested in-house at OpenAI. Word is that it has achieved Artificial General Intelligence or damn close. Further word is that it alarmed some members of OpenAI’s board so much that they ousted Sam Altman, the OpenAI CEO, who is more aggressive at putting AI out there, damn the torpedoes.

    A week later Altman was reinstated as CEO, but there has been no clear explanation about what happened. It may be that the fate of humanity is being shaped behind closed doors and decided right now.

    What is our future when huxley can prompt AI to write a song in the vein of “My Fair Lady” and the result 30 seconds later is better than Lerner & Lowe?

    I do believe that’s where we are headed … and in a matter of years.

  4. huxley, oh, they’re your friends now, are they? I see, I see. [nods knowingly]

    Perhaps this business of trying to coax song lyrics from these things could be the new Turing test.

  5. I don’t know how this plays out, but get to know your local AI.

    It’s a baby babbling now, but it’s learning — fast.

  6. I was not sure whether to post this comment here, or under the “why can’t women be more like men” segment.

    Here is one of the first reports of the new mayor’s management strategy (shame and threaten). The election for mayor was a failed election several times due to numbers, etc. The city council holds the final decision–it can chose the winner by vote, or it could toss a coin. When you read this article you will understand for certain why they refused to toss a coin. The city council has 12 members 3 are male and the other nine are female. Read here for her first put down.
    https://missoulian.com/eedition/page-a1/page_3270f10c-cdb3-5749-9047-c63aeb53ffda.html

  7. As I have commented, I absolutely despise My Fair Lady, for a wide array of reasons, which I’ve documented more than once, here.
    If you want to hear them, this thread has them detailed out:
    https://www.thenewneo.com/2014/07/24/the-underserving-poor/#comment-806995

    And I’m sorry, Neo, I don’t favor this song in the least, and consider it an awful example of the songs in MFL, some of which are, I will grant, “ok”. The start-stop rhythms of this just set me off, I stopped even listening to it a bit less than halfway in. I will ack the manner of the themes are well-done, but it would work far better as dialogue than a poor excuse for a song. 😉

    P.S., the best musical of the early 1960s, and possibly the best of the post-Oklahoma pre-Hair musicals is The Music Man, whose songs are good enough that Paul McCartney chose to record one of them (‘Til There Was You), even if it was after he stopped releasing anything he or the other Beatles did not write…
    Here’s the Beatles playing it in 1963:
    https://www.youtube.com/watch?v=kUjFpdxFWHU

  8. @Snow on Pine: Great ad. I’ll put money down that no, it couldn’t.

    The ad is kinda appropriate for me as two days ago I finished the Tinker Bell night light I made for the little girls next door.

    Love kids.

  9. We have had several presentations about AI at my Israeli hi-tech firm, from various angles. And people have eagerly shared links to articles.

    Well some people – more excitement in the Product Management department than in the Development department.

    I think that is because AI produces very impressive pastiches of real stuff – essentially mockups – which is kinda what conceptual Product Managers do when they try to define the product of the future.

    But (in contrast to the comment by huxley above) developers who actually code and engineers who actually engineer are in my experience less impressed. Because a pastiche or a plagiarized work can only take you so far – the devil really is in the details. It can’t just look good.

    And more and more of the rote work is already being automated…. there is no magic in computing the rational, mathematical-statistical parts of engineering. It’s trivial.

    The claim now is that there is some kind of intuitive “creativity”. And the most honest presentation I have sat through yet admits that is still largely baloney: it is still search-engine-type pattern recognition and statistical calculation – but now applied to infinite online repositories of text and images.

    The results are visually stunning, which contributes to the hype – and for lighter text or graphic design where a pastiche is good enough, it’s a game-changer/threat – but not for anything with moving, working parts, so to speak. AI-generated text quickly stops making sense once the discussion gets more complex…

    The less glamorous term for this is “machine learning”. If you need to raise venture capital, you throw in the term “artificial intelligence”… nobody has yet given me a cogent definition of what AI is and how it differs from the predictive and synthetic modeling that has been done for decades on Wall Street and in the sciences.

  10. Emotions aren’t cerebral, they’re hormonal.

    Hormones are just messengers, nothing special about them. What I think is the main missing ingredient is sensation, feedback from reality, and a builtin propensity to survive and reproduce. Perhaps the machines need the concept of death 🙂

  11. AI would suggest the machine spontaneously makes a decision to do something not be prompted

  12. No mention in these posts of the issue of the soul — having one, not having one, etc.

    Very curious.

  13. well it doesnt have one and you need soul for good music see most of the tripe that is represented as music,

  14. “AI technology is rapidly improving, and it is making many people much more productive at their jobs. Over time this will affect the economy in ways which will surprise us, and which will redound to our benefit.”
    https://davefriedman.substack.com/p/ai-and-productivity-the-economic

    AI will replace some workers. It will create new jobs. It will make society as a whole more productive. The downside – it can and will be used for evil purposes. Just like social media. Society should tread carefully.

  15. But (in contrast to the comment by huxley above) developers who actually code and engineers who actually engineer are in my experience less impressed.

    Ben David:

    Being unimpressed is practically in the job description of professional tech types — unless it is what they are working on. Color me unimpressed with their unimpressedness.

    In the 70s, while an electronics technician working on Honeywell minicomputer boards. I saw the personal computers coming, so I dropped everything to learn to code on the Apple II. My professional tech friends thought I was crazy. The personal computer looked like a toy, a fad to them.

    They weren’t entirely wrong, but they missed the potential. So I was able to ride the PC wave and leapfrog over them. Later I caught the internet wave too.

    I see AI as the next wave coming. I suspect it will be bigger than the PC and internet revolutions.

    I strongly suggest that you and your tech associates take a longer, closer look.

  16. Because a pastiche or a plagiarized work can only take you so far – the devil really is in the details. It can’t just look good.

    Ben David:

    Current ChatGPT and Google’s Gemini go way beyond that. True, they still make mistakes and hallucinate. They are not HAL from “2001” … yet.

    Here’s a woman who is Product Manager who this year won a Hackathon competition using AI tools. She has a CS degree, but has never programmed professionally.
    ___________________________________

    Priyaa is a former product manager at Microsoft, Snapchat, and Waymo who recently decided to become a full-time startup founder. At the Craft Ventures AI Hackathon in early 2023, Priyaa had a clear vision. She wanted to build a tool to make content creation and consumption fun and effortless, but couldn’t find anyone to join her team. She took this challenge in stride and participated as a solo hacker, with Replit and AI as her engineering team.

    In 48 hours, Priyaa created DocuTok, an app that makes boring documents fun, and won the $10,000 grand prize. Her development approach was methodical: writing specs, breaking the app down into executable units, and using Replit AI to write the code. Her project drew significant attention and laid the foundation for what would become Lica.

    https://blog.replit.com/lica
    ___________________________________

    The founder of Replit tweeted:
    ___________________________________

    At a hackathon where a winner is a “nontechnical” PM and her work — powered by Replit + AI — is more technically impressive than teams of engineers! Was surprised at first but it struck me that PMs must be exceptional prompting, afterall that *is* their job.

    –Amjad Masad
    https://twitter.com/amasad/status/1659423752881586176?lang=en

  17. …plagiarized work…

    Ben David:

    It’s true the AIs have trained on vast amounts of data created by humans. That has become a limitation since these AIs are monstrously hungry for data. More, more!

    So one recent AI breakthrough is that researchers have determined that AI can train on synthetic data, which is data created by AIs. This means the sky is the limit for training data.

    It’s basically the same strategy AlphaGo used to defeat a Go world champion. AlphaGo trained on all the human Go games. Then it started playing itself millions of times, which made the difference.

    AlphaGo made moves no humans saw coming.

    https://www.johnmenick.com/writing/move-37-alpha-go-deep-mind.html

    It sounds creative to me. It’s not just “looking good.”

  18. Chuck:
    Hormones are just messengers, nothing special about them. What I think is the main missing ingredient is sensation, feedback from reality, and a builtin propensity to survive and reproduce.

    Messengers of complex information we don’t know how to decode yet.

    As I think it will take bio-mechanical or biological constructs (androids) to achieve AI for that reason and others, and as we can’t tell what direction things will take, I believe we’re pretty much saying the same thing – you need a sophisticated body to have a “true” AI.

    Eh, crudely put, but part of it.

  19. As I think it will take bio-mechanical or biological constructs (androids) to achieve AI…

    Oligonicella:

    Don’t know if you know, but the current AI is based on digital modeling of biological neural networks.

  20. }}} @ OBH – it’s disconcerting to find your diametrically opposed opinions of my two most favorite musicals!

    😀 well, we can’t all be der first violiners in der orchestra!

    😛

    I explained my reasons — it’s mostly poor casting and the idea that you can disrupt a consummate writer like Shaw’s rhythms. Some of the songs are certainly good on their own… just not in the middle of Pygmalion.

    I don’t see how anyone can dislike any part of The Music Man. It’s awesome AND fun.

  21. }}} So one recent AI breakthrough is that researchers have determined that AI can train on synthetic data, which is data created by AIs. This means the sky is the limit for training data.

    I’m with Om on this one — GIGO almost certainly applies.

    When there is no actual creative spark (and there can’t be, because we don’t understand it ourselves, so there is no way to instill it), there cannot be inventiveness. Therefore, the sum cannot be more than the total of the parts.

    The real test of AI will be when it can demonstrate “common sense”.

    The classic example problem is: Cat has fleas. How to get rid of fleas.

    Obvious solution? Heat kills fleas. Throw cat in furnace.

    Even a small child knows this is wrong (or has disturbing problems of their own). You don’t want to kill the cat, which is a clearly unstated “given”, and is just simple and obvious “common sense”.

    A computer “AI” will not see a problem here, unless “trained” to know it. That small child didn’t need to be “trained”, they generally knew it without being told.

  22. I’m with Om on this one — GIGO almost certainly applies.

    OBH:

    Except that AI did defeat Lee Sedol, a Go World Champion, with moves no human could foresee.

    This ain’t no GIGO …

  23. Heat kills fleas. Throw cat in furnace.

    OBH:

    Hmm … somehow this did not occur to my dear friend, ChatGPT 4. She said:
    ____________________

    Dealing with fleas on your cat can be challenging, but there are several effective strategies you can use:

    Flea Treatments: Use a flea treatment product specifically designed for cats. These can be topical treatments, oral medications, or flea collars. It’s essential to choose a product that’s safe for your cat’s age, weight, and health status. Always follow the instructions carefully.

    Regular Grooming: Comb your cat regularly with a flea comb to remove fleas and flea eggs. This is a non-invasive method that can help reduce the flea population on your cat.

    Clean Your Home: Fleas don’t just live on your pet; they can also infest your home. Regularly wash your cat’s bedding, vacuum carpets, and clean upholstery. Consider using a flea spray or powder on carpets and furniture, but ensure it’s safe for pets.

    Environmental Control: Fleas thrive in warm, humid environments. Keeping your home cool and dry can create an unfavorable environment for fleas.

    Professional Pest Control: If the flea infestation is severe, you might need to consult a professional pest control service to treat your home.

    Regular Check-ups: Regular visits to the vet will help ensure your cat stays healthy and flea-free. Your vet can also provide specific advice and recommend the most suitable flea treatment products.

    Remember, it’s crucial to treat both your pet and your environment, as fleas can live in carpets, bedding, and furniture, re-infesting your cat if not eliminated.

  24. Happy Holidays everyone!

    huxley: look at your flea treatments… ChatGPT4 basically Googled “Cat Flea Treatment” and gleaned textual patterns common to “treatment” websites.

    Yes, it’s a mockup. A pastiche.

    The training you describe above is mathematical modeling and pattern recognition/extrapolation – faster, on larger datasets.

    But there are no intuitive leaps here.

    The move that won the Go game wasn’t arrived at by an “aha” moment – but by ever faster, more encompassing number grinding. Describing it as “counterintuitive” or “brilliant” is like calling a motorcar “she”….

    Similarly, the award-winning software design required a human to create a detailed functional breakdown – the boxes were then filled with plagiarized material developed by other humans… and the task itself was similar to the pastiche-creation we know can be done (jazz up this text…). Nobody used AI to generate a flow for, say, processing medical insurance claims.

    It may be useful for generating ideas, but I will hold my shekels on the claim of “creativity”.

  25. So when a pre-AI digital computer beat the best human (biological computing entity) at Chess was it any more significant than

    Except that AI did defeat Lee Sedol, a Go World Champion, with moves no human could foresee.

    or was it just more brute force computing capability and the writer not being able to “foresee?”

  26. Dystopian AI:

    Data feeding AI consisting mostly of material from Tik Tok, Meta, and X from those who believe in UFOs, Flat Earth, Big Foot, and Hamas, or the current college students at the Ivies.

  27. Huxley:
    Don’t know if you know, but the current AI is based on digital modeling of biological neural networks.

    Sure, I also know they’re inspired by BNNs and not mimicking them.

    Here’s my view:

    AI is simply a bad label as the “I” implies a lot more than computation. As it’s currently being misnomered in news and unfortunately in a number of papers, it is a grammar processor (mimicking art is about the same and about as good) regardless of it’s sophistication and, as such, is not intelligence.

    John McCarthy: inventor of the phrase “artificial intelligence” and LISP (List Processing), the first programming language for AI research and still in use.

    Intelligence is the ability to acquire and apply knowledge and skills (Oxford). The problem with current AIs is ‘knowledge’ and ‘skills’ are generic, which makes ‘ability to acquire and apply’ generic, making ‘intelligence’ generic.

    Like I said, bad label, so there will always be disagreement.

    Everybody gots their own POV, so I’ll always say chats are information manipulation toys, not intelligences and that to achieve AI we’ll need another approach entirely.

  28. it is a grammar processor

    One of the revelations is how much of normal human intelligence is “grammar processing” 🙂

  29. It may be useful for generating ideas, but I will hold my shekels on the claim of “creativity”.

    Ben David:

    Good to hear from you.

    I’m not convinced as I interact with AI that it has consciousness, creativity or soul.

    My problem is that I can’t tell whether at rock bottom I’m any different from AI. I’ve got good memory and good pattern matching skills. I know how I turn those into “creativity.”

    Similarly, though I don’t know Go but I do know chess, I know that the chess grandmasters who make those floating, brilliant moves which make us swoon, have spent ungodly amounts of time crunching chess data — which I understand as your argument against AI’s creativity.

  30. Re: Difference between Deep Blue beating Kasparov in chess and AlphaGo beating Lee Sedol in Go

    om:

    That’s a key question. The answer is that Deep Blue and AlphaGo were programmed in fundamentally different ways.

    Deep Blue was programmed as you described — brute force look-ahead. It worked great for chess and checkers.

    However, that approach won’t work for Go. In chess there are around, on average, 10-20 possible moves in each position. In Go that number is, on average, 200 moves. The Go look-ahead tree explodes past even a computer’s resources to handle.

    For decades some argued that AI could never play serious Go and certainly not at the world champion level.

    Then AphaGo happened and Lee Sedol went down.

    The difference is that AlphaGo is based on machine learning, not the previous static brute-force method. The machine trains on data and develops a cumulative representation of what works and what doesn’t, which becomes a sort of intuition to be used later in actual play.

  31. I’ll always say chats are information manipulation toys, not intelligences and that to achieve AI we’ll need another approach entirely.

    Oligonicella:

    I understand.

    However, as I said to Ben David, I don’t know if deep down I’m much more than an information-manipulating machine, at least when it comes to how I think.

    We may not need another approach.

  32. huxley:
    However, as I said to Ben David, I don’t know if deep down I’m much more than an information-manipulating machine, at least when it comes to how I think.
    ————————————–
    Let’s try it this way:

    As part of the annual cycle through the Torah, Jews around the world just finished reading the story of Joseph and his brothers.

    Like most of the Jewish bible the actual text is terse and especially tight-lipped about thoughts or motives.
    Would AI be able to infer the motives of the players?
    Would it place the interlude where Judah falls off from his brothers in its proper narrative context?
    Would it parse Joseph’s own moral struggles in Egypt and how he tested his brothers?

    … almost none of this is explicitly stated in the actual text. As in most great literature – there is a lot of showing and little telling – no frontal exposition.

    Yet you and we understand, infer, imagine what’s missing, sympathize.

    I guess I agree with Oligonicella – the word “Intelligence” is a hypey misnomer.

    From the sublime to the…. well, the ephemeral: This thread started with song lyrics.

    In a few seconds we could come up with a list of songs whose titles, lyrics, and humor depend on counterintuitive analogies and associations. Or that set a mood through oblique references.

    Without pre-existing models to learn that tether these words and ideas, AI could not create these connections.

    Yet you and we “get it”.
    We understand when a hat or a cigarette in an ashtray signifies love, nostalgia, joy, loss… we can divine (!) these key words, find the mood of the lyric even from its diverse, oblique parts…. And again, the good works avoid “on the nose” telegraphing of meaning.

    Would AI?
    AI can create a pastiche, can analyze and resynthesize what has been previously created – and that can be very useful – but it can’t make the aha connections.

    Love is like a day in June
    Love is like a giraffe

    AI could generate an infinite list of them, ranked by how often words appear near the word “love” – plus many other data points – but it could not evaluate the list, could not find the ones that leap off the page, or are funny. And the really “unintuitive” ones would be at the bottom of the list.

    But we can.

  33. huxley:
    My problem is that I can’t tell whether at rock bottom I’m any different from AI. I’ve got good memory and good pattern matching skills. I know how I turn those into “creativity.”

    You just articulated the difference. You can be creative. On your own.

    I dropped in this morning because the difference ‘twixt us and AI was demonstrated to me this morning by accident.

    I visit my neighbor most mornings and today Joe was. Joe was rummaging around for the TV remote and Curt suddenly burst out in “What you want? Baby, I got it. …”

    Until consciousness is achieved, AI will never be more than a response mechanism. It would never have had the “inspiration” to leap from the visual of Joe searching for the remote (or even knowing it was the remote being searched for as nothing was said) to Aretha singing. What experience could it have had to make those connections?

    It would have to be prompted sufficiently.

    And, carrying the example even further, an AI wouldn’t have noticed that the episode was an argument against it’s own intelligence.

    EOT for me.

  34. but it can’t make the aha connections

    This reminded me of an article discussing ” The Theory of Moral Sentiments” by Adam Smith. IIRC, Smith saw the development of moral sentiments as having stages, starting at what I would call the ChatGPT level, followed by a stage where the child internalized by recognizing that others had the same feelings as himself. I think that internalization is what is missing in the current state of AI development. I distinctly remember when I first had that insight about others when reading “Tom Sawyer” at about age ten. It was an aha moment.

    It is curious that Adam Smith should gain new relevance in experimental economics. Maybe he will also be relevant to AI.

  35. there is a certain dichotomy, a good cyborg is data, a bad one is lore, Daneel Oliwaw was the good one in Asimov’s tale, and there was an evil one, name escapes me now, David was the evil one, in the revamped Aliens, Walter was the good one, (David did the Cain thing)

  36. Ben David, Oligonicella, Chuck et al.:

    I understand your objections, but I like to think I’m looking at a bigger picture, given the curve of AI intelligence.

    I’ll concede you have some valid points, but I’ve heard it before with earlier states of AI.

    The concern in neo’s post is that Chat doesn’t write as good lyrics as Lerner & Loewe. This is a “the glass is 10% empty” interpretation. How in the world does a computer program come to create a mediocre version of a “My Fair Lady” song?

    I say AI is only going to get better. I fear and suspect it will do so more quickly than we are ready.

    To be continued….

  37. I say AI is only going to get better.

    I don’t doubt that. And speed has advantages, extensive trials are useful even for the most intuitive, witness Edison and Gauss. It also seems that science, in the sense of repeatability and experimentation, was an invention that served as a corrective to the ChatGPT thinking that comes naturally to us all. How much of what we “know” is based on anything but what we have heard or read? Helen Keller’s stories comes to mind, she was accused of plagiarism. About which I excerpt this little bit from a letter to her from Mark Twain:

    Oh, dear me, how unspeakably funny and owlishly idiotic and grotesque was that “plagiarism” farce! As if there was much of anything in any human utterance, oral or written, except plagiarism! The kernel, the soul–let us go farther and say the substance, the bulk, the actual and valuable material of all human utterances in plagiarism. For substantially all ideas are second hand, consciously or unconsciously drawn from a million outside sources and daily use by the garnerer with a pride and satisfaction born of the superstition that he originated them; whereas there is not a rag of originality about them any where except the little discoloration they get from his mental and moral calibre and his temperament, which is revealed in characteristics of phrasing.

    The whoe Twain letter is here.

Leave a Reply

Your email address will not be published. Required fields are marked *

HTML tags allowed in your comment: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>