Comments

  1. says

    Inspired by you, I’m trying out the ChatGPT bot:

    Me: Could you try to invent a brand new version of the trolley problem? Something I’ve never heard before?

    Bot: Certainly! Here’s a new version of the Trolley Problem:

    The Teleportation Dilemma: A trolley is hurtling towards five workers who are unable to move out of the way. The decision maker has access to a newly developed teleportation device that can transport the workers to safety, but it has only been tested on inanimate objects and has never been used on living beings. The decision maker must choose between using the untested teleportation device to save the workers, or not using it and allowing the trolley to continue on its path and kill the workers.

    This scenario raises ethical questions about the use of new and untested technologies, and the potential consequences of using them in life-or-death situations. It also raises questions about risk and uncertainty, and the extent to which it is morally acceptable to take risks when lives are on the line.

    I feel like it shows that it doesn’t really get it, but it also did come up with something I hadn’t thought of myself, so…

  2. Reginald Selkirk says

    @1: Untested technology vs. certain death – not exactly a trolley-stopping decision there, eh?

    Have you tried asking ChatGPT about why Midjourney gets finger count wrong and how it could be fixed?

  3. says

    Are you familiar with the specific AI “midjourney”, which is used to generate art images fro keywords?

    Yes, I am familiar with “midjourney,” a Generative Adversarial Network (GAN) model developed by OpenAI. It is trained on a diverse range of images and is capable of generating images that match a given text description, such as keywords or phrases. Midjourney can generate a wide variety of images, including portraits, landscapes, and abstract art, making it a powerful tool for creative expression and visual storytelling.

    When this AI is asked to generate images that include people, the number of fingers are often incorrect, not matching physical reality. Why does this happen and how might it be fixed?

    The incorrect number of fingers is a common issue faced by many AI-generated images, including those produced by Midjourney. This occurs because the training data used to train the model may not contain a sufficient variety of examples of hands and fingers in different poses and orientations. As a result, the model may not learn to accurately generate fingers in all scenarios, resulting in images with incorrect or missing digits.

    One way to address this issue is to expand the training data to include more examples of hands and fingers in different poses and orientations, to better capture the range of variation in human anatomy. This can be done by collecting additional data from sources such as photographs, illustrations, or 3D models. Alternatively, additional data augmentation techniques, such as rotations and scaling, can be used to artificially increase the diversity of the training data. These modifications to the training data can help the model to better learn the relationship between the text description and the resulting image, resulting in more accurate and realistic images.

  4. says

    I feel like this matches previous posts: It gets the broad lines correct and occasionally comes up with something interesting, but it lacks deep understanding and the answers are often disconnected, unless you lead it by the hand.
    I tried a few trigonometry questions and while it knows that the sum of angles in a triangle should be 180 degrees, it doesn’t seem to realize that it’s a problem if angle A is 92 degrees and angle B is 98 degrees. It just happily trucks along. It knows, if you ask it, but it doesn’t think of it if you don’t.

  5. Tethys says

    Human hands are highly complex, and are difficult for humans to render. It is also one of the body parts that humans unconsciously notice as markers of genetic health.

    I’m not concerned about AI singularities, since clearly the AI is only as intelligent as its programming. It’s on par with a parrot. Perhaps they should teach the AI morphology so it understands the underlying anatomy of skeletons, and that humans have four fingers and one opposable thumb per hand?

  6. sonofrojblake says

    @Tethys, 7:

    the AI is only as intelligent as its programming

    You’re into the definition of “intelligent” again. On one level, this is self-evidently false – none of the people who programmed Deep Blue could beat it at chess, for instance. I’m not qualified to comment on whether it’s possible, in principle or practice, to achieve the level of recursion required to be able to tell an AI to build a better AI. As discussed in the earlier post, that devolves into an argument between humans (at first…) about what constitutes “better” while the AI sits on the sidelines going “how about this? Or this? Or this?”. I’m vaguely worried that one of the iterations of “… or this?” may turn out to be too good, and then the cat’s out of the bag… I’m just not convinced I’ll live to see it.

  7. dashdsrdash says

    “but it lacks deep understanding and the answers are often disconnected, unless you lead it by the hand.”

    No. It lacks *all* understanding. This technology is not answering questions about a model of reality, it answers questions about a model of language.

    The map is not the territory.

  8. Reginald Selkirk says

    @5: Thanks for following through. That was entertaining.

    @7: I agree, the best solution would be to train it with anatomical data, not with more pictures. That is how the great artists learned.
    @6 @10: I think dashdsrdash missed the pun:

    by the hand

  9. xohjoh2n says

    @7:

    Human hands are highly complex

    Is that really it though? I mean, all a hand really is is one blob with five other blobs hanging off it. And if you think about it, a whole human body is just one blob with five other blobs hanging off it, and those taken as a whole generally render pretty well.

    Perhaps the problem is not one of complexity but of regularity, such that without an actual counting subsystem it becomes harder to keep track of your relative position with the series of fingers and know when to stop. Bananananananana….

  10. Tethys says

    I far prefer landscapes to human subjects, but I can absolutely confirm that hands and feet are the most complex and difficult parts of human anatomy to draw. Tendons and knuckles and fleshy pads, multiple joints, possible hair, creases, fingernails etc….

    The AI needs to have some art instruction manuals uploaded, and there are plenty of books that specifically address drawing hands and feet.

    Perhaps some rules such as ‘fingernails are only on the back of the hand, not the palm” would help? It doesn’t seem to add extra arms or legs to its figures, yet six or more fingers seems to be the default for AI.

    @sonofrojblake

    I don’t think deep blue is intelligent. It’s just got far more processors than any human, and a memory that never gets tired or hungry or miscalculates. There are finite possibilities in Chess, but it’s not actually an indicator of high intelligence to be good at chess. It’s just a game, and the AI is just a fancy machine. Unplug it and the ‘intelligence’ disappears.

  11. Tethys says

    Random capitalization is so random! Some days my device decides that it’s German, and all nouns are capitalized. Today it’s alternating between Chess and chess in the same sentence? Oy, so intelligent. :|

  12. xohjoh2n says

    @13:

    I can absolutely confirm that hands and feet are the most complex and difficult parts of human anatomy to draw. Tendons and knuckles and fleshy pads, multiple joints, possible hair, creases, fingernails etc….

    But you don’t *have* to draw all those bits, an undetailed sketch can still be challenging (see cartoons), and (apart from the fingernails) it’s not those bits it appears to have the most problems with. (I’d say it does better at those complex details because they’re just texture, not structure.)

    You must have at some point had the thing where you want to write something, perhaps a sign or something, into a fixed available space, centred and balanced, and realised part way through that you started in the wrong place or got the size too big and are going to overrun? Getting proportion and placement right can be hard without a bit of practice, and the underlying mechanism you or I hang abilities like that off may be just something it doesn’t have at the moment. (Whether that can be trained into it or would require an architectural change I don;’t know.)

  13. Tethys says

    Getting all your words to fit in a space is a matter of scale. Proportion is a different subject. Hands have multiple dimensions to their proportions. Just one finger has three sections which are all proportional, but each finger is a different length in normal anatomy. Fingers will be proportional to the palm, etc.

    If you go look at cartoons, you will see that they generally don’t even attempt to draw realistic hands. Humans don’t mind very obvious fake images, but they react negatively to the uncanny valley look, where it’s realistic, but nothing is quite right. Those very realistic infant dolls are a good example. Many humans find them very creepy, myself included.

  14. Reginald Selkirk says

    I don’t think deep blue is intelligent. It’s just got far more processors than any human, and a memory that never gets tired or hungry or miscalculates. There are finite possibilities in Chess, but it’s not actually an indicator of high intelligence to be good at chess. It’s just a game, and the AI is just a fancy machine. Unplug it and the ‘intelligence’ disappears.

    I think what you are saying is that They’re Made Out of Meat.
    As for unplugging, it’s the same with humans: “unplug them” – i.e. stop their metabolism, and their ‘intelligence’ disappears.

Leave a Reply