Raw LLM Responses

You should really try making a video, or some images, using AI. It's an eye opening experience that reveals how bad it really is, and how it is definitely *not* taking everyone's job anytime soon. Ask it to make a very specific image, and you'll see what I mean after one or two tries. In case you're not willing to try it, you'll almost certainly encounter something like the following (spoilers): Scenario: we are making a commercial for lexapro or something, and we need an image of a man and his daughter throwing a frisbee in the park, where the kid is wearing a pink dress, appears about 7 or 8 years old, with curly auburn hair, and the man appears to be middle-aged, is wearing tan khakis, a brown shirt, has dark brown hair in a sort of retro bowl-cut and glasses. A totally random but very specific image that is very simple. Surely not a problem for this highly complex technology that requires massive data centers to power. Right? 1st result: we get an image of a man wearing a pink shirt and dark brown khakis, walking through a park hand in hand with a young girl, each holding frisbees in their free hands. They're both wearing glasses, and have bowls over their heads. 1st correction: we abandon the word bowl from the prompt and go with a celebrity that famously sported the sort of style we're after instead. We emphasize that there is only one frisbee and that its being tossed between the two. We add emphasis on the clothing and hair details, and that only the father has glasses. 2nd result: we get an image of Paul McCartney wearing a dark brown dress, walking through a park with a young girl, also in a brown dress, piggy backing over his shoulders. The girl is tossing a frisbee up into the air above her, and they're both still wearing glasses. On the plus side, their hair does more-or-less match the prompt this time. 2nd correction: we abandon our vision for their hair styles, cut out the reference to the celebrity, and emphasize that this is a father-daughter pair standing on their own feet. We emphasize that the frisbee is being tossed back and forth between the pair. We re-emphasize that only the daughter is wearing a dress, only the father has glasses, etc. 3rd result: we see a man with a buzz cut and young girl with pigtails in a park, and this time their bodies are oddly merged such that they seem to be standing on each other's feet somehow, their bodies twisted in freakish ways for this to happen and it's actually sort of disturbing to look at. On the plus side it got the clothing and glasses correct this time, so there is a sense of progress somehow, despite this clear setback. 3rd correction: we reuse the same prompt but add in extra emphasis that there are two distinct people, standing well apart from each other, adding things into the negative prompt like 'merged bodies' to avoid another weird result like the last one. 4th Result: we see a father-daughter pair standing side by side with a wide gap between them, but facing the same direction, both tossing a frisbee ahead of them as if competing in a simultaneous discus throw. There are minor flaws in the clothing, but it's close, and it correctly displays just the father in glasses. 4th correction: we focus on the specifics of how they should be arranged for it to appear like they are tossing a frisbee between them, describing how they should be facing each other, standing about 10 meters apart, with one of the father's arms extended out and a frisbee leaving his hand toward the daughter who has her hands raised to catch it. We feel our heart rate picking up because we can't understand why it can't seem to understand this very simple concept, and it's annoying that we have to be this granular. 5th result: finally we get an image of two people tossing a frisbee to each other. Unfortunately, its a mom and daughter pair this time for some reason, and their faces are virtually identical. 5th correction: in all our focus on tossing the frisbee, perhaps we sort of diluted the part about it being a man and a girl. We repeat that its a father and daughter several times in the prompt to correct this. We feel our face flushing. 6th result: it comes back with a pair of fathers and daughters, like some sort of doublemint gum commercial. We blurt out an expletive and are immediately embarrassed because our co-workers all stopped what they were doing to glance at us. This can go on, and on, and on. We can either (A) keep spiraling down this drain hoping that we'll luck out and it will get it right eventually (which it might), or (B) we can cut our losses, take a deep breath, and start over with a dramatically simplified prompt that abandons the specifics we had in mind and just emphasizes the parts we absolutely must have. Conclusion - if you don't really need something specific you can just do (B) from the start. But usually you need something specific for a reason, so you have to do (A). It's not always that bad, to be fair; a good chunk of the time you get lucky and sometimes fairly early on, but most of the time there will be a struggle to get there, and maybe 1 in 3 or 4 of those struggles will be... very painful. Given the environmental, financial, and societal costs to run these things, is it really worth it just to have a tool that can sorta sometimes get you a version of something you want? How useful is a tool that you can't really control? These quirks are also not really improving. The only thing that seems to be changing is just the speed at which you get your result.