20 Comments
User's avatar
Jeff's avatar

It's a jeopardy machine. You give it the clue, and it gives you the obvious answer.

Expand full comment
SP&EX's avatar

Exceptional formulation! Another commenter suggested it’s the ETF of stocks

Expand full comment
Lee Coursey's avatar

Try generating a photo-realistic image of a troll living under the sink of an employee break room. No matter what I tried, I could never get ChatGPT to do so. It was willing to generate a warning poster reminding employees to feed the troll, but it simply would not generate either a photo or a painting of a troll. You can read the entire, long, fruitless exercise here: https://chatgpt.com/share/67ef130a-47b0-8001-b4f6-6788f7ad23cd

Expand full comment
firtoz's avatar

Yep mine also failed, very odd

Expand full comment
Nicholas Rook's avatar

Weird. It also hit some strange block for me, but I managed it eventually.

Expand full comment
Andy Baio's avatar

Weird, worked for me on the first try with the exact same prompt you used. https://chatgpt.com/share/67ef5843-ad0c-8007-a82c-0847f15e4460

Expand full comment
CRP's avatar

Is this not using diffusion methods rather than an LLM? I get that it's the same interface and chat but an LLM isn't making these images.

Expand full comment
Chris's avatar

No, it's a different method than diffusion

Expand full comment
Benjamin Cordier's avatar

Memory’s a hit hazy, but I believe the new version may have moved away from a diffusion model.

Expand full comment
Chembe Banda's avatar

I enjoyed this post.

Expand full comment
Diogo Gomes's avatar

This is the problem with iconic figures... they are iconic. Here, fixed your copyright infringement: https://files.catbox.moe/uy7mh6.webp

Expand full comment
Derek Martin's avatar

Great article. Loved it.

Small correction, though: He-Man lives in Castle Greyskull (pictured), while Skeleton lives in Snake Mountain. Interestingly, the LLM knew what you meant and generated the correct image regardless.

Expand full comment
Lukas Schreibt Shite's avatar

Wow those pictures. Great read!

Expand full comment
Andy the Hibee fae Tuscany's avatar

There is a little forgotten issue in this whole thing: building up such movies, characters, etc. IS SOMEONE'S JOB. That, if we make it cheaper or simply more convenient to be done by AI, this means that no one will ever do anymore.

Why should I pay a pro when AI can do the same job for free or cheap?

But if someone won't raise a wage anymore from doing a job, no one will ever do that job again.

Is this what we really want?

"I want AI to make my home works easier so I have much more time for studying, writing, reading, etc. Not AI would do my studying, writing, reading, etc. to provide me more time to do my home works!" (AI Polish researcher, sorry, I forgot her name!)

Expand full comment
Hans Elate Gretel's avatar

IMHO it's not stealing and never was - at least in the abstract, as the law bought with money may claim otherwise

Expand full comment
Paul's avatar

Tried with Grok and it didn't have any issues infringing all those precious IPs.

Expand full comment
Stephen Jones's avatar

You've given very specific prompts that, intentionally, represent IP characters outside of the top 10 largest franchises. So, when enough of those {generic but specific description} match {a specific IP}, you're going to get a mashup/blending of images that (mostly) match that IP.

I suspect that a large sample of the top 10 or so IPs have been tagged with a secret {do not generate} label and if a high-enough percentage of those images/features are included in the AI output include that label, the prompt is rejected.

So, an opt-out system would allow IP holders to add the {do not generate} label for their works, too. That's tricky because it would be too easy, just like it is with false content-strikes on YouTube, to maliciously tag other works as your own IP.

It's a tough problem. As we've seen with experiments with "jail-breaking" textual AIs, it is almost always possible to hack around such blocks and should you really be able to block a "style" anyway? An image generator might conscientiously block "Ghibli-style" images directly but "Dreamy Minimalist Japanese Animation" might generate that style anyway. Should that be forbidden?

Expand full comment
Francois Rejete's avatar

When you prompt for the image of an Italian plumber, it starts generating the image without question, but at some point I think the system just plugs back the image it generated asking for a description of it and when the description comes back as "Mario" it hits the blacklist of IPs that you shouldn't touch and cancels your request.

Expand full comment
Tom Kinniburgh's avatar

I disagree that most of those prompts are truly specific. They all end up being from movies, that doesn't mean it's been trained to find multiple archeologists in hats. It's scraped multiple MEDIA i.e Marketing images from the internet and then labelled them up.

This is copywrite infringement in clear sight.

Imagine I asked an artist on fiverr to draw me an archeologist with a hat, they certainly wouldn't give me a photo realistic Harrison Ford. They would look at that image and then interpret it in their own style. That's what the other photo algorithms do, but not the new open AI one. Likeness is one thing, direct like for like comparisson is another.

Expand full comment
B Griffiths's avatar

Does this mean AI has no free will? 😅

Expand full comment