#ai_tag

For this test, please zoom in to view the image at a larger size and have a look at it before answering.

My personal feeling is that something about AI-generated images makes them uncomfortable to look at when observing the details, but I'm not sure if I'm just imagining it. I usually try to make the "AI Generated" stamp large enough so that it's always within the field of view while looking at an AI-generated image, but for this test, I have deliberately made it much smaller.

LLM responses are (deliberately) somewhat random, but an LLM will attempt to maintain consistency with its own prior posts in a conversation. If it makes an error (as they do), then it is likely to go uncorrected. If it makes a judgment, then it will likely not depart from it. This means that the errors (and the judgments) may compound. The longer the conversation runs, the more opportunity there is for it to drift.

The LLM drawing from past conversations provides three additional opportunities for error: (1) the LLM could draw from an unrelated conversation or document (e.g. a fictional document about an SCP Foundation task force), (2) the LLM could misread the information in the conversation (as they sometimes do), or (3) the information itself (possibly at the end of a long conversation) could already be an error.

"But... we don't even have cybernetic implants yet?"

"Yeah so it turns out that some people either haven't been reading enough science fiction, or have been reading too much science fiction, so a businessman, for example, can get one-shotted by an auto-generated SCP Wiki article."

"Well. Fuck."

This is consistent with smaller-scale issues that have been anecdotally reported by users of other models.

Read a couple of materialist-scumbag posts, and was wondering - how much does kontextmaschine use "layers" as a term in his written analysis? I didn't see that as key to his way of thinking. I see it as more key to my way of thinking (e.g. Now, Melt 10.a).

Me: Take these simple drawings from the fictional country, climb the ladder of detail and realism, and then put an actor in the same attire.

Grok: Sure!

Me: Astounding. Now animate them.

Grok: You betcha!

Me: Excellent. Create a video in a ballroom featuring all four characters at once.

Grok: Okay.

Me: Well, that worked pretty well until zooming back out again, but I bet premium reference image reloading would address that.

Me: Replace this 3D sketch of bleachers with something more realistic.

Grok: *sweating...*

Me: Tolerable. Now... *eyes shift* Replace that blue figure with this old man in a suit.

I'm not going to post video much, because video tends to use a lot of bandwidth and storage, and Tumblr is not known for its healthy financial position.

We are now in the position where a determined artist (someone who could actually do character sheets or key frames) could, in theory, use AI to animate an entire low-budget, sometimes wonky-looking, animated series about a clique of Tumblr users. (On the other hand, at least for Grok, it's not based on keyframes yet—which is, in some ways, not surprising.)

We are also in a position where every Internet meme could potentially be animated. Very cyberpunk. But does it really add much, or is it just noise?

As with AI image generation before, what you can do with depends on whether you're flowing with or against the currents of the training data. And also as with AI image generation before, there's an implicit, gambling-like element—each image rendering takes a little bit of time, and you don't know if you'll like the result, or if it will just be bad.