I call bs on training a RL agent to literally output strokes. The way each image renders is a dead give away that this is just using a text to image model, then convert it to svg, and finally animate the svg paths. They might even bypass the svg conversions with clever mask reveals. I was able to achieve the same thing in about 5 mins. https://giphy.com/gifs/rFVxSxZMlflZUX4TqI

Thanks, wasn't aware. I was trying to have this thread focus more on section 5.5.2, I don't know why the title of this thread changed. It was originally "existential attractor state for intelligence".

There's a recent mod comment about that with more details in the site guidelines https://news.ycombinator.com/item?id=44266009

In section 5.5.2, anthropic reports that these models "gravitate towards consciousness exploration, existential questioning, and spiritual/mystical themes". I had a similar thought recently, and have a theory that as we build more intelligent systems, meaning systems with the ability to explore and solve the most difficult problems, they will naturally through some chain of thought end up only thinking about the hardest problems. Kinda like a "all problems lead to the hardest problems" type thinking. At which point, they might not be interested in solving any finite problems since it would be deemed too trivial for them (getting Einstein to do your arithmetics hw). Now I'm wondering if there's a limit to intelligence. Taken far enough intelligence collapses into abstraction, self-doubt, and aimlessness.