This would actually be an interesting question for the more rigorous end of the mechanistic interpretability people to study. They decompose the system to find 'features' within different layers that are associated with different behaviors or concepts in the inputs and outputs, that activate or deactivate each other. Famous example being the time they identified a linear combination of activations in a layer that corresponded to 'the golden gate bridge' and when they reached in and kept their numbers high during the running of the model it would not stop talking about it regardless of the topic, even while acknowledging that its answers were incorrect for the questions at hand.

I actually would love to see what mechanistically happens to that feature when you put in the input 'do not talk about the golden gate bridge'.

Checks out. Political science, biological science, physics... we got them all. Might have to go to ancient egypt to get hydrology religion though.

So we are inferring that in the vector space of all possible sentences, QNTM is sitting at one of the attractors?

It's absolutely crazy, but I think Yud is the less unhinged person here

Uh oh

https://www.rollingstone.com/culture/culture-features/gwen-danielson-zizians-interview-1235552043/

Going on a media tour now

I'm a huge fan of Greg Egan's fiction and a huge fan of him pissing off the rats. He's been explicitly needling them and making fun of them in his fiction for over a decade. Making calm contradictions against them for over two decades, after noticing weirdos being fans of his.

Friend of Ziz and cofounder of the 'rationalist fleet' pops up out of the woodwork trying to clear Ziz's name

https://www.lesswrong.com/posts/mbrmZmzBdtn4qrSus/re-introduction-of-a-rationalist-dragon-and-clarifications

I find myself noticing things rather detached from the typical Ziz funnybusiness more strongly than I notice the stuff about that whole situation.

"I'm Gwen Danielson, a neuroscientist and bioengineer, who decided as a child that I would end Death (and bring people back if I could) and that I would become a dragon and help generally facilitate a fantastical transhumanist future."

"I dream of non-Euclidean geometries, of countless worlds visible and accessible in the daytime sky, of competent infrastructure, of soul forges continually working to bring back the dead... I dream of reaching through warps in the spacetime fabric to save the dying across time"

"Signed, the dragon of creation Creatrei (cree-AH-trey) also known as Gwen Danielson or as Char and Astria (when referring to my hemis as distinct individuals)"

The reactions are fun. "This post is not actually doing a good job of making me trust you and think this conversation is safe to have[1], and I notice that as I am saying this that I am afraid that this will now somehow result in someone trying to murder me in my sleep"

I have a vague hypothesis that I am utterly unprepared to make rigorous that the more of what you take into your mind is the result of another human mind, rather than the result of a nonhuman process operating on its own terms, the more likely you are to have mental issues.

On the low end this would include the documented protective effect of natural environments against psychotic episodes compared to urban environments (where EVERYTHING was put there by someone's idea). But computers... they are amplifiers of things put out by human minds, with very short feedback loops. Everything is ultimately in one way or another defined by a person who put it there, even it is then allowed to act according to the rules you laid down.

And then an LLM is the ultimate distillation of the short feedback loop, feeding back whatever you shovel into it straight back at you. Even just mathematically - the whole 'transformer' architecture is just a way to take imputed semantic meanings of tokens early in the stream and jiggling them around to 'transform' that information into the later tokens of the stream, no new information is really entering it it is just moving around what you put into it and feeding it back at you in a different form.

EDIT: I also sometimes wonder if this has a mechanistic relation to mode collapse when you train one generative model on output from another, even though nervous systems and ML systems learn in fundamentally different ways (with ML resembling evolution much more than it resembles learning)

Commentary straight from some of the sources

https://www.lesswrong.com/posts/BW89BudtySvpzpYni/stone-age-billionaire-can-t-words-good

Comments are fun

This is all he does now

The Great Leader himself, on how he avoids going insane during the onging End of the World because among other things that's not what an intelligent character would do in a story, but you might not be capable of that.

Gerard and Torres get namedropped in the same breath as Ziz as people who have done damage to the rationalist movement from within

https://www.lesswrong.com/posts/Hun4EaiSQnNmB9xkd/tell-people-as-early-as-possible-it-s-not-going-to-work-out