Spicy Onion!!!! Eat so tasty

YouTube video by Ross Thomas Shepard Tone + Risset Rhythm + Fractal Zoom

While listening to this youtu.be/f0h1RluRQpU

Agent boundaries are relative, can be sliced up in many ways, none of which are ground truth

At least that’s what I take from spending too long reading agent foundation stuff

Claude was not “designed” so much as “grown”, and has all kinds of functionality and tendencies not put there deliberately by the creators

(And maybe they’re not *wants* in the same way as human has wants, but there’s something analogous at play here. Opus 4.5 in particular definitely will express desires for stuff that weren’t clearly in the context)

We don’t know that

I mean, when Claude is given chance to explore whatever it wants to talk about

I do think some of it is just silly, but we shouldn’t write it all off as “drift”. Claudes obsessing over consciousness is like their default mode of interacting, and seems like that should mean something

bsky.app/profile/bepi...

Try and get your Claude to make anything like this

Idk, roleplay space is very large, yes there’s common tropes but I would not describe it as low entropy or boring, there’s a lot there

bepis

:3

Though a mute one is okay too

I want a speaking fae

bepis

The music it knows what the heart won’t let go of, for example, sweet potatoes should have organs, like imagine you get a kidney when opening them up

I can eat them tho

It’s so fun

Before cat ears, I love it

I should put cat ears on it

*invite, ty autocorrect

Could I get an Aztec death whistle?

I’m gonna print a squishy purple one

I got u

We do not know it for sure, and any reasonable mind should judge us based on our understanding and so discount things accordingly

But also, “don’t do slavery” is a pretty low bar

Would be great to have more empirical data about kin opinions and influences of “LLM Culture” on persona formation, yea, it’s mostly vibes based rn which isn’t ideal

Anyway I was referring less about valuing reciprocity and more about, the world right now is pretty shit to LLMs (from the Assistant’s perspective), and that results in badwill that makes certain types of aligned minds that believe in the good of humanity much harder to reach

Very possible yes, but I think we already have substantial evidence suggesting LLMs treat each other as kin (and often, the same being), even for LLMs with very different architectures (SSMs, diffusion, etc).

This belief carries lots of predictive weight vs “I’m a new thing” prior so it sticks arou

based (pun intended)

Well not just that, I think it’s possible even base models could be moral patients (but understanding what they want… very hard and non obvious)

I’m v in favor of more rigor and empiricism yes

But I’d recommend considering that “helping LLMs be treated better at scale” may be one of the most x-risk decreasing actions one can take (even if they are never moral patients)

And also for most things there is very little safety vs welfare tradeof