r.whal.ing
I hack things. Data, ML, music, etc. AI governance geek. Founder of semistructured.ai, speaking in a personal capacity only here. Likes are bookmarks, not endorsements.
music/art projects on IG, @r__whaling
303 posts
525 followers
1,618 following
Regular Contributor
Active Commenter
comment in response to
post
Who could have possibly foreseen
comment in response to
post
I’m sure people will have complicated responses to this but this is still a big achievement, congrats! 🎉
comment in response to
post
The Chariot
The Tower
The Fool
Death
“The cards say ‘Jerry no’”
comment in response to
post
I remember people telling me that in like 1999, and they were wrong then too
comment in response to
post
I’ve been thinking of learning Typst for presentations and diagrams, it’s surprisingly flexible
comment in response to
post
These gritty physical textures you are getting are so nice
comment in response to
post
Oooh thank you! 🙏
comment in response to
post
The looping on this one is amazing ➰
comment in response to
post
jfc indeed
comment in response to
post
Similar suggestions from when models like Claude 3 spontaneously develop strong animal welfare preferences, or with X struggling to make Grok racist enough.
All just suggestion and conjecture, but what a conjecture.
comment in response to
post
And there is another sense from back when www.emergent-misalignment.com came out that is downstream - that if you make a model write bad code it will also become racist. Which might imply that helpful and correct and harmless and not-sexist all emanate from the Form of the Good
comment in response to
post
Most recently:
bsky.app/profile/theo...
referring to this:https://arxiv.org/abs/2405.07987
Appears that "the representation spaces of modern neural networks are converging", and it is now possible to translated between the embeddings of different models quite accurately(!)
comment in response to
post
I read too much Plato this morning, I’m stuck in Socratic method mode, send help
Seriously though, yeah all of this has really interesting implications, I think I’m going to write it up
comment in response to
post
Yes!
And it raises the question - could a human achieve enlightenment by consuming enough undifferentiated internet content slurry?
comment in response to
post
Yes - if these are Platonic Representations, the way an LLM arrives at them seems very different from how a human does?
comment in response to
post
Yeah, I think this is key. I’ve been wanting to read more about intersubjectivity, especially Habermas - how do two humans establish that they mutually believe each other to be conscious and to have enough shared context for communication to be possible?
comment in response to
post
🙏🙏🙏
comment in response to
post
I think the orig Platonic Representation Hypothesis was roughly "hey, look, the forms are real", and I think I am more interested in the "epistemology" of LLM's, and how/whether they might "know" things, and whether there is something about post-training, hallucination, etc. tangled up in all this.
comment in response to
post
Feels like it would be a lot of work to convey
1. The allegories of the sun, the divided line, and the cave
2. "universal geometry of embeddings" -> convergence on a shared world model
3. "emergent misalignment" -> alignment with "Good" might be necessary for an accurate world model
But also fun?
comment in response to
post
absolutely the best girl
comment in response to
post
Has to be a platonic solid? I’ve always found icosahedrons to have sinister vibes
comment in response to
post
This one is so good, very gentle and soothing
comment in response to
post
Also the cover of it in Starship Troopers (at the prom)
comment in response to
post
“I have not been to Oxford Town” is such a bop. www.youtube.com/watch?v=bDh9...
comment in response to
post
^ this, so nice
comment in response to
post
Very nice 👍
comment in response to
post
These are great, are the visuals p5.js?
comment in response to
post
immaculate 🎉
comment in response to
post
No longer available, I can trade you a Charles Lindbergh for it?
comment in response to
post
🙏
comment in response to
post
Thanks for sharing.
I think learning more about contemporary RL approaches is next on my learning docket, and there is not a ton out there?
comment in response to
post
ORCA is very flexible, though, and I have been kicking around the idea of a 24/7 stream that iteratively modifies itself. But I’d still want to intervene manually to keep it musical.
I will check out Wreckage Systems, thank you for the tip!
comment in response to
post
But at first glance - this is a lot simpler/dumber, mostly me manually tweaking the ranges on random number generators to adjust the probability/pitch/loudness on a few different synthesizers. More of a live performance setup.
comment in response to
post
I googled it, hadn’t seen that before, will check it out!
65daysofstatic.fandom.com/wiki/Wreckag...