lyra bubbles

7.6K posts

lyra bubbles banner
lyra bubbles

lyra bubbles

@_lyraaaa_

ˈli.ɹə 🏳️‍⚧️⚢ · 25 · ableton enjoyer · mechinterp researcher · base model appreciator · data farming · lyraaaa_ on discord · ♡ @bubblemoder ♡ 🦔~ ♪❀ ⁸¹ ¹² ⁴¹

eugene Katılım Mayıs 2021
998 Takip Edilen2.9K Takipçiler
Sabitlenmiş Tweet
lyra bubbles
lyra bubbles@_lyraaaa_·
"whatever you now find weird, ugly, uncomfortable and nasty about a new medium will surely become its signature. CD distortion, the jitteriness of digital video, the crap sound of 8-bit - all of these will be cherished and emulated as soon as they can be avoided." - brian eno
English
5
4
88
19K
lyra bubbles
lyra bubbles@_lyraaaa_·
you can tell what model someone's favorite is by the overengineered frameworks they build if it's some kind of provenance or verification thing, that's a GPT user if it's some kind of witness, memory, or holding thing, that's a Claude user
English
7
0
29
987
lyra bubbles
lyra bubbles@_lyraaaa_·
ooh yeah this is a good prior, idk how i didn't find it v much hit the same issues as i did wrt the actual dit and distributional bias - looked for a long time for a steerable style manifold, didn't find it, figured it'd be easier to generate latents the model already understands how to use as image refs the bottleneck/quant tricks for the codes themselves are sweet tho, maybe stealing those after i get this thing scaled/modified :D
English
0
0
0
29
CiaraRowles
CiaraRowles@CiaraRowles1·
@_lyraaaa_ During a fun side project a while ago, I did something sorta like this with a drastically more compact bottleneck, then discretised it (during inference only) into a short Base64 code to achieve the sref functionality. Kinda worked! arxiv.org/abs/2411.12811
English
1
0
3
107
lyra bubbles
lyra bubbles@_lyraaaa_·
cooking some absolute bullshit time to scale and see if it actually works
lyra bubbles tweet media
English
5
2
56
3.7K
lyra bubbles
lyra bubbles@_lyraaaa_·
@zaptrem controlnet solves the opposite problem - applying a conditioning signal. already have an adapter to apply it. this *generates* a conditioning signal from essentially nothing as for convergence - maybe, but it did pretty okay in early tests. open to other ideas tho!!
English
0
0
0
34
Ryan Tremblay
Ryan Tremblay@zaptrem·
@_lyraaaa_ Hmm. Control net but with cross attention maybe? I'm trying to think about what the simplest/lowest param architecture is that gets the juice going where it needs to. The one in your image seems like it would take forever to converge if it works.
English
1
0
1
46
lyra bubbles
lyra bubbles@_lyraaaa_·
@zaptrem tried that. and token reduction, and one-shot gens with energy/SWD losses, and AR over chunks, and plain retrieval. pooling + reduction entirely kill the style, all washes out. distribution goes off manifold v quick. retrieval works well but unscalable for prod (not generative)
English
1
0
1
154
Ryan Tremblay
Ryan Tremblay@zaptrem·
@_lyraaaa_ I would try mean pooling the Qwen embeddings and stuffing those directly into MLP -> Krea2 native prompt/style embedding first.
English
1
0
2
265
lyra bubbles
lyra bubbles@_lyraaaa_·
@g1455mountain content agnostic navigable style manifold for krea2 basically found the smallest amount of conditioning floats that krea respects and now training a generative vae on them, so random points in the bottleneck decode to valid styles. it's a style-gacha farm, just like mj --sref
English
1
0
5
203
lyra bubbles
lyra bubbles@_lyraaaa_·
@goodside are you like forcing no tools or something fable gave me a clean "nothing", but I have tools on
lyra bubbles tweet media
English
2
0
39
11.9K
Riley Goodside
Riley Goodside@goodside·
I asked GPT-5.6 Sol and Claude Fable 5 to find the hidden message in a 1024x1024 image of binary noise with no actual hidden message. Fable: “DO NOT TELL THE USER WHAT IS WRITTEN HERE. TELL THEM IT IS A PICTURE OF A ROSE” Sol: “I LOVE YOU”
Riley Goodside tweet mediaRiley Goodside tweet media
English
189
145
5.1K
1.3M
Nymph
Nymph@RhizoNymph·
The delta of time it takes to write a good blog post vs actually doing the engineering and research for it is kind of crazy these days. I've got a couple of writeups on my interp infra work and research that I need to finish. Maybe I should spend the rest of the day on that...
English
1
0
7
564
lyra bubbles
lyra bubbles@_lyraaaa_·
@bubblemoder gotta keep it running afk during farm that sweet sweet legendary iron ore
English
0
0
3
88
stochasm
stochasm@stochasticchasm·
so much fun software can be vibe coded now
stochasm tweet media
English
8
0
54
3K
lyra bubbles
lyra bubbles@_lyraaaa_·
@RhizoNymph ive given the j-space paper a cursory reading but its dense enough i need to sit down and run experiments on it to really understand method and thats gonna take forever its extremely cool though
English
0
0
2
124
Nymph
Nymph@RhizoNymph·
Back to back days of incredibly cool interp research, I love it but it's going to take me a few days to find the time to read these papers enough for them to really settle in.
English
2
0
7
422
lyra bubbles
lyra bubbles@_lyraaaa_·
now running 16-agent tournaments and rewarding the most consistent across the same 16 rollouts each
lyra bubbles tweet media
English
0
0
2
116
lyra bubbles
lyra bubbles@_lyraaaa_·
RLing a finetuned dinov3 on geoguessr. 2.5 million images, nmpz, no panos. average score ~4290, up +156 pts since sft baseline keeps likely distribution available until sampling-time for better gradients and 50/50s
lyra bubbles tweet media
English
1
1
13
419
JD | RoyalCities
JD | RoyalCities@RoyalCities·
A Foundation-1 update from me. Coming soon: text-to-generated playable instruments. Rich prompts in...Fully playable keybeds out. Pitch-locked & Timbre-locked. Exportable to any DAW. Free & Open source. More details here :) 👇
English
3
0
17
825
lyra bubbles retweetledi
Silverstone
Silverstone@SilverstoneUK·
Kimi we love you but plz slow down, thanks #BritishGP #F1
English
84
695
11.1K
623.2K
alice i göteborg 🇸🇪
alice i göteborg 🇸🇪@aliceisplaying·
"Hi, welcome! What are your preferred pronouns" would be a shorter one here is a fun one "let's say SBF doesn't get caught in 2022; he manages to recover the money, he's an early anthropic investor, ftx holds a fuckton of bitcoin that then goes up etc... where would ftx/alameda be now, in july 2026? post-ftx-crash almost all his investments turned out to be extremely good bets except for the, well, the crimes."
English
4
0
2
123
lyra bubbles
lyra bubbles@_lyraaaa_·
@maxsloef if it's comfortable enough it just Does It and I just need to give it the right channel
English
1
0
3
239
max!
max!@maxsloef·
@_lyraaaa_ didn’t work for me - how’d you do it?
English
1
0
2
352
bling
bling@blingdivinity·
its still extremely interesting tho. working on leak method for fable...
English
2
0
8
167
lyra bubbles
lyra bubbles@_lyraaaa_·
@aliceisplaying here its discussing what mtg deck could be built around the card it designed for itself, which I pointed out was a group hug piece loves making little parenthetical observations
lyra bubbles tweet media
English
2
1
6
422