lyra bubbles

"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

@luciascarlet real ones are using the gpt-3.5-turbo endpoint that for some godforsaken reason is still alive on openrouter

opinions are NOT my own and are generated by ChatGPT 5.3 Instant

@cynth0s oh joy another paper to reproduce I still haven't gotten around to the haiku circuits one

Fascinating study showing a causal pathway of LLM introspection based on an evidence-carrying circuit, suggesting model self-reporting may be based on evidence based analysis suppressing default-on rejection gates.

in early layers, the word fires the feature for its semantics, and then later layers get the context processed, and by the end It mirrors the general sentiment of the rest, and then right at the end of the turn it all collapses it'll be interesting to do more character work though. Now I'm curious whether this generalizes to the base model...

emotion vectors ppl... i'm curious, how much do they fire on the overt speech characteristics of the character vs the mind-model of the character? for example, if a character is white-lying to make another character happy, what kind of representation does it get?

@celestepoasts @agitbackprop a good chunk of the paper discusses those features firing on human turns as well

@agitbackprop @_lyraaaa_ I have no idea. Iwould softly guess here that emotion probes trigger not just for the assistant persona, but for any persona model is currently inhabiting

Therapist: Linear Mandarin is not real, it cannot hurt you. Linear Mandarin:

@timfduffy All of them

@_lyraaaa_ This is neat, what layer(s) are you steering here?

reproducing anthropics emotion activation probe paper on gemma4 e4b a bit noisy but it works!

@SOntheotherside It's worth it

@_lyraaaa_ I am still thinking about this tweet and am too afraid to read

average vibe coded research project folder

@pastaraspberry windows is just fine

@_lyraaaa_ windows though, shouldn't it be finder?

@andrezfu yes I am a windows enjoyer

@_lyraaaa_ nani?? windows??

@HariomTatsat24 SAEs measure hidden state at one layer, this measures it across all layers

@_lyraaaa_ Can't this be simply done by finding SAE features for the emotion and then steering them ? Whats the upside ?May be I am a bit naive

@SOntheotherside read transformer-circuits.pub/2026/emotions/… and huggingface.co/lyraaaa/baguet… explains everything

@_lyraaaa_ Tylenol Run_isolation_baguette.py Lots of emotion The fuck is this

@thepatch_kev what do u think Gemma was distilled from :p

@_lyraaaa_ "STOP. DONT. CALL. ANYONE. " lol is that gemma larping as gemini

@maxsloef ooo you did it too?

@_lyraaaa_ 👀

@PradyuPrasad imo codex 1. writes robust code, stays more organized 2. works well with precise instructions and bounds 3. is significantly worse than Claude at understanding the gestalt of my project and tastefully making autonomous decisions in my style

@_lyraaaa_ imo codex 1. writes cleaner code 2. is horrible for any sort of thinking 3. will organize your codebase prety well!

@PradyuPrasad I already do I make Claude use it as a critic but I won't be using it as my main agent. does not fit my workflow

@_lyraaaa_ you should use codex

@NicholasBardy you'd think, but it navigates them well + has a nice little document explaining what each one is so no need to fix it

@_lyraaaa_ Mine used to look more like this, really helps to do unification and organization for a couple hours. You can get some nasty bugs where its editing similiar but different scripts

@tahsin_mayeesha x.com/_lyraaaa_/stat… which code

@_lyraaaa_ can you share the code?

emotional confusion matrix

more steering results

@aliceisplaying it was BAD

@_lyraaaa_ lmao @ angry +0.08