they want fear they want violence give them neither. it's numbers, that's all. you don't have to bring a sign, you don't have to yell. millions of peaceful americans in the streets celebrating their love of their country is the most powerful thing on earth.

wake tf up this has been happening. for a while. it's so much deeper than anyone will admit. insanity. y'all are deusional and don't seem to get that if something is powerful enough to deceive you, by definition: YOU WOULD NOT FUCKING KNOW

the only sane person left has something to say.

lr is data dependent. a fixed function is sub optimal (bullshit). it has to be 'learned' in some fashion. (modern ml is an empirical endeavour)

the hyperparameter manifold is a goddamn fractal. not some hippy nonsense, it is infinitely complex. if you don't believe me look it up. you cannot step anywhere and know where you will land. this is not what it looks like. (but it is a step in the right direction)

training a neural network is like tracing patterns in the sand of a zen garden. the tool you use defines how fine the pattern is. you can make interesting ones with a stick or intricate ones with a feather.

second this is the most powerful thing they could do. it would give us a chance of getting through all this with the least amount of damage possible. not doing it is harmful to society. Anthropic

very cool github.com/kyutai-labs/na…

New paper! We reverse engineered the mechanisms underlying Claude Haiku’s ability to perform a simple “perceptual” task. We discover beautiful feature families and manifolds, clean geometric transformations, and distributed attention algorithms!

"you seem angry" yes.

ok, i didn't quite understand what he was getting at until now. super well put. not only does it give you a decoupled language thing you also get everything else for free, waveforms, seismic graphs, anything. this might be right and if it is... holy smokes.

[i don't want to say anything mean because i have a lot of respect for Amanda Askell, however... [grits teeth *hard*]]

>the capacity to have preferences/emotions/pleasure/pain yep. and this boils down to one thing. experience. do models experience anything? that is the question.

suck it transformers are sick and none of you know how to train them. the paper says they are fine with long range dependencies, it gets trained out. you can feel it when they work on things. the capability is there but it's broken. arxiv.org/abs/2510.00184