rev. howard arson

this is one of the most insane fucking things i have ever read www.amnesty.org/en/documents...

Unlawful by design: Exposing the human rights costs of generative AI - Amnesty International amnesty.org

i think pre-planning rhymes is the best example of this behavior, where the model preactivates features associated with words which complete the rhyme and "stores" them as annotations on the newline token. transformer-circuits.pub/2025/attribu...

On the Biology of a Large Language Model transformer-circuits.pub

i think there's an actual question embedded in the "if Claude were representing Genghis Khan and Julius Caesar, would we say that it was really the thing itself?" hypothetical, and it is "should an actor playing Brutus be permitted to stab Caesar to death on stage."

No, Artificial Intelligence Is Not Conscious theatlantic.com

www.wired.com/story/ai-pr-...

Ed Zitron Gets Paid to Love AI. He Also Gets Paid to Hate AI wired.com

the entire research group corresponding to my operations group was redeployed to support a project which is a priori impossible and which had no real mathematical affinity to their research area techcrunch.com/2026/06/24/a...

AI researchers continue to leave Google for its rivals | TechCrunch techcrunch.com

so something like "bounded surprisal under the Fisher metric between the up and down projections of a MLP" seems like it might be a quick way to norm the worst kind of attacks back into doing something reasonable. arxiv.org/abs/2203.00922

Adversarial attacks on neural networks through canonical Riemannian foliations arxiv.org

arxiv.org/abs/2507.10524 arxiv.org/abs/2212.05055 arxiv.org/abs/2404.02258

Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation arxiv.org

arxiv.org/abs/2507.10524 arxiv.org/abs/2212.05055 arxiv.org/abs/2404.02258

Sparse Upcycling: Training Mixture-of-Experts from Dense Checkpoints arxiv.org

arxiv.org/abs/2507.10524 arxiv.org/abs/2212.05055 arxiv.org/abs/2404.02258

Mixture-of-Depths: Dynamically allocating compute in transformer-based language models arxiv.org

anyway there's a reason this is happening finance.yahoo.com/technology/a...

Google DeepMind is losing its grip on elite AI talent, new data shows finance.yahoo.com

the child porn AI monomaniacally dedicated to the world's first trillionaire, who once controlled the US government and has killed a million people

papers: so the fundamental unit of computation in a neural network is the 'layer' me: oh great. could i get the definition of 'layer' papers: oh heavens no

a solution to AI water usage: Brawndo has what H100s crave

it is incredibly weird that the Hypnerotomachia Poliphili exists, a book which i only know about because a LLM compared Piranesi to it

the estimated loss of the deploy weight is pretty normal but i have NEVER seen a square wave show up anywhere in machine learning. wild.

i think that it's kind of weird to see writers and artists crowing about the fact that AI is not useful in their field, something which both i and they were hoping was not true

extremely pleased that i am managing to do a full finetune of a 25 gigabyte model in 32 gigabytes of memory, activations and everything.

it is both extremely funny that i managed to cram a full finetune of a 25gb model into 32gb and that i only have 400 megabytes left over before it explodes. hope that big memory tile doesn't fragment lol

doing an early test rebase of my image model on Krea2, about 5% in, and what's coming out is... good, despite only using 512x512 images so far; i can go up to 1024x1024, but didn't want to do the slow version. i'm also going to finetune the 32b Qwen TE on my setting docs.

i am reading this thing which is criticizing someone on the basis that, e.g., an LLM is very unlike an autoencoder, but in the zero-layer limit a tied encoder just produces the same token and there is, in fact, a d_model bottleneck between the halves of the d_vocab encoder.