bilal

it's Space Quebec, because they speak a language with the same phonology as French which is incomprehensible to Francophones

Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction We present Visual AutoRegressive modeling (VAR), a new generation paradigm that redefines the autoregressive learning on images as coarse-to-fine "next-scale prediction" or "next-resolution prediction...

there’s been some interesting work lately on multiscale autoregressive image modeling arxiv.org/abs/2404.029...

city of stairs and the tainted cup both very good 👍

once again coming crawling back to AdamW after every paper published after 2015 has failed me again

i think they’re just starting to realize what they unleashed w the tweet

from the other app xd

Love to log on to the Horrors app to catch up on today's Horrors

we’re doing ai exorcisms in 2025 huh

like one of the big things about 404 media, Brian Merchant, Paris Marx, Ed Zitron, etc., is that they neither know nor care how the subject of their criticism actually works

I wish academic ML was a bit more skeptical of papers and less skeptical of industry. I get that it sucks to not have visibility on details, but it doesn’t invalidate the results. On the flip side, there are too many papers whose message are parroted despite sketchy experiments.

we’re going shopping

an LLM that uses streetview to pre-drive down the route and assemble comments like "at the big red barn, turn left" "when you get to the sorta squiggly road, take the exit" like a farmer would

if you squint hard enough everything in ml is either a special case of the KL div or newton’s method

a lot of machine learning research is about discovering which parts of mathematics are actually L2 regularization and which parts of mathematics are actually Adam

justine tunney the libc mutex micro optimizations person??

This guy needs to read Manufacturing Consent! You’re not supposed to do this yourself you gotta hire editors who already agree with you, this is amateur hour shit…

pov: post training researchers learning what pretraining researchers do while waiting for the model to train

accidentally typed rm -fr and i’m using that now

congrats!!

thanks for cleaning it up

ai generated slavoj zizek voice on slop video of some bizarre rural chinese cooking

incredible new forms of postings emerging

interesting is there anywhere i can read more about this

its all approximating numbers w other numbers all the way down. everything else is an implementation detail! 😛

if your values do matter replace them w values similar to them aka ones (parameter sharing / shared kv cache / factorizing a large matrix into two small ones / lora / adafactor)

i love how every efficiency advance in machine learning is approximating [expensive operation] by either ones (just pass it straight through) or zeros (doesn’t matter, just don’t compute it/sparsity)