←
guille28,065 posts
guille avatar

guille

@angeris

ceo @ ? | partner @ ??

◷ Joined April 2011▧ 1,241 Photos and videos
guille avatar
guille@angeris·13 Mar 2022

Math is just tautologies and I still haven’t made my mind up on whether that’s based or cringe

guille avatar
guille@angeris·Nov 10

also incredibly reasonable pricing ? what’s going on here

guille avatar
guille@angeris·Nov 10

it’s crazy, ordered billet yesterday at noon, arrived today in the morning??

guille avatar
guille@angeris·Nov 9

you already have the console of your teenage dreams in your pocket (your phone)

guille avatar
guille@angeris·Nov 9

btw there will probably be an insane number of sick and cool webgpu games coming out soon that you can play on your phone get some ar glasses and a very portable bluetooth gamepad

Andrija Novakovic avatar
⟳ guille retweeted
Andrija Novakovic@AndrijaNovakov6·Nov 9

Happy to announce new work with @kobigurk and @ronrothblum . We prove unique decoding radius proximity gaps for expander codes with sufficient expansion. Expanders codes are used in prover efficient SNARKs like Brakedown and Bolt.

Post attachment
guille avatar
guille@angeris·Nov 5

god i love the internet

this is madness. you have an iOS 27.2 chain and you only use it to fact check MacRumors? can you use it to check the time as well?

will brown avatar
⟳ guille retweeted
will brown@willcb·Nov 4

people are like “why do LLMs always think they’re in an eval lol” and then have bad dreams about forgetting to turn in a homework assignment 10 years ago

guille avatar
guille@angeris·Nov 4

idk man, I think we haven't yet converged on architecture if you can train this for ~$20K

We’re a 2-person lab researching more efficient architectures for generative image and video. linum.ai/field-notes/pjit-21… Last month, we published Pyramid-JiT, a pixel-space architecture that converged 11x faster than its latent-diffusion counterpart. Today, we're releasing a 2B Pyramid-JiT text-to-image model as a milestone towards more affordable video models. Trained from scratch for only ~$20K. See blog for details & weights (Apache 2.0).

0:24
guille avatar
guille@angeris·Nov 2

i guess it shouldn’t surprise me that oai/anthropic have been rl-ing their models to be able to interpret/modify weights but this is crazy work

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…

Post attachment
guille avatar
guille@angeris·Nov 1

ok I might be an idiot but why does the american AC coloring standard have black for hot ?? totally fucks with my dc brain

guille avatar
guille@angeris·Oct 29

today’s (yesterday’s really) short post on Everything Is Broken And That’s Very Annoying But We Barely Notice

Post attachment
guille avatar
guille@angeris·Oct 27

funniest possible outcome is that P vs NP is independent of ZFC because of some silly thing like “the exponent in any P algorithm for SAT must be too large to be expressible”