@Aryvyo Uhhh does your country have rare earths?

big cooling fan doesn't want you putting your 4090s in the freezer

@BillyZelsnack @foundation_robo Pretty cringe

@Teknium1 they're already doing it, they're just not very good at it yet

@AlbyHojel @rami_mmo yoooo you guys got into YC? 🔥

@main_horse @tphuang if it's that slow i guess they aren't even making the full R1 model a priority. together is the best in the world at specdec (along with maybe fireworks), i'm surprised they didn't apply their usual approach to r1

With CoT models like R1, we allow the model to solve only a small part of the problem per inference pass, much like generative diffusion models. But the price is inference latency. I wonder what chain-of-thought distillation would look like

Are there any decent r1 vision-language models finetunes out there?

what do sparse autoencoders even do man

o3 is okay I guess (it just did my job for me)

RT @kanyewest: Please remember I’m only 35 days into the year

RT @physical_int: Many of you asked for code & weights for π₀, we are happy to announce that we are releasing π₀ and pre-trained checkpoin…

big moves in the big 2025

@_aidan_clark_ lol

@elyxlz it will provide value, it just won't be written by you or me

RT @shxf0072: mange to hack grpo with unsloth on free colab its painfully slow but works :p

@nisten did nobody else implement an MLA kernel yet? 💀

1000 bouncing yellow sphere AGI benchmark

Inference is where it's at 😎