@kalomaze Satya to Sam: "No more AGI hype, focus on products"

@kalomaze @kalocide they can just rent those gpu here, and do betterr

@qtnx_ will check, thnx.

@qtnx_ you can try those 1/N threads may be. but can understand the struggle with perfection on writing part.

@qtnx_ would you consider writing any blogs on what you built? I think you do ship cool stuff (and fast too)

@_trish_07 starting to like zig for simple bump up over c. tried rust but nope, too many things going on.

@Teknium1 intelligence explosion doesn't mean inf time abundance as well imo. you have got limited time, and what you prioritize to build and sell on that, will still matter.

reading this today

@nearcyan captcha but to weedout scammers from DM.

RT @abacaj: Why does qwen2.5 0.5b base work and llama 3.2 1b base not work nearly as well? Qwen models are effectively pretrained with SFT…

RT @kellerjordan0: New NanoGPT-Medium speedrun record: 2.92 FineWeb val loss in 28.1 minutes on 8xH100 Previous record: 29.3 minutes Chang…

@tokenbender limbo is such an exp. only thing close, i have is unravel and then may be it takes two.

RT @unixpickle: Trying Muon for a hobby project. Blue is Adam, green is Muon. Code is pretty simple too. https://t.…

@justinskycak so that's where ive been wrong all this time.

RT @helloiamleonie: 4 tutorials that helped me learn LLM fine-tuning in 3 weeks: 1. “How to Fine-Tune LLMs in 2024 with Hugging Face” by @…

RT @novasarc01: use leetgpu to practice writing cuda kernels...great ui, no gpu required, completely free!

RT @kalomaze: on today's episode of "you can just do things" did you know you can split the MLP projections of a dense Transformer by a gi…

@remstack @theirtestuser is he talking about kache lmao?

RT @kalomaze: this is an old paper but i believe unfamiliar people should read it

RT @kalomaze: the elites don't want you to know this but you can train your base LLM into a classifier without initializing new lm_head pa…