normal texts from normal friends

honey have you seen my box of amd gpus??? — oh never mind, found it

I am asked to reach 4096 github stars for our CUTLASS project before GTC'24. It is 3853 now. Please help me. github.com/NVIDIA/cutlass…

guy who only uses LLMs to answer questions found in the MMLU test set: 'dunno man, model works great for me'

started the night with 'i need to shave 85ms off of this program runtime'... four hours later i am 17ms down, 70ms to go. if i stay up all night, surely, i am going to write. the fastest program, ever

this week in sf tech events: i'm going down to ocean beach to scream into the pacific, does anyone wanna come with

2024 get drunk fast: take a shot every time someone mentions synthetic training data

i feel like we didn't talk enough about how inductor + triton generates faster kernels than flashattention2??? that deserved more....... attention 🥁

the pytorch compiler stack being written in python is a huge win for ml research types too chickenshit to debug xla/c++

the girl who cried mismatched grad tensor types in the backward pass

the pytorch decision to commit to propagating the made-up verb 'fakeify' throughout the codebase is a linguistic travesty but a fun one. e.g. 'tensors have been fakeified', 'automatic fakeification'

there's actually two types of research engineers: research research engineers and research engineer engineers. they're different paths and every research engineer has to pick one or the other eventually

this entire time my friend thought servicenow was service ONTARIO. he thought service ontario was building code language models for some unfathomable reason

this week i've learned that i was 'born in the 1900s' and also that torch.distributed is 'legacy' so kindly bury me under the sand at ocean beach

if you're the personality hire and i'm the personality hire then who tf is compiling this graph???