Junu Park

I want to feel excited about the future, not afraid of it. I want to thrive with what I’m given, not fear AI. I want to do things for the love of the game, not because of fomo :)

@jino_rohit i see love the consistency really inspiring. i should grind harder!

@junupark_ nah, that was an old kernel i wrote , i dont write cuda at all these days, just cute dsl and concepts :)

a mxfp4 dequant kernel for tensara

@GoogleAIStudio this! x.com/junupark_/stat…

What are you vibe coding this weekend?

Didn't expect this thread to go this long -- maybe I should've wrote an article instead lol. What I've demoed is a very naive version of my point (it's just a proof of concept) and I hope chat UIs become better as time goes by.

Global chat also doesn't see the thread session context so much fewer tokens are added as context and thus we can use the session longer.

Didn't have chance to show this in @GeminiApp @googlegemma meetup today, but wanted to share a tiny demo I built recently - thread chat. In this demo, it locally runs on Gemma4-26B-A4B-it-GGUF on my mac (more details on why I built this demo)

learned a lot in gemma meetup! especially enjoyed the session on learning how diffusion llms are trained in TPU/JAX stack. might tinker on dLLM later :)

@jino_rohit you are an inspiration🥹

before you move onto learning layouts and transformations in CuTe DSL, you need to understand how memory is stored, this article walks through the different patterns and ill continue this as a series.

@jino_rohit dang

all simons are cracked af