heya, i'm guangyu chen, aka nathan - about me writings

i am currently working on ml research at kimi moonshot. prior to this, i worked on applied interpretability research at tilde.

research interests: model architecture, efficient attention, continual learning

here are some things i believe in:

• agency is a superpower
• taste is learnable
• write things down
• everything is figureoutable with friends
• the world is hackable in wholesome ways
• you'd probably love ssh config with named hosts

here are some things i like:

• mangoes, cold showers, trees, excalidraw, google docs, elegant kernels, mixing lego pieces, stargazing, snowboarding, meditation, learning from my sensei gemini, green, intuitively understanding things, light mode, open source, cli, spicy food, blocks, connecting the dots (click around the empty places on this page!), "twitter" - the name, twitter's rec algo

things i hate:

• twitter's rec algo, autotuning, short convolution, shared experts (rms matching limits variants), pp communication, indeterministic init patterns, inductive bias, pre-norm, weight clip, powerpoint figures and all office tools

i love people and write back. send me a message! (twitter probably works the best)