Arun Kumar

I write primary-source explainers on transformers and LLM inference — reading the papers and checking what they actually measured. More at diffstudy.com

Joined on Sep 6, 2026

Arun Kumar

Sep 18

GRPO doesn't remove the reward model. It removes the critic.

#ai #machinelearning #llm #computerscience

3 min read

Want to connect with Arun Kumar?

Create an account to connect with Arun Kumar. You can also sign in below to proceed if you already have an account.

Create Account

Already have an account? Sign in

Arun Kumar

Sep 14

Muon doesn't replace AdamW. Every Muon run still has AdamW in it.

#ai #machinelearning #deeplearning #computerscience

3 min read

Arun Kumar

Sep 10

"RoPE extrapolates to longer contexts" — someone finally measured it, and it doesn't

#ai #machinelearning #llm #computerscience

3 min read

Arun Kumar

Sep 6

Speculative decoding won't change your model's distribution. It might still change your output.

#computerscience #ai #machinelearning #llm

4 min read

DEV Community

Arun Kumar

Badges

Writing Debut

GRPO doesn't remove the reward model. It removes the critic.

Want to connect with Arun Kumar?

Muon doesn't replace AdamW. Every Muon run still has AdamW in it.

"RoPE extrapolates to longer contexts" — someone finally measured it, and it doesn't

Speculative decoding won't change your model's distribution. It might still change your output.