difficultyang

Credit where credit is due, GPT 4.5 came up with this joke

it's midnight and I have just realized I actually know nothing about MoE

Apparently I'm a DTensor noob but so is everyone else

A few weeks ago our 4yo told us to pack her lunch on Fridays because she didn't want pizza. This was puzzling, but we didn't inquire too much. Today she told us that her mouth sore is better and she wants to eat pizza again. BALANCE HAS RETURNED.

Not very impressed with Codex's code explaining chops

You commune with ancient spirits. I read git commit logs. We are the same.

So like, agentic context building is sort of like literate programming on the fly

it's a fucking list monad. they put a list monad in my sharding propagation rules

You've heard of Moe Shop but what about MoE shop

Expert parallelism actually just tensor parallelism on the batch dimension

I hate that git rebase -i doesn't support recursive interactive rebases

so I'm like working on a deep dive of DTensor implementation and for some reason it's looking awfully like an agentic coder trajectory 😂 docs.google.com/document/d/1fC…

So tell me why is it I am tired all evening and so I decide to go to bed and OOPS I am perky as porky pig

so I was using claude code and I rejected a change it made, did it myself, and then pushed it to a PR. and then CLA bot said claude hadn't signed the cla. so I open .git/config and what should I find but a big steaming turd

"Mom, can we have token choice?" "We already have token choice at home" The token choice at home:

Asked Claude to refactor some printing code that I vomited out last week. It did a good job. Good job Claude.