Incredibly somatically sick (there might be a multiple day migraine going on), but also excited to be in the bay this weekend + meet people in the area up until Monday

this isn't me trying to misuse the claude-code api either... claude-code really wants to use dozens to thousands of 'subagents' at once, and simply needs to be wired up to apis which facilitate this interaction pattern

mother of god...

NO CLAUDE CODE DON'T DO IT THE QUOTA CAN'T TAKE MUCH MORE OF THIS

it's incredible how this one is going

yes, i will propagate a subagent calling api across two repositories by telling claude code 'hey so we already wrote this. why not review how it worked in the other operating context and figure out what modifications you need to make to the subagent <-> external api caller?'

https://t.co/erYjJ19Ax7 crosslink

so i'm dual wielding two claudes right now

they can't keep me away forever

types are cool but you should learn how to metaprogram the good old fashioned way, by blowing up your computer, before talking about safety a lot, and still blowing up your computer,

this text is not nonsense. but it is somewhat trite and draws too much from the strict narrative events and beats of the reference media. it is also several grades of literary capability above tinystories or basically what gpt-4 could do a few years ago.

aiming for 100% claude use (even if it means i have to run two claude code projects at the same time and real time strategy gamer between them)

posthumanity's first exam

for my haikus, parsers rights, for my sonnets, trans rights

this is rather big: you needed to hire like 30 devops people to get one nine of reliability, then fire 20 of the devops people and hire 2 more to get to three nines of reliability in the old days, and this was super slow and tedious and expensive in staffing afaict.

it looks like claude opus can handle the responsibility of writing async code which dispatches extremely heavy work to extremely rate limited apis.

what if you put an extra credit free response field into the rl environment

it's really wild that the first pass version of this (that was the claude code session *testing* the interfaces they wrote, btw!) worked at all. there's kinds of in-context-learning going on here that quite surprise me as an observer

they actually wrote up their process notes with the title 'First Wet Claude Dry Run'

"i think this is actually one of the strongest arguments for doing 'wet claude dry runs', wherein the actual dataset work task launch interface and actual first-order-subagent orchestration interface are observed directly by a claude code session."

"claude code sessions are therefore a pretty reasonable proxy and first pass debugging tool for how to deploy and apply NLP tools in a broader sense (because we can ask a claude code to *do* some NLP task, or to query a 'subagent'..."