Coetzee "wrote programs for a computer that used an algorithm to select words from a set vocabulary and create repetitive lines. Coetzee never published these results, but edited and included phrases from them in poetry that he did publish."

i'm not staring into the void i'm studying it

so far opus 4.5 in claude code has launched 2,281 haiku 4.5 task agents (in batches of 50), each with 12k initial context and 2-4 tool calls, without interruption it's going for 14,000

update: 3,193 haiku subagents launched by opus in claude code without interruption / intervention ✵ Galloping

x.com/i/article/1993…

github.com/UW-Madison-Lee… Please find our preprint & code here. Any feedback would be greatly appreciated!

update: 5,748 total haiku subagents launched by opus in claude code without interruption / intervention ✷ Galloping ...

literally

gonna read this in my prime intellect hat

There is a uniquely Promethean vibe in Wenfeng's project. Before DS-MoE, only frontier could do efficiency. Before DS-Math/Prover, only frontier could do Real math. Before DS-Prover V2, only frontier could do Putnam level. Before DS-Math V2, only frontier could do IMO Gold…

update: 13,722/14k total haiku subagents launched, nearing the end. there was one human intervention needed when I hit a claude code 5-hr limit around 7k, and I wrote 'continue'. opus has otherwise survived many compactions and 24+ hours of delegating work and verifying. note:

oh dang

it’s actually a miracle how well it works out when you follow the fun

3.3M parameters thinker It increased my internal valuation from $1 to at least $2

I strongly condemn dunking on Prime Intellect, they're doing the exact right thing. Post-training Chinese base models to the frontier level is in fact *more important* right now than learning to pretrain our own bases. I basically don't care what PI, Arcee and others can

I usually stay out of takes like this on here. I am not going to change the original poster’s mind, but I might give a bit more context for everyone else reading. And I saw firsthand how hard this was to pull off and how much work went into Intellect-3. I know RL at scale is

do what now