what makes you say this? as far as i can tell GCP docs + Scaling Book combine to paint the picture of 2x MXUs, each being 256x256 systolics docs.cloud.google.com/tpu/do… jax-ml.github.io/scaling-boo…

pretty straightforward they used a pre-existing formalization from Google's Formal Conjectures project (basically a bank of community-reviewed formalizations of problems) they did make a few seemingly inert code hygiene modifications (inlining definitions etc), but looks legit

i don't think (3) will matter much? w/ humans out of the loop you only really care about aggregate throughput of your swarm of agents (assuming enough work is parallelizable) and batching economics mean you get superlinear gains in throughput per $ as you trade away tok/s/user

taking reservations for my LEGO 5nm wafer capacity, volume production starts Q1 2027

X makes me show government ID to report impersonation :/

orcd.mit.edu/news/orcd-newsl… orcd.mit.edu/running-jobs/av… ~hundred H100s and something like 80 H200s

alright done fyi in general the person creating your mailing list can easily add themselves to the list and get a copy of all the emails that pass through it (i won't, but stay vigilant)

everybody wants to call their rust code blazing fast, but nobody wants to look at no big-ass flamegraphs

fun little branchless @fiveoutofnine and i put into seaport that i still think about

came across this after changing a variable from being in an Arc to static tanked performance turns out it was a false sharing issue with some unrelated memory. apparently this is a really common bottleneck!