few problems in this world can't be solved by throwing GB200s at them.

If anyone wants to know what it’s like in the luminal office, this is a perfect example.

bitter lesson matmul search in Luminal!

RT @matthewjgunton: hugely excited to be part of s25 😎

joined yc s25. rewriting the whole ai stack.

kernel search in luminal

This is the beginning of the Models of Life timeline predicted by @owl_posting.

not done until our kernels are pretty. obviously still needs some work.

i can't square this with the other hazy post about how tensor cores make up 95% of the flops on a GPU. wouldn't it be massively beneficial?. or is it because matvecs are so bandwidth constrained compared to matmuls?. if someone from hazy follows me or knows someone from hazy lmk!

RT @matthewjgunton: Link here:

RT @joefioti: Since we’ve got a lot of new people following Luminal's progress, I figure we should go over where we are and where we’re goi….

If you're still reading this far, star the repo!.

Post on discord:

We’re on track to have a prototype of this cloud in the next few months. If you’re interested in using this, helping contribute, or know someone who would be, please reach out.

The most valuable thing we can bring to market is a Vercel-style deployment cloud that allows devs to go from fast local to fast production workloads easily. This lets us keep the focus of the repo the same, while monetizing in a purely complementary way.

I kept this as a hobby project of mine because of this question. I disagree with a lot of open-source monetization schemes that involve consulting or changing the repo license down the road.

One concern I always have with seeing open-source projects take on ambitious goals is the question of how they will make money?. To be honest, this stumped me for a long time.