Joe Fioti
@joefioti
Followers
2K
Following
653
Media
122
Statuses
1K
it's not possible, it's necessary. building a compiler @luminal_ai (yc s25) to solve inference.
San Francisco, CA
Joined June 2015
i’m hiring cuda people in SF. we’re solving compilers and deleting 90% of the ML stack. hmu if you’re good
40
44
1K
You show me a compiler that generates a single fused megakernel for this entire model, I quit my job right now and I work for you.
7
8
125
as many others have said this is the single best place on earth to build the future. if you’re applying shoot me your application, happy to give feedback.
Applications for the YC Winter 2026 batch are now open! Apply by November 10: https://t.co/gNl84El3BS
0
0
1
“pcie for scaleup” idk about you but i prefer carrier pigeon from one card to the next.
2
0
8
every single time I’ve eased up on the activation reference checking tolerance I’ve regretted it later on. always stay on 1e-5.
1
0
9
“the hardware lottery” today is sort of a myth. currently it’s “the hardware + optimized software lottery” which is a much smaller target to hit. we’re working to bring it back to the pure “hardware lottery”
0
0
1
i do wonder how many of these exist. our bet is the compiler unlocks this + a lot more since it makes many more architectures viable.
hybrid mamba architectures (to me, subjectively) are a weird case of "yes, this is objectively better than pure attention, but the additional implementation complexity and absolutely terrible ecosystem support makes it not worth the squeeze"
1
0
5
benchmarking inference servers is, as the kids say, *non-trivial*.
3
1
34
No worse feeling than realizing all these clever ideas I had last week more clever people had 6 months ago…
6
0
19
when i look up my current dumb coding problem and it has a cs theory wikipedia page with "no known P solution exists"
3
0
29