Eric Alcaide (@eric_alcaide) 's Twitter Profile
Eric Alcaide

@eric_alcaide

common prosperity

ID: 774723604412071936

linkhttp://hypnopump.github.io calendar_today10-09-2016 21:37:55

679 Tweet

997 Takipçi

694 Takip Edilen

Eric Alcaide (@eric_alcaide) 's Twitter Profile Photo

I remember discussing an early concept of this over a nice pizza and iterating concepts over a hike. Congrats on getting this out ! 🥳🚀

I remember discussing an early concept of this over a nice pizza and iterating concepts over a hike. 

Congrats on getting this out ! 🥳🚀
Eric Alcaide (@eric_alcaide) 's Twitter Profile Photo

However the proposition 2 is a bit confusing, 2x linears that are supposed to produce complex numbers awont have a sine-cosine relationship. Furthermore, the complex eigenvalues can do some modulo algebra but are not as general as householder products🤔

However the proposition 2 is a bit confusing, 2x linears that are supposed to produce complex numbers awont have a sine-cosine relationship.

Furthermore, the complex eigenvalues can do some modulo algebra but are not as general as householder products🤔
Eric Alcaide (@eric_alcaide) 's Twitter Profile Photo

FlashAttn, register maxxing, block tile scheduling, warp specialization, exp-poly approx, etc. Many such cases ! All signs of work to do in ML compilers

Joey (e/λ) (@shxf0072) 's Twitter Profile Photo

we need automated way to test operator (and its grad effect) over long time for different and mixed precision some key ops need to be in high precision (float 32) this should be solve with automated search no debuging

we need automated way to test operator (and its grad effect) over long time for different and mixed precision 

some key ops need to be in high precision  (float 32)
this should be solve with automated search
no debuging
Eric Alcaide (@eric_alcaide) 's Twitter Profile Photo

>In the Foundation series, there’s a phase where societies lose nuclear tech and return to coal >It’s a marker of collapse

Eric Alcaide (@eric_alcaide) 's Twitter Profile Photo

In case you're wondering, CANs = Chebyshev-accelerated Newton-Schulz (CANS) Using Chebyshev's alternance theorem to find better coefficients. Concurrent work to PolarExpress but I would bet slightly worse, although there's no direct comparison. arxiv.org/abs/2506.10935