Hasan (@hasanunlu9) 's Twitter Profile
Hasan

@hasanunlu9

Tesla AI

ID: 621120184

linkhttp://hasanunlu.com calendar_today28-06-2012 16:21:31

41 Tweet

209 Followers

592 Following

Neuralink (@neuralink) 's Twitter Profile Photo

We're excited for our update event livestream tomorrow at 3pm PT. We'll tweet out a link to the livestream in the early afternoon.

Hasan (@hasanunlu9) 's Twitter Profile Photo

Here is the starter code for Neuralink compression challenge. It includes Huffman encoding and decoding in cpp, along with some heuristics to optimize data encoding. Currently, it achieves a compression ratio of 2.59x and successfully passes their script. It is far from their

Tony Duan (@tonyduan_) 's Twitter Profile Photo

Lots of hard work and many late nights went into the making of FSD 12.5, from across the team. Many ideas were simplified and re-worked from first principles. Hope everyone has a chance to try it out. It's a release we're proud of.

Lots of hard work and many late nights went into the making of FSD 12.5, from across the team. 

Many ideas were simplified and re-worked from first principles.

Hope everyone has a chance to try it out. It's a release we're proud of.
Hasan (@hasanunlu9) 's Twitter Profile Photo

I installed a mini jet engine on a ball-bearing cart I had built many times when I was a kid. Now, it no longer requires anyone to push it :)

Hasan (@hasanunlu9) 's Twitter Profile Photo

It seems that torch.kron is just a subset of convtranspose2d with the kernel size equal to the stride. I did a simple benchmark for the following tensor: torch.kron took 41 us, while convtranspose2d took 26 us. torch.kron could be replaced with convtranspose2d. PyTorch

It seems that torch.kron is just a subset of convtranspose2d with the kernel size equal to the stride. I did a simple benchmark for the following tensor: torch.kron took 41 us, while convtranspose2d took 26 us. torch.kron could be replaced with convtranspose2d.  <a href="/PyTorch/">PyTorch</a>
Hasan (@hasanunlu9) 's Twitter Profile Photo

I think there is a minor error in flash-attention-3 paper. CUDA code no issues but algorithm-1 description in O_i scaling has wrong adjustment. There should be no reciprocal.

I think there is a minor error in flash-attention-3 paper. CUDA code no issues but algorithm-1 description in O_i scaling has wrong adjustment. There should be no reciprocal.
Yun-Ta Tsai (@yuntatsai1) 's Twitter Profile Photo

Insane product push on the week of my birthday. We refactored the entire system to drastically simplify the pipeline — direct photons to control — yet providing a lot more functionality under the same unify framework. It is probably one of the biggest rewrite in years when we