Name retweeted
After Belgium scored their 4th goal vs USA, the whole team mocked President Donald Trump by doing his signature dance move
Pretty savage troll, especially on US soil & after Trump fought to have FIFA review the bogus red card suspension on Balogun.
Well played 🤝
Name retweeted
5 completely unrelated people on this train, of all ages, blasting music/videos/tiktoks from their phones.
We are in a nationwide acting the cunt crisis
Name retweeted
You’ve found me at a very Luke Littler vs James Wade at the UK Open Butlin’s Minehead on ITV4 time in my life
Name retweeted
Might read a book. 📕
Only joking , going to look at my phone while a shit film plays in the background that I’ve seen at least 3 times and eat sugar until the anxiety is so overwhelming I have to go to bed. 😉
Name retweeted
My partner bought a CD player that we can plug into our car so we can play our CDs and the Amazon listing features the following images. 💀💀💀
Name retweeted
One built-in UI/UX feature of LLM interfaces I'd love is proof. I almost always do this manually - for example if the LLM recommends running some commands with some switches, I manually look up and verify the API in the docs to make sure those switches are correct and that I understand what they do. i.e. I want to double check the LLM's recollection. A feature that automatically brings in original material / reputable sources and highlights relevant sections as proof alongside factual generations would be very cool.
Highly amusing update
Highly amusing update, ~18 hours later:
llm.c is now down to 26.2ms/iteration, exactly matching PyTorch (tf32 forward pass). We discovered a bug where we incorrectly called cuBLAS in fp32 mathmode 🤦♂️. And ademeure contributed a more optimized softmax kernel for very long rows (50,257 elements per row, in the last logits layer).
But the fun doesn’t stop because we still have a lot of tricks up the sleeve. Our attention kernel is naive attention, not flash attention, and materializes the (very large) preattention and postattention matrices of sizes (B, NH, T, T), also it makes unnecessary round-trips with yet-unfused GeLU non-linearities and permute/unpermute inside our attention. And we haven’t reached for more optimizations, e.g. CUDA Graphs, lossless compressible memory (?), etc.
So the updated chart looks bullish :D, and training LLMs faster than PyTorch with only ~2,000 lines of C code feels within reach. Backward pass let’s go.
Name retweeted
as a longtime fan of Conway's "Game of Life" algorithm i never once considered visualizing the time axis physically -- this is breathtaking and elegant to me
credit:
instagram.com/reel/C2hoRnFsm…
Name retweeted
Fun fact: If you have a good musical ear, you can tell the speed of a passing vehicle by listening to the pitch interval it makes as it goes by. You don't even need perfect pitch since it only depends on the ratio.
If you hear more a major third or more, they're speeding!
Name retweeted
Did you know that Apple offers a white noise generator on iPhone iOS 15 or later, to help you sleep or simply relax and focus. You can also choose Rain, Stream, Ocean and more too. Very handy indeed!
Name retweeted
Episode 1379 - Armando Iannucci — WTF with Marc Maron Podcast >> Recorded 120 seconds after Liz Truss resigned! open.spotify.com/episode/6C6…