Where wants to knock a trillion dollars of value off of NVIDIA with tinygrad.
Preamble
IMHO: AMD's , the compiler infrastructure a huge amount of modern software quietly runs on. So when he talks, I try to keep up. His point wasn't only about software. It was about how the hardware we run works with the applications we write, and why you can't reason about one without the other.
He said he'd been told his approach wouldn't work, because play at 's whole thing, so that one I could at least place. Matching the best models to the right hardware: CPU, NPU, GPU.
I'm not going to pretend I can explain all of that yet. I can't. But it's in the notebook for when I get to the "kernel" level of my studies. Lol.
Hassan Akbari tied it together
, that got a cease-and-desist letter from the DMV. Got hired to hack a company's own systems to keep them safer. I could not tell if I was watching an inspirational documentary or a cautionary afterschool special. But I already said he was awesome, and I stand by it.
His talk was deep. Graphs, lines of code, evaluations. I have pictures, if anyone wants them. At one point he said the , which I looked up, is an open-source neural network framework and deep learning library, and on stage he put it at 25,000 lines of pure Python, about 9,000 of that the core, meaning the essential engine with everything else built around it.
. Insert audience laughter here. Same moat Chris Lattner was talking about. George just comes at it from the scrappy open-source end.
His pitch is GPUs for the middle class. He talked about growing up in New Jersey with a parent who was a teacher. He doesn't care about data centers. What he cares about is the highest development velocity, and not the out-of-the-box-fast kind. His argument, the way I understood it: an out-of-the-box and
What's next
Full disclosure: I did day one as a day trip, LAX to SFO and back. Day two happens without me, and I'm bummed to miss it. I'll have to watch
SOCIAL SHARE CARD GENERATOR