The ventures he has started (I can think of tinygrad and comma ai) all seem like half finished tech demos.
The ventures he has started (I can think of tinygrad and comma ai) all seem like half finished tech demos.
His AMD rants were a valuable warning about the quality of their hardware. I wish he'd done that maybe 10 years ago when I was buying AMD cards thinking that they might work with pytorch in a year or so. I knew they had problems but if I'd realised how bad the situation was I'd have held my nose and gone with Nvidia.
The rants weren’t breaking news to anyone who was familiar with PyTorch or adjacent communities. He seized upon a weak moment for AMD to try to launch his own company. Unfortunately he launched his effort with an attack on the company he was effectively trying to partner with, making the entire venture DOA.
It’s too bad, too, because it would have been interesting to see if anything could have been accomplished with a more friendly offer of cooperation. He’s obviously talented as a developer, but effectively going on the attack for the company that forms the foundation of the business you’re trying to build is obviously not going to end well.
How many people in the PyTorch and adjacent communities are trying to port the stack to AMD cards? I'd guess less than 100. That is a pretty small community. They don't do that much in the way of publicity (and, case in point, George does and he attracts a certain number of haters).
George is probably the first public reference I saw who wasn't pointing at CUDA as the problem but tagging kernel & hardware bugs on AMD cards as the blocker. Those are really different things; CUDA isn't that complicated to implement as a half-baked thing. I can implement naive matrix multiplication if I need to and it'd be a fun month of reading academic papers to learn better techniques. But that does nothing to work around hardware and firmware bugs which are an order of magnitude harder to deal with and not something I want to deal with.
I'm happy to buy in to bad software libraries. I don't want a bar of hardware bugs.
He definitely writes in a below-HS level.
Edit: you edited your comment after I told you he made comma.ai.
Bit dishonest, but whatever, I wouldn't describe comma.ai as a "half finished tech demo" but you're allowed to your own opinion about it.