georgehotz

8 ms·

91.
by georgehotz 4y ago
Yea, benchmark based compilation, that's already happening in the tinygrad compiler we use for openpilot to determine the local group size. https://github.com/geohot/tinygrad/blob/caea34c52996cde2ed46...
92.
by georgehotz 4y ago
Fused into one operation since the Tensor isn't resolved until I call .numpy() kafka@tubby:/tmp$ cat fuse.py from tinygrad.tensor import Tensor x = Tensor.zeros(1) for i in range(5): x += i print(x.numpy()) kaf
93.
by georgehotz 4y ago
We could have NEG instead of SUB, but with the constant folding it's a wash. DIV is already an HLOP with reciprocal (used to use POW, but that was slower. And what would you implement RELU in terms of?
94.
by georgehotz 4y ago
> they're dynamically shaped and hit near peak While this is true for most common GEMM looking ops, if you tread off the beaten path things get slow (odd channel sizes, batch sizes, etc...). Right now in PyTorch, GroupNorm is 2x slo
95.
by georgehotz 4y ago
Min is an HLOP. From: https://github.com/geohot/tinygrad/blob/master/tinygrad/tens... def min(self, axis=None, keepdim=False): return -((-self).max(axis=axis, keepdim=keepdim)) All folded together,
96.
by georgehotz 4y ago
That CONV is only used on the older backends. The GPU and LLVM backend rewrite CONV as MUL+SUM, to be fused later, and thus only use the 4 OpTypes. https://github.com/geohot/tinygrad/blob/master/tinygrad&
97.
by georgehotz 4y ago
Still working on this! Unlike self driving, which is conceptually solved, things like cooking and cleaning aren’t. There’s a very cool challenge remaining here.
98.
by georgehotz 4y ago
There’s only one outcome for hedonists. Wireheading. I suspect many people will go this way, I won’t.
99.
by georgehotz 4y ago
We are happy to wire $1 million tomorrow. I think the scale that might make them care is 10-100x that, if it even is a scale thing.
100.
by georgehotz 4y ago
The patent in the complaint doesn't even have anything to do with what we do as a company! It describes a system that calls 911 when it detects a crash (like onstar). Ignore the (very generic) title and read on in the actual patent. h
101.
by georgehotz 4y ago
Yes I believe you want the "Ford Co-Pilot360 Assist 2.0", not to be confused with the "Ford Co-Pilot 360™" or "Ford Co-Pilot 360™ 2.0" that won't work. The "Ford CoPilot 360 Assist+" is probably
102.
by georgehotz 4y ago
Supporting the body allows us to find bugs in openpilot that wouldn't be easy to find in cars. For example, cars barely pitch and roll, so if you have something subtly off there, it's a lot easier to spot on the body. The body is
103.
by georgehotz 4y ago
Do you want to throw your life away selling ads? Or do you want to build open source software and solve self driving cars?
104.
by georgehotz 4y ago
Ford has two issues: 1. Older cars have a lockout on steering after 10 seconds baked into the EPS firmware. 2. While newer cars don't have this lockout, adaptive cruise control (ACC) is only on high trims. Therefore low trims lack rada
105.
by georgehotz 4y ago
The safety driver isn't in the car physically, they "work remotely" https://www.wired.com/story/cruises-robot-car-outages/
106.
by georgehotz 4y ago
There's already 20 million supported cars on the road, that number is increasing year over year, and the derivative of that number is increasing year over year. Why would we worry about this?
107.
by georgehotz 4y ago
We are about 30, and about to move into a new 20k sq ft office with production and a datacenter on site. Yes!: https://comma.ai/jobs And one more: https://twitter.com/comma_ai/status/15480302953407
108.
by georgehotz 4y ago
No need to partner, comma doesn't offer that. The code is open source, they can just run it. They still can (and probably should from what I've heard about DreamDrive). The port to their hardware shouldn't be hard.
109.
by georgehotz 5y ago
I would merge this PR.
110.
by georgehotz 5y ago
It depends what your opinions of those agencies are, and if you believe they have technical competence or not. After the 737 MAX fiasco, I've begun to seriously doubt it. "you would not understand that DMV issuing these permits is
111.
by georgehotz 5y ago
Once software capable of driving at a superhuman level (say 10x the average) exists, the regulatory side will be trivial. What regulator wouldn't want to save 30,000 lives? Sadly, that software is still a ways away. Don't fall for
112.
by georgehotz 5y ago
Tesla has tons of videos on YouTube from users showcasing similar technology, including in San Francisco. While FSD beta may still have issues, I see no reason to believe that this doesn't also have similar issues.
113.
by georgehotz 5y ago
Full disclosure, I founded a somewhat competing company in the space, comma.ai Why is this progress? Evaluating progress in self driving cars is extremely complex. Why do people trust that the California government is able to do so? A miles
114.
by georgehotz 5y ago
Did you hear about the startup founder who gave themselves a bajillion shares of their company? Now, if I dumped into entry liquidity SPAC style...that's a different story. But please be more precise in your criticism, I'm not sur
115.
by georgehotz 5y ago
This kind of crap is exactly why we stopped using recruiters. I hate e-mails like that too, and they aren't even effective.
116.
by georgehotz 6y ago
The MLCompute API supports three options "CPU", "GPU" or "any", which is GPU sometimes and CPU sometimes. Not the neural engine. There's some MLComputeANE code in the framework, but afaik there's no w