I avoid new technology, exactly because I'm an engineer.
I wonder if that is just me, but when new technology is introduced and hyped, I usually take a quick look at implementations, research, and talks just to get an idea of what the state of the art really is like.
As a result, I have become the late adopter among my group of friends because the first iteration of any new technology usually just isn't worth the issues.
You know this effect when you open a new book and immediately spot a typo? I felt that way looking at state of the art AI vision papers.
The first paper would crash, despite me using the same GPU as the authors. Turns out they got incredibly lucky not to trigger a driver bug causing random calculation errors.
The second paper converted float to bool and then tried to use the gradient for training. That's just plain mathematically wrong, a step function doesn't have a non-zero gradient.
The third paper only used a 3x3 pixel neighborhood for learning long-distance moves. Doesn't work, I cannot learn about New York by waking around in my bathroom.
That gave me the gut feeling that most people doing the research were lacking the necessary mathematical background. AI is stochastical gradient descent optimization, after all.
Thanks to TensorFlow, it is nowadays easy to try out other people's AI. So I took some photos of my road and put them through the state of the art computer visible AIs trained with KITTY, a self-driving car dataset of German roads. All of them couldn't even track the wall of the house correctly.
So now I'm afraid to use anything self-driving ^_^
I wonder if that is just me, but when new technology is introduced and hyped, I usually take a quick look at implementations, research, and talks just to get an idea of what the state of the art really is like.
As a result, I have become the late adopter among my group of friends because the first iteration of any new technology usually just isn't worth the issues.
You know this effect when you open a new book and immediately spot a typo? I felt that way looking at state of the art AI vision papers.
The first paper would crash, despite me using the same GPU as the authors. Turns out they got incredibly lucky not to trigger a driver bug causing random calculation errors.
The second paper converted float to bool and then tried to use the gradient for training. That's just plain mathematically wrong, a step function doesn't have a non-zero gradient.
The third paper only used a 3x3 pixel neighborhood for learning long-distance moves. Doesn't work, I cannot learn about New York by waking around in my bathroom.
That gave me the gut feeling that most people doing the research were lacking the necessary mathematical background. AI is stochastical gradient descent optimization, after all.
Thanks to TensorFlow, it is nowadays easy to try out other people's AI. So I took some photos of my road and put them through the state of the art computer visible AIs trained with KITTY, a self-driving car dataset of German roads. All of them couldn't even track the wall of the house correctly.
So now I'm afraid to use anything self-driving ^_^