13 years ago, my team at NVIDIA was helping Audi develop vision systems for self parking at that time. The video itself still looked impressive, even today, especially considering the compute was tiny at that time. Lots of CUDA graph optimization went into it to make it run — like separable convolutions that serve as bedrocks of inference engine today. It sounds trivial today but backend we had to rearchitect how GPU reads camera frames isochronously without causing back pressure and jitters. Yet, after the demo we stopped working on it. It didn’t go anything except making video and playing it at all-hands. Back then, I had a code repo for “cool-demo-but-never-shipped”. It’s a graveyard. That prompted me to decide joining Google to at least ship a thing or two since I felt bored at writing papers. Shipping is hard, but at least my family can use it.