I'm driven by a simple conviction: the best way to make technology matter is to put it into products people actually use. For several years that meant autonomous driving — taking AI models from research into vehicles that ship at scale. It's a problem that isn't solved yet, but the gap between a demo and a product narrows every year, and you can see the result on roads worldwide.
Now I'm turning to embodied AI and robotics. Different form factor, same fundamental challenge: getting intelligence out of the lab and onto hardware that has to work every day, on a real power budget, in the real world. It's the hardest kind of engineering, and it's where I want to spend the next chapter.
Totally agree and we see some many good ideas/techniques are all in PyTorch/4090 workstation but as we all see in Autonomous Driving industry, we need to deploy every good things to a resource constrained computing system. And it needs to be cost effective, reliable and implement those good ideas. As long as physical AI and robotics industry go for production, we definitely need to solve these Physical AI Deployment issue.
@lexfridman What's next after transformer? I'm interested about the latest papers from Google DeepMind: https://t.co/LtoZuz4ZLz
Also any good direction or path for AGI Robotics?