Inspirational talk by Andrej to a young audience on the importance of building AI agents. An interesting line snuck in between- “and I got distracted a bit with self-driving…” he later adds self-driving belongs to a class of problems that appears easy but is very hard to actually make into a product - perhaps will take a decade. That observation is the most honest and damning statement on self-driving @Tesla to date. As a Tesla user for 7 years now, while I am uncomfortable driving any car other than a Tesla given how much it reduces the cognitive overload of driving, the mass in the long tail of boundary conditions seems too hard to solve by just more sampling.

That is an existential threat to any company involved in NLP tasks. Perhaps it will not be too long when generative models for images solve discriminative tasks too - text/image in - text out - a chimera of image captioning and text conditioned generative models