Chain of Thought (what he's describing on his first point) is NOT the same thing as infinite context. It's not even a case of "I'm simplifying this for non-technical people", he's just straight up conflating two completely different things.
His explanation of agents is also wrong in that they don't "learn" anything. They have their internal model weights from training and they have the stuff in their context window and that's it. Agents don't have any magical "learning" ability. Finish one task with an agent and it will immediately forget what it just did.
These are the people making the big decisions and guiding the industry and they a) don't even fucking know how the technology works, b) are blatantly lying about it, or c) both.