I think he's constantly trying to time the peak (of AI improvement, of compute capabilities, of profitability) and keeps having to readjust his predictions every time it's proven wrong. Timing the peak is a fool's errand unless you have extremely hard and unavoidable data that everyone will have to face by a deadline. But contrary to what AI boosters often imply in response to skeptics being wrong, failing to time the peak doesn't mean the peak doesn't exist. So far it has existed for every new technology.