> As good as GPT-3 is today it will be twice as good within a year, and thousands of times as good within a decade. When they figure out how to marry it with even basic reasoning and memory, it's hard to imagine how that will enhance its powers.<p>This is Kurzweil's fundamental mistake, which undermines basically everything he is saying: Extrapolating exponential growth trends into the future.<p>There is absolutely no reason to believe exponential growth trends continue indefinitely. In fact, every single such trend in the past has come to a stop. Exponential growths are always exceptions, not rules.<p>GPT-3 will not be "thousands of times as good" within a decade. It is a very limited model, and it happens to be doing well at the moment, but it doesn't just keep improving. This approach will hit its fundamental limits at some point, and then the growth just stops.