To be more precise: within five years of any AI having broadly human-level intelligence, AI will be much more cost-effective than a human for virtually all tasks
That is to say, not long after AI can do roughly what a human can do at some cost, it will be able to do at least what a human can do, much cheaper than a human.
This could be supported by
Normal rates of AI progress should bridge a gap from ‘human-level at some cost’ to ‘much more cost effective than a human for virtually all tasks’ in less than five years
Human level AI will bring about - or at least coincide with - fast progress in AI. This progress will probably be fast enough to make human-level AI much more cost effective than a human within five years of its development.
[put in stuff from #place1]
There are various reasons this could happen. AI itself could be helpful with AI progress, though this seems somewhat unlikely if the AI in question does the tasks that a human can do at a much higher cost. But ‘human-level’ isn’t necessarily so simple. An AI which is not cost-competitive with a human on most tasks could be very good at some specific, well defined task. Such as developing AI, perhaps.
Another reason progress might be expected to be fast just after human-level AI is developed is that at that point, AI would be on the brink of becoming much more economically important. This might be expected to spur interest and investment.
Further, we might just expect human-level AI to be achieved at a time when AI is experiencing fast growth, as disproportionately many achievements will be reached in such times.
That is, the fast AI progress needed to make AI much cheaper is very unlikely to exist when human-level AI is developed, but then quickly disappear. If things are moving fast just after human-level AI is developed, they will continue to do so for some time.
[xxx]
Suppose that the amount of thinking that AI researchers do, and the number of clever insights they come up with as a result, remains fixed. It might still be that creatures who are only a tiny bit better than us - in terms of design effort - are much more useful, in terms of their capacities for valuable work.
By pushing past human level then, researchers would automatically have fast progress in AI, at least in practical terms.
There will be faster and slower bouts of progress, and disproportionately much research will happen during fast bouts.
[xxx]