20 years is far too long a time horizon to make a sensible prediction, to even say if the other post is an overestimate or an underestimate by either time or percentage.
20 years ago, the state of the art in AI was so bad that it predates reCAPTCHA being a functional business model for how to get labelled training data for an OCR to read scans of books. In cars, the DARPA Grand Challenge (2004) had no cars able to complete the route, and the winner only got 11.78 km (of 240 km) before getting stuck, and the subsequent Grand Challenge had yet to happen because that was October 8, 2005.
I can't sensibly forecast past 2032, because that's when various things currently under exponential growth reach absolute levels that I think suggest the exponential assumption may stop working. But also, I don't expect wild AI-induced economic disruption* before then, because one of those exponential growths is renewable energy getting to near-parity with current global electricity supply by about then, and the fact that the current global electricity supply is "only" 250 watts/capita, which means** there's just not enough electricity to run enough computers to disrupt that many jobs.
But if the question is "when will the AI be good enough to do ${task}?", rather than "how fast can we deploy the AI when this happens?", I have seen both (1) people claim we're decades away from AI doing things mere months before AI did those things, and also (2) Musk repeatedly saying FSD is 0.5-2 years away every year for the last decade, and when he finally soft-launched his robotaxi it was 6.5 years behind Waymo's.
* by which I mean international >20% unemployment that's not just due to a war or a global financial crisis or a pandemic
** absent algorithmic breakthroughs; I don't expect enough of those to matter, but I do expect some. I am also accounting for computers getting more energy efficient, because that's still happening and we're a long way from the theoretical limits.
20 years ago, the state of the art in AI was so bad that it predates reCAPTCHA being a functional business model for how to get labelled training data for an OCR to read scans of books. In cars, the DARPA Grand Challenge (2004) had no cars able to complete the route, and the winner only got 11.78 km (of 240 km) before getting stuck, and the subsequent Grand Challenge had yet to happen because that was October 8, 2005.
I can't sensibly forecast past 2032, because that's when various things currently under exponential growth reach absolute levels that I think suggest the exponential assumption may stop working. But also, I don't expect wild AI-induced economic disruption* before then, because one of those exponential growths is renewable energy getting to near-parity with current global electricity supply by about then, and the fact that the current global electricity supply is "only" 250 watts/capita, which means** there's just not enough electricity to run enough computers to disrupt that many jobs.
But if the question is "when will the AI be good enough to do ${task}?", rather than "how fast can we deploy the AI when this happens?", I have seen both (1) people claim we're decades away from AI doing things mere months before AI did those things, and also (2) Musk repeatedly saying FSD is 0.5-2 years away every year for the last decade, and when he finally soft-launched his robotaxi it was 6.5 years behind Waymo's.
* by which I mean international >20% unemployment that's not just due to a war or a global financial crisis or a pandemic
** absent algorithmic breakthroughs; I don't expect enough of those to matter, but I do expect some. I am also accounting for computers getting more energy efficient, because that's still happening and we're a long way from the theoretical limits.