I broadly agree with a 3-20 year timeline for a majority of office work. But some important qualifying statements I would add:

- some jobs will stay with humans even when AI would be better at it. We already see a lot of this with even with pre-AI automatisation. Neither markets nor companies are perfectly efficient

- at the point where AI is better than the average human, half of all humans are still better than AI. For companies or departments built around employing lots of average people the cutover point will be a lot earlier than for shops that aim to employ the best of the best. Social change is inevitable long before the best are out of work

- the actual benchmark for " replacement" is not human vs machine, but human plus machine vs machine alone. But the difference doesn't matter much because efficiency increases still displace workers

- I don't think robots will advance enough to meet this timeline. This is not just a software issue. Humans have an amazing suite of sensors and actuators. Just replicating a human hand is insanely complex. Walking, jumping robots are crude automatons in comparison. We can cover a lot with specialized robots, but we won't replace humans in physical jobs in 20 years

I agree that robots are much further off than people expect, in raw technical terms. As you point out, the sensors and actuators in a human hand are far beyond the state of the art.

But all of that is assuming a world where research is being done by humans, or by some mix of humans and something like current LLMs. The bottlenecks would ultimately come down to human judgement and human oversight, and that's a significant limiting factor. Plus, you have to push matter around, which takes time, and you have to extract a lot of information out of limited experiences, which LLMs are bad at.

But if someone is reckless and clever enough to build AIs that can completely replace engineers, or that only need humans as hands, then I don't think we can count on robotics remaining intractable for more than a decade or so. In a wide variety of circumstances, it's possible to make do with worse actuators than the human hand, or with specialized actuators. We can already build incredibly precise motors and specialized sensors. The trouble comes with trying to pack enough of them together to replicate the full generality of the human hand. (I have actually helped build task-specific actuators that did quite well with a single motor and a single visual sensor, before.)

So to put my position more precisely: we cannot automate manual labor robotics without having previously automated creative intellectual labor. But conditional on automating creative research, then I expect worryingly rapid advances in robotics.

To be clear, I think that developing fully-general replacements for human intellectual and physical labor would potentially be the biggest disaster in all of human history.