More than a quarter of U.S. GDP lies with industries that are physical-labor-intensive and likely to remain so. In turn, the realization of AGI as defined here depends a lot on this claim: "Once AI can handle most knowledge work, highly capable physical robots will follow within a few years at most." How plausible is this claim?
I don't have a good guess about its plausibility; I'd like to know more.
It's a great question, and I don't really have a lot to say. My intuition is based on three things:
1. There's already quite a bit of progress in both hardware and software for robots. Software, in particular, seems to be making rapid strides using the same sorts of techniques that have been so successful in training LLMs (the "transformer" architecture, gathering large amounts of training data, etc.). Previously, it was extremely difficult to train robots for anything but the most repetitive, stylized activities in highly controlled environments, but it seems this barrier is finally starting to crack. However, I'm unclear on the details and even less clear on the state and ongoing pace of progress on the hardware side.
2. AI that can handle "most knowledge work" would likely be able to further accelerate progress on the software side of robotics, and possibly on the hardware side as well.
3. If (1) and (2) combine to make it seem that the goal of highly capable / general robots is within reach, the financial incentives to push over the line will be stupendous, and there will be a tsunami of investment.
But of course this is all handwaving. I'm hoping at some point to be able to convene some discussions to shed more light on all of this.
More than a quarter of U.S. GDP lies with industries that are physical-labor-intensive and likely to remain so. In turn, the realization of AGI as defined here depends a lot on this claim: "Once AI can handle most knowledge work, highly capable physical robots will follow within a few years at most." How plausible is this claim?
I don't have a good guess about its plausibility; I'd like to know more.
It's a great question, and I don't really have a lot to say. My intuition is based on three things:
1. There's already quite a bit of progress in both hardware and software for robots. Software, in particular, seems to be making rapid strides using the same sorts of techniques that have been so successful in training LLMs (the "transformer" architecture, gathering large amounts of training data, etc.). Previously, it was extremely difficult to train robots for anything but the most repetitive, stylized activities in highly controlled environments, but it seems this barrier is finally starting to crack. However, I'm unclear on the details and even less clear on the state and ongoing pace of progress on the hardware side.
2. AI that can handle "most knowledge work" would likely be able to further accelerate progress on the software side of robotics, and possibly on the hardware side as well.
3. If (1) and (2) combine to make it seem that the goal of highly capable / general robots is within reach, the financial incentives to push over the line will be stupendous, and there will be a tsunami of investment.
But of course this is all handwaving. I'm hoping at some point to be able to convene some discussions to shed more light on all of this.