It is mentally healthy to have an informed perspective especially when the more rational and informed perspective gives us a reason for more hope. In case you did not notice, there is not much room to shrink the feature size of transistors (TSMC is making 2nm features now, and atoms are about 0.1 nm in size, so there is not much room to shrink stuff). Furthermore, if the transistors are too small, they won’t work because of quantum tunnelling. There is also a limit to the energy efficiency of irreversible computation because in order to reliably delete information, one must overcome thermal noise. We are approaching this energy efficiency limit, so I wish TSMC good luck the progress in the performance of irreversible computation, since they are going to need it.
We can get beyond these limits using reversible computation, but reversible computation is a difficult technical challenge. Furthermore, reversible computation comes with a computational complexity overhead. It takes more time/space and parallelism to compute reversibly than it does to compute irreversibly. We may therefore have some time before we get sufficient hardware improvements that make AI an existential threat.
On the other hand, it looks like most people who are talking about AI do not know about the limits of irreversible computation and the promise and challenges of reversible computation. This does not appear to be very mentally healthy to me. This is a complete turn-off. I hope the AI community learns to do better.
The physical limits mainly apply to irreversible computation. But it seems like powerful reversible computation is attainable. Once we get well-optimized reversible computation, I will not make any bets against AGI. But building reversible computing technologies will probably be exceedingly difficult since we have to deal with things like a computational complexity overhead with reversible computation. This means that we probably have some time left before an AI apocalypse to try to get a good solution to the AI alignment problem or to just have fun.
It is mentally healthy to have an informed perspective especially when the more rational and informed perspective gives us a reason for more hope. In case you did not notice, there is not much room to shrink the feature size of transistors (TSMC is making 2nm features now, and atoms are about 0.1 nm in size, so there is not much room to shrink stuff). Furthermore, if the transistors are too small, they won’t work because of quantum tunnelling. There is also a limit to the energy efficiency of irreversible computation because in order to reliably delete information, one must overcome thermal noise. We are approaching this energy efficiency limit, so I wish TSMC good luck the progress in the performance of irreversible computation, since they are going to need it.
We can get beyond these limits using reversible computation, but reversible computation is a difficult technical challenge. Furthermore, reversible computation comes with a computational complexity overhead. It takes more time/space and parallelism to compute reversibly than it does to compute irreversibly. We may therefore have some time before we get sufficient hardware improvements that make AI an existential threat.
On the other hand, it looks like most people who are talking about AI do not know about the limits of irreversible computation and the promise and challenges of reversible computation. This does not appear to be very mentally healthy to me. This is a complete turn-off. I hope the AI community learns to do better.
Are you saying people should be more skeptical of AGI because of the physical limits on computation and thus more hopeful?
The physical limits mainly apply to irreversible computation. But it seems like powerful reversible computation is attainable. Once we get well-optimized reversible computation, I will not make any bets against AGI. But building reversible computing technologies will probably be exceedingly difficult since we have to deal with things like a computational complexity overhead with reversible computation. This means that we probably have some time left before an AI apocalypse to try to get a good solution to the AI alignment problem or to just have fun.