But will it ever be relevant to human beings? I mean, will it ever be relevant in however long we have, before superhuman AI emerges?
Quantum computing is a kind of reversible computing that already exists in limited form. But will classical reversible computing ever matter? We already have the phenomenon of “dark silicon”, in which parts of the chip must go unused so it can cool down, and we have various proposals for “adiabatic computation” and “adiabatic logic” to push the boundaries of this, but it’s unclear to me how much that overlaps with the theory of reversible computing.
My guess is probably not, but the probably matters here, though it depends on when superhuman AI emerges, and when reversible computers become practical.
And in any case, you’re responding to a different scenario than the answer was focusing on, as you added the condition of reversible computers appearing before AI taking all our jobs.
Well, AI has already received a lot of funding, and a lot of progress has been made. But we are near the limit of the roadmap for our conventional irreversible CMOS technology. In other words, since we already have chips with features about 3 nm long, there is not much room to shrink the features of these chips. We are running out of improvements of irreversible computation. Maybe we do not see as much progress in reversible computation development because of a lack of funding and interest and not because reversible computing is somehow infeasible. If there were as much interest in reversible computing as there were in AI, the promise of energy efficient reversible computing may seem much closer and a problem which is actually solvable. I therefore do not think it is fair to compare reversible computing with AI this way because of the lack of interest in reversible computing.
So are you saying that AI will be efficient enough to replace humanity before reversible computing becomes practical? I do not believe that this is likely since it would require superhuman AI to be possible on inefficient hardware without making any use of any reversible-friendly algorithms like backtracking. Any superhuman AI will be able to figure out reversible computation, so we should expect superhuman AI to be reversible.
And quantum computing is good for only specific algorithms doing very specific tasks, while reversible computing will be good for all forms of computation.
There is absolutely no reason for quantum computation and AGI to obtain a free pass while everyone suddenly becomes skeptical of reversible computation. Yes, we should be skeptical, but we should apply our skepticism equally to reversible computation, quantum computation, robotics, and AI.
I do not believe that potentially profitable energy efficient reversible computation is that far off. After all, one can design cryptocurrency mining algorithms for reversible computing hardware; these mining algorithms can be made to be 99.9% reversible without any time/space overhead that is typically incurred from using reversible instead of irreversible computation. If we wanted to, we can make profitable reversible ASICs for cryptocurrency mining in the foreseeable future, but this will only work if people stop ignoring reversible computation. After all, the energy used for cryptocurrency mining is otherwise wasted, so we can use it to accelerate reversible computing development. If and when we use cryptocurrencies with mining algorithms that are designed for reversibility, the development of reversible computing hardware will not stop, and it will replace all computing hardware.
But if you are afraid of reversible computing since it would make it that much easier for AI to take over, then you should invest in a cryptocurrency with a reversibility resistant mining algorithm.
I once knew someone who had planned to create superintelligence by building a dedicated computer network on the moon. Then the Internet came along and he was able to dispense with the lunar step in his plan. Reversible computing seems like that—physically possible, but not at all a necessary step.
From where I stand, we’re rushing towards superhuman AI right now, on the hardware we already have. If superhuman AI were somehow 30 years away, then yes, there might be time for your crypto bootstrap scheme to yield a new generation of reversible chips. But in a world that already contains GPT-4, I don’t think we have that much time.
I agree that having a computer network on the moon is impractical. I cannot really see the purpose of putting a computer network on the moon unless it is a part of a lunar station, and I certainly cannot see how that will result in a superintelligence. But I can certainly imagine how reversible computation will result in a superintelligence since the superintelligence will eventually incorporate reversible computation, so the only real question is whether AGI will come before reversible computation.
AI is often hyped, and a lot of this hype happens for a good reason. We are currently in an AI spring or an AI summer where people are excited and worried about AI, and during an AI spring, people may tend to overestimate the abilities of AI and the future progress of AI. The reason why some people may overestimate the future progress of AI is because we need to be cautious about AI since it is a potentially very dangerous technology, and this caution is healthy. But the previous AI springs were followed by AI winters, and this time I do not believe we will have another AI winter, but we may have another AI fall where AI has a more proper amount of hype.
The difference between previous AI winters and my predicted future AI fall is that in previous AI winters, there was a lot more room for progress with irreversible hardware, but we are at a point where there is not much room to improve our irreversible hardware. Rest in peace, Gordon Moore (1929-2023). This means that in order to get AGI without reversible computation, there better be very great algorithmic improvements. Therefore, on a hardware regard, my predicted AI fall will see improvements that are not quite as great as they used to be. Are you sure you want to base your AGI on mainly just software improvements and minor hardware improvements along with hardware improvements that are orthogonal to the energy efficiency per logic gate operation? If not, then we will also need reversible computation.
So if we have AGI before energy efficient reversible computation, do you think that the AGI will not use cryptocurrency mining to accelerate the development of reversible computing hardware?
P.S. One of the main reasons to use cryptocurrency mining to accelerate the development of reversible computation is that there is no downside except for how reversible mining has no market capitalization (The cryptocurrency community has an education problem.). People will still be mining cryptocurrency, so one might as well use that cryptocurrency mining in order to try to solve the problem of reversible computation as well as possible, and Bitcoin would have been better if it had a mining algorithm that was specifically designed for reversible computation at the beginning.
I cannot really see the purpose of putting a computer network on the moon [to create superintelligence]
Probably the scenario involved von Neumann machines too—a whole lunar industrial ecology of self-reproducing robots. This was someone from Russia in the first half of the 1990s, who grew up without Internet and with Earth as a geopolitical battlefield. Given that context, it makes visionary sense to imagine pursuing one’s posthuman technolibertarian dreams in space. But he adjusted to the Internet era soon enough.
if we have AGI before energy efficient reversible computation, do you think that the AGI will not use cryptocurrency mining to accelerate the development of reversible computing hardware?
You may be aware that Robin Hanson and Eliezer Yudkowsky have debated a few times over differing scenarios for the AI future. One of the differences is that Robin envisages a kind of pluralism and gradualism, a society and an economy where humans and human uploads and autonomous AIs are interacting as peers for quite some time. On the other hand, Eliezer predicts that the AGI era yields a superintelligent agent quite rapidly, one which, in the words of Bill Joy, “doesn’t need us”.
I think an AGI using a crypto bootstrap to develop reversible hardware, really only makes sense in a future like Robin’s. In Eliezer’s scenario, the AI just directly appropriates whatever resources it needs for its plans.
It will probably be easier to make self reproducing robots in a lab instead of on the moon. After all, in a laboratory, you can control variables such as the composition of minerals, energy sources, and hazards much better than you can just by sending the robots to the moon. But by the time we are able to have self-reproducing robots, we probably would have made reversible computers already.
But if your and Eliezer’s predictions come true, you will need to not only get superhuman AGI running before we have energy efficient reversible computation that is profitable for many purposes, but you will also need this superhuman AGI to be able to reproduce itself and take over the world without anyone noticing before it is too late.
Are you sure that your superhuman AGI will be able to figure out how to take over the world without even having efficient reversible hardware in the first place? It is one thing for a superhuman AGI to take over the world without anyone being able to do anything about it, but it is another thing entirely for that superhuman AGI to begin with limited and inefficient hardware resources.
P.S. You are also making an assumption that superhuman AGI will not have any use for a currency. In order for this assumption to be reasonable, there must be only a few (like 3 or 4) instances of superhuman AGI where they all know each other. This also seems unlikely. Obtaining currency is one of those instrumentally convergent goals that all these superhuman AGIs with goals would have.
But will it ever be relevant to human beings? I mean, will it ever be relevant in however long we have, before superhuman AI emerges?
Quantum computing is a kind of reversible computing that already exists in limited form. But will classical reversible computing ever matter? We already have the phenomenon of “dark silicon”, in which parts of the chip must go unused so it can cool down, and we have various proposals for “adiabatic computation” and “adiabatic logic” to push the boundaries of this, but it’s unclear to me how much that overlaps with the theory of reversible computing.
My guess is probably not, but the probably matters here, though it depends on when superhuman AI emerges, and when reversible computers become practical.
And in any case, you’re responding to a different scenario than the answer was focusing on, as you added the condition of reversible computers appearing before AI taking all our jobs.
Well, AI has already received a lot of funding, and a lot of progress has been made. But we are near the limit of the roadmap for our conventional irreversible CMOS technology. In other words, since we already have chips with features about 3 nm long, there is not much room to shrink the features of these chips. We are running out of improvements of irreversible computation. Maybe we do not see as much progress in reversible computation development because of a lack of funding and interest and not because reversible computing is somehow infeasible. If there were as much interest in reversible computing as there were in AI, the promise of energy efficient reversible computing may seem much closer and a problem which is actually solvable. I therefore do not think it is fair to compare reversible computing with AI this way because of the lack of interest in reversible computing.
So are you saying that AI will be efficient enough to replace humanity before reversible computing becomes practical? I do not believe that this is likely since it would require superhuman AI to be possible on inefficient hardware without making any use of any reversible-friendly algorithms like backtracking. Any superhuman AI will be able to figure out reversible computation, so we should expect superhuman AI to be reversible.
And quantum computing is good for only specific algorithms doing very specific tasks, while reversible computing will be good for all forms of computation.
There is absolutely no reason for quantum computation and AGI to obtain a free pass while everyone suddenly becomes skeptical of reversible computation. Yes, we should be skeptical, but we should apply our skepticism equally to reversible computation, quantum computation, robotics, and AI.
I do not believe that potentially profitable energy efficient reversible computation is that far off. After all, one can design cryptocurrency mining algorithms for reversible computing hardware; these mining algorithms can be made to be 99.9% reversible without any time/space overhead that is typically incurred from using reversible instead of irreversible computation. If we wanted to, we can make profitable reversible ASICs for cryptocurrency mining in the foreseeable future, but this will only work if people stop ignoring reversible computation. After all, the energy used for cryptocurrency mining is otherwise wasted, so we can use it to accelerate reversible computing development. If and when we use cryptocurrencies with mining algorithms that are designed for reversibility, the development of reversible computing hardware will not stop, and it will replace all computing hardware.
But if you are afraid of reversible computing since it would make it that much easier for AI to take over, then you should invest in a cryptocurrency with a reversibility resistant mining algorithm.
I once knew someone who had planned to create superintelligence by building a dedicated computer network on the moon. Then the Internet came along and he was able to dispense with the lunar step in his plan. Reversible computing seems like that—physically possible, but not at all a necessary step.
From where I stand, we’re rushing towards superhuman AI right now, on the hardware we already have. If superhuman AI were somehow 30 years away, then yes, there might be time for your crypto bootstrap scheme to yield a new generation of reversible chips. But in a world that already contains GPT-4, I don’t think we have that much time.
I agree that having a computer network on the moon is impractical. I cannot really see the purpose of putting a computer network on the moon unless it is a part of a lunar station, and I certainly cannot see how that will result in a superintelligence. But I can certainly imagine how reversible computation will result in a superintelligence since the superintelligence will eventually incorporate reversible computation, so the only real question is whether AGI will come before reversible computation.
AI is often hyped, and a lot of this hype happens for a good reason. We are currently in an AI spring or an AI summer where people are excited and worried about AI, and during an AI spring, people may tend to overestimate the abilities of AI and the future progress of AI. The reason why some people may overestimate the future progress of AI is because we need to be cautious about AI since it is a potentially very dangerous technology, and this caution is healthy. But the previous AI springs were followed by AI winters, and this time I do not believe we will have another AI winter, but we may have another AI fall where AI has a more proper amount of hype.
The difference between previous AI winters and my predicted future AI fall is that in previous AI winters, there was a lot more room for progress with irreversible hardware, but we are at a point where there is not much room to improve our irreversible hardware. Rest in peace, Gordon Moore (1929-2023). This means that in order to get AGI without reversible computation, there better be very great algorithmic improvements. Therefore, on a hardware regard, my predicted AI fall will see improvements that are not quite as great as they used to be. Are you sure you want to base your AGI on mainly just software improvements and minor hardware improvements along with hardware improvements that are orthogonal to the energy efficiency per logic gate operation? If not, then we will also need reversible computation.
So if we have AGI before energy efficient reversible computation, do you think that the AGI will not use cryptocurrency mining to accelerate the development of reversible computing hardware?
P.S. One of the main reasons to use cryptocurrency mining to accelerate the development of reversible computation is that there is no downside except for how reversible mining has no market capitalization (The cryptocurrency community has an education problem.). People will still be mining cryptocurrency, so one might as well use that cryptocurrency mining in order to try to solve the problem of reversible computation as well as possible, and Bitcoin would have been better if it had a mining algorithm that was specifically designed for reversible computation at the beginning.
Probably the scenario involved von Neumann machines too—a whole lunar industrial ecology of self-reproducing robots. This was someone from Russia in the first half of the 1990s, who grew up without Internet and with Earth as a geopolitical battlefield. Given that context, it makes visionary sense to imagine pursuing one’s posthuman technolibertarian dreams in space. But he adjusted to the Internet era soon enough.
You may be aware that Robin Hanson and Eliezer Yudkowsky have debated a few times over differing scenarios for the AI future. One of the differences is that Robin envisages a kind of pluralism and gradualism, a society and an economy where humans and human uploads and autonomous AIs are interacting as peers for quite some time. On the other hand, Eliezer predicts that the AGI era yields a superintelligent agent quite rapidly, one which, in the words of Bill Joy, “doesn’t need us”.
I think an AGI using a crypto bootstrap to develop reversible hardware, really only makes sense in a future like Robin’s. In Eliezer’s scenario, the AI just directly appropriates whatever resources it needs for its plans.
It will probably be easier to make self reproducing robots in a lab instead of on the moon. After all, in a laboratory, you can control variables such as the composition of minerals, energy sources, and hazards much better than you can just by sending the robots to the moon. But by the time we are able to have self-reproducing robots, we probably would have made reversible computers already.
But if your and Eliezer’s predictions come true, you will need to not only get superhuman AGI running before we have energy efficient reversible computation that is profitable for many purposes, but you will also need this superhuman AGI to be able to reproduce itself and take over the world without anyone noticing before it is too late.
Are you sure that your superhuman AGI will be able to figure out how to take over the world without even having efficient reversible hardware in the first place? It is one thing for a superhuman AGI to take over the world without anyone being able to do anything about it, but it is another thing entirely for that superhuman AGI to begin with limited and inefficient hardware resources.
P.S. You are also making an assumption that superhuman AGI will not have any use for a currency. In order for this assumption to be reasonable, there must be only a few (like 3 or 4) instances of superhuman AGI where they all know each other. This also seems unlikely. Obtaining currency is one of those instrumentally convergent goals that all these superhuman AGIs with goals would have.