3 ms·

I mean, sure, but the value proposition of a slow simulation containing slow minds is a little dubious, isn't it? It would be neither cheap nor useful, so why w

I mean, sure, but the value proposition of a slow simulation containing slow minds is a little dubious, isn't it? It would be neither cheap nor useful, so why would it be made?

naasking 2y ago
There's considerable value actually. You could simulate a whole economy and test economic theories if you could simulate a country full of minds, all with different interests and values. There's no need for this kind of simulation to be real-time. That there's value in simulating lots of minds is the core premise of the simulation hypothesis.
- breuleux 2y ago
  We can't simply assert that this is valuable, we have to analyze it properly. Mass simulation will happen if the expected value of the simulation exceeds the opportunity cost of, well, everything else we could do with these resources instead. First, the expected value... it is, frankly, questionable. Certainly it's great in the best case scenario where we stumble upon a good system and then put it in application, but how likely is that? I can see several cases where the entire value would be sunk: * You will need to run not just one, but many simulations. However, the same argument that implies the simulation will be slower than reality also implies that it will be bigger than reality. The problem isn't merely that this simulation won't occur in real time, it's also that it's going to be physically enormous. Even running a single one would be quite the undertaking. * It is quite possible that an economic system would work very well at first, only to collapse 50 years later for a variety of difficult-to-foresee reasons. If the simulation runs too slowly, we may have to commit to a system before the simulation can demonstrate its collapse. Or wait centuries, but then the initial conditions may have diverged far too much for the experiment to translate. * We're not even guaranteed to glean any important insights from such a simulation. Civilization is a chaotic system that involves a ton of smart agents who constantly come up with paradigm-shifting inventions or new ways to bend and manipulate the system. It won't take much time before all you're doing is explore some random alternate future, although it's slower than real time, so it's actually some random alternate past. It's not useless, but it's middling. * And then you have to implement changes. Very sensitive to the current conditions, which change faster than the simulator can keep up with. Good luck. As for the second part of the equation, the opportunity cost: likely humongous. Insofar that simulation hardware is a general purpose substrate, there is a mind-boggling number of uses for it. If you can simulate minds faster than real time, sure, that's useful. If not, it's unclear to me how a poor imitation of a real thing would even crack the top 10 of things you'd want to do with this hardware. Sorry for all the text. I think the whole debate can be summarized like this: the simulation hypothesis makes sense if simulation is cheap, it falls apart if it isn't. And I don't think simulation is going to be cheap (unless it's quite crude and approximate, in which case it won't tell us much about reality.)
  - naasking 2y ago
    > * You will need to run not just one, but many simulations. However, the same argument that implies the simulation will be slower than reality also implies that it will be bigger than reality. You don't need to simulate reality, you only need to simulate the minds that perceive reality. Thus cuts computational costs by many orders of magnitude over a reality simulation. We don't have the understanding yet to be able to precisely quantify it, but just eliminating the dependency on quantum mechanics is huge. > * It is quite possible that an economic system would work very well at first, only to collapse 50 years later for a variety of difficult-to-foresee reasons. And climate models could diverge from reality 100 or 200 years later too, we still run those because reasonably accurate short to medium term information is better than none. > (unless it's quite crude and approximate, in which case it won't tell us much about reality.) Even our crude and approximate simulations of quantum chemistry, protein folding, weather, climate, orbital dynamics etc. tell us a lot.
    - breuleux 2y ago
      > You don't need to simulate reality, you only need to simulate the minds that perceive reality. Thus cuts computational costs by many orders of magnitude over a reality simulation. Ehhhhh... let's think about it. First, that does not eliminate the speed bottleneck, which is determined by what you do decide to simulate accurately. It is... unclear whether human minds can be simulated efficiently. I would not be surprised if the best accurate simulated brain we can muster is ten times slower and ten times bigger than a real one. Or worse. Second, surely you need to simulate more than minds. You need some kind of environment to put them in. But if you're giving them some pale ersatz of reality in which they cannot develop effective technology, because stuff stops working when they stop looking at it, I'd argue you're wasting them. Minds act very differently depending on their capabilities --- simulating minds without simulating reality is a bit like studying the behavior of birds with clipped wings. > just eliminating the dependency on quantum mechanics is huge. What makes you think it can be eliminated? I'm not saying minds require quantum mechanics in principle, but evolution doesn't try to make its designs robust to changes in the laws of physics. There's a sizable chance that simulation of a human brain fails if you ignore quantum effects, simply because that violates its calibration. You would therefore need to reimplement minds from first principles, but that's a whole other problem. > And climate models could diverge from reality 100 or 200 years later too, we still run those because reasonably accurate short to medium term information is better than none. These models run far, far, far faster than real time. You can't forecast a phenomenon using a simulation that's slower than the phenomenon itself.

I mean, sure, but the value proposition of a slow simulation containing slow minds is a little dubious, isn't it? It would be neither cheap nor useful, so why would it be made?

naasking 2y ago

There's considerable value actually. You could simulate a whole economy and test economic theories if you could simulate a country full of minds, all with different interests and values. There's no need for this kind of simulation to be real-time. That there's value in simulating lots of minds is the core premise of the simulation hypothesis.

breuleux 2y ago
We can't simply assert that this is valuable, we have to analyze it properly. Mass simulation will happen if the expected value of the simulation exceeds the opportunity cost of, well, everything else we could do with these resources instead. First, the expected value... it is, frankly, questionable. Certainly it's great in the best case scenario where we stumble upon a good system and then put it in application, but how likely is that? I can see several cases where the entire value would be sunk: * You will need to run not just one, but many simulations. However, the same argument that implies the simulation will be slower than reality also implies that it will be bigger than reality. The problem isn't merely that this simulation won't occur in real time, it's also that it's going to be physically enormous. Even running a single one would be quite the undertaking. * It is quite possible that an economic system would work very well at first, only to collapse 50 years later for a variety of difficult-to-foresee reasons. If the simulation runs too slowly, we may have to commit to a system before the simulation can demonstrate its collapse. Or wait centuries, but then the initial conditions may have diverged far too much for the experiment to translate. * We're not even guaranteed to glean any important insights from such a simulation. Civilization is a chaotic system that involves a ton of smart agents who constantly come up with paradigm-shifting inventions or new ways to bend and manipulate the system. It won't take much time before all you're doing is explore some random alternate future, although it's slower than real time, so it's actually some random alternate past. It's not useless, but it's middling. * And then you have to implement changes. Very sensitive to the current conditions, which change faster than the simulator can keep up with. Good luck. As for the second part of the equation, the opportunity cost: likely humongous. Insofar that simulation hardware is a general purpose substrate, there is a mind-boggling number of uses for it. If you can simulate minds faster than real time, sure, that's useful. If not, it's unclear to me how a poor imitation of a real thing would even crack the top 10 of things you'd want to do with this hardware. Sorry for all the text. I think the whole debate can be summarized like this: the simulation hypothesis makes sense if simulation is cheap, it falls apart if it isn't. And I don't think simulation is going to be cheap (unless it's quite crude and approximate, in which case it won't tell us much about reality.)
- naasking 2y ago
  > * You will need to run not just one, but many simulations. However, the same argument that implies the simulation will be slower than reality also implies that it will be bigger than reality. You don't need to simulate reality, you only need to simulate the minds that perceive reality. Thus cuts computational costs by many orders of magnitude over a reality simulation. We don't have the understanding yet to be able to precisely quantify it, but just eliminating the dependency on quantum mechanics is huge. > * It is quite possible that an economic system would work very well at first, only to collapse 50 years later for a variety of difficult-to-foresee reasons. And climate models could diverge from reality 100 or 200 years later too, we still run those because reasonably accurate short to medium term information is better than none. > (unless it's quite crude and approximate, in which case it won't tell us much about reality.) Even our crude and approximate simulations of quantum chemistry, protein folding, weather, climate, orbital dynamics etc. tell us a lot.
  - breuleux 2y ago
    > You don't need to simulate reality, you only need to simulate the minds that perceive reality. Thus cuts computational costs by many orders of magnitude over a reality simulation. Ehhhhh... let's think about it. First, that does not eliminate the speed bottleneck, which is determined by what you do decide to simulate accurately. It is... unclear whether human minds can be simulated efficiently. I would not be surprised if the best accurate simulated brain we can muster is ten times slower and ten times bigger than a real one. Or worse. Second, surely you need to simulate more than minds. You need some kind of environment to put them in. But if you're giving them some pale ersatz of reality in which they cannot develop effective technology, because stuff stops working when they stop looking at it, I'd argue you're wasting them. Minds act very differently depending on their capabilities --- simulating minds without simulating reality is a bit like studying the behavior of birds with clipped wings. > just eliminating the dependency on quantum mechanics is huge. What makes you think it can be eliminated? I'm not saying minds require quantum mechanics in principle, but evolution doesn't try to make its designs robust to changes in the laws of physics. There's a sizable chance that simulation of a human brain fails if you ignore quantum effects, simply because that violates its calibration. You would therefore need to reimplement minds from first principles, but that's a whole other problem. > And climate models could diverge from reality 100 or 200 years later too, we still run those because reasonably accurate short to medium term information is better than none. These models run far, far, far faster than real time. You can't forecast a phenomenon using a simulation that's slower than the phenomenon itself.