= 27d567bc2d48d9f40e9ccc20ea370b15)
Desrtopa comments on Theists are wrong; is theism? - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (533)
There was already a thread on this. The general consensus seems to be that it isn't practical, if possible.
Hmm from my reading of the thread it doesn't look like much of a consensus.
I may want to revive this - the arguments against practicality don't seem convincing from an engineering perspective.
From a high quality upload's scanned mind one should get a great deal of information about the upload's closest friends, relatives, etc. The data from any one such upload many not be overwhelming, but you'd start with a large population of such uploads. People who were well known and loved would be easier cases, but you could also supplement the data in many cases with low-quality scans from poorly preserved bodies.
This should give one prior generation. Going back another previous generation would get murkier, but is still quite possible, especially with all the accessory historical records.
The farther back you go, the less 'accurate' the uploads become, but the less and less important this 'accuracy' becomes.
For example, assuming I become a posthuman, I will be interested in bring back my grandfather. There a huge space of possible minds that could match my limited knowledge and beliefs about this person I never met. Each of them would fully be my grandfather from my subjective perspective and would fully be my grandfather from their subjective perspective.
There is no objective standard frame of reference from which to evaluate absolute claims of personal identity. It is relative.
But if you simulate anything other than the actual brain states of the people in question, then they won't behave in exactly the same way. No matter how many other people's knowledge of me you integrate, for example, you won't have the data to predict what I'll eat for breakfast tomorrow with any accuracy (because I almost invariably eat breakfast alone.) Tiny differences like this will quickly propagate to create much larger ones between the simulation and the reality. Jump forward a few generations and you have zero population overlap between the new generation of the simulation and the next generation that was born in reality. If you're attempting historical recreation, this would be a pretty useless way to go about it.
If you wanted to create a simulation that was an approximation of a particular historical period at one point, but quickly divorced from it as it ran forward, that would be much more plausible, but why would you want to? Everything I can think of that could be accomplished in such a way could more easily be accomplished by doing something else.
Sure, but that's not relevant towards the goal. There are no 'actual' or exact brain states that canonically define people.
If you created a simulation of an alternate 1950 and ran it forward, it would almost certainly diverge, but this is no different than alternate branches of the multiverse. Running the alternate forward to say 2050 may generate a very different reality, but that may not matter much - as long as it also generates a bunch of variants of people we like.
This brings to mind a book by Heinlein about a man who starts jumping around between branches - "Job: a comedy of Justice".
Anyway, my knowledge of my grandfather is vague. But I imagine posthumans could probably nail down his DNA and eventually recreate a very plausible 1890 (around when he was born). We could also nail down a huge set of converging probability estimates from the historical record to figure out where he was when, what he was likely to have read, and so on.
Creating an initial population of minds is probably much trickier. Is there any way to create a fully trained neural net other than by actually training it? I suspect that it's impossible in principle. It's certainly the case in practice today.
In fact, there may be no simple shortcut without going way way back into earlier prehistory, but this is not a fundamental obstacle, as this simulation could presumably be a large public project.
Yes the approach of just creating some initial branch from scratch and then running it forward is extremely naive. If you'd like I could think of ten vastly more sophisticated algorithms that could shape the branch's forward evolution to converge with the main future worldline before breakfast.
The first thing that pops to mind: The historical data that we have forms a very sparse sampling, but we could use it to guide the system's forward simulation, with the historical data acting as constraints and attractors. In these worlds, fate would be quite real. I think this gives you the general idea, but it relates to bidirectional path tracing.
Such as?
We can get to that if you can establish that there's any good reason to do it in the first place.
Your justifications for running such simulations have so far seem to hinge on things we could learn from them (or simply creating them for their own sake, it appears that you're jumping between the two,) but if we know enough about the past to meaningfully create the simulations, then there's not much we stand to learn from making them. Yes, history could have branched in different ways depending on different events that could have occurred, we already know that. If you try to calculate all the possibilities as they branch off, you'll quickly run out of computing power no matter how advanced your civilization is. If you want to do calculations of the most likely outcomes of a certain event, you don't have to create a simulation so advanced that it appears to be a real universe from the inside to do that.
Excellent!
The two are intertwined - we can learn a great deal from our history and ancestors while simultaneous valuing it for other reasons than the learning.
Thinking is just a particular form of approximate simulation. Simulation is a very precise form of thinking.
Right now all we know about our history is the result of taking a small collection of books and artifacts and then thinking alot about them.
Why do we write books about Roman History and debate what really happened? Why do we make television shows or movies out of it?
Consider this just the evolution of what we already do today, for much of the same reasons, but amplified by astronomical powers of increased intelligence/computation generating thought/simulation.
This is what we call a naive algorithm, the kind you don't publish.
Calculations of the likely outcomes of certain events are the mental equivalents of thermostat operations - they are the types of things you do and think about when you lack hyperintelligence.
Eventually you want a nice canonical history. Not a book, not a movie, but the complete data set and recreation. As it is computed it exists, eventually perhaps you merge it back into the main worldline, perhaps not, and once done and completed you achieve closure.
Put another way, there is a limit where you can know absolutely every conceivable thing there is to know about your history, and this necessitates lots of massively super-detailed thinking about it - aka simulation.
This is the kind of naive forward extrapolation that gets you sci fi dystopias. Most of the things we do today don't bear extrapolating to logical extremes, certainly not this.
No I don't. I think you should try asking more people if this is actually something they would want, with knowledge of the things they could be doing instead, rather than assuming it's a logical extrapolation of things that they do want. If I could do that, it wouldn't even bottom the list of things I'd want to do with that power.
The simulation doesn't teach us more than we already know about history. What we already know about history sets the upper bound on how similar we can make it. Given the size of the possibility space, we can only reasonably assume that it's different in every way that we do not enforce similarity on it. The simulation doesn't contribute to knowing everything you could possibly know about your history, that's a prerequisite, if you want the simulation to be faithful.
This would be true if we were equally ignorant about all of history. However, there are some facts regarding history we can be quite confident about- particularly recent history and the present. You can then check possible hypotheses about history (starting from what is hopefully an excellent estimation of starting conditions) against those facts you do have. Given how contingent the genetic make-up of a human is on the timing of their conception and how strongly genetics influences who we are it seems plausible a physical simulation of this part of the universe could radically narrow the space of possibilities given enough computing power. Of course parts of the simulation might remain under-determined but it seems implausible that a simulation would tell us nothing new about history as a simulation should be more proficient than humans at assessing the necessary consequences and antecedences to any known event.
Radically narrow, but given just how vast the option space is, it takes a whole lot more than radically narrowing before you can winnow it down to a manageable set of possibilities.
This post puts some numbers to the possible configurations you can get for a single lump of matter of about 1.5 kilograms. In a simulation of Earth, far more matter than that is in a completely unknown state and free to vary through a huge portion of its possibility space (that's not to say that even an appreciable fraction of matter on Earth is free to vary through all possible states, but the numbers are mind boggling enough even if we're only dealing with a few kilograms.) Every unknown configuration is a potential confounding factor which could lead to cascading changes. The space is so phenomenally vast that you could narrow it by a billion orders of magnitude, and it would still occupy approximately the same space on the scale of sheer incomprehensibility. You would have to actively and continuously enforce similarity on the simulation to keep it from diverging more and more widely from the original.
Said reference post by AndrewHickey starts with a ridiculous assumption:
This is voodoo-quantum consciousness: the idea that your mind-identity somehow depends on details down to the quantum state. This can't possibly be true - because the vast vast majority of that state changes rapidly from quantum moment to moment in a mostly random fashion. There thus is no single quantum state that corresponds uniquely to a mind, rather there is a vast configuration space.
You can reduce that space down to a smaller bit representation by removing redundant details. Does it really matter if I remove one molecule from one glial cell in your brain? The whole glial cell? All the glial cells?
There is a single minimal representation of a computer - it reduces exactly down to it's circuit diagram and the current values it holds in it's memory/storage.
If you don't buy into the idea that a human mind ultimately reduces down to some functional equivalent computer program, than of course the entire Simulation Argument won't follow.
Who cares?
There could be infinite detail in the universe - we could find that there are entire layers beneath the quantum level, recursing to infinity, such that perfect simulation was impossible in principle .. and it still wouldn't matter in the slightest.
You only need as much detail in the simulation as . . you want detail in the simulation.
Some details at certain spatial scales are more important than others based on their leverage casual effect - such as the bit values in computers, synaptic weights in brains.
A simulation at the human-level scale would only need enough detail to simulate conscious humans, which will probably include simulating down to rough approximations to synaptic-net equivalents. I doubt you would even simulate every cell in the body, for example - unless that itself was what you were really interested in.
There is another significant mistake in typical feasibility critique of simulationism: assuming your current knowledge of algorithmic simulation is the absolute state of the art for now to eternity, the final word, and superintelligences won't improve on it in the slightest.
As a starting example, AndrewHickey and you both appear to be assuming that the simulation must maintain full simulation fidelity across the entire spatio-temporal field. This is a primitive algorithm. A better approach is to adaptively subdivide space-time and simulate at multiple scales at varying fidelity using importance sampling, for example.
Good point. I'm reconsidering...
I wonder what kind of cascade effect there actually is- perhaps there are parts of the simulation that could be done using heuristics and statistical simplifications. Perhaps that could be done to initially narrow the answer space and then the precise simulation could be sped up by not having to simulate those answers that contradict the simplified model?
I wonder how a hidden variable theory of quantum mechanics being true would effect the prospects for simulation- assuming a super intelligence could leverage that fact somehow (which is admittedly unlikely).
Most of the things we do today are predictable developments of what previous generations did, and this statement holds across time.
There is a natural evolutionary progression: dreams/daydreams/visualizations -> oral stories/mythologies -> written stories/plays/art -> movies/television->CG/virtual reality/games->large scale simulations
It isn't 'extrapolating to logical extremes', it is future prediction based on extrapolation of system evolution.
Of course it does. What is our current knowledge about history? It consists of some rough beliefs stored in the low precision analog synapses of our neural networks and a bunch of word-symbols equivalent to the rough beliefs.
With enough simulation we could get concise probability estimates or samples of the full configuration of particles on earth every second for the last billion years - all stored in precise digital transistors, for example.
This is true only for some initial simulation, but each successive simulation refines knowledge, expands the belief network, and improves the next simulation. You recurse.
Not at all. Given an estimate on the state of a system at time T and the rules of the system's time evolution (physics), simulation can derive values for all subsequent time steps. The generated data is then analyzed and confirms or adjusts theories. You can then iteratively refine.
For a quick primitive example, perhaps future posthumans want to understand in more detail why the roman empire collapsed. A bunch of historian/designers reach some rough consensus on a model (built on pieces of earlier models) to build an earth at that time and populate it with inhabitants (creating minds may involve using stand in actors for an initial generation of parents).
Running this model forward may reveal that the lead had little effect, that previous models of some roman military formations don't actually work, that a crop harvest in 32BC may have been more important than previously thought .. and so on.
With the help of hindsight bias.
As wedrifid says, in the light of hindsight bias. Instead of looking at the past and seeing how reliably it seems to lead to the present, try looking at people who actually tried to predict the future. "Future prediction based on extrapolation of system evolution" has reliably failed to make predictions about the direction of human society that were both accurate and meaningful.
Or you could very easily find them removing the lead from their pipes and wine, and changing their military formations. If you don't already know what their crop harvest in 32BC was like, you can practically guarantee that it won't be the same in the simulation. This is exactly the kind of use that, as I pointed out earlier, if you had enough information to actually pull it off, you wouldn't need to.
I'll just reiterate my response then:
Any information about a physical system at time T reveals information about that system at all other times - places constraints on it's configuraiton. Physics is a set of functions that describe the exact relations between system states across time steps, ie the temporal evolution of the system.
We developed physics in order to simulate physical systems and predict and understand their behavior.
This seems then to be a matter of details - how much simulation is required to produce how much knowledge from how much initial information about the system.
For example, with infinite computing power I could iterate through all simulations of earth's history that are consistent with current observational knowledge.
This algorithm computes the probabilities of every fact about the system - the probability of a good crop harvest in 32BC in Egypt is just the fraction of the simulated multiverse for which this property is true.
This algorithm is in fact equivalent to the search procedure in the AIXI universal intelligence algorithm.
I do not believe this is correct. In particular the 'just a' is not accurate. Approximate simulation is a particular kind of thinking not the reverse.
I'm willing to try on your taxonomy but don't quite understand it.
The term thinking certainly covers a wide variety of computations, but perhaps the most important is prediction.
Does this sound more accurate:
Cortical-forward-simulation is just a particular form of approximate simulation. Simulation in general encompasses all the most precise forms of prediction.
More accurate, but still not right. Simulation just doesn't have special privileges. Again, the general, absolute claim of "all the most" invalidates the position. You can make and even logical prove precise predictions without simulating.
How? Got an example?