رونوشت کامل A Human Brain Will Run on a Computer Within a Few Decades | Anders Sandberg

was good at playing Doom. I don't think it beat the game. Mark Fielding (03:01) Quantum consciousness. Jeremy Gilbertson (03:00) Consciousness. So David Chalmers, consciousness, the hard problem. I've read about, I know a lot of people have heard about that in context. I read Fall or Dodge and Hell, Neil Stevenson that talks about scanning brains and all of that stuff. What is the date? You mentioned a date in the paper, I think, 2064, I think. Talk to me about the context of that date. Anders Sandberg (03:28) So the way we got dates, and generally as somebody is doing future studies, I tend to point out that whatever date you put in, that's going to be wrong and embarrassing. But in this case, we were kind of taking the various levels of resolution you could possibly simulate a brain on from a very, very crude, simple model using very simplified neurons over to each individual neuron, over to every little compartment of the neuron. down to the molecular and atomic level. And then you can kind of count up how much computation you need for that. And then you take Moore's law and then you just take a ruler and just continue it very far. And then you can get dates that way. If you want to be a bit more careful, you can analyze it by saying maybe Moore's law slackens off after a while. We're a bit uncertain about that. We're really uncertain about what resolution you need to scan the brain on. So you can combine that and get some probability distribution. Of course I have a paper doing this. But you end up something like, okay, mid-century when we did this, I think the peak probability was something around 2060 or so. My updated probabilities these days as technology has advanced, after all, we wrote this back in the stone age back in 2008, is that yeah, it's going pretty swimmingly right now. We're getting enormous amount of compute, partially thanks to all the... AI companies, I think we can take some hand me down data centers and play around with. People have been much better than expected at scanning brain tissue in all sorts of weird ways we had no inkling about. So things are happening, but it could of course turn out that actually we're not getting the right information at all here. We might be getting parts of it, but missing the important thing. So nothing will happen. So we're still pretty uncertain. But generally I would imagine that in the few decades, we're definitely going to be able to run human sized brains in computers. And we've got data sets that are probably also going to be human sized. So then the question is, can we interpret that data set and make it run well? And that is an open question. We have some encouraging results for a fruit fly. Which is also kind of funny because we believe that of course you're going to start with the little nematode worm C elegans, 303 neurons. How hard can that be? That can run almost on a pocket calculator. Well, at least a current smartphone can easily do it. But it turned out that it's very hard to get all the data out of that tiny little brain. The strength of the synapses is really hard to measure just because we're too small. So then it turned out that the fruit fly, who's kind of the workhorse to mix metaphors for a lot of biology, Well, various people at the Janelia and Google, they scanned that brain and made a complete connectome. And then a while later, somebody asked an intern, could you try to estimate the strength of the synaptic connections? That was mostly a test to see if they could learn the software that the system was using. And he got a decent result. That is really surprising. I had not expected that because this is kind of a picture of a brain rather than the actual electrochemical squishy stuff. And then another guy used a very simplified neural model and found that some behaviors, very simple behaviors, but still fly behaviors could be emulated in this. So now there is a company which I'm advising that is working on this. And there is a bit of controversy about how much of a fly brain have they actually uploaded. But you certainly get a simulated body of a fruit fly that is kind of walking in a very wobbly way. and it's partially controlled by parts of a fly brain, we get there. Although a fly is not terribly smart. It's about 30 or 50,000 neurons and we got about 90 billion neurons. So plenty of work to expand. Mark Fielding (07:22) That's a big jump. Anders Sandberg (07:23) It's a big job. But one of the things that hit me here was we used a very crude model and got surprisingly good results. It's probably going to turn out that we need a lot of finnicky details to get everything. All the subtle, beautiful details of fruit fly behavior that other fruit flies really care about might be missing. But even the crude model got surprisingly far and that's encouraging. After all, real brains need to function in this messy material world with noise and the cells dying for no good reason and us shaking our heads and getting very rough sensory inputs. They can't be too finicky. If they broke down, if a single molecule was out of a place, we would never be able to do anything. Mark Fielding (08:10) 2064 did you say the year was Jeremy? That's not... Anders Sandberg (08:13) That was in the original paper, it's crept down. My mental model is that, yep, it might be much closer now. Still, Jeremy Gilbertson (08:21) Well there's Anders Sandberg (08:22) think AGI might be getting there before we get to the brain emulation. It's a bit like the tortoise versus the rabbit. So you have a brain emulation tortoise, so it's of plodding along the big microscopes and a lot of simulations. Meanwhile, we have an AI rabbit kind of running around all over the place and sometimes doing really well. and sometimes just stagnating for no good reason. Jeremy Gilbertson (08:45) Mark, check this out. for, so for just Mark Fielding (08:46) Jamie, we, shall we go for it? Jeremy Gilbertson (08:48) real quick, just for

context, 2064, someone out there listening right now, that's 25 years old. When they're our age, that's what we're talking about. So just, it's, it's, it's, it's a bit of a blink. Um, before we train, before we transition, it sounds like you wanted a subject change. just, in a world where we have 8 billion emulated minds, like What does that look like and what does it do for agency and meaning and all like the squishy stuff that we talk about? Anders Sandberg (09:20) Yeah. So it's interesting because from an outside perspective, of course, it might look very alien. Robin Hanson in his book, The Age of EM was talking about these gigantic data center cities. But of course, from the inside, I think most people would want an environment that makes sense to them, which is probably going to be a three-dimensional virtual reality that looks quite a lot like the normal world. But the really interesting thing is the sociological part. So if you're living in a virtual environment, You can of course interact in different ways. You can move around in different ways. You can change your body and you can look like whatever you want. But if you want to be taken seriously at the business meeting, you better have a suit on even if you could look like whatever. And presumably the family structure, you might still want to have your loved ones and you need to negotiate that. If I make a copy of myself or both me and my copy married to my spouse, That's an interesting question and given that my spouse happened to be a lawyer, this is going to be very important. Mark Fielding (10:24) Is this double consciousness? I multiply myself more than once so I can have multiple me's in experimenting in different virtual worlds? Am I uploading my consciousness with that? And since we change from Anders Sandberg (10:36) Well, people have... Mark Fielding (10:39) second to second, the lag between uploading my consciousness and what happens, it's not the same in me. It's a different me. There will never be an exact replica of my consciousness. Anders Sandberg (10:50) Well, the question is whether we are exact replicas of ourselves some few moments ago. And I personally Mark Fielding (10:56) you Anders Sandberg (10:57) think that personal identity is very much of an illusion. It's a good story that makes a lot of sense in a practical way, especially as we're living right now in an uploaded world. It might make much less sense, but you still have your thread of memories. You might still have your goals and your values. And if you're branch out or not, your relationship to the other branches. might depend quite a lot on different takes. So I think that I'm the equivalence class of all sufficiently Anders-like processes. So the Anders consortium, we might just be saying that, yep, we are all Anders's. And I have other friends saying, no, no, no, there can only be one of me. If there is a copy, he's somebody else. And of course, the copy is going to say that too. So I don't

know how we're going to settle this. perhaps highlander style with virtual swords or whatever. It gets interesting because you could frame it in a lot of different ways and bet people are going to handle it in very different ways. So the real question is, of course, what are you going to run your society on and how you organize it. Robin's book is very much kind of a minimum model, which is very much a strong economic competition. Well, I'm kind of somewhat neoliberal. I'm also my mom and my mother-in-law were both union people. So I immediately suggested in your world, I would totally form an upload union. There is kind of a collective bargaining actually is quite powerful and useful. You can organize, you can actually create societal structures. And that's of course what we normally always do. If marriage gets really weird when you can copy, you're going to invent rules. Maybe we're going to be local to different groups. maybe we're going to be a universal rule that we define marriage in particular society in a particular way. Jeremy Gilbertson (12:46) If my brain was scanned right. Mark Fielding (12:46) Maybe there's organic barriers to this. Maybe you have two consciousnesses in superposition in one on the virtual world, one in the real world. Or maybe there's a law that prevents that from ever happening. Anders Sandberg (13:01) In a lot of science fiction stories, they try to prevent this weirdness because it makes storytelling so hard. And I always felt like, wait a minute, this is one of the selling points. I would love to run in parallel. I always believed also that once you branch, you can never rejoin. But actually we seem to be getting better and better merging neural networks. So it might actually be that, yeah, you can go off and do things. So one of you stays at home, one goes off to the moon, comes back. and then you merge the memories. And at that point, the narrative thread gets very weird because you had remembered that weekend from two different places. And some people are going to say, that's totally reasonable. Others are going to say, wait a minute, this is way too wild trip. Mark Fielding (13:46) fascinating because it's in almost Jeremy Gilbertson (13:47) energy. Mark Fielding (13:47) every science fiction story we're kind of obsessed with the notion of it as a species. Jeremy Gilbertson (13:52) I don't know if you're seeing the smoke come out of my ears as we're like thinking, thinking through this. It's, it's so we as a society and you, all of your, lot of your research points back to like, how does this map out and the history of societies and how people organize and what that's going to look like and how that's going to change. We have, so there's three of us in this discussion right now, as we're talking, we may in the future have the ability to create a hundred of ourselves that almost has that have this many worlds interpretation and move through the world in different ways, bringing that stuff back. Are we capable of managing that noise and getting a through line out of all that? Like, how does that how do you think about that from a human condition, our ability to just process all of that? What does that look like? Anders Sandberg (14:44) Yeah, a lot of it also depends on how well you can get along with yourself and other people. So I like to imagine that if there is a dozen Andersons around, we would all very happily form a team or maybe go off and do our research and get along really well. That's what I like to think. I don't know whether that's actually true or whether we're going to rub each other in the wrong way and get really annoyed at each other. And that is, of course, just the easiest part. We're forming organizations that are full of weird pathologies, whether that is endless meetings and the careerists taking over and things and missing the point of why we have made the organization. We have all the traditional pathologies when you have selfish people involved in organizations that are supposed to solve collective problems. But then again, we also get new kinds of solutions. Reputations, for example, they have, of course, been around since forever. all ancient talk about honor, it's all about reputation management, except that these days, reputation management might be more about search engine optimization and having a good LinkedIn profile or whatever works in your subculture. And we also can notice these reputation in other ways. I'm very fond of the Stack Exchange sites, for example, where you get little points for asking good questions. People can upvote you or downvote you depending on how good your question is. And answers similarly get upvoted and downvoted. And suddenly we have a reputation economy. It's slightly addictive to get upvotes. You want to give good answers. So for a long while, this worked really well. Now we have a lot of people jumping in and they want that rush of getting a high score. So they use AI to write answers that sound good, but actually slightly dodgy. So... It's facing a really interesting problem. And I think this is going to happen in all human society. We're creating structures and then somebody is going to try to game them. The best structures adapt to that. And we learn their lessons. Sometimes they're harsh lessons. And then we update it. That's how we're done with the political structures. Okay, we have a complex society. The strong man cannot run everything. He needs to delegate. Gradually, that delegation turns into committees, feudalism. We realized that actually we might want to have buy-in from people. Okay, we need to have a parliament. Actually, the parliament has too much to do. We need to have committees and departments and agencies. And how do we keep them on track? Well, we have auditors. This is of course why traditional fantasy novels are showing in such a simple, easily to understand society. Real medieval feudalists was super complicated and would rarely have made it for good fantasy fiction because nobody can keep track of all the different roles. And similarly, modern society is even more complex, bigger than we can think. But part of that complexity is offloaded. I don't know how to run the society, but there are political scientists, there are politicians, there are a lot of bureaucrats that do that. And now we're going to get the AI counterpart too. And presumably also the AI counterpart of auditors and checks and balances. to avoid that somebody is spamming the system or somebody who's uploaded a copy of himself in a million copies and that way wins the election. Jeremy Gilbertson (18:07) This mark. Mark Fielding (18:07) With just on uploading at the start of all this, you said uploading to a virtual world, but I've got one word for you, humanoids. So why can't we, we'll be able to upload ourselves into organic replicas of ourselves rather than virtual worlds in a, in an internet, whatever that might look like. But there's, could be multiple me's walking around on the earth, not in a virtual world. Anders Sandberg (18:31) Yeah, and I think many of us would like that. After all, it's nice to be in the real world. It has many beautiful advantages. It is also kind of important. After all, he who controls the power supply to the data center controls the data center.