I did just read Accelerando. Why do you ask?

June 11, 2008 7:38 AM

What kind of computing power would be needed to simulate consciousness?

My question to you all is inspired by the news of Roadrunner smashing the supercomputing record. Roadrunner, of course, will simulate the behavior of the weapons in the first fraction of a second during an explosion. We also have supercomputers designed to simulate the global weather systems. All this surely is quite complicated.

And so I wonder: what type of computing power would be required to simulate human consciousness (or, perhaps if we wish to avoid walking before we can run, the consciousness of a lesser mammal/primate? In so doing I mean creating a nearly perfect physical model of every atom, its momentum, etc. of both the subject and it's environment.

Of course, actually determining the composition, location, and momentum of every atom poses its own challenges. This feeds nicely into my second question. What about something more simple: building a physical model of human egg and sperm, or a zygote, plus the surrounding environment and then growing the simulated embryo to maturity?

posted by prunes to religion & philosophy (21 comments total)

No one really knows. Its a software problem first and a hardware problem second. No one has figured out the software part. Arguably, you could say "well, we need x amount of power to simulate a neuron and the human brain has x amount of neurons, so we need x * y." I think that's still pretty flawed. You can simulate the functions of a car in software but it really has very little to do with how an actual car is made. The software and our understanding of consciousness just isnt there to make this guess.
posted by damn dirty ape at 7:48 AM on June 11

...or the consciousness of a lesser mammal/primate

Forget the level of animal. It's the word "consciousness" that is a problem here.

Forget mammals and primates. Stick to, even, a goldfish, and it's still an unsolved problem because we don't even know what consciousness is, so how can we simulate it yet?

Also, is it even possible to "simulate" consciousness without it actually BEING consciousness? These questions are as much philosophy as comp sci.
posted by rokusan at 8:00 AM on June 11

Much discussion on this subject has been produced in the recent Singularity threads on the blue.
posted by ghost of a past number at 8:02 AM on June 11

For the purposes of my question, I'm assuming that achieving a faithful physical model of the entity plus its environment would be akin to having achieved consciousness from the perspective of that entity. Whether that entity is truly conscious is a question for philosophers given that the computer is by necessity deterministic (although I would call the entity conscious.)
posted by prunes at 8:05 AM on June 11

I agree with rokusan. I think a better goal would be "intelligence", for which there exists a simple test.

This paper might be helpful.
posted by mpls2 at 8:09 AM on June 11

Wikipedia claims at least 10 petaflops
posted by uandt at 8:12 AM on June 11

What about something more simple: building a physical model of human egg and sperm, or a zygote, plus the surrounding environment and then growing the simulated embryo to maturity?

That's definitely more complicated than simulating human intelligence/consciousness because we know much less about how a zygote becomes a fully-functioning human than about how the brain works.
posted by mpls2 at 8:13 AM on June 11

First we have to have a computational model of something, then we can simulate it on a supercomputer. Neuroscientists are already running their models on systems, for example Anatomy of a Cortical Simulator gives an overview of an attempt to run a mouse brain simulation on BlueGene/L. The simulation ran for 9 sec in real time for every 1 sec in virtual mouse brain time. Next year's computer will be able to run it closer to real time.

For your question, if we had a "faithful physical model of the entity plus its environment" we could already simulate it on a supercomputer, just not in real-time.
posted by demiurge at 8:15 AM on June 11 [2 favorites]

That's definitely more complicated than simulating human intelligence/consciousness because we know much less about how a zygote becomes a fully-functioning human than about how the brain works.

But the beauty of just getting all atoms and their relative states is that you don't need to know how things work. You just begin the simulation and let a process that was already in motion play out.
posted by prunes at 8:30 AM on June 11

for example Anatomy of a Cortical Simulator gives an overview of an attempt to run a mouse brain simulation on BlueGene/L.

I dont think thats a mouse brain. That's just x random neuron simulations. The structure, development, and position and interconnects of that brain is missing. Not to mention senses and an environment. If the poster wants a HAL-like machine then this method wont produce it. If you did this with billions neurons for a human you'd have an interesting research project, not something out of a Stross book.

This is like playing random notes and calling it a symphony.
posted by damn dirty ape at 8:53 AM on June 11 [1 favorite]

I agree that this isn't like a real mouse brain. But my point was that, given the model, we can simulate it already. The computational power isn't the main problem.
posted by demiurge at 8:58 AM on June 11

Eliza "simulates" human consciousness.
posted by caddis at 10:59 AM on June 11

This article is about a group of people that are taking an interesting approach to that. It discusses how much each part of the brain might take, how much electricity they project to need if/when they have the entire brain simulated, etc.
posted by Nattie at 11:40 AM on June 11

Are you asking how much computer power would it take to accurately predict the checmical reactions in a human brain?

Or asking about simulating consciousness, because that's an entirely different problem. A Simulation generally has two properties: 1) it's a simulation, not reality. They take shortcuts whenever possible. That doesn't mean that they are inaccurate, but you don't need to simulate every atom in a bridge to get a reliable structural assessment, and 2) Simulations are made with some point in mind. This helps direct the shortcuts. Using the bridge example again, for a structural assessment, type of steel used is important, but color of paint isn't. However when you're simulating the bridge for marketing materials, color is a primary concern, what it's made out of is largely unimportant.

So you need to deeply define what you're after. It could change the results bu several orders of magnitude. Do you want a real time brain simulation, something you could scan your brain into and achieve immortality? Or do you just want an ethical way to do skinner box experiments?

(And I'm not getting into the efficiency of programming such a beast. The cleverness of an algorithm can also make orders of magnitude difference in how much computer power you need.)
posted by Ookseer at 11:54 AM on June 11 [1 favorite]

If i'm interpreting your question correctly, you're asking "how much computing power would it take to completely simulate a slice of the universe, oh, say, about the size of a basketball.

The consciousness/intelligence/AI thing is something of a red herring. If your simulation is perfect, it'll support any physical process the real universe can.

So... as damn dirty ape says, it's a software problem not a hardware problem. Until we figure out what the rules of physics are, we can't simulate them. And there are no guarantees that those rules will ever be accessible to us.

And even then... your question doesn't matter. If we only have half the required computing power, we'll just run the simulation at half speed. You can run the simulation on the entire Google cloud or on a 1Ghz Pentium - the end result will be the same, it'll just take longer.
posted by Leon at 12:44 PM on June 11

Read this for a run-through of why people are balking on the idea of replicating consciousness. Really, you just need to read up to the end of "The easy problems and the hard problems." You're arguing about the 'easy problems': why, if we just replicate everything you find in a brain, boom, you'll have consciousness, because you'll be able to process things, etc.

But, we can never know if our replicate is actually experiencing consciousness, if there's something we missed in our Grand Biological Layout. Also, this presumes that biology is destiny, which, eh.

Basically, the problem of consciousness is still so poorly-understood you can't necessarily replicate consciousness at this stage. Replicate the brain's processing power, sure, but consciousness is a whole new can of worms.
posted by flibbertigibbet at 3:28 PM on June 11

I would recommend some of Susan Blackmore's writing on this topic. For example this article in which she examines what she offers a meme-bases theory on what would be required for a machine to have human-like consciousness. To quote "The illusion that we are a conscious self having a stream of experiences is constructed when memes compete for replication by human hosts. Some memes survive by being promoted as personal beliefs, desires, opinions, and possessions, leading to the formation of a memeplex (or selfplex). Any machine capable of imitation would acquire this type of illusion and think it was conscious."

I think what she is saying here is that, if you assume you are in the business of trying to re-create something that is an artfully constructed illusion, you can feel free to simplify and cut corners providing you get the feedback loops right.
posted by rongorongo at 4:14 PM on June 11

For the purposes of my question, I'm assuming that achieving a faithful physical model of the entity plus its environment would be akin to having achieved consciousness from the perspective of that entity.

I think the problem is, a "faithful physical model" can be unbelievably complex. Consider the existence of folding@home -- simulating, to an acceptable degree of accuracy, what goes on in a single protein molecule over the course of a few milliseconds is so complex that we need to harness thousands of computers over a period of years to get reasonable traction on the problem.

I read several years ago about an attempt to simulate some subatomic process -- chemical bonding or some such, occurring billions of times per second -- in which effective simplifications allowed the calculations to be performed in just a few years, whereas a "naive" calculation would have taken something like 10,000 years of processing time.

Simulations are all about simplification -- many, many orders of magnitude of simplification -- and this requires understanding the problem to a high degree. Needless to say, we aren't even close to having this level of understanding of the brain.
posted by bjrubble at 11:20 PM on June 11 [1 favorite]

But, we can never know if our replicate is actually experiencing consciousness, if there's something we missed in our Grand Biological Layout.

I would argue that, if you are simulating a human brain, and you ask it, "what is it like to see the color red," and it doesn't reject the question as meaningless, it's conscious. How would a zombie even process such a question?
posted by bjrubble at 11:34 PM on June 11

I would reject this question as meaningless - I can't think of anything to answer that would convey any sort of information. Am I some kind of zombie, philosophical or otherwise?
posted by ghost of a past number at 11:05 AM on June 12

I think a few simple calculations show that we won't be simulating human brains at the atomic level anytime soon. Atomic-level simulations are much more difficult than the neuron-level simulations discussed above, because there are many many many more atoms than neurons.

Some websites say that the human brain is about 500g, others that it's about 1500g. Let's go with the lower bound, 500g. 500g is 2.5x10²⁵ carbon atoms. (if the brain is primarily water, then it would contain about twice that many atoms)

Let's make up some other rules that are likely to gravely underestimate the difficulty of simulating a physical system at the atomic level:

• Each atom is affected by at most 10,000 other atoms, and the identity of those 10,000 atoms can be determined effortlessly
  
• Each atom-atom interaction can be simulated in 1 floating-point computation
  
• A timestep of 1 nanosecond will give an acceptable simulation quality

That means that the computing power to run one second of atom-level brain simulation is: 2.5x10²⁵ * 10000 * 1x10⁹ = 2.5x10³⁷ FLOPs.

A 1000 petaflop computer (that's 2000 times faster than the fastest supercomputer listed on Wikipedia today, the IBM Blue Gene/L with 478.2 TFLOP-- 478 trillion floating-point computations per second) makes about 3x10²⁵ floating-point calculations in a year. At this rate, it would take 8.3x10¹¹ years to run the 2.5x10³⁷ calculations needed to simulate the atoms in a human brain for one second.
posted by jepler at 7:36 AM on June 13