I think we need to separate the concept of whole brain emulation, from that of biology-inspired human-like AI. This actually looks pretty bad for Robin Hanson’s singularity hypothesis, where the first emulations to perfectly emulate existing humans suddenly make the cost of labor drop dramatically. If this research pans out, then we could have a “soft takeoff”, where AI slowly catches up to us, and slowly overtakes us.
CNRG_UWaterloo, regarding mind uploads:
Being able to simulate a particular person’s brain is incredibly far away. There aren’t any particularly good ideas as to how we might be able to reasonably read out that sort of information from a person’s brain. That said, there are also lots of uses that a repressive state would have for any intelligent system (think of automatically scanning all surveillence camera footage). But, you don’t want a realistic model of the brain to do that—it’s get bored exactly as fast as people do.
So we should expect machine labor to gradually replace human labor, exactly as it has since the beginning of the industrial revolution, as more and more capabilities are added, with “whole brain emulation” being one of the last features needed to make machines with all the capabilities of humans (if this step is even necessary). It’s possible, of course, that we could wind up in a situation where the “last piece of the puzzle” turns out to be hugely important, but I don’t see any particular reason to think that will happen.
Robin’s economic model for growth explosion with AI uses a continuum of tasks for automation. The idea is that as you automate more tasks, those tasks are done very efficiently, but the remaining ones become bottlenecks, making up more of GDP and limiting growth. Think Baumol’s cost disease: as our manufacturing productivity has increased, economic growth winds up more limited by productivity improvements in service sectors like health care and education, computer programming, and science that have been resistant to automation.
As you eliminate the last bottlenecks where human labor is important, you can speed the whole process of growth up to computer-scales, rather than being held back by the human weakest link. One can make an analogy to Amdahl’s law: if you can parallelize 50% of a problem you can double your speed, at 90% you can get a 10x speedup, at 99% a 100x speedup, and as you approach 100% you can rush up to other limits.
Similarly, smooth human progress in automating elements of AI development (which then proceed very quickly, bottlenecked by the human-limited elements) could produce stark speedup as automation approaches 100%.
However, such developments do make a world dominated by whole brain emulations rather than neuromorphic AI less likely, even if they still allow an intelligence explosion.
I think we need to separate the concept of whole brain emulation, from that of biology-inspired human-like AI.
This seems completely true. Part of the problem is that the media hype surrounding this stuff drops lines like this:
Spaun can recognize numbers, remember lists and write them down. It even passes some basic aspects of an IQ test, the team reports in the journal Science.… the simplified model of the brain, which took a year to build, captures many aspects of neuroanatomy, neurophysiology and psychological behaviour… They say Spaun can shift from task to task, “just like the human brain,” recognizing an object one moment and memorizing a list of numbers the next. And like humans, Spaun is better at remembering numbers at the beginning and end of the list than the ones in the middle. Spaun’s cognition and behaviour is very basic, but it can learn patterns it has never seen before and use that knowledge to figure out the best answer to a question. “So it does learn,” says Eliasmith.
Basically: to explain this stuff to normal readers, writers anthropomorphize the hell out of the project and you end up with words like ‘intuition’ and ‘understanding’ and ‘learn’ and ‘remember’ - which make the articles both sexier and way more misleading. The same thing happened with IBM’s project and, to my understanding, the Blue Blain Project as well.
Actually I’m not sure if any of that is a problem. Spaun is quite literally “anthropomorphic”—modeled after a human brain. So it’s not much of a stretch to say that it learns and understands the way a human does. I was just pointing out that the more progress we make on human-like AIs, without progress on brain scanning, the less likely a Hansonian singularity (dominated by ems of former humans) becomes. If Spaun as it is now really does work “just like a human”, then building a human-level AI is just a matter of speeding it up. So by the time we have computers capable of supporting a human mind upload, we’ll already have computer programs at least as smart as humans, which learn their knowledge on their own, with no need for a knowledge transplant from a human.
If Spaun as it is now really does work “just like a human”, then building a human-level AI is just a matter of speeding it up.
As I explained in this comment, Spaun can only perform tasks that are specifically and manually programmed into it. It is very, very far from working just like a human. It’s definitely incapable of learning new skills or concepts, for example. What the original article said was:
They say Spaun can shift from task to task, “just like the human brain,” recognizing an object one moment and memorizing a list of numbers the next.
Well gosh, my desktop computer can also shift from task to task, just like the human brain, mining bitcoins one moment and decoding MPEGs the next. This is either PR or (perhaps unintentional) hype by the reporter, saying something that is literally true but gives the impression of much greater accomplishment.
(Which isn’t to say that Spaun might not continue with or inspire further more interesting developments, but a lot of people seem to be overly impressed with it in its current state.)
Hi, it’s Terry again (one of the researchers on the project)
The interesting thing (for me) isn’t that it can shift from task to task, but that it can shift from task to task just like the human brain. In other words, we’re showing how a realistic neural system can shift between tasks. That’s something that’s not found in other neural models, where you tend to either have it do one task or you have external (non-neural) systems modify the model for different tasks. We’re showing a way of doing that selecting routing and control in an entirely neural way that maps nicely onto the cortex-basal ganglia-thalamus loop.
Oh, and, since we constrain the model with a bunch of physical parameters influencing the timing of the system (reabsorption of neurotransmitter, mostly), we can also look at how long it takes the system to switch tasks, and compare that to human brains. It’s these sorts of comparisons that let us use this sort of model to test hypotheses about what different parts of the brain are doing.
According to my understanding, Spaun is only able to shift between a fixed set of tasks, and according to a fixed algorithm (if the first two inputs are “A1”, route the information one way so that it ends up doing one task, and if the first inputs are “A2“, route the information another way, etc.) that was manually designed. You haven’t explained yet (or emulated) how human brains are able to switch fluidly between an ever changing set of possibe tasks, and without having to be prompted by specific codes such as “A1” and “A2”.
If my understanding is correct, I think a clearer and fairer description of your accomplishment might be that you’ve demonstrated task shifting “on a (simulated) neural substrate that is structually similar to the human brain”, rather than task shifting “just like the human brain”.
Yup, I’d say that’s a fair way of expressing it, although I think we take “neural substrate that is structurally similar to the human brain” much more seriously than other people that use phrases like that. It’s a similar enough substrate that if fixes a lot of our parameter values for us, leaving us less open to “fiddle with parameters until it works”.
We’ve also tried to make sure to highlight that it can’t learn new tasks, so it’s not able to work in the fluid domains people do. It also doesn’t have any intrinsic motivation to do that switching.
Interestingly, there are starting to be good non-neural theories of human task switching (e.g. [http://act-r.psy.cmu.edu/publications/pubinfo.php?id=831] ). These are exactly the sorts of theories we want to take a close look at and see how they could be realistically implemented in spiking neurons.
I think we need to separate the concept of whole brain emulation, from that of biology-inspired human-like AI. This actually looks pretty bad for Robin Hanson’s singularity hypothesis, where the first emulations to perfectly emulate existing humans suddenly make the cost of labor drop dramatically. If this research pans out, then we could have a “soft takeoff”, where AI slowly catches up to us, and slowly overtakes us.
CNRG_UWaterloo, regarding mind uploads:
So we should expect machine labor to gradually replace human labor, exactly as it has since the beginning of the industrial revolution, as more and more capabilities are added, with “whole brain emulation” being one of the last features needed to make machines with all the capabilities of humans (if this step is even necessary). It’s possible, of course, that we could wind up in a situation where the “last piece of the puzzle” turns out to be hugely important, but I don’t see any particular reason to think that will happen.
Robin’s economic model for growth explosion with AI uses a continuum of tasks for automation. The idea is that as you automate more tasks, those tasks are done very efficiently, but the remaining ones become bottlenecks, making up more of GDP and limiting growth. Think Baumol’s cost disease: as our manufacturing productivity has increased, economic growth winds up more limited by productivity improvements in service sectors like health care and education, computer programming, and science that have been resistant to automation.
As you eliminate the last bottlenecks where human labor is important, you can speed the whole process of growth up to computer-scales, rather than being held back by the human weakest link. One can make an analogy to Amdahl’s law: if you can parallelize 50% of a problem you can double your speed, at 90% you can get a 10x speedup, at 99% a 100x speedup, and as you approach 100% you can rush up to other limits.
Similarly, smooth human progress in automating elements of AI development (which then proceed very quickly, bottlenecked by the human-limited elements) could produce stark speedup as automation approaches 100%.
However, such developments do make a world dominated by whole brain emulations rather than neuromorphic AI less likely, even if they still allow an intelligence explosion.
This seems completely true. Part of the problem is that the media hype surrounding this stuff drops lines like this:
Basically: to explain this stuff to normal readers, writers anthropomorphize the hell out of the project and you end up with words like ‘intuition’ and ‘understanding’ and ‘learn’ and ‘remember’ - which make the articles both sexier and way more misleading. The same thing happened with IBM’s project and, to my understanding, the Blue Blain Project as well.
Actually I’m not sure if any of that is a problem. Spaun is quite literally “anthropomorphic”—modeled after a human brain. So it’s not much of a stretch to say that it learns and understands the way a human does. I was just pointing out that the more progress we make on human-like AIs, without progress on brain scanning, the less likely a Hansonian singularity (dominated by ems of former humans) becomes. If Spaun as it is now really does work “just like a human”, then building a human-level AI is just a matter of speeding it up. So by the time we have computers capable of supporting a human mind upload, we’ll already have computer programs at least as smart as humans, which learn their knowledge on their own, with no need for a knowledge transplant from a human.
As I explained in this comment, Spaun can only perform tasks that are specifically and manually programmed into it. It is very, very far from working just like a human. It’s definitely incapable of learning new skills or concepts, for example. What the original article said was:
Well gosh, my desktop computer can also shift from task to task, just like the human brain, mining bitcoins one moment and decoding MPEGs the next. This is either PR or (perhaps unintentional) hype by the reporter, saying something that is literally true but gives the impression of much greater accomplishment.
(Which isn’t to say that Spaun might not continue with or inspire further more interesting developments, but a lot of people seem to be overly impressed with it in its current state.)
Hi, it’s Terry again (one of the researchers on the project)
The interesting thing (for me) isn’t that it can shift from task to task, but that it can shift from task to task just like the human brain. In other words, we’re showing how a realistic neural system can shift between tasks. That’s something that’s not found in other neural models, where you tend to either have it do one task or you have external (non-neural) systems modify the model for different tasks. We’re showing a way of doing that selecting routing and control in an entirely neural way that maps nicely onto the cortex-basal ganglia-thalamus loop.
Oh, and, since we constrain the model with a bunch of physical parameters influencing the timing of the system (reabsorption of neurotransmitter, mostly), we can also look at how long it takes the system to switch tasks, and compare that to human brains. It’s these sorts of comparisons that let us use this sort of model to test hypotheses about what different parts of the brain are doing.
According to my understanding, Spaun is only able to shift between a fixed set of tasks, and according to a fixed algorithm (if the first two inputs are “A1”, route the information one way so that it ends up doing one task, and if the first inputs are “A2“, route the information another way, etc.) that was manually designed. You haven’t explained yet (or emulated) how human brains are able to switch fluidly between an ever changing set of possibe tasks, and without having to be prompted by specific codes such as “A1” and “A2”.
If my understanding is correct, I think a clearer and fairer description of your accomplishment might be that you’ve demonstrated task shifting “on a (simulated) neural substrate that is structually similar to the human brain”, rather than task shifting “just like the human brain”.
Yup, I’d say that’s a fair way of expressing it, although I think we take “neural substrate that is structurally similar to the human brain” much more seriously than other people that use phrases like that. It’s a similar enough substrate that if fixes a lot of our parameter values for us, leaving us less open to “fiddle with parameters until it works”.
We’ve also tried to make sure to highlight that it can’t learn new tasks, so it’s not able to work in the fluid domains people do. It also doesn’t have any intrinsic motivation to do that switching.
Interestingly, there are starting to be good non-neural theories of human task switching (e.g. [http://act-r.psy.cmu.edu/publications/pubinfo.php?id=831] ). These are exactly the sorts of theories we want to take a close look at and see how they could be realistically implemented in spiking neurons.