The idea is that human-like AI is intrinsically more safe and can be used to control AI development
As there are no visible ways to create safe self-improving superintelligence, but it is looming, we probably need temporary ways to prevent its creation. The only way to prevent it, is to create special AI, which is able to control and monitor all places in the world. The idea has been suggested by Goertzel in form of AI Nanny, but his Nanny is still superintelligent and not easy to control, as was shown by Bensinger at al. We explore here the ways to create the safest and simplest form of AI, which may work as AI Nanny. Such AI system will be enough to solve most problems, which we expect the AI will solve, including control of robotics, acceleration of the medical research, but will present less risk, as it will be less different from humans. As AI police, it will work as operation system for most computers, producing world surveillance system, which will be able to envision and stop any potential terrorists and bad actors in advance. As uploading technology is lagging, and neuromorphic AI is intrinsically dangerous, the most plausible way to human-based AI Nanny is either functional model of the human mind or a Narrow-AI empowered group of people.
Yes, unfortunately I think you are right avturchin. I have come to similar conclusions myself. See my comment elsewhere on this thread for some of my thoughts.
Edit: I read your article, and we have a lot of agreements, but also some important disagreements. I think the main disagreement is that I spent years studying neuroscience and thinking hard about intelligence amplification via brain-computer interfaces and genetic enhancement of adult brains, and also about brain preservation and uploading. For the past 7 years I’ve instead been studying machine learning with my thoughts focused on what will or won’t lead to AGI. I’m pretty convinced that we’re technologically a LOT closer to a successfully dangerously recursive self-improving AGI than we are to a functionally useful human brain emulation (much less an accurate whole brain scan).
Here are some recent thoughts I’ve written down about human brain emulation:
Our best attempts at high accuracy partial brain emulation so far have been very computationally inefficient. You can accept a loss of some detail of the emulation, and run a more streamlined simulation. This gives you many orders of magnitude speed up and removes a lot of need for accuracy of input detail. There’s lots of work on this that has been done. But the open question is, how much simplification is the correct amount of simplification? Will the resulting emulated brain be similar enough to the human you scanned the brain of that you will trust that emulation with your life? With the fate of all humanity? Would a team of such emulations be more trustworthy than copies of just the best one?
In the extreme of simplification you no longer have something which is clearly the brain you scanned, you have a system of interconnected neural networks heavily inspired by the circuitry of the human brain, with some vague initialization parameters suggested by the true brain scan. At this point, I wouldn’t even call it an emulation, more like a ‘brain-inspired AGI’. If you are going to go this route, then you might as well skip the messy physical aspects of step 1, and the complicated data analysis of step 2, and just jump straight to working with the already processed data publicly available. You won’t be likely to get a specific human out the end of this process even if you put a specific human in the beginning. Skipping the hard parts and glossing over the details is definitely the fastest easiest route to a ‘brain-like AGI’. It looses a big chunk of the value, which was that you knew and trusted a particular human and want an accurate emulation of that human to be entrusted with helping society. Yet, there is also still a lot of value you retain. This ‘brain-inspired AGI’ will share a lot of similarities with the brain and make us able to use the suite of interpretability tools that neuroscience has come up with for mammalian brains. This would be a huge boon for interpretability work. There would be control-problem gains, since you’d have a highly modular system with known functions of the various modules and the ability to tune the activity levels of those modules to get desired changes in behavioral output. There would be predictability gains, since our intuitions and study of human behavior would better carry over to a computational system so similar to the human brain. In contrast, our human-behavior-prediction instincts seem to mainly lead us astray when we attempt to apply them to the very non-brain-like systems of modern SotA ml systems like large language models. Many of the groups who have thought deeply about this strategic landscape have settled on trying for this ‘brain-like AGI’ strategy. The fact that it’s so much faster and easier (still not as fast or easy as just scaling/improving SotA mainstream ml) than the slow highly-accurate brain scan-and-emulation path means that people following the low-accuracy high-efficiency ‘brain-inspired AGI’ path will almost certainly have a functioning system many years before the slow path. And it will run many orders of magnitude faster. So, if we could conquer the Molochian social challenge of a race to the bottom and get society to delay making unbrainlike-AGI, then we certainly ought to pursue the slower but safer path of accurate human brain emulation.
I don’t think we have that option though. I think the race is on whether we like it or not, and there is little hope of using government to effectively stop the race with regulation. Slow it down a bit, that I think is feasible, but stop? No. Slow it down enough for the slow-but-safe path to have a chance of finishing in time? Probably not even that.
If this viewpoint is correct, then we still might have a hope of using regulatory slowdown plus the lossy-approximation of ‘brain-like AGI’ to get to a safer (more understandable and controllable) form of AGI. This is the strategy embraced by the Conjecture post about CogEms, by some leading Japanese AI researchers, by Astera:Obelisk (funded by Jed McCaleb), by Generally Intelligent (funded in part by Jed McCaleb), by some of the researchers at DeepMind (it’s not the organization-as-a-whole’s primary focus, but it is the hope/desire of some individuals within the organization), and a long list of academic researchers straddling the borderlands of neuroscience and machine learning.
I think it’s a good bet, and it was what I endorsed as of Fall 2022. Then I spent more time researching timelines and looking into what remaining obstacles I think must be cleared for mainstream ml to get to dangerously powerful AGI… and I decided that there was a big risk that we wouldn’t have time even for the fast lossy-approximation of ‘brain-like AGI’. This is why I responded to Conjecture’s recent post about CogEms with a concerned comment of this apparent strategic oversight on their part: https://www.lesswrong.com/posts/ngEvKav9w57XrGQnb/cognitive-emulation-a-naive-ai-safety-proposal?commentId=tA838GyENzGtWWARh
....
A terminology note: You coin the term human-like AI, others have used the term brain-like AI or brain-like AGI or CogEms (cognitive emulations). I believe all these terms point at a similar cluster of ideas, with somewhat different emphasis.
I don’t really have a preference amongst them, I just want to point out to interested readers that in order to come up with a broader range of discussion on these topics they should consider doing web searches for each of these terms.
A nice summary statement from avturchin’s article, which I recommend others interested in this subject read:
‘We explored two approaches where AI exists and doesn’t exist simultaneously: Human uploads based AI and secret intelligence organization of a superpower based AI (NSI-AI). The first is more technically difficult, but better, and NSI-AI is more realistic in near term, but seems to be less moral, as secret services are known not be not aligned with general human values, and will reach global domination probably via illegal covert operation including false flags attacks and terrorism.’
I hate to do it, but can’t resist the urge to add a link to my article First human upload as AI Nanny.
The idea is that human-like AI is intrinsically more safe and can be used to control AI development
As there are no visible ways to create safe self-improving superintelligence, but it is looming, we probably need temporary ways to prevent its creation. The only way to prevent it, is to create special AI, which is able to control and monitor all places in the world. The idea has been suggested by Goertzel in form of AI Nanny, but his Nanny is still superintelligent and not easy to control, as was shown by Bensinger at al. We explore here the ways to create the safest and simplest form of AI, which may work as AI Nanny. Such AI system will be enough to solve most problems, which we expect the AI will solve, including control of robotics, acceleration of the medical research, but will present less risk, as it will be less different from humans. As AI police, it will work as operation system for most computers, producing world surveillance system, which will be able to envision and stop any potential terrorists and bad actors in advance. As uploading technology is lagging, and neuromorphic AI is intrinsically dangerous, the most plausible way to human-based AI Nanny is either functional model of the human mind or a Narrow-AI empowered group of people.
Yes, unfortunately I think you are right avturchin. I have come to similar conclusions myself. See my comment elsewhere on this thread for some of my thoughts.
Edit: I read your article, and we have a lot of agreements, but also some important disagreements. I think the main disagreement is that I spent years studying neuroscience and thinking hard about intelligence amplification via brain-computer interfaces and genetic enhancement of adult brains, and also about brain preservation and uploading. For the past 7 years I’ve instead been studying machine learning with my thoughts focused on what will or won’t lead to AGI. I’m pretty convinced that we’re technologically a LOT closer to a successfully dangerously recursive self-improving AGI than we are to a functionally useful human brain emulation (much less an accurate whole brain scan).
Here are some recent thoughts I’ve written down about human brain emulation:
Our best attempts at high accuracy partial brain emulation so far have been very computationally inefficient. You can accept a loss of some detail of the emulation, and run a more streamlined simulation. This gives you many orders of magnitude speed up and removes a lot of need for accuracy of input detail. There’s lots of work on this that has been done. But the open question is, how much simplification is the correct amount of simplification? Will the resulting emulated brain be similar enough to the human you scanned the brain of that you will trust that emulation with your life? With the fate of all humanity? Would a team of such emulations be more trustworthy than copies of just the best one?
In the extreme of simplification you no longer have something which is clearly the brain you scanned, you have a system of interconnected neural networks heavily inspired by the circuitry of the human brain, with some vague initialization parameters suggested by the true brain scan. At this point, I wouldn’t even call it an emulation, more like a ‘brain-inspired AGI’. If you are going to go this route, then you might as well skip the messy physical aspects of step 1, and the complicated data analysis of step 2, and just jump straight to working with the already processed data publicly available. You won’t be likely to get a specific human out the end of this process even if you put a specific human in the beginning. Skipping the hard parts and glossing over the details is definitely the fastest easiest route to a ‘brain-like AGI’. It looses a big chunk of the value, which was that you knew and trusted a particular human and want an accurate emulation of that human to be entrusted with helping society. Yet, there is also still a lot of value you retain. This ‘brain-inspired AGI’ will share a lot of similarities with the brain and make us able to use the suite of interpretability tools that neuroscience has come up with for mammalian brains. This would be a huge boon for interpretability work. There would be control-problem gains, since you’d have a highly modular system with known functions of the various modules and the ability to tune the activity levels of those modules to get desired changes in behavioral output. There would be predictability gains, since our intuitions and study of human behavior would better carry over to a computational system so similar to the human brain. In contrast, our human-behavior-prediction instincts seem to mainly lead us astray when we attempt to apply them to the very non-brain-like systems of modern SotA ml systems like large language models. Many of the groups who have thought deeply about this strategic landscape have settled on trying for this ‘brain-like AGI’ strategy. The fact that it’s so much faster and easier (still not as fast or easy as just scaling/improving SotA mainstream ml) than the slow highly-accurate brain scan-and-emulation path means that people following the low-accuracy high-efficiency ‘brain-inspired AGI’ path will almost certainly have a functioning system many years before the slow path. And it will run many orders of magnitude faster. So, if we could conquer the Molochian social challenge of a race to the bottom and get society to delay making unbrainlike-AGI, then we certainly ought to pursue the slower but safer path of accurate human brain emulation.
I don’t think we have that option though. I think the race is on whether we like it or not, and there is little hope of using government to effectively stop the race with regulation. Slow it down a bit, that I think is feasible, but stop? No. Slow it down enough for the slow-but-safe path to have a chance of finishing in time? Probably not even that.
If this viewpoint is correct, then we still might have a hope of using regulatory slowdown plus the lossy-approximation of ‘brain-like AGI’ to get to a safer (more understandable and controllable) form of AGI. This is the strategy embraced by the Conjecture post about CogEms, by some leading Japanese AI researchers, by Astera:Obelisk (funded by Jed McCaleb), by Generally Intelligent (funded in part by Jed McCaleb), by some of the researchers at DeepMind (it’s not the organization-as-a-whole’s primary focus, but it is the hope/desire of some individuals within the organization), and a long list of academic researchers straddling the borderlands of neuroscience and machine learning.
I think it’s a good bet, and it was what I endorsed as of Fall 2022. Then I spent more time researching timelines and looking into what remaining obstacles I think must be cleared for mainstream ml to get to dangerously powerful AGI… and I decided that there was a big risk that we wouldn’t have time even for the fast lossy-approximation of ‘brain-like AGI’. This is why I responded to Conjecture’s recent post about CogEms with a concerned comment of this apparent strategic oversight on their part: https://www.lesswrong.com/posts/ngEvKav9w57XrGQnb/cognitive-emulation-a-naive-ai-safety-proposal?commentId=tA838GyENzGtWWARh
....
A terminology note: You coin the term human-like AI, others have used the term brain-like AI or brain-like AGI or CogEms (cognitive emulations). I believe all these terms point at a similar cluster of ideas, with somewhat different emphasis.
I don’t really have a preference amongst them, I just want to point out to interested readers that in order to come up with a broader range of discussion on these topics they should consider doing web searches for each of these terms.
A nice summary statement from avturchin’s article, which I recommend others interested in this subject read:
‘We explored two approaches where AI exists and doesn’t exist simultaneously: Human uploads based AI and secret intelligence organization of a superpower based AI (NSI-AI). The first is more technically difficult, but better, and NSI-AI is more realistic in near term, but seems to be less moral, as secret services are known not be not aligned with general human values, and will reach global domination probably via illegal covert operation including false flags attacks and terrorism.’