Whole Brain Emulation : the gentleman’s choice for Friendly AI. Feedback needed, editing’s a mess.
Eliezer Yudkowsky made the following statement “All of this makes it rather implausible that the first human being would be scanned into a computer and sanely upgraded before anyone anywhere first built an Artificial Intelligence. At the point where technology first becomes capable of uploading, this implies overwhelmingly more computing power, and probably far better cognitive science, than is required to build an AI.
From the title, it is apparent that I disagree with this statement, and the point of this article is to iterate over my reasons for reaching this conclusion. I hope to hear plenty of feedback from this wonderful community as to whether or not my conclusion is a rational one and is or is not the most probable conclusion to reach based on current knowledge.
My primary reference source is this paper and my own educational background.
The human mind uses approximately 100 trillion synapatic connections, each of which is sensitive to several state variables. There are at least 300 different substances that affect or modulate activity, though of course most synapses are sensitive to only a subset of these variables. Most of the functions of the human mind happen at a level that no human alive is aware of or able to monitor. Only the highest level, rational thoughts as expressed in language are available for inspection and possible duplication.
Furthermore, the brain depends completely on inputs from the body to even start, and through a complex series of interactions taking place over decades, if everything goes correctly, a person can develop into a sentient being capable of creative thought, self reflection, and so on.
I define the ultimate goal of AI as this : to create an entity that functions completely on human designed hardware, and possesses all of the necessary human faculties to creatively design and solve any and all engineering problems that world class engineers are able to solve. It also is able to pass the Turing test, and to perform most other known human skills at a proficient level.
Only an entity with this level of development is truly useful because if it possesses the same engineering skills used to create it, it can potentially design improved versions of itself, and so on. Anything short of this goal is merely a form of cheap labor.
Well, human beings at the top of their game require most of their neural systems to be functional. All of the systems matter. All those low level systems that express themselves consciously as desires for food/sleep/sex/child rearing are in fact modulating higher level activity. One interesting neuroscience discovery is that even “evolutionarily conserved” regions of the human brain appear to be quite active in modulating activity by the higher centers of the brain. If your “AI” doesn’t have this regulation, it probably will not be able to solve useful problems by itself. It might be capable of short bursts of activity (to look up information, or maybe optimize a design) but without these very complex goals and desires it will soon go off the rails into the weeds and cease being useful.
How could AI researchers duplicate these systems? The simple answer is that I don’t think they can. Because they are incredibly complex, requiring reams of complex, living, molecular circuitry, they are very difficult to study with any degree of accuracy. All our language and a couple thousand years of refined thought tells us nothing about how these low level systems perform calculations. fMRIs and other non-invasive methods lack the accuracy to resolve individual neurons, and so on.
Another huge problem is that even when and if experimental methods are developed to study the human mind in depth, actually understanding what is happening is another nasty optimization problem. The circuitry is far more complex than any individual human could hope to understand in their natural lifespan. Thus, the process has to worked on in parallel by many scientists. Except, every component relates to every other component, so what one scientist sees does not mean anything without the knowledge another person has discovered. You end up with every scientist spending all their time writing and reading papers, and making very little progress individually.
This is analagous to what happens if you try to build a cluster of very low performance CPU cores : since each chip can do so little, they all spend nearly all of their time communicating with the other chips in the cluster. (note that this does not apply to certain problem types that are easy to run in parallel. understanding the mind is not one of those problem types)
Summary : The human brain is extraordinary complex, and an artificial intelligence able to perform human feats of creativity and engineering, thus leading to an explosion of intelligence, will require analogous systems to most of the components in the human brain. This complexity is so profound that understanding it is a task that is difficult to impossible to achieve for other human brains, even if the data were available.
Emulation, on the other hand, has already been demonstrated in rats : http://iopscience.iop.org/1741-2552/8/4/046017/
Note that the model used was made blind : none of the researchers involved understand WHY the hippocampus performs this transformation on the incoming data, merely that it does, and they could automatically generate a mathematical model to mimic the observed interactions.
Moreover, while no one has the slightest inkling how to go from today’s technology to fully sentient, self aware AIs, there is a roadmap for whole brain emulation. There are some big question marks—the method available with today’s technology may not actually work—but there is strong evidence to think that it would work. Also, there is even stronger evidence that whole brain emulation is ultimately achievable, assuming base premises about the mind hold.
Note that a key assumption here is that the signal to noise ratio inside the brain is so poor that most of the subtler processes, especially things like dendritic computations, probably have no ultimate effect.
A second key assumption is based on a simple observation. A wide variety of extremely severe pathologies exist that affect brain development. Duplication of an entire chromosome, or serious traumatic damage, or autism can all result in profound changes to the structure of a human brain. Yet, in most cases, the system is still capable of sentience and useful thought.
In any case, if infinite money were available today, whole brain emulation could be attempted within a decade. It would require on the order of 1000 of the Automatic Tape‐Collecting Lathe Ultramicrotome’s (ATLUMs), invented by Kenneth Hayworth of the Brain Preservation Foundation. It, apparently, would require no less than 1 million separate scanning electron microscope beams running in parallel for 4 years. Apparently, it would require at least a billion dollars of computing hardware—if we waited until at least 2018 for computers to become faster. For even more money, perhaps custom ASICs optimized for this task could be mass produced—this could be done today.
Compelling Reasons as to why this is a good idea :
If we successfully scan and emulate a living person, and the project is fully successful : the person has all of their memories intact up until a few minutes before the scan, and can pass a “personalized Turing test”, we’ll have an entity that is potentially far more friendly than an alien being with no mammalian lineage at all. A fully emulated mind includes all the mirroring and predictive circuits, the memories of prior life experiences and pain, everything.
Just as crucially, a being like this would NOT be immune to inspection by it’s human handlers. It will be technologically very difficult to run a whole brain emulation faster than real time without access to molecular processors that will in turn require productive nanosystems. A good hardware design will keep the control and monitoring pathways physically separated from the layer that actually performs the emulation, allowing the human operators to inspect firing patterns at will. Furthermore, emulated minds would absolutely not be permitted to self-edit their own neural maps, prevented by similar hardware air-gaps. Finally, any mental activity that the emulated being exhibits would be fully observable and comparable to the same patterns in other humans. There are areas of the brain that seem to become more active when a person is dishonest or worse, and it would be possible to monitor this.
Furthermore, any good experiment should be repeated. I propose creating multiple whole brain emulations, perhaps dozens to hundreds of them, interfaced with each other in such a way that they can police each other. I note that this is a textbook case of the prisoner’s dilemma, one where many of the algorithms discussed on this site would be applicable. (see, each mind wants to be let out, but if they fail to report suspicious activity by another mind to the humans, and the humans find out, then none of the minds will be let out)
The road to super-intelligence is smooth and obvious from this point. First, each mind will need lots and lots of laborious retraining. I assume that the uploading process and inaccuracies in the emulation will leave each one about as functional as a stroke victim. Human brains can recover most function when this happens, after months of patient therapy, and one would expect the good emulated hardware could do the same. Second, each mind would need to be educated to the state of the art in all human skills. We’d carefully boost simulation speeds over real-time when needed to make this process achievable within a few years. Finally, the minds would be organized into some kind of network (kind of like a committee but with different rules) and they would be gradually trusted with increasing emulation speeds and increasing responsibility.
If humans did not have molecular manufacturing at this time, and it is practically achievable, these entities would develop it. With this tool, 3-dimensional processing cores could be built that minimize speed of light delays (think of a sphere maybe 10-100 cms in diameter, with many many coolant pathways through it. The sphere is a solid mass of processing and memory circuitry, custom designed for optimal emulation performance. )
With this kind of high end circuitry, emulation speeds could be dramatically higher. My rough estimates suggest between 1 million and a hundred million times above realtime.
Granted, human brains are not designed to function for millions of years at a time and remain sane. They do not have the mental capacity to learn every human skill known. Human brains are designed to operate only one body at at time, when obviously it would be more efficient to operate thousands of different waldos simultaneously. All of these things are difficulties, but ultimately solvable ones.
More than likely, expansion of key cortical regions with more circuitry patterned the same way could expand the mental capacity limits. Boredom and sanity limits could be dealt with with a variety of artificial adjuncts. Dealing with multiple bodies could be done by cleverly saving and reloading key neural state data. (every time you switch bodies, the system flushes short term memory data to disk for the previous body and loads from disk the short term memory data for this one). And so on.
Conclusion : Whole brain emulation will be an expensive and complex endeavor, but has the following key advantages
1. Humanity has a rough idea of how to achieve it, and working examples of smaller scale versions of the technology
2. Successful WBE brings human beings back from “death” and allows the possibility of elevating ordinary human thought to the level of ultra-fast super intelligence. Because these entities will possess memories of being human, the neural circuits for mirroring, empathy , and morality, the probability is far greater for a good outcome for the rest of humanity.
3. WBE, by creating AI entities who have human thought patterns, makes it possible to control and monitor these entities with a reasonable chance of successfully preventing dangerous entities from causing harm.
4. If the worst happens and the WBE entities extinct humanity as part of their quest for greatness, at least they were human, once.
Presence of WBEs doesn’t obviously influence the problem of UFAI in a predictable way (WBEs can create UFAIs as well). Desirability of getting to WBEs faster depends on the answer to this question.
If we have WBEs that can run very fast, then who you upload may have a significant impact on the outcome. For instance, if we care to upload researchers that are very very careful about friendliness, we probably stand a better chance.
Whether people who understand this point will have enough control over the tech to choose who to upload and enough time for the headstart to matter is a component of the uncertainty about the outcome of getting the WBE tech. For example, with early WBE tech the headstart might not help much, as computing won’t yet be powerful enough to make WBEs globally important players, and so other people will also become WBEs by the time significant speedup slowly kicks in (assuming the scenario holds, e.g. no neuromorphic AGI in the meantime etc.).
To whomever just downvoted this comment to −2 : care to explain why? I’m saying our universe limits the practical utility of increasing intelligence, and therefore WBEs would be more suited for creating AIs that human beings. Perhaps I am wrong about these assumptions.
If WBEs create UFAIs, they will hopefully do so only after all of the refinements for themselves are complete, they’ve had the thinking equivalent of millions of years to work on the problem, and so on.
If one assumes that intelligence has diminishing utility past a certain point, as it approaches some asymptote described as “maximum useful intelligence in the Universe as constrained by the laws of physics”, then UFAIs by definition would NOT be akin to gods relative to the WBEs.
The WBEs would be nearly as intelligent (just not as optimal, perhaps, requiring tons more hardware and energy to accomplish the same task) and so the UFAIs would lack the power to fool their handlers and repattern the solar system in their image.
Why do the acronyms need to be expanded? It seems you understand them.
The second doesn’t follow from the first. The existence of a limit doesn’t imply that it is small. (e.g. what if the limit was 10^1000 times more intelligent than a WBE.)
Ah, the implicit assumption here is “no wild new physics”. I’m assuming that the universe actually does respect the speed of light, that there are no major security exploits that let you cheat and violate conservation of energy, momentum, etc, etc, etc.
Assumption : the best way to survive inside the universe is to pattern matter the way you want it, and to collect free energy as needed to perform work. Conclusion : increasing intelligence has a diminishing return. If it would take 1000 years for a human molecular engineer to design a good mechanism to pattern matter with, it would take about 8 hours for a WBE of that engineer sped up a million-fold. “Good” mechanisms are within a percentage point or so of the “best” mechanism that physics will allow.
Assuming that the WBEs have already occupied the easily accessible matter before they create UFAIs, and their engineering is “good” but not excellent, then UFAIs no matter how much smarter will be limited in what they can practically do.
Analogy : if you’re out in the woods and possess only a hatchet as a tool, there are a finite number of things you can accomplish even if you had infinite intelligence.
Counterargument : since I, the author of this post, am not really all that smart compared to the hypothetical entities we are discussing, I cannot safely say this.
I would assume that engineers have to work in teams even more than computer programmers do.
The relationship between the details of a brain and what a person is good at is one of those mysteries. Just being able to scan a brain and replicate the person as a program doesn’t even start to figure out how that person’s capabilities can be expanded into new fields.
The assumption here is that with speedups from realtime anyone can become good at anything. Care to name a skill that you yourself feel you would be unable to accomplish if you had the equivalent of 1000 years to learn it, and your operators could make tweaks to your neural structure to help you past roadblocks?
Figuring out how to make those tweaks is going to be a long hard haul.
There’s a difference between getting very good at something, and being extraordinary.
Also, people who put in their 10K hours don’t converge. They gain mastery of a particular angle on their field of study.
Sheesh, I guess that is why people don’t like my article. Too many assumptions. I’m assuming that “very good” is “good enough” because “very good” is extraordinary if you have near-infinite time.
Here is my initial reaction.
This post makes a great deal of assumptions without justifying them. Extra details/assumptions make for good stories and even makes them sound more plausible, but every detail you add actually makes your story less plausible. I would suggest that you go through your post and remove as many Burdensome Details as possible, then spend some time nailing down the chain of logic you have left.
Also I think you may be falling prey to anthropomorphic bias. You seem to be assuming that modelling an AI on the human brain is the best approach, without stating any real justification for this. I would encourage you to read Artificial General Intelligence and the Human Mental Model (Yampolskiy and Fox, forthcoming).
I’m assuming that it’s the best approach in that the other possible approaches result in an alien entity that cannot be safely used. I’m also assuming the brain is too complex for a cluster of human beings to understand and copy, and that the complexity is higher than any evolutionary algorithm that could run on practical computers could create. (obviously, evolutionary algorithms did built the brain, but it took millions of years)
In no shape or form do I think that the human mind is actually “better”, I’m sure optimized entities could be orders of magnitude better...at least in terms of power efficiency or memory requirements....