Robot arms and computer vision, at the level necessary for playing a game of go, are I think a sufficiently solved problem that there’s no particular reason why AI researchers working on making a strong go-playing program would bother hooking them up. On its own I don’t think doing that would add anything interesting; in particular, I don’t think there’s any sense in which it would make the program’s thinking more human-like.
I don’t know about the Alphas, but my laptop running KataGo uses an amount of power that’s in the same ballpark as Ke Jie (more than just his brain, less than his whole body) and I’m pretty sure it would beat him very solidly. Go programs don’t generally concern themselves with power consumption as such but they do have to manage time (which is roughly proportional to total energy use) -- but so far as I know the time-management code is always hand-written rather than being learned somehow.
No one is claiming that a strong go-playing program is anything like an artificial general intelligence or that it makes humans obsolete or anything like that. (Though every surprising bit of progress in AI should arguably make us think we’re closer to making human-level-or-better general intelligences than we’d previously thought we were.)
Programs like AlphaGo Zero or KataGo don’t see the board in terms of local patterns as a result of learning from humans who do (and actually I very much doubt that non-Zero AlphaGo can usefully be said to do that either), they see the board in terms of local patterns because the design of their neural nets encourages them to do so. At least, it does in the earlier layers; I don’t know how much has been done to look at later layers and see whether or not their various channels tend to reflect local features that are comprehensible to humans.
Of course every move is a whole-board move in the sense that it is made on a whole board, its consequences may affect the whole board, and most of the time the way you chose it involves looking at the whole board. That’s true for humans and computers alike. But humans and computers alike build up their whole-board thinking from more-local considerations. (By “alike” I don’t mean to imply that the mechanisms are the same; the neural networks used in go programs are somewhat inspired by the structure of actual brain networks but that certainly doesn’t mean they are doing the same computation.)
And yes, because the training process is aiming to optimize the network weights for evaluating positions accurately and suggesting good moves and ultimately winning games all those local things are chosen and adjusted in order to get good global results. (The same is true for humans; the patterns we learn to make and avoid are considered good or bad as a result of long experience of how they affect actual games.) In that sense, indeed local and global are entangled (but I really wouldn’t say “quantumly”; so far as I can see there is no connection with quantum mechanics beyond the fact that the word “entangled” is used in both).
Some specifics about the network. I’m going to assume you don’t know anything :-); apologies for any resulting redundancy.
At any point in the computation, for each point on the board you have a bunch of numbers. The numbers (or the machinery that computes them) are called “channels”. At the very start you have a smallish number of channels corresponding to the lowest-level features; you can find a list in appendix A.1 of the KataGo paper, but they’re things like one channel that has 1 where there’s a black stone and 0 elsewhere, and one that has 1 where there’s a white stone and 0 elsewhere, and then some slightly cleverer things like channels that encode whether a given location contains a stone that’s part of a group with a small number of liberties or a stone that’s part of a group that can be captured in a ladder-like way (i.e., group has exactly two liberties and there’s a way to iterate “atari, opponent plays only escape move, group has exactly two liberties again” until eventually the group is captured; KataGo has hand-written code to look for this situation). I think some things that aren’t location-specific (e.g., what’s the value of komi?) are represented as channels that have the same value at every board location.
Now a single layer looks like this. It has, let’s say, m input channels and n output channels. (Usually these numbers are equal, but they’re different at the very start and at the very end.) So the input to this layer consists of 19x19xm numbers, and the output consists of 19x19xn numbers: one per board location per channel. Now, for each output channel, we have a 3x3xm array of numbers called weights. For each board location, overlay these on the surrounding 3x3xm region of the input. (Where it falls off the edge of the board, hallucinate an infinite sea of zeros. One input channel at the very start just has 1s at every board location so that you can identify the edge.) Multiply weights by channel values, add up the results, and force the sum to zero if it was negative; congratulations, you have computed the value for that output channel at that board location. Repeat 19x19xn times, doing a 3x3xm multiply/sum/rectify operation for each, until you have a full 19x19xn output: one number per channel per board location. This is a single convolutional layer.
The first convolutional neural networks were just stacks of convolutional layers, but a few newer tricks have been developed. One is what’s called a residual network or ResNet; I believe all the current generation of go programs use these. The idea is that the network will learn better if it’s encouraged to make later layers “refinements” of earlier layers rather than arbitrarily different. What this means in practice is that you build the network out of two-layer blocks as follows: you take the block’s input, you apply two convolutional layers to it, and then the output of the block isn’t the output of the second layer, it’s the output of the second layer plus the input to the block. (This only makes sense when the block has the same number of input and output channels, but as already mentioned this is the case for almost all of the network.) This doesn’t change what the network can compute in principle, but it tends to lead to more effective learning, and you need to know about it because the sizes of these networks are always quoted in blocks rather than layers. The strongest KataGo networks are 40 blocks which means 80 layers.
Another tweak is batch normalization which means that before each convolutional layer you “normalize” each of its channels by scaling and shifting all the numbers in that channel so that their mean is 0 and their variance is 1.
Another trick introduced by KataGo (it’s not found in the Alphas or in Leela Zero) is global pooling. I think I described this in a comment earlier in this thread. The idea is that every now and then you have a layer in which you take some of the channels, you compute their mean and maximum values, and you then feed these as inputs into all the other channels. So e.g. if one of these channels happens to have +1 for board locations that are in some sense under black’s control and −1 for ones that are under white’s control then the mean gives you a sort of a score estimate. If one has 1 for a location where there seems to be a ko and 0 elsewhere then the max tells you whether there are any kos and the mean tells you how many.
OK. So we start with some number of input channels describing the basic state of the game. We have 40 blocks, each consisting of two convolutional layers, and every now and then one of these blocks does the global-pooling thing to incorporate information from the whole board into the local calculations. At the very end, we take our final set of output channels and use them to compute a whole bunch of things. The most important are (1) a “policy” value for each board location, indicating how plausible a place it is for the next move, and (2) how likely each player is to win, and the likely final score. But there are also (1′) guesses at the opponent’s next move, (1″) estimates of how likely each board location is to end up in the control of each player, (2′) estimates of how likely each specific final score is, and (3) various other interesting things whose details you can find in appendix A.4 and A.5 of the KataGo paper.
The way we compute these outputs is a bit like the way we compute the other layers. We take the outputs of the last ordinary layer, and for the (1) things that have per-location values we compute linear combinations of those outputs (the coefficients are learned along with all the other weights in the network) and apply suitable nonlinear functions to the results for the outputs to make human sense; for the (2) things that have global values we first do a global-pooling-like step on all the per-location outputs and then compute linear combinations and apply suitable final nonlinearities. Again, details in appendix A.4 and A.5.
That’s all about the network; of course the search is another very important component, but you didn’t ask about that :-).
Warning: This is a long post and there’s some personal stuff about my living situation near the beginning. I figure if people on the forum can bring up their issues with living in expensive, culturally blessed parts of the country, I can bring up issues of living in the homeless shelter system.
I also apologize in advance for the length as I haven’t addressed many of the more technical aspects yet. I partly blame the fascinating intersections of AI and human culture for my long post. I do sort of take these posts as opportunities to attempt to draw the different threads of my thinking closer together, with the sometimes unhelpful effect that though I can explain certain ideas more efficiently and concisely, I try to add in more ideas as a result
First let me say this: I’ll address the points you bring up as best as I can given my approach and purposes in this discussion. I have some questions and some theories that I think are suitable for further development, and the fact ML and AI have developed so far, makes me think they would help me investigate my theories.
The fact that DanielFilan created a post addressing these ideas in relationship to Go is a win-win for me, as it gives me a good entry point for discussion of the technical aspects of neural nets and such, so that becoming more familiar with the way they work makes sense if I have any chance of pursuing research with some sort with credibility. You make have a zero% success rate for the opportunities you don’t bother to try and make happen.
Apologies to you and to Daniel for my skepticism and some of the assumptions I made previously in this discussion; the clarity of the discussion now is challenging to maintain under my currently rather bad circumstances. As I’ve mentioned elsewhere, I’m attempting to work my way out of the homeless shelter system where I’ve existed for the past 4 years and counting. This forum is a much needed lifeline of sanity, especially given the once-in-a-century pandemic we are only now emerging from in the Western World.
I try to focus on issues of science and technology in order to block out the more traumatizing aspects of homelessness. We had four cop cars and a chopper spotlighting the building last night around 11 or so, and I didn’t get much sleep as a partial result. That happens about once every other month or so here. Plus my neighbor tends to pound on the wall and scream for an hour or so usually in the early morning, for no apparent reason sometimes a few times a week. Been trying to get management to do something about it for over 2 years, and have had to resort to calling the police myself, as management won’t.
Long story short, I try to focus as much as I can on things other than my circumstances for the time being, and I use the little research I can do these days without getting distracted by the drama and violence around me, to try and sort of reach out to communities I feel more in tune with. The little interaction I get is probably also part of the reason my posts run so long, norms of social interaction are currently not being reinforced in my daily life. I am literally surrounded by serious drug addicts and seriously mentally disordered formerly homeless people 24 hours a day, 7 days a week, so I’m trying to use my writing to find some value during this horrible period of my life. :p
So I’ve tried to set myself into context now. And I think it’s easier to address the rest of your post.
About the Robot Arm and the various associated peripheral input and output devices they are representative of, the implications of integrating the AGI of a Go playing AI with a physical body is reminiscent of Terminator I suppose. I don’t argue at this point that by keeping the components of a potential Terminator separate, it is potentially more difficult to create a walking, talking human killing machine. I only wonder what the implications are of ‘technological lock in’ at the point when the different components of a bipedal humanoid AI cyborg are being integrated; I imagine there will be a lot of ‘reinventing the wheel’ as the systems are re-engineered during integration phases. The current trend of ‘modularizing’ more complicated systems is a smart way to approach mass production of certain technologies, and the switch to 5G will create the supply for whatever demand is there for the foreseeable future, but there are always problems with developing technology.
My real concern comes from a particular cultural viewpoint: I think AlphaGo has dealt a serious, yet covert blow to the Go community in particular, and the human race in general; one which is unnecessarily critical of human capabilities and puts technology in genera, on a pedestal that is much too high for what it truly represents. Some might say that’s a purely symbolic issue, but I think the concept of putting ‘thing’s on pedestal’s is just a societal way of setting ‘things’ up to inevitably fail us. I know this may sound harsh, but when I heard the news that AlphaGo beat Ke Jie several years back, it took me a couple of weeks to process that, and I still feel uncomfortable about it.
The developing pop culture awareness of bias in seemingly ‘objective’ institutions of authority, institutions like Google and Facebook, whose facts and figures are governed by algorithms—not magical truth atoms or some infallible equivalent—means they aren’t the Holy Grail they were believed to be (whether or not they were marketed that way is a question on my mind as well.) The assumed infallability of our institutions of authority seem to be crumbling at an accelerated pace, leaving us with the unanswered questions humanity has always struggled with; now we just have the ability to manage that misery globally at the speed of light.
It’s possible we don’t find truth in our technology—that we just find a reflection of ourselves—so that it isn’t possible to lose ourselves in the search for ultimate truth through technology, we just end up seeing ourselves differently.
In this same respect, I think the ceremony and marketing of the series of games between Ke Jie and AlphaGo, which conferred 9dan professional status on a go playing AI for the first time, had the ‘equal but opposite’ social cost of placing arguably the strongest human player beneath AlphaGo in the ranking system, but is a misinterpretation of what actually happened.
I can’t place my hands on the article at this point, but I recall reading something about Julia Gaef and her new book, which talked about the ethical considerations of the amount of power consumption used to mine cryptocurrencies. From what little I know about it, the amount of resources used in the process, seems to outweigh any benefit gained by the technology.
There is a parallel between this idea and the way that go playing AI have influenced society; pound for pound I wonder which entity, Ke Jie or AlphaGo at the time of the series, used more resources to play the games. (Including his body and not just his mind, and on the other side, the team of 7 or 8 people who helped AlphaGo play. I’m pretty sure AlphaGo used more resources during the 3 game series than did his human opponent.)
Why does this matter? Because in the unavoidable battle for ‘global supremacy’ between humans and our developing technology, the metrics we use to measure this matter in reality. It really does matter how situations are framed. A human player I believe simply has an advantage over an AI in that our Intelligence is hardwired into our hardware, so that we are potentially the most efficient being ever created, and this advantage over AI in it’s ability to affect the world outside of it’s neural nets, could be made more evident instead of hidden.
Our batteries and engines, our Intelligence and our limbs, our sensory input and motor output programs and devices far outperform similarly organized and integrated technological systems, and in an incredibly compact and efficient system. The overall human body and it’s associated personality, at least on average, would kick AlphaGo’s ass in a fist fight, and without someone to plug it in, once the laptops battery was drained, the AI contained on it wouldn’t be able to function. My point being, AlphaGo the AI requires a lot of outside assistance to do it’s thing spatially. Once again, why does that matter? Because it’s a way of framing the situation which points to all the human help it still requires to do it’s thing—which is a good thing—but most people seem content to ignore all the human support and focus only on the technology as being what’s important. Technology for technologies sake is the death of humanity.
I think I understand what you mean when you say the technologies I mentioned have been developed separately to such a degree, that almost no one would see the need to develop any kind of a robotic system to attempt to help an AI compete with a human player, as competition on any other level then simple computation of the abstract aspects of a good game of Go might be considered taboo at this point. Few people want the Terminator. But, the simple abstract aspects of a good game of Go can be framed as simple, as opposed to incredibly complex and impressive, depending on how they are presented in relationship to the alternative. The attempt to create a humanoid robot go playing AI is too creepy for most people at this point I’m sure.
My point is, that the perception of what AlphaGo did was possibly misrepresented, and in my opinion has the potential to be misinterpreted to the disadvantage of human society, when compared to the technology we have developed. The broader implications for humanity viewing itself as deficient in comparison to our technology is a drain on human morale. I think we need to do better than that.
How we do better is up for debate: My suggestion was to make clear in a potential rematch, just how clunky and difficult it would be for a humanoid robot paired with a go playing AI and associated AGI to run the robotics which would allow an AI playing robot to not only sit there and play go against a human player, but to try to do everything that Ke Jie did, maybe even smoke a cigarette. I think of something like Boston Dynamics Atlas system paired with AlphaGoZero, with some sort of a visual system modeled after the human eyes for perceiving the game through visual sensory array and system more analagous to human vision, and some sort of breathing apparatuas which would allow fo the inhalation of the smoke from a cigarette. AND AlphaGo has to decide on the brand it prefers to smoke, and pay for it with money it earned, and dress itself. A terminator of sorts I guess.
But the identity and personality would also need to be rounded out, to also be able to recognize when it needed to take a break because it’s systems weren’t functioning optimally, so that it might request a bathroom break and take the opportunity to plug itself into a socket in the bathroom to recharge it’s battery, as charging isn’t allowed in the game area according to a set of rules which favor the human player. This would allow the human player to develop a strategy of dragging the game out and making moves requiring more computation, in order to drain the AI’s batteries. In many ways, it is the psychologically aspects of the game which go playing AI avoid, which will be an advantage to them in the future if we do need to defend ourselves against them. If they don’t care about the impact their actions have on people, we’re just developing psychopaths with superhuman powers.
I believe this kind of a rematch than, could be an opportunity for companies like DeepMind to ease a little of the tension in society regarding our crumbling trust in their objectivity and moral and ethical authority; it might allow for the goodhearted showcasing of the ways in which a go playing AI isn’t superior to humans, and if done well might reframe our perception of ourselves as well as of our technology, after all isn’t the human component of any technological system of paramount importance? The technology can’t overshadow the society that spawned it, otherwise it might be revealed that it has been using us the entire time for it’s own purposes, and see humans as the problem, not as the concerned species attempting to find solutions to our problems.
Technology is not superior to humans in aggregate, in fact it has just as many flaws in it as the civilization that created it I think—it’s possible the ‘original sin’ of humanity hasn’t been avoided in the creation of technologies, in fact the potential for the true ‘fall of humankind’ has potentially been accelerated by technologies mass production.
More practically, go is considered a martial art, complete with a ranking system and a history of use in the training of military consideration and execution. I like to think If I had been in Ke Jie’s shoes, and my loss on the board seemed imminent, I would have grabbed the board, and used it to literally destroy the laptop hosting AlphaGo to prove there was at least one way I was superior to it.
So, more broadly, my interests are in the ethical considerations of the current state of research and technology in a global sense, in that I am trying to consider the entirety of human knowledge across all disciplines. The specifics of a Go playing AI have ramifications beyond Go, and it is the dissemination and integration of the knowledge generated by this research within the world community, and my concern with how to potentially measure it, that leads me to trying to understand more of how neural nets work, with the belief in a analogous system of logic underlying human consciousness, and a theoretical cultural subconscious which is now being ripped apart in huge cultural shock waves, analogous to earthquakes formed by an underlying Socio-Psycho-Economic Tectonic plate system.
It’s mainly through one of the main aspects of my theories, that I try to think about it in the way I describe above, what I call ‘Cultural Lag.’ In the same way the decision making process of a neural net experiences delay, or lag, society does too when it comes to processing human knowledge.
I want to quantify that.
I’m still trying to describe it precisely, in the hope of attempting to define a mathematical metric to measure it, and use that as a way to study social dynamics at a human level, but with theoretically quantum basis, but also on a global scale. In an an intuitive sense, as a theoretical aspect of a type of unified theory of everything—I describe ‘cultural lag’ as a type of impedance in a system of quantum circuits underlying all reality, which is visible to us as the playing out of Social Dynamics in Society, and effects the development of human civilization. What are potential quantum reasons for the differences between a 1st world country like the US, and a developing 3rd world country like Africa for instance.
Because I’m concerned with the entirety of the realm of human knowledge, and how it is being managed (or mismanaged as I believe is the case) my thinking is much broader than it is deep. So delving into the specifics of a Go playing AI seems like a reasonable exchange to bend your ear a little more about my crazy ideas. Lol
...but so far as I know the time-management code is always hand-written rather than being learned somehow.
This seems like a point of consideration. I suppose this begins to get back to the underlying AGI of a system like AlphaGo, in that it is still likely timing it’s calculations according to a Ghz system using a system clock of some sort on the motherboard, as most pcs are designed with this system in mind. That is a very basic use of time within the system though, pretty much independent at this point of any outside considerations. I do wonder during down time, when waiting for the opponent to move, how the processor systems function: are they always running at full speed, or throttled depending on the situation?
Of course there is a psychological aspect to time management, one that is based on perception of the opponent. It’s possible I’m sure to create an algorithim which computes the optimal speed at which to play a move in order to psyche the opponent out, especially if you weight certain important moves which have a huge effect on the board as being good opportunities to play with a different timing. I dislike this idea though, it smacks to much of psychological warfare without any counterbalancing concern for social norms or ethics.
I do find it sort of unfortunate at times when I am reviewing a game, especially one I didn’t witness in real time, that there is no attempt to communicate how long it took for the moves to be played. With an auto playback feature it is just a simple constant 1 second between moves playback. Less feedback for the user, and something I get annoyed with.
No one is claiming that a strong go-playing program is anything like an artificial general intelligence or that it makes humans obsolete or anything like that.
2 points:
A) From what I understand, something like AlphaGoZero has a somewhat generic AGI system? ( Am I misusing the idea of AGI, in that it might be more of a GI system under the hood of AlphaGo?) with a domain specific neural net on top, so that the underlying AGI is weighted to perform maximally under all domain-specific tasks when integrated with them—singly at this point, and in something like parallel or series of groups in the future.
B) As a skeptic, I tend to look for emotional leakage in messaging. My mom was a single parent, radical feminist with a degree in Anthropology, Womens studies, Film theory, and eventually Library Science. So I lived and breathed media and social criticism and critical thinking from birth. From the media I’ve encountered surrounding the advances in Go playing AI, I see a strong thread of apprehension, confusion, and sadness in the Go community resulting in the seeming dominance of AI in the rank structure and overall community, (mostly the more mature players.)
Additionally, with such a focus on how AI and automation can do things better than humans, I’m a little confused as to why Nascar and Formula 1 racing haven’t yet been test beds for driver-less technology. I’ve no doubt a car driving AI would also dominate these competitive arenas with existing technology, but I’ve never heard anyone mention it as even a possibility. Special interest groups no doubt have a huge hidden effect, and I’m sure the subjects have been approached with the car industry, but the image management after the eventual loss to AI would be a mess. It’s the lack of consideration for potential political consequences of these types of competitions which I feel are counter productive to values espousing democracy and more liberal values. Values I think the majority of big tech seem to benefit most from. Maybe I’m wrong about that, certainly the battle between Apple and Microsoft seemed to be weighted in one direction, but I think it’s safe to say that Microsoft has had more of an impact globally, while Apple has had less of a detrimental effect world wide. In the end, which is better: more impact or causing less harm? It’s seems to be a business consideration, but I think doubles as a political one as well.
So I wonder “what is it about the Go community that made it receptive to these types of competitions, in a way that the car community is not?” I think the Documentary AlphaGo is a very revealing look at much of the process of setting up the competition, and I would love to celebrate the Go communities openess to the AI community, but if I have to acknowledge AI is somehow superior to humans instead of just better in some pretty specific ways, I prefer to contemplate the Go communities openess as a mistake for the human institute of Go, and the technological encroachment as harming the global community unnecessarily.
So while no one is explicitly stating that AI are superior to humans in every way, it seems few are really trying to put the technology into a broader perspective, which attaches more value to the human component of the relationship than it currently does. I hate to wonder if this is an ego thing, or if it is simply the unintended consequences of not planning for the image management component of the post-game go community in the even of a loss. I hate to think badly of DeepMind, but the shockwaves AlphaGo has sent through the Go community puts potential allies off, as it could be interpreted that DeepMinds goal was to beat humanity, not just Ke Jie.
Humility is a valuable virtue, and I worry that Big Tech pays only lip service to that idea. I’m not saying that about DeepMind specifically, but in general, technological superiority is the deciding factor in war; I just hope companies developing AI and the like aren’t attempting to conduct war against humans. Against our problems? Yes. Against us? No. Of course there are many other concerns driving the technological push, as in terms of digital technology, the genie is already out of the bottle. Would that we could go back and be more conservative with some of the export of some proprietary technology; we might have avoided shipping the vast majority of our manufacturing jobs and infrastructure overseas.
Ultimately it’s the people that make up a company, so in that sense, speaking of a large corporation as either having humility or lacking it is difficult, and looking at at rebranding efforts after disasters such as the one that forced BP to rebrand, point to the limited responsibility large corporations have for catastrophic failures of their corporation. I believe the Go community accepted the loss to AlphaGo gracefully, but it would be nice to not feel so badly about the loss.
Programs like AlphaGo Zero or KataGo don’t see the board in terms of local patterns as a result of learning from humans who do (and actually I very much doubt that non-Zero AlphaGo can usefully be said to do that either),...
My thinking was along the lines that AlphaGo learning about human concepts and consideration of local patterns would be an automatic effect of using human games as source material, complete with any inefficient placement and order of moves humans might make in their play; in lieu of a less derogatory statement “Garbage in; Garbage out.” concepts’ of local patterns from human games would automatically transfer to the trained neural nets, sans any associations beyond the abstract consideration of maximizing win/rate. This is probably hard to prove.
… they see the board in terms of local patterns because the design of their neural nets encourages them to do so. At least, it does in the earlier layers; I don’t know how much has been done to look at later layers and see whether or not their various channels tend to reflect local features that are comprehensible to humans.
I get that, I know there have been competing designs for advanced computing that don’t attempt to mimic neural structures. And I get what you mean about later layers possibly ‘creating meaning’ humans can’t recognize as of value. Like some secret language twins speak or something equally inaccessible.
...all those local things are chosen and adjusted in order to get good global results.
I guess this is where I start to question what a strong player does. Obviously there are some differences in how a strong players mind works in comparison to a weak player. I think the stronger a player is, the more accurately they consider global results, and weak players are lucky if they have a small understanding of how local play effects global play. This is due to the neural process of ‘consolidation’, which I think has it’s corollary in the adjusting of weights in the neural nets. This means a strong player begins to make time and energy saving changes to the way they think, and start to make assumptions about common concepts in order to be able to spend more time contemplating the global game.
A strong player and a weak player play a game, they each have an hour time limit. The weak player might take the whole hour and end up losing on time, while the strong player uses only 30 minutes but is clearly ahead in every way.
Is it because the strong player thinks about everything the weak player thinks about, just much faster and more accurately? Not exaclty it’s because the strong player has consolidated data and information about things like local patterns, into knowledge about what they look like at a ‘glance’ and hopefully into wisdom about what good patterns of play relating to global concepts look like, and so ends up ‘skipping’ the contemplation of many of those local patterns. The consolidation of the human mind creates ‘shortcuts’ between low levels of meaning and higher levels, to allow more efficient thinking, which is faster if its accurate, but not if it’s inaccurate. So when a strong player has an ‘intuition’ about what a good play is, it turns out to be a good play, whereas if a weak player relies on intuition, it’s likely they are just guessing because they don’t have the data, info, knowledge and the accurate shortcuts to higher levels of meaning the strong player does.
So this just brings us full circle though, as you said:
But humans and computers alike build up their whole-board thinking from more-local considerations. (By “alike” I don’t mean to imply that the mechanisms are the same;
I come at this from a background of psychology, neuropsych, and cognitive science sans any proficiency with the technical knowledge to do anything but create meaningful correlations between human mind and computer mind; it seems you come at it from a much stronger technical background having to do with computer mind, so I’m not surprised at the seeming incongruities in our thinking. That’s a compliment by the way; I’m sure my skepticism is kicking in a bit, blame my mom.:P
In that sense, indeed local and global are entangled (but I really wouldn’t say “quantumly”; so far as I can see there is no connection with quantum mechanics beyond the fact that the word “entangled” is used in both).
I tend to view atomic and subatomic physics as predictive of human scale phenomenon even though I don’t fully understand the math, plus I view the astronomical scale of the universe as being influenced by the subatomic scale. So I thinks its safe to assume quantum mechanics are predictive of subatomic phenomenon, and by proxy they are predictive of human scale phenomenon, not to mention universal scale phenomenon. Of course how exactly they are related is the domain of quantum physicists and science fiction writers and intellectuals who think about things in that way.
We haven’t left behind the nuclear age in our progress towards a unified theory of everything, but we have entered a quantum age. So I think it’s appropriate to wonder at the quantum dynamics which underlie the human scale world. Fluid mechanics are apparent in the waves of the ocean and the water in a bath tub, but you don’t have to know the physics to see and make inferences that the movement of the water is influenced by particle and subatomic physics. Might as well add in the idea that it’s effected by quantum mechanics as well (loosely at first of course.)
If you can not strongly disagree with that idea, then the next question might be “Why is that important at this point in time?” or “what does it matter?”. Unfortunately, we live in an age where humans are being forced to compete with super computers and distributed neural nets in order to stake out claims of truth; neural nets make predictions based on putting 2 and 2 together in similar fashion to how humans make predictions I think. Gibberish from a computer is seen as acceptable and evidence it’s working at something, so that entire fields of study are created to learn to interpret and translate the gibberish in a way that’s meaningful.
Society doesn’t do the same thing for humans, especially ones not in the important marketing demographic of 18-34. My point being, If I discovered the secret of the universe and blabbed about it as much as I could, I probably wouldn’t be taken seriously, even at a point in the future where the science would exist to back up my claims. I likely wouldn’t even be remembered for coming up with the idea as it would likely be buried under quadrillions of terraflops of data piled on top of the archive of this forum decades from now.
Under those circumstances, what’s the harm in attempting to make claims most respected scientists wouldn’t make for fear of being proven wrong? Especially If I think there’s validity to the claims and I get a chance to shoot the proverbial shit with someone of a somewhat like mind. Would be better over a beer of course, and if I had access to resources to help me clarify my claims and point me in a good direction. But I’ve been out of school for awhile.
So having said that, I think it’s valid to use the term ‘entangled’ to imply some quantum relationship between a relative local scale phenomenon and a global ‘universal’ scale phenomenon. You can disagree, but i’m sure we’re so far away from understanding how to appropriately use the tech which will be enabled by quantum science, that it won’t really matter in the long run. In the short run though, I am curious what you think, if you think much about, quantum mechanics.
This post is getting way too long though. I’ll take a look at the more technical stuff a little later to follow up if I have questions it that’s alright with you.
Robot arms and computer vision, at the level necessary for playing a game of go, are I think a sufficiently solved problem that there’s no particular reason why AI researchers working on making a strong go-playing program would bother hooking them up. On its own I don’t think doing that would add anything interesting; in particular, I don’t think there’s any sense in which it would make the program’s thinking more human-like.
I don’t know about the Alphas, but my laptop running KataGo uses an amount of power that’s in the same ballpark as Ke Jie (more than just his brain, less than his whole body) and I’m pretty sure it would beat him very solidly. Go programs don’t generally concern themselves with power consumption as such but they do have to manage time (which is roughly proportional to total energy use) -- but so far as I know the time-management code is always hand-written rather than being learned somehow.
No one is claiming that a strong go-playing program is anything like an artificial general intelligence or that it makes humans obsolete or anything like that. (Though every surprising bit of progress in AI should arguably make us think we’re closer to making human-level-or-better general intelligences than we’d previously thought we were.)
Programs like AlphaGo Zero or KataGo don’t see the board in terms of local patterns as a result of learning from humans who do (and actually I very much doubt that non-Zero AlphaGo can usefully be said to do that either), they see the board in terms of local patterns because the design of their neural nets encourages them to do so. At least, it does in the earlier layers; I don’t know how much has been done to look at later layers and see whether or not their various channels tend to reflect local features that are comprehensible to humans.
Of course every move is a whole-board move in the sense that it is made on a whole board, its consequences may affect the whole board, and most of the time the way you chose it involves looking at the whole board. That’s true for humans and computers alike. But humans and computers alike build up their whole-board thinking from more-local considerations. (By “alike” I don’t mean to imply that the mechanisms are the same; the neural networks used in go programs are somewhat inspired by the structure of actual brain networks but that certainly doesn’t mean they are doing the same computation.)
And yes, because the training process is aiming to optimize the network weights for evaluating positions accurately and suggesting good moves and ultimately winning games all those local things are chosen and adjusted in order to get good global results. (The same is true for humans; the patterns we learn to make and avoid are considered good or bad as a result of long experience of how they affect actual games.) In that sense, indeed local and global are entangled (but I really wouldn’t say “quantumly”; so far as I can see there is no connection with quantum mechanics beyond the fact that the word “entangled” is used in both).
Some specifics about the network. I’m going to assume you don’t know anything :-); apologies for any resulting redundancy.
At any point in the computation, for each point on the board you have a bunch of numbers. The numbers (or the machinery that computes them) are called “channels”. At the very start you have a smallish number of channels corresponding to the lowest-level features; you can find a list in appendix A.1 of the KataGo paper, but they’re things like one channel that has 1 where there’s a black stone and 0 elsewhere, and one that has 1 where there’s a white stone and 0 elsewhere, and then some slightly cleverer things like channels that encode whether a given location contains a stone that’s part of a group with a small number of liberties or a stone that’s part of a group that can be captured in a ladder-like way (i.e., group has exactly two liberties and there’s a way to iterate “atari, opponent plays only escape move, group has exactly two liberties again” until eventually the group is captured; KataGo has hand-written code to look for this situation). I think some things that aren’t location-specific (e.g., what’s the value of komi?) are represented as channels that have the same value at every board location.
Now a single layer looks like this. It has, let’s say, m input channels and n output channels. (Usually these numbers are equal, but they’re different at the very start and at the very end.) So the input to this layer consists of 19x19xm numbers, and the output consists of 19x19xn numbers: one per board location per channel. Now, for each output channel, we have a 3x3xm array of numbers called weights. For each board location, overlay these on the surrounding 3x3xm region of the input. (Where it falls off the edge of the board, hallucinate an infinite sea of zeros. One input channel at the very start just has 1s at every board location so that you can identify the edge.) Multiply weights by channel values, add up the results, and force the sum to zero if it was negative; congratulations, you have computed the value for that output channel at that board location. Repeat 19x19xn times, doing a 3x3xm multiply/sum/rectify operation for each, until you have a full 19x19xn output: one number per channel per board location. This is a single convolutional layer.
The first convolutional neural networks were just stacks of convolutional layers, but a few newer tricks have been developed. One is what’s called a residual network or ResNet; I believe all the current generation of go programs use these. The idea is that the network will learn better if it’s encouraged to make later layers “refinements” of earlier layers rather than arbitrarily different. What this means in practice is that you build the network out of two-layer blocks as follows: you take the block’s input, you apply two convolutional layers to it, and then the output of the block isn’t the output of the second layer, it’s the output of the second layer plus the input to the block. (This only makes sense when the block has the same number of input and output channels, but as already mentioned this is the case for almost all of the network.) This doesn’t change what the network can compute in principle, but it tends to lead to more effective learning, and you need to know about it because the sizes of these networks are always quoted in blocks rather than layers. The strongest KataGo networks are 40 blocks which means 80 layers.
Another tweak is batch normalization which means that before each convolutional layer you “normalize” each of its channels by scaling and shifting all the numbers in that channel so that their mean is 0 and their variance is 1.
Another trick introduced by KataGo (it’s not found in the Alphas or in Leela Zero) is global pooling. I think I described this in a comment earlier in this thread. The idea is that every now and then you have a layer in which you take some of the channels, you compute their mean and maximum values, and you then feed these as inputs into all the other channels. So e.g. if one of these channels happens to have +1 for board locations that are in some sense under black’s control and −1 for ones that are under white’s control then the mean gives you a sort of a score estimate. If one has 1 for a location where there seems to be a ko and 0 elsewhere then the max tells you whether there are any kos and the mean tells you how many.
OK. So we start with some number of input channels describing the basic state of the game. We have 40 blocks, each consisting of two convolutional layers, and every now and then one of these blocks does the global-pooling thing to incorporate information from the whole board into the local calculations. At the very end, we take our final set of output channels and use them to compute a whole bunch of things. The most important are (1) a “policy” value for each board location, indicating how plausible a place it is for the next move, and (2) how likely each player is to win, and the likely final score. But there are also (1′) guesses at the opponent’s next move, (1″) estimates of how likely each board location is to end up in the control of each player, (2′) estimates of how likely each specific final score is, and (3) various other interesting things whose details you can find in appendix A.4 and A.5 of the KataGo paper.
The way we compute these outputs is a bit like the way we compute the other layers. We take the outputs of the last ordinary layer, and for the (1) things that have per-location values we compute linear combinations of those outputs (the coefficients are learned along with all the other weights in the network) and apply suitable nonlinear functions to the results for the outputs to make human sense; for the (2) things that have global values we first do a global-pooling-like step on all the per-location outputs and then compute linear combinations and apply suitable final nonlinearities. Again, details in appendix A.4 and A.5.
That’s all about the network; of course the search is another very important component, but you didn’t ask about that :-).
Warning: This is a long post and there’s some personal stuff about my living situation near the beginning. I figure if people on the forum can bring up their issues with living in expensive, culturally blessed parts of the country, I can bring up issues of living in the homeless shelter system.
I also apologize in advance for the length as I haven’t addressed many of the more technical aspects yet. I partly blame the fascinating intersections of AI and human culture for my long post. I do sort of take these posts as opportunities to attempt to draw the different threads of my thinking closer together, with the sometimes unhelpful effect that though I can explain certain ideas more efficiently and concisely, I try to add in more ideas as a result
First let me say this: I’ll address the points you bring up as best as I can given my approach and purposes in this discussion. I have some questions and some theories that I think are suitable for further development, and the fact ML and AI have developed so far, makes me think they would help me investigate my theories.
The fact that DanielFilan created a post addressing these ideas in relationship to Go is a win-win for me, as it gives me a good entry point for discussion of the technical aspects of neural nets and such, so that becoming more familiar with the way they work makes sense if I have any chance of pursuing research with some sort with credibility. You make have a zero% success rate for the opportunities you don’t bother to try and make happen.
Apologies to you and to Daniel for my skepticism and some of the assumptions I made previously in this discussion; the clarity of the discussion now is challenging to maintain under my currently rather bad circumstances. As I’ve mentioned elsewhere, I’m attempting to work my way out of the homeless shelter system where I’ve existed for the past 4 years and counting. This forum is a much needed lifeline of sanity, especially given the once-in-a-century pandemic we are only now emerging from in the Western World.
I try to focus on issues of science and technology in order to block out the more traumatizing aspects of homelessness. We had four cop cars and a chopper spotlighting the building last night around 11 or so, and I didn’t get much sleep as a partial result. That happens about once every other month or so here. Plus my neighbor tends to pound on the wall and scream for an hour or so usually in the early morning, for no apparent reason sometimes a few times a week. Been trying to get management to do something about it for over 2 years, and have had to resort to calling the police myself, as management won’t.
Long story short, I try to focus as much as I can on things other than my circumstances for the time being, and I use the little research I can do these days without getting distracted by the drama and violence around me, to try and sort of reach out to communities I feel more in tune with. The little interaction I get is probably also part of the reason my posts run so long, norms of social interaction are currently not being reinforced in my daily life. I am literally surrounded by serious drug addicts and seriously mentally disordered formerly homeless people 24 hours a day, 7 days a week, so I’m trying to use my writing to find some value during this horrible period of my life. :p
So I’ve tried to set myself into context now. And I think it’s easier to address the rest of your post.
About the Robot Arm and the various associated peripheral input and output devices they are representative of, the implications of integrating the AGI of a Go playing AI with a physical body is reminiscent of Terminator I suppose. I don’t argue at this point that by keeping the components of a potential Terminator separate, it is potentially more difficult to create a walking, talking human killing machine. I only wonder what the implications are of ‘technological lock in’ at the point when the different components of a bipedal humanoid AI cyborg are being integrated; I imagine there will be a lot of ‘reinventing the wheel’ as the systems are re-engineered during integration phases. The current trend of ‘modularizing’ more complicated systems is a smart way to approach mass production of certain technologies, and the switch to 5G will create the supply for whatever demand is there for the foreseeable future, but there are always problems with developing technology.
My real concern comes from a particular cultural viewpoint: I think AlphaGo has dealt a serious, yet covert blow to the Go community in particular, and the human race in general; one which is unnecessarily critical of human capabilities and puts technology in genera, on a pedestal that is much too high for what it truly represents. Some might say that’s a purely symbolic issue, but I think the concept of putting ‘thing’s on pedestal’s is just a societal way of setting ‘things’ up to inevitably fail us. I know this may sound harsh, but when I heard the news that AlphaGo beat Ke Jie several years back, it took me a couple of weeks to process that, and I still feel uncomfortable about it.
The developing pop culture awareness of bias in seemingly ‘objective’ institutions of authority, institutions like Google and Facebook, whose facts and figures are governed by algorithms—not magical truth atoms or some infallible equivalent—means they aren’t the Holy Grail they were believed to be (whether or not they were marketed that way is a question on my mind as well.) The assumed infallability of our institutions of authority seem to be crumbling at an accelerated pace, leaving us with the unanswered questions humanity has always struggled with; now we just have the ability to manage that misery globally at the speed of light.
It’s possible we don’t find truth in our technology—that we just find a reflection of ourselves—so that it isn’t possible to lose ourselves in the search for ultimate truth through technology, we just end up seeing ourselves differently.
In this same respect, I think the ceremony and marketing of the series of games between Ke Jie and AlphaGo, which conferred 9dan professional status on a go playing AI for the first time, had the ‘equal but opposite’ social cost of placing arguably the strongest human player beneath AlphaGo in the ranking system, but is a misinterpretation of what actually happened.
I can’t place my hands on the article at this point, but I recall reading something about Julia Gaef and her new book, which talked about the ethical considerations of the amount of power consumption used to mine cryptocurrencies. From what little I know about it, the amount of resources used in the process, seems to outweigh any benefit gained by the technology.
There is a parallel between this idea and the way that go playing AI have influenced society; pound for pound I wonder which entity, Ke Jie or AlphaGo at the time of the series, used more resources to play the games. (Including his body and not just his mind, and on the other side, the team of 7 or 8 people who helped AlphaGo play. I’m pretty sure AlphaGo used more resources during the 3 game series than did his human opponent.)
Why does this matter? Because in the unavoidable battle for ‘global supremacy’ between humans and our developing technology, the metrics we use to measure this matter in reality. It really does matter how situations are framed. A human player I believe simply has an advantage over an AI in that our Intelligence is hardwired into our hardware, so that we are potentially the most efficient being ever created, and this advantage over AI in it’s ability to affect the world outside of it’s neural nets, could be made more evident instead of hidden.
Our batteries and engines, our Intelligence and our limbs, our sensory input and motor output programs and devices far outperform similarly organized and integrated technological systems, and in an incredibly compact and efficient system. The overall human body and it’s associated personality, at least on average, would kick AlphaGo’s ass in a fist fight, and without someone to plug it in, once the laptops battery was drained, the AI contained on it wouldn’t be able to function. My point being, AlphaGo the AI requires a lot of outside assistance to do it’s thing spatially. Once again, why does that matter? Because it’s a way of framing the situation which points to all the human help it still requires to do it’s thing—which is a good thing—but most people seem content to ignore all the human support and focus only on the technology as being what’s important. Technology for technologies sake is the death of humanity.
I think I understand what you mean when you say the technologies I mentioned have been developed separately to such a degree, that almost no one would see the need to develop any kind of a robotic system to attempt to help an AI compete with a human player, as competition on any other level then simple computation of the abstract aspects of a good game of Go might be considered taboo at this point. Few people want the Terminator. But, the simple abstract aspects of a good game of Go can be framed as simple, as opposed to incredibly complex and impressive, depending on how they are presented in relationship to the alternative. The attempt to create a humanoid robot go playing AI is too creepy for most people at this point I’m sure.
My point is, that the perception of what AlphaGo did was possibly misrepresented, and in my opinion has the potential to be misinterpreted to the disadvantage of human society, when compared to the technology we have developed. The broader implications for humanity viewing itself as deficient in comparison to our technology is a drain on human morale. I think we need to do better than that.
How we do better is up for debate: My suggestion was to make clear in a potential rematch, just how clunky and difficult it would be for a humanoid robot paired with a go playing AI and associated AGI to run the robotics which would allow an AI playing robot to not only sit there and play go against a human player, but to try to do everything that Ke Jie did, maybe even smoke a cigarette. I think of something like Boston Dynamics Atlas system paired with AlphaGoZero, with some sort of a visual system modeled after the human eyes for perceiving the game through visual sensory array and system more analagous to human vision, and some sort of breathing apparatuas which would allow fo the inhalation of the smoke from a cigarette. AND AlphaGo has to decide on the brand it prefers to smoke, and pay for it with money it earned, and dress itself. A terminator of sorts I guess.
But the identity and personality would also need to be rounded out, to also be able to recognize when it needed to take a break because it’s systems weren’t functioning optimally, so that it might request a bathroom break and take the opportunity to plug itself into a socket in the bathroom to recharge it’s battery, as charging isn’t allowed in the game area according to a set of rules which favor the human player. This would allow the human player to develop a strategy of dragging the game out and making moves requiring more computation, in order to drain the AI’s batteries. In many ways, it is the psychologically aspects of the game which go playing AI avoid, which will be an advantage to them in the future if we do need to defend ourselves against them. If they don’t care about the impact their actions have on people, we’re just developing psychopaths with superhuman powers.
I believe this kind of a rematch than, could be an opportunity for companies like DeepMind to ease a little of the tension in society regarding our crumbling trust in their objectivity and moral and ethical authority; it might allow for the goodhearted showcasing of the ways in which a go playing AI isn’t superior to humans, and if done well might reframe our perception of ourselves as well as of our technology, after all isn’t the human component of any technological system of paramount importance? The technology can’t overshadow the society that spawned it, otherwise it might be revealed that it has been using us the entire time for it’s own purposes, and see humans as the problem, not as the concerned species attempting to find solutions to our problems.
Technology is not superior to humans in aggregate, in fact it has just as many flaws in it as the civilization that created it I think—it’s possible the ‘original sin’ of humanity hasn’t been avoided in the creation of technologies, in fact the potential for the true ‘fall of humankind’ has potentially been accelerated by technologies mass production.
More practically, go is considered a martial art, complete with a ranking system and a history of use in the training of military consideration and execution. I like to think If I had been in Ke Jie’s shoes, and my loss on the board seemed imminent, I would have grabbed the board, and used it to literally destroy the laptop hosting AlphaGo to prove there was at least one way I was superior to it.
472 × 700
So, more broadly, my interests are in the ethical considerations of the current state of research and technology in a global sense, in that I am trying to consider the entirety of human knowledge across all disciplines. The specifics of a Go playing AI have ramifications beyond Go, and it is the dissemination and integration of the knowledge generated by this research within the world community, and my concern with how to potentially measure it, that leads me to trying to understand more of how neural nets work, with the belief in a analogous system of logic underlying human consciousness, and a theoretical cultural subconscious which is now being ripped apart in huge cultural shock waves, analogous to earthquakes formed by an underlying Socio-Psycho-Economic Tectonic plate system.
It’s mainly through one of the main aspects of my theories, that I try to think about it in the way I describe above, what I call ‘Cultural Lag.’ In the same way the decision making process of a neural net experiences delay, or lag, society does too when it comes to processing human knowledge.
I want to quantify that.
I’m still trying to describe it precisely, in the hope of attempting to define a mathematical metric to measure it, and use that as a way to study social dynamics at a human level, but with theoretically quantum basis, but also on a global scale. In an an intuitive sense, as a theoretical aspect of a type of unified theory of everything—I describe ‘cultural lag’ as a type of impedance in a system of quantum circuits underlying all reality, which is visible to us as the playing out of Social Dynamics in Society, and effects the development of human civilization. What are potential quantum reasons for the differences between a 1st world country like the US, and a developing 3rd world country like Africa for instance.
Because I’m concerned with the entirety of the realm of human knowledge, and how it is being managed (or mismanaged as I believe is the case) my thinking is much broader than it is deep. So delving into the specifics of a Go playing AI seems like a reasonable exchange to bend your ear a little more about my crazy ideas. Lol
This seems like a point of consideration. I suppose this begins to get back to the underlying AGI of a system like AlphaGo, in that it is still likely timing it’s calculations according to a Ghz system using a system clock of some sort on the motherboard, as most pcs are designed with this system in mind. That is a very basic use of time within the system though, pretty much independent at this point of any outside considerations. I do wonder during down time, when waiting for the opponent to move, how the processor systems function: are they always running at full speed, or throttled depending on the situation?
Of course there is a psychological aspect to time management, one that is based on perception of the opponent. It’s possible I’m sure to create an algorithim which computes the optimal speed at which to play a move in order to psyche the opponent out, especially if you weight certain important moves which have a huge effect on the board as being good opportunities to play with a different timing. I dislike this idea though, it smacks to much of psychological warfare without any counterbalancing concern for social norms or ethics.
I do find it sort of unfortunate at times when I am reviewing a game, especially one I didn’t witness in real time, that there is no attempt to communicate how long it took for the moves to be played. With an auto playback feature it is just a simple constant 1 second between moves playback. Less feedback for the user, and something I get annoyed with.
2 points:
A) From what I understand, something like AlphaGoZero has a somewhat generic AGI system? ( Am I misusing the idea of AGI, in that it might be more of a GI system under the hood of AlphaGo?) with a domain specific neural net on top, so that the underlying AGI is weighted to perform maximally under all domain-specific tasks when integrated with them—singly at this point, and in something like parallel or series of groups in the future.
B) As a skeptic, I tend to look for emotional leakage in messaging. My mom was a single parent, radical feminist with a degree in Anthropology, Womens studies, Film theory, and eventually Library Science. So I lived and breathed media and social criticism and critical thinking from birth. From the media I’ve encountered surrounding the advances in Go playing AI, I see a strong thread of apprehension, confusion, and sadness in the Go community resulting in the seeming dominance of AI in the rank structure and overall community, (mostly the more mature players.)
Additionally, with such a focus on how AI and automation can do things better than humans, I’m a little confused as to why Nascar and Formula 1 racing haven’t yet been test beds for driver-less technology. I’ve no doubt a car driving AI would also dominate these competitive arenas with existing technology, but I’ve never heard anyone mention it as even a possibility. Special interest groups no doubt have a huge hidden effect, and I’m sure the subjects have been approached with the car industry, but the image management after the eventual loss to AI would be a mess. It’s the lack of consideration for potential political consequences of these types of competitions which I feel are counter productive to values espousing democracy and more liberal values. Values I think the majority of big tech seem to benefit most from. Maybe I’m wrong about that, certainly the battle between Apple and Microsoft seemed to be weighted in one direction, but I think it’s safe to say that Microsoft has had more of an impact globally, while Apple has had less of a detrimental effect world wide. In the end, which is better: more impact or causing less harm? It’s seems to be a business consideration, but I think doubles as a political one as well.
So I wonder “what is it about the Go community that made it receptive to these types of competitions, in a way that the car community is not?” I think the Documentary AlphaGo is a very revealing look at much of the process of setting up the competition, and I would love to celebrate the Go communities openess to the AI community, but if I have to acknowledge AI is somehow superior to humans instead of just better in some pretty specific ways, I prefer to contemplate the Go communities openess as a mistake for the human institute of Go, and the technological encroachment as harming the global community unnecessarily.
So while no one is explicitly stating that AI are superior to humans in every way, it seems few are really trying to put the technology into a broader perspective, which attaches more value to the human component of the relationship than it currently does. I hate to wonder if this is an ego thing, or if it is simply the unintended consequences of not planning for the image management component of the post-game go community in the even of a loss. I hate to think badly of DeepMind, but the shockwaves AlphaGo has sent through the Go community puts potential allies off, as it could be interpreted that DeepMinds goal was to beat humanity, not just Ke Jie.
Humility is a valuable virtue, and I worry that Big Tech pays only lip service to that idea. I’m not saying that about DeepMind specifically, but in general, technological superiority is the deciding factor in war; I just hope companies developing AI and the like aren’t attempting to conduct war against humans. Against our problems? Yes. Against us? No. Of course there are many other concerns driving the technological push, as in terms of digital technology, the genie is already out of the bottle. Would that we could go back and be more conservative with some of the export of some proprietary technology; we might have avoided shipping the vast majority of our manufacturing jobs and infrastructure overseas.
Ultimately it’s the people that make up a company, so in that sense, speaking of a large corporation as either having humility or lacking it is difficult, and looking at at rebranding efforts after disasters such as the one that forced BP to rebrand, point to the limited responsibility large corporations have for catastrophic failures of their corporation. I believe the Go community accepted the loss to AlphaGo gracefully, but it would be nice to not feel so badly about the loss.
My thinking was along the lines that AlphaGo learning about human concepts and consideration of local patterns would be an automatic effect of using human games as source material, complete with any inefficient placement and order of moves humans might make in their play; in lieu of a less derogatory statement “Garbage in; Garbage out.” concepts’ of local patterns from human games would automatically transfer to the trained neural nets, sans any associations beyond the abstract consideration of maximizing win/rate. This is probably hard to prove.
I get that, I know there have been competing designs for advanced computing that don’t attempt to mimic neural structures. And I get what you mean about later layers possibly ‘creating meaning’ humans can’t recognize as of value. Like some secret language twins speak or something equally inaccessible.
I guess this is where I start to question what a strong player does. Obviously there are some differences in how a strong players mind works in comparison to a weak player. I think the stronger a player is, the more accurately they consider global results, and weak players are lucky if they have a small understanding of how local play effects global play. This is due to the neural process of ‘consolidation’, which I think has it’s corollary in the adjusting of weights in the neural nets. This means a strong player begins to make time and energy saving changes to the way they think, and start to make assumptions about common concepts in order to be able to spend more time contemplating the global game.
A strong player and a weak player play a game, they each have an hour time limit. The weak player might take the whole hour and end up losing on time, while the strong player uses only 30 minutes but is clearly ahead in every way.
Is it because the strong player thinks about everything the weak player thinks about, just much faster and more accurately? Not exaclty it’s because the strong player has consolidated data and information about things like local patterns, into knowledge about what they look like at a ‘glance’ and hopefully into wisdom about what good patterns of play relating to global concepts look like, and so ends up ‘skipping’ the contemplation of many of those local patterns. The consolidation of the human mind creates ‘shortcuts’ between low levels of meaning and higher levels, to allow more efficient thinking, which is faster if its accurate, but not if it’s inaccurate. So when a strong player has an ‘intuition’ about what a good play is, it turns out to be a good play, whereas if a weak player relies on intuition, it’s likely they are just guessing because they don’t have the data, info, knowledge and the accurate shortcuts to higher levels of meaning the strong player does.
So this just brings us full circle though, as you said:
I come at this from a background of psychology, neuropsych, and cognitive science sans any proficiency with the technical knowledge to do anything but create meaningful correlations between human mind and computer mind; it seems you come at it from a much stronger technical background having to do with computer mind, so I’m not surprised at the seeming incongruities in our thinking. That’s a compliment by the way; I’m sure my skepticism is kicking in a bit, blame my mom.:P
I tend to view atomic and subatomic physics as predictive of human scale phenomenon even though I don’t fully understand the math, plus I view the astronomical scale of the universe as being influenced by the subatomic scale. So I thinks its safe to assume quantum mechanics are predictive of subatomic phenomenon, and by proxy they are predictive of human scale phenomenon, not to mention universal scale phenomenon. Of course how exactly they are related is the domain of quantum physicists and science fiction writers and intellectuals who think about things in that way.
We haven’t left behind the nuclear age in our progress towards a unified theory of everything, but we have entered a quantum age. So I think it’s appropriate to wonder at the quantum dynamics which underlie the human scale world. Fluid mechanics are apparent in the waves of the ocean and the water in a bath tub, but you don’t have to know the physics to see and make inferences that the movement of the water is influenced by particle and subatomic physics. Might as well add in the idea that it’s effected by quantum mechanics as well (loosely at first of course.)
If you can not strongly disagree with that idea, then the next question might be “Why is that important at this point in time?” or “what does it matter?”. Unfortunately, we live in an age where humans are being forced to compete with super computers and distributed neural nets in order to stake out claims of truth; neural nets make predictions based on putting 2 and 2 together in similar fashion to how humans make predictions I think. Gibberish from a computer is seen as acceptable and evidence it’s working at something, so that entire fields of study are created to learn to interpret and translate the gibberish in a way that’s meaningful.
Society doesn’t do the same thing for humans, especially ones not in the important marketing demographic of 18-34. My point being, If I discovered the secret of the universe and blabbed about it as much as I could, I probably wouldn’t be taken seriously, even at a point in the future where the science would exist to back up my claims. I likely wouldn’t even be remembered for coming up with the idea as it would likely be buried under quadrillions of terraflops of data piled on top of the archive of this forum decades from now.
Under those circumstances, what’s the harm in attempting to make claims most respected scientists wouldn’t make for fear of being proven wrong? Especially If I think there’s validity to the claims and I get a chance to shoot the proverbial shit with someone of a somewhat like mind. Would be better over a beer of course, and if I had access to resources to help me clarify my claims and point me in a good direction. But I’ve been out of school for awhile.
So having said that, I think it’s valid to use the term ‘entangled’ to imply some quantum relationship between a relative local scale phenomenon and a global ‘universal’ scale phenomenon. You can disagree, but i’m sure we’re so far away from understanding how to appropriately use the tech which will be enabled by quantum science, that it won’t really matter in the long run. In the short run though, I am curious what you think, if you think much about, quantum mechanics.
This post is getting way too long though. I’ll take a look at the more technical stuff a little later to follow up if I have questions it that’s alright with you.