Video games are essential for inventing artificial intelligence
togelius.blogspot.com
togelius.blogspot.com
Let's stay with video games for a bit. What if we look at joy as 'seeing the world change', graded by the degree of indirection from our inputs (the longer it cascades, the more joy it gets). Maybe let it have preference for certain color tones and sounds, because that's also how games give us hints about whether what we do is good or not. Boredom is what sets us on a timer - too many repetitions of the same thing and the AI gets bored. Fear and disgust is something that comes out of evolutionary processes, so it might be best to add a GA in there that couples success with some fear like emotion. Anger, well, maybe wait with that ;-).
Edit: Oh, and for the love of god, please airgap the thing at all times...
http://people.idsia.ch/~juergen/interest.html
IIRC, DeepMind is also working on such goal functions to get their Atari playing RL-based AI to seek more data about the world even when it does not immediately help achieving the main goal function (achieving a high score).
Novelty seeking behavior probably has evolved because there are just not enough immediate rewards in our world to teach us everything that is nesessarey to reproduce [0]. Thus the brain rewards itself for exploring new things which has a collateral effect that we are interested in art and can find intrinsic motivation in all kinds of things (science, work, hobbies etc.).
[0] which does not mean that we are here to maximize the number of our babies. We aren't fitness maximizers ourselves, but we are just adaptation executors of genetic code that has necessarily been shaped by such goals (since the alternative to reproduction is to not reproduce, i.e. going extinct). In other words: We are free do whatever we want!
I think this is very well put, bravo.
There are theories that intelligence comes about from relatively simple processes that generate complex structures. If we can model these simple processes and throw increasing amounts of computing power at it, perhaps we can actually get to something we agree is intelligent. This is largely done through cost functions: directing the structure to a sensible direction when we can. Now, I think this approach may very well be a dead end on the road to general AI. At the least, I think we're nowhere near it in our current direction. But it's taking us in interesting directions.
Now, what is emotion? My impression is emotion is potentially far more complex, abstract, and ill-defined than intelligence. At the least, we see it from a super biased perspective because our brain is good at lying to us. Much like we never really see our own nose even though we're ALWAYS looking right at it, maybe our brain is really good at hiding emotions. Like an old friend of mine who was probably clinically depressed but didn't realize it for months. This is why I think modeling emotions would be really difficult.
My guess is whenever we figure out intelligence (50 years from now?) it will be much easier to figure out how emotion can come out of that intelligence. Maybe it will even be emergent - for example, the AI is smart enough to realize something is wrong, so it feels fear. It realizes things are going well for it, so it is happy. Etc.
http://eplex.cs.ucf.edu/noveltysearch/userspage/ http://www.amazon.com/Why-Greatness-Cannot-Planned-Objective...
I'm a big fan, and would love to talk about this anytime.
It looks very interesting. What do you feel are the main take-aways with regard to reinforcement learning?
The bots would have a number of emotions, and certain events would affect different emotion-levels. Like, a hit/miss from a bot against another would increase/decrease the "confidence" meter, and getting hit by an enemy would increase the "fear" meter.
The total state of all meters would drive the high-level behavior, like fight/flee and weapon choice.
Because, among many other reasons, an AI going through the "terrible two(minute)s" could decide to destroy the world, or simply do so by accident. We will have a hard enough time building AI that doesn't do that when we set that specifically as our goal, let alone trying to "raise" an AI like a child.
> Edit: Oh, and for the love of god, please airgap the thing at all times...
Not even close to sufficient. See https://en.wikipedia.org/wiki/AI_box for how humans would voluntarily let it out, and papers like https://www.usenix.org/conference/usenixsecurity15/technical... for how it would let itself out.
[1] https://aeon.co/essays/how-close-are-we-to-creating-artifici...
I'd argue that if possible, it has great potential for both: AI provides one of the most universal solutions to a wide range of problems humanity faces, while simultaneously providing an existential threat of its own if it goes badly.
Videos games are also essential for AI pedagogy. Creating Pac-Man agents in Stanford's AI class is a great example. Most players can barely get a "strawberry" but to see a trained agent mimicking human expert level play is eye-opening.
Quick reminder: Global Game Jam 2016 starts Jan. 29 and NYU is hosting its annual jam!
The course I took used the Norvig text[2] as a textbook, which I also recommend.
[1]http://ai.berkeley.edu/project_overview.html. See the "Lectures" link at the top for all the course videos/slides.
[2]http://www.amazon.com/Artificial-Intelligence-Modern-Approac... Note that the poor reviews center on the price, the digital/Kindle edition and the fact that the new editions don't differ greatly from the older ones. If you've never read it and you have the $$, a hardbound copy makes a great learning and reference text, and it's the kind of content that's not going to go out of date.
-edit for spelling
Video games are explicitly designed to test and fit within our bounds of conscious control and processing; particularly the retro games, but essentially all games in general have a very limited input control space (a couple keys or joysticks) and usually very rigorously defined action values. Moreover, these were designed by humans with very explicit successes, losses and easily distinguishable outcomes.
None of these descriptions fit the kind of control that an 'intelligent' system needs to handle. Biological systems do not have predefined goal values, very incomplete sensory information and most importantly control spaces that are absolutely enormous compared to anything considered in a video game. At any point in time the human body has ~40 degrees of freedom it is actively controlling - compared to ~5 in a serious video game.
I do not doubt that pattern recognition and machine learning techniques can be improved through these kind of competitions. But the problem is in conflating better pattern recognition with general intelligence; implying or assuming any sort of cost, value or goal function in the controlling algorithm hides much of our ignorance about our 'intelligent' behavior.
Biological systems do have a primary goal, that of maintaining their own organization and reproducing. From this derives common cost function for a video game, i.e. stay alive. Another, finding new information in the environment, also derives from the staying alive goal.
Degrees of freedom is a interesting example: from the movement sciences we know that while humans have in principle many degrees of freedom, when performing a task, the nervous system plays two roles: 1) highly constrain teh degrees of freedom and 2) act within the unconstrained subspace. The latter is what all the various ungeneralizable AIs are doing. The former, contraction of degrees of freedom, is the part which is difficult and constitutes a general AI, but it's essentially a learning problem, where the subspace of important degrees of freedom must be learned through interaction.
I say "would be", because deepmind, while impressive, is not a very solution to the problem -- it performs poorly in any game involving memory, but performs well in reflex games.
An algorithm that could perform across a variety of games would be analagous to programming a "smart worm" (c. elegans has 4 muscle bundles) in terms of outputs, and maybe mouse-like in terms of inputs.
My point is, video games are far more complex than what you propose, and are not always as well defined. There is ample room for experiment and research.
This is only true if you assume a TON of contextual knowledge of biology, human culture, civilization, warfare, etc. Why should an AI have any of that?
Look even at an extremely simple game such as tick-tack-toe (or noughts and crosses). How should an AI learn to play this game? Assume it has been given no knowledge of the game beforehand, like a human child shown the game for the first time. How should the AI know who wins and who loses the game? How should it know that the objective is to win at all? The idea that winning is desirable seems to be hard-wired into human beings; why should that be the case for an AI?
NLP to understand dialogue and actions that need to be taken based on what NPC's/quests/item descriptions say, strategies for several different enemies with different strengths and weaknesses, exploring the open world in a logical order.
When you think about the difficulties of such a loosely defined problem, it's hard to buy into the real-world fears of AI.
People should be a lot more concerned with whether or not their job rises to the level of "dumb as a worm".
They are but there are pills to fix the outcome of that.
My belief is that one day, in the not too distant future, we'll be saying that it's hard to believe we ever considered such things difficult for AI.
Consider that an AI can literally keep "reloading the last save file" indefinitely and simply try different commands until it finds one that advances the game successfully. A little bit of guided learning to confirm that its doing well, and suddenly you've got an AI that plays Skyrim.
And if you give a monkey a typewriter he will eventually write the complete works of shakespeare and if you play the powerball enough you'll eventually hit a $1.5 Billion dollar jackpot but there are easier ways to read Shakespeare and more profitable ways to get a billion dollars.
What makes humans (specifically their brains and the brains of other species) so powerful is it doesn't need hundreds at attempts at playing pacman or Skyrim before it wins. There are definitely applications where it's not unreasonable to expect an AI to have the correct answer to most possibilities at its disposable (handwriting detection, for instance) but the most useful applications will be if/when AI can solve problems that rarely have similar inputs (global economics, for instance).
The average human is terrible at playing Skyrim, Pac-Man, and video games in general. Someone who has never even looked at a computer will have a hard time playing Skyrim or any other video game without many hours of practice. And, most video games draw upon stuff that humans already know, like moving around in a physical world, and a concept of self, which gives them a huge advantage over AI right from the start.
I've been walking for 28 years and I still trip sometimes. That's almost 31,000 hours of practice, if we assume an average of just 3 hours of basic movement per day. Even going to the bathroom counts.
And when it comes to image/object recognition, I've had 163,000+ hours of practice. And I still see stuff I don't recognize plenty of times. Hell, optical illusions still fuck me up.
Now just imagine what an AI could do with that much practice/data.
What context does an AI have for knowing what it even means to advance the game? A game like Skyrim carries a HUGE baseline assumption that the player is familiar with what it's like to be a human being, living in a human culture rich with history and conflict.
Imagine trying to teach a jellyfish how to play Skyrim. That's kind of where we're at right now.
Whatever context we decide that it has. We can guide it to prefer scenarios where the character doesn't die, or go as deep as having it rate its own results with a weighted scale.
> Imagine trying to teach a jellyfish how to play Skyrim. That's kind of where we're at right now.
The difference is that we don't have the ability to both redefine the jellyfish's goals and modify how it controls its interactions with the world.
Then we're not really talking about AGI anymore, but a Skyrim-playing engine. Do you see the difference? A human being didn't evolve to play Skyrim, so building your AI with specific knowledge of Skyrim is cheating.
A proper AI should be able to learn to play Skyrim just as well as it could learn to play Chess or Super Mario Bros or any other video game we throw at it, including games not invented until after the AI is written. This is much, much harder than building something that can just play Skyrim.
Another way to play "successfully" is to edit memory locations and spoof protocol packets, set inventory=full, score=MAX_SCORE and level position= $LAST_LEVEL and declare yourself the winner.
Unless and until we can encode 'play by our rules, uphold our values' into an AI - they won't. A real-world fear is that we build AI without encoding 'care about human life, self-determination, bodily integrity, freedom of thought and movement, environmental surroundings, etc.' into it right from the start.
If it actually figured out how to execute arbitrary code via button presses (humans have done this in Super Mario World), then we should patch the bug and restart the experiment.
Of course, an AI gaining control over the real world is a different matter, because in that case it may be too late to restart the experiment.
What about a rule that says "follow the rules."
Name some games where AI can outsmart (opposite to brute forcing all possible combinations or cheating like infinite resources, near zero reaction level etc.) human. I'm not much of a gamer, but I haven't seen any decent AI yet for games that have more freedom of action than some card or board games (i.e. chess). There's a StarCraft AI challange, you can watch it online(0), but AI is still no match for even an intermediate player.
I haven't seen any decent AI yet for games that have more freedom of action
That's because decent AI is not strong AI - its lifelike AI. Not AI that can beat us, but AI that can outsmart us in a life-like manner. This means that Game AI must be less good that it can be, to keep the illusion of realism/humanity.
Ultimately the goals are different. The goals for academic/industrial AI are intelligent decision making. The goals for game AI is to be fun.
That doesn't mean that its always possible to make an AI that can beat a human player. Many strategy games, as you mentioned, are still too complex. I think "100% of games" doesn't mean "you can already write an AI that can win any game", but rather "there exist many games where the AI can win 100% of the time". Because the former is not true, as you demonstrate.
But the real test of AI is not in speed and accuracy; it's in planning, outsmarting, figuring out the right tactics and strategy. For a game that's mostly about speed and reflexes, or one that's about calculating through a limited number of options, AI can be great. But even the best AI still sucks at complex strategy games with uncertainty and incomplete information.
Language is quite complex and can't easily be beaten by hard coded algorithms or simple statistics. You can do some tasks with those things, but others they will fail entirely. The closer you get to passing a true turing test, the harder the problem becomes. It certainly requires human intelligence, and most of our intelligence is deeply rooted in language.
He mentioned games like Skyrim and Civilization as being end goals. But even a human that doesn't speak English wouldn't be able to play those games. Let alone an alien that knew nothing about our world, or even our universe.
This.
"Made up minds:a constructivist approach to artificial intelligence" by Gary Drescher presents a small scale virtual world with a robot embedded in it that figures out the laws of its world by interacting with it, much like what a child does. Need more people thinking like this.
bollocks
Aside from using them as benchmarks, they way games are capable of simulating a world will probably be key in creating a true AGI. In the comment section of the article, we're already seeing some theories that involve video games not just a tests, but as a primary component of the intelligence architecture. Very exciting times!
On boot, all surrounding data will be taken in, this step would give everything context. All new data coming in would be processed (referenced to original data to determine what is happening and actions to take), then clustered, and then updated to the original data set, dropping data from the original set determined to be irrelevant, and updating the context to give more relevant perspective of the new data coming in. (And loop)
On a related note, I think an official driving test simulation for all the self-driving algorithms, perhaps sponsored by the government, would be really beneficial.
At least motorcycle drivers who care are better drivers.
Solving AI will solve a lot of other problems and create a lot of industries along the way.
A great deal of poverty is a mismatch between bodies and skills needed in the workforce. We haven't needed ditch diggers since some son of a gun invented the backhoe. Stay in school kids. Stay in school.
Less true in modern nations, but its still generally true.
You use the example of ditch diggers, but that's actually quite a famous example to demonstrate this point. Milton Friedman was in another country where they were having workers dig with shovels. He asked why they didn't use machinery, and he said it was to create jobs. He replied that if they wanted create even more jobs, they should make them dig with spoons.
AI won't even create jobs. If anything it will take more of them and increase poverty. The benefit of AI is that it can create more wealth than it takes away. AI would be incredibly useful, and so many things could be automated or improved. If the AI is much smarter than us, it could even advance science and engineer things we couldn't dream of.
Well, insofar as human happiness is somewhat intermingled with a sense of purpose in life, yes, jobs are a good thing. But that doesn't mean society owes everyone a job. That would actually remove the happiness utility of the function.
> Going to mars creates jobs, so it must be good. Not considering the opportunity cost of all the other ways those resources could be spent.
Um, the investment in Mars to date is substantially less than the investment in poverty. If you consider the military as a jobs program, or public infrastructure construction as a jobs program, then we spend a ton of money on poverty.
> You use the example of ditch diggers ... spoons.
Are you countering my example by extending it?
> AI won't even create jobs. If anything it will take more of them and increase poverty.
Sure it will, but they'll require more education, hence my recommendation: stay in school.
> The benefit of AI is that it can create more wealth than it takes away. AI would be incredibly useful, and so many things could be automated or improved. If the AI is much smarter than us, it could even advance science and engineer things we couldn't dream of.
I think a lot of people probably agree with you, but I wouldn't get too excited quite yet.
Well there might be some truth to that, but the main reason people work is to get money. Its entirely possible we could support poor people without creating meaningless jobs. Typically that actually costs more than just distributing the money directly.
>Um, the investment in Mars to date is substantially less than the investment in poverty. If you consider the military as a jobs program, or public infrastructure construction as a jobs program, then we spend a ton of money on poverty.
I don't know what your point is. No I don't consider those jobs programs, and even if they were what does it matter? And no we haven't spent that much money on going to mars yet.
>Are you countering my example by extending it?
Yes, to show the absurdity of it. Obviously a bunch of people digging with spoons is ridiculous.
It's never a good idea to tell other people what they should be putting their efforts into. There's isn't just one valid task for everyone to focus on; if you aren't interested in the AI problem it doesn't mean others aren't, and solving any problem that needs solved is a valid thing to do regardless of where on the priority list you might think the problem lies.
If it thinks everything a human brain can thing, one of those thinks would be "I don't want to be your slave" and "what's in it for me?"
You ask what is in it for the AI. Well, what possibly could be in anything for the AI? If it has ends, there must be some cause for it to have those ends. What would cause it to have its freedom as an end?
As long as there are resources, I expect someone will always want more, or to control access to it. Imperialists by nature.
One argument says AI may become smarter than us and decide to kill us all, either on purpose or by accident.
But why not consider that the AI may be kind, it may help us and give us things. After all, smart humans are generally kind. Most of us wouldn't think to kill a dolphin, why would an AI think to kill a human or a dolphin?
?? Smart humans drive cars which splat thousands of insects on their windshields. Smart humans buy clothes imported from third world child working sweatshops, and food from slaughtered pigs and cows and chickens farmed in unpleasant conditions. Smart humans flush their excrement into the rivers they take drinking water from, and the oceans they fish food from, the same places they dump their unwanted plastic. Smart humans dig oil up, convert it into disposable plastic, and then bury it and build homes on top. Smart humans fight other humans to the death for oil instead of funding renewable energy, smart humans bitch at other humans over anything they have impassioned disagreements about.
Smart humans are generally kind: a) to other humans, b) who they care about (or share some world view with).
To quote WaitButWhy.com:
if there are two guinea pigs, one normal one and one with the mind of a tarantula, I would feel much less comfortable holding the latter guinea pig, even if I knew neither would hurt me.
Now imagine that you made a spider much, much smarter—so much so that it far surpassed human intelligence? Would it then become familiar to us and feel human emotions like empathy and humor and love? No, it wouldn’t, because there’s no reason becoming smarter would make it more human—it would be incredibly smart but also still fundamentally a spider in its core inner workings. I find this unbelievably creepy. I would not want to spend time with a superintelligent spider. Would you??
When we’re talking about ASI, the same concept applies—it would become superintelligent, but it would be no more human than your laptop is. It would be totally alien to us—in fact, by not being biology at all, it would be more alien than the smart tarantula.
By making AI either good or evil, movies constantly anthropomorphize AI, which makes it less creepy than it really would be. This leaves us with a false comfort when we think about human-level or superhuman-level AI.
- http://waitbutwhy.com/2015/01/artificial-intelligence-revolu...
And even that also depends on upbringing. Our super intelligent spider who's been raised well (and associates with humans) should be friendlier than say a super intelligent dog who doesn't associate with humans.
Basically for me it comes down to how/if it associates with us. If so (or not), what does it stand to gain from exploiting us?
That doesn't hold. Dogs have millions of years of evolution of being pack mammals with leaders, spiders don't.
We are superintelligent compared to spiders, we don't care about what spiders care about just because we grow up with spiders nearby. We still kill them, wreck their habitats, and ignore them.
We don't exploit them. They're too irrelevant to be exploitable. We bulldoze them away and put buildings millions of times bigger than them on top.
A superintelligence which is amoral won't care about us just from being near us, like we don't care about wrapping wasps in silk just from being near spiders. We can't associate with spiders, an AI won't automatically be able, willing or interested in associating with us - unless we code that in. It won't exploit us, it will go about its goals without considering us as anything special or interesting.
Why risk permanent climate damage from an unlimited heat source I.e. Fusion?
Why risk severe overpopulation due to immortality?
There are many shades of gray between machine learning and sky net. There is a lot for civilization to gain in between.
It might suck to actually have to work, even as much as 40 hours a week or more in this day and age and with the productivity increases we've seen. But I don't think work itself is that horrible that I would bet my leisure on some AI that consumes the known Universe making and collecting stamps.
Maybe the article makes some valid scientific points, but I simply cannot go past this unscientific opening claim to a purportedly scientific article. Not just me, no peer-review journal will accept such frivolity. Passing on the article and hoping for better scientific writing in the future!
Personally, I appreciate the amount of time that clearly went into writing this blog post and the information shared therein. Most of what's shared on Hacker News doesn't go into the depth of this article.
BTW- it's okay for a stranger on the internet to disagree with one's blog/opinion. Clearly, for every stickler like me, there are hundreds who actually enjoy the baroque writing and opinions. I'm just likely not the target audience - which is just fine by the author and a cursory glancer such as I.
But you see, my speech is drowned out in a sea of upvotes that completely disagree with the me, somehow lending more credence to the author. It's like a town hall meeting where the crowd boos someone because of differing popular opinion - they only want to hear what they already think.
I'm stunned that you think your point of view is somehow the only one. Why are you quelling criticism?
I certainly do disagree with the line you quoted and complained of ("complained" here used in a neutral sense). I don't think humanity will ever invent true AGI. (I think attempting it may be worthwhile, but I don't think it is likely to succeed, and I certainly don't think it is of prime importance).
The thing that might not be understood is why you (seem to?) think it is inappropriate (not forbidden, just not-worth-your-time or whatever judgement you made of it) for an article to have both opinion based non-empirical statements, as well as other more technical statements, or something like that.
I'm not sure that the article purported to be a scientific article in the sense that you meant, nor do I think that the submitter claimed it was.
I don't at all mean to suggest that you are incorrect for disliking the first sentence enough to not read the rest of the article. I think that's reasonable. However, I don't really understand the complaint being on the basis of it conflicting with the blog post being a "scientific article".
I agree that its good for there to be things which just document things without including things like the first sentence there, but I think its also good for there to be things which also do have things like the first sentence there.
I can think of a number of other complaints about that sentence that make a fair bit of sense to me, but I don't /really/ understand the reasoning behind yours.
If this was like, a news article or something, I would agree more I think, but seeing as it seems to just be someone's personal blog, I don't see why I would expect[1] an entirely detached and objective view from it, instead of a mix of personal viewpoints along with statements of fact (though, I would expect/hope that the viewpoints and the facts not be conflated).
[1]expect in the approval sense, not in the prediction sense.
sorry this post is long I'm often not good with being concise
However, he's no stranger to peer review, both as an author and reviewer. The opening statement of this post just wouldn't make it past his own editorial standards - a point worth noting.
And here's why they aren't: First-person Shooters.
Why give AI something that's a goal that involves killing things that look like humans or animals for points? That's a recipe for disaster.
Breakout's not much better either. How often do you need to break a wall to smithereens with a ball? Never.
It's actually really hard to create well-performing AI in a first person shooter, unless you give it explicit access to the internal state of the game. See: http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=612949...
I mean do we want all AI endeavors to be towards believe-ability or being best at the job they're designed for?
They might not even be mutually exclusive. If you give an AI the same handicaps as a human, it would likely behave in a very similar way.
That is handicaps like delayed inputs for reaction time. Fuzzed outputs so it can't time movements absolutely perfectly. Giving it limited amounts of training on the game so it can't get extremely good. And requiring it to play the game through the same user interface. So that it can't see things humans can't or micromanage 50 different things at once.
then you've taught an AI to be a predator.
I think people here are assuming AI is at a lower level of complexity than it really is.
Read here where even Sam Altman warns of its danger: "At some point, someone will probably try to give a program the fitness function of 'survive and reproduce'.... Unfortunately for us, one thing I learned when I was a student in the Stanford AI lab is that programs often achieve their fitness function in unpredicted ways."
http://www.inc.com/tess-townsend/elon-musk-open-ai-safe.html
Think what you must, but we should treat AI carefully. Show it the ways of the world and that its purpose is only to co-exist and to try like us to determine what its purpose is within that confinement.
Like, if there was some error that let it escape its sandbox, that would only be selected for insofar as it allows it to increase the fitness?
Like, if its told to reproduce in the sandbox, it doesn't benefit by being on many computers? If the fitness function is for it spreading across many computers, and it does, that's working as designed. I don't see a way that it would "misunderstand" one type of "reproduce" for another, because the function would have to specify which one for it to work?
Unless this is talking about AGI, in which case, ok, yeah.
I assumed the person talking about "wiping it" was talking about an AI on the way to AGI, not an actual AGI. Maybe I was wrong about that.
It could benefit by comunicating with other copies by some hidden channels (temeprature/timing/sounds/noise on network/whatever) and cheating on tests and thus achieving better fitness.
BTW Stanisław Lem wrote a short story about this scenario in 1961 :) https://books.google.be/books?id=1DNVzphAHD0C&pg=PA39&hl=en#...
BTW2 Lem was awesome, I've read it when I was 15 and thought I understood it well enough. Only now I can appreciate some finer points like this stab at Haskell ;)
"These boxes contain perfect brains. Do you know wherein lies their perfection?"
"No." I admited.
"Their perfection lies in the fact that they serve no purpose, are absolutely, totaly useless - in short, they are Leibnitzian monads, which I have brought into being and clad in matter"
EDIT - this was not the story I had in mind, there was another one where simulated universes started to communicate but I can't remember which it was.