For AI researcher Game AI is like porn. It's cheap tricks and obviously fake but oddly fascinating. Sometimes you find a new trick you want to try in real life.
Marvin Minsky once said "I bet the human brain is a kludge." If I had to bet, I would say that human brain is full of dirty tricks, incomplete solutions, shortcuts and artificially limited problem spaces evolved to pick berries and avoid tigers, not to understand the world. Combining many tricks together can create illusion of generality that is very convincing.
> If I had to bet, I would say that human brain is full of dirty tricks, incomplete solutions, shortcuts and artificially limited problem spaces evolved to pick berries and avoid tigers, not to understand the world.
Wikipedia has an interesting list of cognitive biases [1]. Going through these, I tend to think of all of them as heuristic failures, where those shortcuts and incomplete solutions are pushed to edge cases.
I’d go further and conjecture that our cognitive biases are necessary, that the “theoretically perfect” models that these biases are compared against when being called biases are actually spherical cows in a vacuum.
However, I’m not a neuroscientist, and my knowledge and use of A.I. is limited to hobby projects.
Necessary to what end? It could be necessary to maximize survival for the species, but that doesnt make it necessary to maximize survival for an individual.
It's perfectly possible our heuristics are muddied beyond necessity in order to generate variety in action, so as to reduce risk to the unknown for the species - even though it would cause a minority of individuals to consistently make sub-optimal choices. From a speculative standpoint, it's easy to find examples of people doing things that we consider 'stupid' but it pays off because of some unlikely event occurring in coincidence.
Reality has a lot of unknowns. There is no perfect model that could account for that. It's possible being hyper-intelligent (beyond our current ability) is (or was) a disadvantage for the species
Take depression and nihilism for example. Great intelligence can overcome the very drives that makes us want to keep living - which is an arbitrary cause, and an extremely tedious activity.
Yeah, that's the thing - the brain and body evolved for a much different environment than the one we find ourselves in today.
So some aspects which used to be advantages are now disadvantages that we have to actively manage with things like exercise, healthy diet, meditation. Which the brain fights against.
I’d take the other side of that strongly, if it was possible. The human brain is definitely not a kludge by any definition. And a lot of the tricks people think they know are not tricks at all.
Neuroscientists are the new doctors of the 50s. We thought the appendix was useless turns out it has many uses. We thought priming and all these “tricks” were things and then the crisis and Kahneman et al were debunked.
I bet the brain is just as elegant and powerful as it has to be to do the incredible complex things we do, and we’re just so far from really understanding it that we run around appendicizing all sorts of things we just don’t really know well yet.
Not at all. Kludge being the operative word. I claim the brain is fantastically not kludgy. It has incredibly flexibility, adaptiveness and no duct tape or shortcuts. It “has to be” incredibly good at so many things - it’s basically the perfect general computation machine. It “has to be” not kludgy. Minsky is part of an era where scientists were all about showing how humans were easily fallible and much of that was debunked.
Sorry but you’re not following my point or his if you think they agree.
>I claim the brain is fantastically not kludgy. It has incredibly flexibility, adaptiveness and no duct tape or shortcuts.
Why can't it be both? Maybe the duct tape and shortcuts are what give the human brain it's fantastic adaptability. Duct tape and shortcuts aren't a bad thing necessarily. Personally, I think that's what gives the human race as a whole it's fantastic adaptability. You've got millions of people each with their own duct tape and shortcuts to the same problems, meaning each of us does things just a little bit differently, we see other humans with their shortcuts and slap them onto our own with some duct tape and we get better at things or learn something new. Do this over millions of years and generations and you've got a pretty damn capable brain that's slapped together millions of years worth duct taped together solutions and skills that keep growing as we hand our giant ball of duct tape to successive generations.
I just don’t think in millions of years of evolution the shortcuts are what worked. I think our brains are the result of non shortcuts winning over a long period of time.
It is funny how willing computer scientists are to want to use a duct tape analogy. I think it’s because programming, which is basically the polar opposite of a brain (precise and unintelligent) requires so much damn duct tape if you want to get anywhere useful. Meanwhile a brain literally requires as little duct tape as possible if you want it to be generally good.
>It is funny how willing computer scientists are to want to use a duct tape analogy.
I'm not a computer scientist. I'm just going off of the way I learn and the way i've watched other people learn or the way i've seen people learn while i've been teaching them or training them. That and the general way everything kind of works. As amazing as everything in the world seems, when you break everything down to the smallest components, they all rely on the same skills and techniques we've been using since we were picking berries and hunting mammoths. The materials, accuracy and scales of our work have changed and improved dramatically, but fundamentally, most of what we do can be traced back to the same old things we've always done. We're just really good at building on layers and layers of things and applying things and knowledge to novel concepts and ideas.
Take music for an example, we just keep making the same music over and over and over again, yet we still find new and novel ways to make it sound different and new to the point where most people don't realize they've been listening to the same few songs in new forms for the last 100 years or so at least.
Your music example is just wrong. Music has changed consistently, constantly and honestly miraculously and in such beautiful ways over the years. From all the great classical geniuses to jazz to the extraordinary advances in sound design and mixing, avante garde and experimental new styles that show up year after year, I mean that’s your example? Sorry, but disagree, strongly.
> we just keep making the same music over and over and over again, yet we still find new and novel ways to make it sound different and new to the point where most people don't realize they've been listening to the same few songs in new forms for the last 100 years or so at least.
What do you mean by this? Oftentimes I see people reaching the conclusion that because a lot of modern music is built around the same major scales and largely homogenous chord progressions, it must be the same, but this simply isn't the case.
>a lot of modern music is built around the same major scales and largely homogenous chord progressions, it must be the same, but this simply isn't the case.
But it is, to the point where there's songs I learn only somewhat and i'll get confused to which lyrics are which and I'll sing a mix of the two songs when playing them, imagine and no woman no cry jump are two it happens to me with all the time. Same with truly madly deeply and kryptonite oddly enough. Then there's the whole pachabel's canon meme which is entirely true. There's a few songs, the most recent is some maroon 5 and i think a jonas brother's song which I was confused as hell as to why I liked until I realized they're just another rework of canon.
That's not even getting started on the direct ripoffs such as the 1000'@ of songs based on the Amen break or the thousands of songs that are basically a simple 8 or 12 bar blues progression or the tangled web of constant remakes that is reggae and dancehall and every song that lifts a bassline or melody from them without credit all the way up to such obscure things as the friendly neighbours tune from earthbound being a relick of the real rock riddim.
The cognitive biases linked in another comment and other general failure of our reasoning and logic (or even how our pattern matchers work too well, seeing patterns where there are none, looking at you, clown-shaped-cloud) tells me that there are still shortcuts, errors and less-than-elegant. We make mistakes in reason and thought all the time, sometimes fatally. I don't think that diminishes from how incredible the brain is and how adaptable and generalised its function is, I just don't believe that it doesn't take shortcuts given all of the errors we make.
I'm a hobbyist at sleight-of-hand magic and therefore like to read and watch lots of material on the matter (eg the book "Sleights of Mind" by neuroscientists Stephen Macknik and Susana Martinez-Conde, and science writer Sandra Blakeslee) and there are an incredibly many ways to fool or trick the brain into believing untruths, even obvious ones.
Sleights of hand are not proof the brain is failing in any way. That our visual processing is limited is not a sign that the brain is somehow taping together systems in poor ways.
Cognitive biases were also incredibly exaggerated and if you’ve read into the replication crisis and various pushback Kahneman has gotten on his studies you’ll find much of what we call cognitive bias was wrong or exaggerated.
That we figure out super inventive ways to trick others into thinking we’re not smart is just more proof of how clever we really can be.
A great example is loss aversion which, if you read the debate around, is a totally smart strategy given we have absorption barriers.
That I pay attention to things in a certain way that allows magicians to fool me in predictable ways doesn’t mean our brains are failing in some way - it means they are making a trade off that helped us survive better with the given processing power we have. But that trade off is almost certainly the smart trade off, and labeling it “kludgy” is lazy thinking.
> Sleights of hand are not proof the brain is failing in any way
You missed the point of what I was saying, which is that because of my interest in sleight of hand, I spend a lot of time reading and watching different material on "magic", the book Sleights of Mind being such material. Its not about sleight of hand, but rather that it's easy to trick the mind into believing things. Also, most sleight of hand isn't really about what our eyes see at all, but more about masking actions -- hiding one action with a bigger action or through misdirection, making use of the fact that our minds find it hard to focus on multiple things at once. I would say most sleight of hand tricks are about curating what the other person perceives. Sure, sometimes its by hiding what is being done from the eyes, but often its doing something in plain sight but in a way that the spectators don't connect the dots.
From the wikipedia synopsis of Sleights of Mind: "Macknik and Martinez-Conde say that magic tricks fool us because humans have hardwired processes of attention and awareness that are hackable. Good magicians use our inherent mental and neural limitations against us by leading us to perceive and feel what we are neurologically inclined to."
I'll give you a simple example:
In card magic, using a concept known to magicians as "time misdirection" (that is, by putting time between cause and effect), spectators often retell the performance as having happened in their hands, even when they never actually touched the cards.
The book and other material has more. The point is that our brains piece together incomplete information and make assumptions. We see what we expect to see, hear what we expect to hear. Our brains pattern match and see patterns in noise.
I'm not saying that this isn't useful to our survival, if you look at the outer ring of this image from wikipedia[1], it shows that there are very good and valuable reasons why we do these things, but they still lead us to make mistakes, errors, wrong assumptions or act illogically, sometimes to disastrous effect. I'm not and didn't say it was a "kludge" exactly, but I do think that the brain is taking less-than-perfect shortcuts (to save time, memory or make up for too much/too little/too noisy information).
I definitely understood that, I just don’t find it convincing. The pattern matching is as good as it can be, and as suited to survival as can be, and kludgy is never the term I’d use for that. It’s just an oddly negative way to frame the most competent and complex thing we don’t understand yet.
Do you agree that there are shortcuts, though, which often lead to erroneous outcomes? Like I said, I didn’t personally use the word kludge just that I think the brain is taking shortcuts and those shortcuts lead to imperfect decisions/results at times. I do agree that it’s a super complex system which we don’t yet understand, so why kludgeyness that we see may well just be something we don’t see the true nature or reason for yet.
Sounds like the right way to do things. I mean people aren’t perfectly intelligent, but being a bit stubborn about things even when you’re wrong sounds like a safe strategy in world of incomplete information. If people were less wrong-headed-confident they may not defend important other things they don’t know 100% to be right.
It’s a heuristic that can be wrong, but the heuristic itself isn’t kludgy just the occasional specific use case can be wrong.
Here’s an analogy: any individual ant does many things that look stupid for its own survival. But would you call the ants thinking “kludgy”?
> We thought the appendix was useless turns out it has many uses.
Can you elaborate?
I recently learned that current appendectomy procedure (at least in Europe) calls for the removal of the appendix even if a surgery should reveal that the organ is in fact completely healthy – because the procedure from that point is considered so unrisky, and the organ considered so useless, that the potential future risk of a medical professional misinterpreting the existence of an operational scar as an indicator of a previously performed appendectomy is deemed more problematic.
The kludginess of brains is textbook neuroscience, from instincts to reflexes, to illusions and more.
Indeed, Game AI puts the "Game" first -- as it should for Games, but to the detriment of anyone who cares more about AI and wants Games to be a fun place to study AI.
If we split Game AI into "problem solving" (like pathfinding) and "opponent personality", then we can recover a lot of good AI that generalizes beyond games, without being misled by the parts that only useful for tricking people in a toy environment.
Instead of downvoting me without a stated reason, can you provide a refutation? I'm open to changing my mind, but this is where it stands right now. I think we need to choose those we hold in esteem.
Because when you actually read it, you will download 1 article per 30 minutes. If you batch-download all of their content now, you'll create much more peak load and, hence, costs for them.
You can use --limit-rate=500k if you want to limit the bandwidth you're consuming.
And load does not translate to cost for everybody. If you saturate the connection to my VPS, I don't pay more, it just gets slower for everybody in contention. I can spin up mirrors but if I'm offering a free resource like this, I'd be more likely to limit the bandwidth-per-client-IP or just actively let it run slow. They could even limit the bandwith to the subdirectory with...
On the other hand, if you download it all at once you aren't constantly reloading a page when you get back to it and you aren't limited to reading it when you're online. I do see the point you're making, though I think it depends on the author's perspective.
Very cool. I always expect a link like this to either be some super basic examples (e.g. how to implement flocking) or articles detailing techniques used in games from ~20 years ago.
Very cool how recent and modern these are (along with super reputable authors)
Is this directly from the authors? If yes, I'm a bit shocked given the prices for the book when searching for ed 1, ed 2 and ed 3 on Google. Please add a donation button to the site.
I just finished the first four sections and I love it. Thanks a lot!
They always do this when the next version of the book is close to coming out. The first 2 have been available for free for a while. It is expected with this that the 4th is coming soon.
I've been thinking a lot about trying to make an AI for a turn based 4X game. I believe an AI that could defeat the strongest human players in (for example) Civilization would be more impressive than AlphaStar and the Dota AIs.
I think it might give the gaming industry a kick in the pants to start utilizing more advance AI techniques in general, since it seems almost all discussions of strong AI in games are dominated by apologists explaining why it's not practical. Just one example of strong AI in a successful game would change the industry.
After strong AIs are common, we can persue the even more interesting task of dumbing them down in fun ways.
Having an AI defeat the strongest human players (without cheating) isn't often the hard part, its making it fun, believable, interesting or, indeed, beatable (its no fun if you can never ever win) is often the hard part. A perfectly minmaxing AI that makes perfect decisions isn't very interesting to players. Outsmarting players in interesting ways is, well, interesting, but just always playing the best move in any given scenario (like what deep blue did with its search-based "AI") can create perfect play if the search space is within the bounds of time/memory of the AI, but that's not very interesting to play against or watch.
I don't know about that, I much prefer playing chess video games against an artificially limited AI. Granted you did say "often", and I agree this is the exception.
Starting with an AI that is weaker than human players, and then artificially limiting it even further does not produce a great AI. Although, some players still choose to play against these easiest of AIs.
I believe starting with an AI that is far stronger than any human, and then artificially limiting it will be fun. This is a fundamentally different situation. This is like a chess engine, no human can hope to win without some artificial limitations on the AI. This is a type of AI we have never experience outside of a few abstract board games like chess. I look forward to seeing AIs like this come to 4X style games such as Civilization.
Would you rather play chess against a computer you were bound to be crushed by 100 out of 100 times, or a still formidable opponent you could eke out a 25% winrate against?
> Unfortunately, the time between seeing a decision acted out and the actual act of making that decision can mean that all relevant information has already been discarded. Ideally if the entire game simu-lation could be rewound to the exact moment in time when the error occurred, it would make notoriously difficult problems to debug, trivial to understand why they occurred.
> Game engines have typically made reproducing these types of problems easier using deterministic playback methods (Dickinson 2001), where the entire state of the game simu-lation can jump back in time and resimulate the same problem over and over (Llopis 2008).
Imagine if you could do this for all programming? [from chapter 6]
The RVO chapter, and that concept I general is an amazing one because they really created a method for 2 autonomous characters avoiding collisions on a natural way - with code that is easy to understand
For AI researcher Game AI is like porn. It's cheap tricks and obviously fake but oddly fascinating. Sometimes you find a new trick you want to try in real life.
Marvin Minsky once said "I bet the human brain is a kludge." If I had to bet, I would say that human brain is full of dirty tricks, incomplete solutions, shortcuts and artificially limited problem spaces evolved to pick berries and avoid tigers, not to understand the world. Combining many tricks together can create illusion of generality that is very convincing.