55 replies to this topic

[James Trotter]

Quote:

Original post by Metahawk I am writing a dissertation on games AI and am interested in frequent gamers opinions on the current state of game AI.

Very cool! Will I be able to read it when you're done?

I think that most people are in agreement here. Graphics have indeed seen amazing advancement in recent years. A lot of focus is on the graphical aspect of a game, more so than AI. So, yes, AI has been neglected in a lot of games.

I've played alot of RPG games, and the AI in that genre is mostly concerned with NPC interaction. I recall specifically two games which I played, the one being Morrowind, the other being Gothic 2. Morrowind had stunning graphics, and was enjoyable to play... for a little while. Gothic 2, at first, was disappointing when comparing the visuals with Morrowind. However, at length, Gothic 2 was much more enjoyable, (this is likely a matter of taste, though), and really served to will the suspension of disbelief. This was all because of the NPCs interacted in a much more believable way.
Let me give you a simple example: In Gothic 2, while walking around in the city, you would notice people mingling in groups, actually having conversations with each other. In morrowind, there is absolutely no interaction between NPCs, whatsoever. The environment in Gothic 2 is much more believable.

Does anyone else have any good examples?

[Zodiak]

Well, I just think that when one DOES create an awsome AI, the gaming industry will get that boost like with physics. What I mean, is if you haven't heard of a certain game, and then it comes out with average graphics etc. but with some awsome AI system, it will immediately be noticed... It's just a matter of making an excellent AI system, that's all. Because above all, people tend to admire AI the most when it's there. So the game will be a huge success, with millions of people going like 'Oh, Joe, check it out... Yesterday that <beep> had actually outsmarted me... it's so fun! I created a line of defence in the north, but that <beep> anticipated it and attacked me from the rear..." something like that. It's way way more fun to play with good AI than with graphics + physics because without AI they mean nothing
(unless we are talking about multiplayer).

[James Trotter]

Quote:

Original post by Zodiak Well, I just think that when one DOES create an awsome AI, the gaming industry will get that boost like with physics. What I mean, is if you haven't heard of a certain game, and then it comes out with average graphics etc. but with some awsome AI system, it will immediately be noticed... It's just a matter of making an excellent AI system, that's all. Because above all, people tend to admire AI the most when it's there. So the game will be a huge success, with millions of people going like 'Oh, Joe, check it out... Yesterday that <beep> had actually outsmarted me... it's so fun! I created a line of defence in the north, but that <beep> anticipated it and attacked me from the rear..." something like that.

Well, this is debatable. I get the shit kicked out of me if I play against anything other than an 'easy' computer opponent in Warcraft 3. Does that mean the game has good AI? I think RTS games generally are not the games which are in need of improvement in their AI.

Quote:

(unless we are talking about multiplayer).

That's a very good point. I can't imagine that a lot of MMORPGs, for instance, would require the same kind of AI as would a single player game. I suppose the other players make up for it, and give you all the 'intelligence' you need.

[AlbertoT]

"Well, I just think that when one DOES create an awsome AI, the gaming industry will get that boost like with physics "

I definitly agree with you, in my opinion A.I. is, better ,should be more important than phisics.
The point is,it it just a matter a good will ?

"I created a line of defence in the north, but that <beep> anticipated it and attacked me from the rear..." something like that."

Cool, but is possible to do something like that?
If you can not distinguish a real opponent from a virtual one you have passed the Turing test.
Do not forget that no machines passed the Turing test, yet

[Zodiak]

No, no, no! See, that's what I tried to explain. Nowdays a lot of games just make your foes healthier, aiming better, faster etc. when increasing the difficulty level. They don't change the AI, they just make it harder for you to fight your enemies.

What I am talking about is creating a great AI SYSTEM, which will enable your computer opponents to actually 'THINK'. So they might flank, they might concentrate on defence... It's really hard to explain. The ultimate goal is to make a challenging AI, such AI that a player won't 'know' that he is playing with AI rather than real-human. Believe, it's not some sci-fi movies, it's possible. Just make a human AI. Humans make a lot of mistakes, that's just a matter of keeping the AI real. For example, in many shooters AI actually know you are there even though you are hiding, so that it (AI) can plan his actions on certain circumstances... Something like that.

It's a really big topic, and I hate games with unrealistic AI. You know, there's a good example. My dad is absoultely not a computer guy, so when he sees me playing a game (about war, for example) he makes absolutely logical suggestions on what should I do next. But the problem is: his real-life exp. => suggestions WON'T WORK in that game! They just won't! It's no use, FOR EXAMPLE, building extensive defence lines, flanking your opponents (just an example), because I know it won't make any difference. So to play and beat games we actually have to be GAME-AI-SMART rather than real-life smart, even in realistic games. *I* know how they work, but make a soldier play something like FSW or BiA or any other *realistic* game and he will fail, because he uses real-life exp. and logic, while I know that the enemy WILL NEVER in *that* window... Something like that, you know...

We need to make a human AI which does not mean we has to be as smart as us (I think it will never ever be possible), but as *real* and somewhat even *stupid* as us. That way it'll be fun.

[BrianL]

Quote:

They don't change the AI, they just make it harder for you to fight your enemies.

This is frequently intentional. Game companies have a limited amount of time to create behaviors/assets/sounds for games. Any time you design content only a subset of the audience will see, a strong argument can be made that you are wasting time that could better be spent on content everyone will see.

Quote:

The ultimate goal is to make a challenging AI, such AI that a player won't 'know' that he is playing with AI rather than real-human

Minor detail, but those are really two very different goals. A challenging AI is much different than a human-like AI.

Quote:

while I know that the enemy WILL NEVER in *that* window... Something like that, you know...

This is a great example of a game design decision/game tech limitation than an AI limitation though. There are a lot of very valid issues people here have with gameplay; I think many come down to game design more than AI tech limitations.

Guys, yes, graphics have made massive advances, but so has AI. Graphic advances are simply easier to see. Does anyone remember when Quake2 AI was marketed as advanced because the AI was capable of ducking? Constrast that with Halo 2 for instance. There is a massive AI tech improvement there. The fact that graphics improvements are easier to see is a strong argument for continued work in that area.

Many game companies have just as many AI programmers as graphics programmers (1 or 2 people working for 1 to 4 years). It isn't a lack of focus, its a lack of low hanging fruit. AI improvements take a lot of time, are higher risk, and are typically higher risk to game design.

I also don't believe it is a lack of peoples knowledge about AI technologies. Technologies are just tools. In first person games for instance, the bigger issue is how the AI interacts with 'external' systems such as animation, sound, etc. We can't do completely procedural content generation, so we are restricted to use of assets built by artists, animators, sound designers, and level designers. These define the boundaries of what the AI is capable of, which rapidly becomes a buisness issue.

Take a look at: http://stuffo.howstuffworks.com/halo2-ai.htm This is a pretty article about where game AI is at right now, particularly how it relates to buisness decisions and game design.

[uutee]

My opinion, as an experienced gfx/physics progammer and as a less-experienced AI programmer, is that AI is more hard to program. It really requires your "own thought" based on the exact game you're programming - there are no "cookbook recipes" as there are for gfx/physics things.

Therefore, it might be that it's easier for developers to concentrate on graphics. Want outstanding, state-of-the art graphics? Just buy the GPU gems book and start implementing the recipes! Want believable, intelligent and creative AI behavior? "It's up to you."

-- Mikko

P.S. Yes, I know certain algorithms such as ANNs, GAs and such, are fairly established, but in the end they are just tools - applying them to your exact problem requires more thought.

[frostburn]

I'm not sure if AI is being neglected because of graphics, but I feel that a lot of work could be put into AI to make games better. Here are my observations and ideas:

I agree with Steadtler that graphics sells the game, but the AI is what makes it great. Game boxes has pictures of the graphics on the back, and in features it's listed "Advanced AI". What do people see? IMO the pictures. It's when they play the game they see how the AI is. If a game has good graphics (sub-HL2/Doom3), but a very realistic and life-like AI that retreats and calls for backup and such if faced with overwhelming odds, then I would like that better than a game with photorealistic graphics, but enemies that just swarms me.

I also agree with James Trotter, in that the non-enemy NPCs in Gothic 2 looks more realistic than those in Morrowind. I'm not saying the AI is better, as I can't remember them walking around and actually doing stuff, but they stand and talk with each other which makes them more life like than those in Morrowind who just stand in a spot or patrols an area all day long.

A FPS where the enemy's AI analyses a "render" of what he sees and identifies the player (or his team-mates) visually would be great. AFAIK in most games just let's the AI know where the player is even if he's perfectly camoflagued and standing still and otherwise made no action to reveal his position. Their advantage can be compared to the enemies being rendered with a pure red color and a pointing arrow on the player's screen. Some games, most notably the Thief games, Splinter Cell games and some of the Delta Force games is taking cover and concealment into concideration, but more often than not it's too black and white. If the player stands still in the shadow in front of a white, lit wall the enemies still won't see him, even though he should have a very distinct silhouette. I doubt making "AI-sight" would be easy, but it's certainly something that would add realism. Even in a simplified form that takes into account camoflague, silhouette, light, movement and other things like that would be better than what is in most games. Sound is also a thing the enemy should be aware of, but not necesserily be enough to pin-point the player's position (it might at higher difficulties).

Enemy cooperation and needs:
In all genres the enemy should be able to retreat and look for cover or assistance if he feels threathened. I don't think it would be hard to implement (at least calling for help), and would add realism to games where they today mostly just runs heedlessly towards the player, guns blazing. Again, there are games that are better than others, but in most cases the only need the enemy has is to kill the player. If he at the same time tried to stay alive then there wouldn't have to be an overwhelming and sometimes respawning number of enemies to make the game challenging. Team tactics with cover fire and advancing to cover would make it more challenging to hold a position. Most games make this act a shooting gallery where the only thing the player needs to to is to avoid running out of ammo. Adding realism could be as easy as placing the enemy behind corners and having them pop out and fire a burst once in a while. Then the player needs to aquire the targets and shoot, and at the same time avoid being shot from one of the other corners. Keeping the crosshairs in a single spot and hope to catch the enemy as he pops out is simple, while doing the same when the enemy has the player pinned from multiple spots is foolhardy.
Speaking of ammo, it seems that enemies in FPS games has an infinite amount of it. They hardly ever fire bursts or single shot, even though in real life that would increase accuracy and lower ammo consumption. Most games penealies the player by increasing spread if he keeps the trigger depressed, but I'm not sure if it's the same with the AI enemies. IMO it should, and the AI should compensate by making every shot count (of course there are times for full auto as well). If the AI has to find more ammo if he runs out would also increase realism. This would be easy enough by setting waypoint and a path to the armory or a stockpile (which doesn't have to be accessible to the player if it should be hard for the player to find ammo).

AI in RPGs:
For the most part cities in RPGs tend to be populated by a lot of people just standing around having nothing to do. In Gothic they stand around talking all the time, but they don't stop talking after a while and go to do some work or running errands. In Morrowind they stand around and waits for someone to walk up to them and talk, or in the case of guards patrol an area. They don't ever walk up to another NPC and talk to them, or walk from a building to another. Most other RPGs are like this as well. Giving the people a list of places to go, and letting them stop and talk about rumors to people they meet wouldn't be very hard to do, and would add a lot of realism. It doesn't have to be extensive, just enough to give impression that they are real people. Farmers would get up at sunrise and go milk the cows, eat breakfast, tend the fields, eat dinner, go to bed, etc. People in cities would get up, go to their store and open it up, go to the market, go and eat, go home and go to bed etc. Guards could get out of bed, have an inspection, guard change (perhaps exchange some words), tend their equipment, go to a tavern, go to bed etc. Townspeople could wander around to various stores and chat with the store-keeper and perhaps buy stuff, to the market etc.
In most cases it could be a simple list of what to do at what time. They could have a list of people they know, and rumors they know about, and stop for a chat if they meet. In most cases the player wouldn't notice it if everything is the same each day, but if they do nothing at all each day, then it would be noticable.
For even more realism, the shopkeepers could walk around in their store and arrange their goods and other things like that until the player walks up to the counter. Of course this depends on the type of store. Pharmacists or Alchemists could be working in their labs, armorers and smiths could work in their smithy, scholars and mages could read books etc.
At night the thieves could go out of the thief guild and sneak around in the town. At day pickpockets could be roaming the streets and target passers by. If the player notices them and follows them to the thieves guild it would be a lot more realistic than if he just asks a person on the street where the hidden guild is located.

Monsters in RPGs tend to be incridibly stupid. They mindlessly attack the player who walk in on them and never try to flee if they're wounded. In some cases it's ok, but in some cases it would be more realistic if they decide to run if they're out of their league, and just fight if escape isn't an option. Even if they're initially aggressive, they might want to flee if the player is more powerful. I'm sure you agree that fighting a pesky Lvl 1 rat worth 10 XP isn't interesting if you're level 20 and need 10000XP to reach next level. Undead creatures and mindless or extremely hostile enemies would fight to the death.

AI in RTS:
I'm not into RTS, but a friend of mine mentioned the Close Combat series in regards to AI. In his opinion the AI in CC1: A bridge too far is about as great as it gets, but in the newer games of the series the AI has been lacking. Even so far as that if he destroys a column in an ambush on an open field, the enemy keep sending more units in the same way just to be destroyed as well (more or less swarming). In the first game they learned where ambushes was placed so that a single tactic couldn't be used all the time. In real life this would be done by using radio, or in lower tech scenarios, with observers following the column that returned to warn the rest of the force.
Also, in the first game (not sure about the others), the player controlled units had a morale meter that controlled how they reacted to the player's orders. If their morale was very poor they would refuse to enter dangerous areas or even try to desert. In complete panicked frenzy they could do weird things like running towards a machine gun nest to destroy it or other such things. One of his soldiers once destroyed 3 nests inside some buildings without the player's command because he panicked and acted on "instinct". CC1 is a pretty "old" game and it's graphics can't measure up the the graphics of modern RTS games, but because he feels the AI was so spectacular he considers it one of the best strategy game he's ever played. A great example of Steadtler's comment.

[WeirdoFu]

Lol...when I saw the thread I had to say something....

The way I see it is that AI is still truly an after-thought in the industry. Why do I say that? Well, for starters, most of the games out there that have some sort of "AI" is based on just a fixed heuristic algorithm. That's probably like the stuff you see in the first 3 or 4 chapters when you take a college AI course. A* pathfinding has been around for like...forever and its still being used to death, and sometimes not because its useful or good or required, just because everyone else seems to use it and implementations can be easily found. Some "smarter" games, from what I hear, do user modelling, but that still takes time and still quite prone to exploits, etc. In general, most AI in games out there are nothing more than plug-ins. They are just plug-in methods, classes, or even scripts that are on the outer ring of an engine.

For true AI, AI must not be just a plug-in, it must be a pivotal and fundamental piece in the engine design. Then you move beyond plug-in and heuristic based AI and can move into agent-based AI. From there, you can acquire and build emergent behavior. Let's say for example the power of building some sort of memory into a monster entity on the engine level. Then you can write a simple set of behavior based on this "memory" that it poses. It can then manipulate and alter behavior based on experience. In short "Evolutionary Programming" or "Genetic Programming" where each monster/NPC, over time, evolves its own behavior defining finite state machine. That would create some interesting results. Imagine an RPG where you don't randomly encounter monsters and when you do, they will try to run away when they feel they are about to lose. Then later on, they'll come back at you again and try to take you on with a different strategy. Or even specifically ambush you when he knows that your party's hp and mp are lower and more vulnerable. Then we will have a different game altogether where you have a fixed number of monsters and xp is earned truly through "experience" fighting them.

The point is, AI must move from the outter ring of game development down towards the inner ring.

One of my ideas is to set up an NPC server. Kind of like a boot camp server for all NPCs in RTS games. Every so often, they come back to the server and their deaths and kills are tallied and they either get killed permanently or have their strategies combined with other well performing NPC, then sent off into your games agains. In that scenario, even your boring single player game can constantly evolve and change as the NPCs start learning new strategies and behaviors that they've learned from you, other players, or just other NPCs.

[CpMan]

Well, I'm not an AI programmer and I don't frequent this forum all too much, but I think that AI is one of those things that has slowly advanced in the past, much like physics did, and will probably make a large jump in the next couple of game "generations". The thing is, right now, super advanced AI is not really NEEDED for the games which get most of the press, shooters. But the seeds are already being sowed. For instance, take games like Halo 2 and Fable. Halo 2's easy and medium difficulties sport fairly standard AI, though it's more intelligent than most. Legendary, however is a whole new animal. Not only is it harder to survive, but the AI is quick on it's feet and does fairly unpredictably interesting things. Fable had some significant interest in that a lot of the people acted very much like people, even if they were a little one-sided.

The desert isn't all that barren, but it has yet to bloom.

[BrianL]

I wonder at times if people confuse complex implementations with behavior and functionality. Given enough time (and space!), anything could be implemented as a finite state machine. That doesn't mean a finite state machine is the best solution (though it can be a useful one), but as it is turing complete.
Similarly A* is a powerful tool, known to produce good solutions to problems. If it meets the requirements put forth by the game design, why look for something more complex?

In a modern game, everything is a 'plug-in'; for instance, the renderer and networking systems can be abstracted and replaced. This is just basic software engineering; minimizing dependencies makes components more reusable and less fragile. Many commercial games already have AI who sense events in the world (or the AIs internal state) and change behavior based on these. This isn't something that requires anything fundemental in the engine.

I know I am repeating myself here, but much of what we are talking about is design decisions, not technical limitations. Tactical AI behavior, as described by frostburn in his enemy cooperation section, is applied by many games which deem it appropriate. Other games, like Serious Sam and Doom 3 intentionally avoid these behaviors as they aren't what the game is meant to be about. Better vision is definitely an issue. Doing something like rendering the scene and analysing it is still a tough AI program when response time doesn't matter. We may do this in games eventually, but this is a long way off.

Game AI definitely has space to improve; I would never suggest otherwise. But to say that AI is an after-thought is ignoring a huge body of work by a large number of programers (for instance, just check out the AI Wisdom books -- most of the authors have implemented the systems they describe in games).

I think the closest event to SIGGRAPH would be the AAAI conferences (for instance, http://www.aaai.org/Conferences/AIIDE/2005/aiide05.html). The primary reason this event does not receive as much attention in the game community as SIGRAPH is possibly direct applicability. AI research arguably covers more ground than graphics reseach, which reduces the density of useful information. Another issue is the response time required by game AI. A rapidly changing environment makes many of the more 'academic' solutions less useful, as many solutions are more targeted at less dynamic situations (though many can be adapted).

[BrianL]

A design issue example:

People who study such things have stated that the average PC gamer doesn't finish the single player campaign, let alone replay it. The result is a deemphasis on replayability. What this means is that each player having a unique experience/the game adapting is not seen as something worth investing time in given the market demand.

On the other hand, replayability in console games is perceived as more important due to market demand. At the same time, people don't necessarily want the game to get harder (or really vary) when they replay it; they want to dominate the game or have a less intense, more 'fun' experience, different ending, etc.

Again, none of this is meant to suggest that learning based solutions are worthless -- they may be a massive feature for some game some day. Until the consumers start demanding it though, pushing these through as a primary feature worth risking millions of dollars on (assuming we are talking about a large scale commercial title) will be a hard sell, as it isn't very marketable.

[WeirdoFu]

Thanks for the information. I guess I came off in the wrong way from being over-excited.

Finite state machines are just a small beginning, which is where evolutionary programming got its start trying to do sequence predictions. However, finite state machins are usually what governs game AI. As for the adapting of behavior, I guess I can't say too much about that because I don't know what they are doing in industry, but in most cases, the hardest thing is AI is to remove all personal bias of the programmer. Even in the academic field, many researchers are having trouble figuring out what is required to create "emergent behavior." I guess my dream of game AI is one in which the behavior is not actually hard coded in any way. Some rules are just given and then the behaviors just emerge through the testing phase of the game and the actual playing of the game by players. In this sense, the game gains an element of unpredictability. This will effectively create vastly more dynamic worlds that may make MMORPGs that much more realistic.

From a single player point of view, I can see how often times many people don't finish the game before moving on to the multiplayer. I personally am biased towards good, interesting, and well written stories, which aren't really a big focus in most games designed for multiplayer, which is another reason I stick with RPGs.

However, I can say that the future of AI in gaming and academia will be genetic programming. Proper implementation of genetic programming will let the system rewrite and recompile itself for better performance, behavior, or whatever. Years ago when it was first proposed, most thought it was too processor intensive to try, but lately, it has become a feasible option again. For game developer, genetic programming will make it possible to write one AI engine, give it a bunch of operators, and external information and have it write its own behavior. But of course, that is just my dream. :p

Time to get my nose out of books and see what the industry is really doing. :D

[AlbertoT]

Zodiac and other friends

I got your point and I do hope you are right, but...
I read an excellent AI book, last summer.
The author explained the "future of AI for game programming" admitting ,however,that no real game has actually implemented these techniques yet.
I was going to send him an E-MAIl.
"Why don't you prepare a demo for the main software houses?"
A simple level would be enough.
No graphics, no sounds, some GDI icons and text based dialogues,nothing more.
Just to show the advantages of new and sophisticated AI techniques vs traditional low level ones.
I do not suppose that commercial software houses have something against AI , as a prejuidice , I mean.
I do not suppose they have budget problems,either.
Unfortunatly in my opinion the truth is different.
AI can be of use for specific(but important) tasks , such as pathfindings, chasing\evading algo's , steering etc.
The game design, as a whole ,is still far beyond the possibility of a machine.
Deep Blue play chess using "brute force" it does not use "AI"
No computer can play the chinese game "Go" which is very close to a war\strategy game

[JD]

I think game AI needs to be controllable and thus provide an experience for the player. I think the kind of AI we're discussing here is more in tune with AI brain not game AI. I think that explains the lack of motivation to pursue a brain AI in game field where we need a more focused or specialized AI based on our game requirements. Given a brain AI we could make it be specialized but why when we can apply hard coded behaviors to achieve that as well. Lot less work than developing a brain AI and lot less resouces to eat up as well.

[Extrarius]

Quote:

Original post by BrianL Before discussing reasons for neglect, what deficiencies are you refering to?

The inability of characters in tactical games to behave as intelligently as William van der Sterren's quake 2 bots apparently did on machines with many times the power is a big issue.

The inability of game AI in virtually any game to adapt in any way is pretty bad. I hate it when I figure out some pattern that causes the AI to decide poorly consistantly, especially when the pattern is very simple (such as "get enemy's attention->hide behind corner->wait behind partial cover for enemies to walk into your bullets" in many FPSes). Some very simple 'overlord AI' could adopt strategies on a large scale - if chasing the player down always results in a quick death, perhaps the units should instead wait for the player to come out? Such simple adaptability (try case X and Y and see which works better) would make modern games much more interesting.

[WeirdoFu]

From a pure gamer prospective, I think good game AI should contribute two things to a game.

1. Good game AI should be used to increase the entertainment value of the game by being able to present players with more challenging and dynamic situations. This can reduce the dullness of repetitive tasks, like say levelling up in an RPG game, by making the enemies less predictable and have more interesting behaviors. For me, after a while, levelling up in an rpg game is like a no brainer, since there really is no more thinking involved as soon as you figure out the attack patterns. Of course, for games like Contra where pattern is important when trying to beat the game, variation may not be as good.

2. Constantly evolving AI should also be able to increase re-playability of a game. Why would we want to replay a game once we've already beaten it? Games like Devil May Cry just have differing difficulty levels where enemies are just harder to kill. Some just raise the HP of the enemy. What if everytime you play a game, the enemies get smarter, and not just tougher? Whether or not that will increase replayability is questionable, but there are all sorts of possibilities.

[Maleficus]

Where I wish games would focus attention is on multi-player AI (or "bots").

Quake III, Unreal Tournament and Counter-Strike:Source have all been helped by having really
great AI available out of the box for people who want the multi player experience without all of the downsides (lag, jerks, cheaters, lack of players, too many players, unbalanced skill levels, etc).

Investing a lot of time and energy in single player AI COULD be argued as a waste, but I'll say right now that any game I came across that had good AI for the single player game, I ALWAYS finished and enjoyed.

But, for the bean counters who want to save the money, then I'd argue that bots that simulate a multi-player experience would be an acceptable cost that extends the games life and play.

btw: to the original posters comments - it has been neglected in many games, but not ALL of them. There are some that are really making impressive strides, and then there are those where it was never a priority or didn't have the time.

And yes, I think that just because your a great GFX coder doesn't mean you'll write great AI code - really fun, intelligent AI is hard to create (I know from personal experience), and is a TOTALLY different field than GFX.

[abcdefg]

I think the reason AI hasnt been bigger than it is so far is that AI requires such a huge amount of CPU power. When you take away the processor cycles you lose to graphics, your left with an amount too small to run real AI on the 20 or so guys you need to run it on. Most AI algorithms need lots of iterations to actually achieve anything, and as fast as computers are they just arent fast enough to do real AI without resorting to tricks.