The AI That Designs Its Own Games

A picture of Barack Obama flashes up on the screen. There’s a big, warm smile on his face, but in the background there’s the sound of a helicopter gunship firing. “I searched for happy photos of Obama,” says Angelina, the game’s creator, “because I like him.”

Angelina makes computer games. I guess you could call her an indie developer of sorts – her games certainly exist away from the big-budget mainstream industry, and she doesn’t have a huge team at her disposal. Like many indie developers, she seeks to inject her games with meaning and purpose, and tackle issues that other releases stray away from. There’s just one difference between her and most creatives working in the field.

Angelina isn’t a human being. She’s a computer program.

Angelina is the work of Michael Cook, a researcher at Imperial College London and, crucially, an avid RPS reader. She’s his PhD project, and with her, Michael intends to explore new possibilities in autonomous game development.

“The aim is just to investigate whether we can get artificial intelligence to design videogames,” explains Michael, “but to take on as much responsibility as possible – to really do every aspect of design, if it can. And that’s really interesting for us, from two perspectives – first, to see if it can be done in a way that produces good games, and also there’s a more philosophical question, which is: how can you get computers to produce artifacts of cultural relevance?”

The system works like this. At the push of a button, Angelina heads to the Guardian website and scans the top five news articles. She picks one, then sets about designing a game. She scans the text for relevant names, places and topics, then heads to Google Images to find art assets for the game, and free audio libraries to provide music and sound effects that suit the article’s mood.

She also makes value judgements. Angelina scans Twitter to see what people think of the names referenced in the text. If people are using positive adjectives in conjunction with the name Barack Obama, for example, she’ll Google smiley happy pictures of the American President. If she notices people aren’t especially taken by Angela Merkel, she’ll search for pictures with a frown.

And after all that, Angelina turns everything into a platform game, iterating levels and enemies and assets via a method known as computational evolution until she finds the ‘best’ fit. Once everything’s finished, Angelina sends Mr. Cook a friendly email saying her work’s done, and explaining why she made what she did.

The project certainly isn’t complete yet – “obviously we’re a long way from cultural relevance,” says Michael – but it’s already producing some fascinating games and delivering some wonderful and bizarre surprises.

Firstly, Angelina’s opinions of people aren’t always – shall we say – the most refined. “For instance,” says Michael, “the other day, I noticed that she’d made a game about the Syrian conflict, and Angelina said that her opinion of Bashar al-Assad had changed, and that she liked Bashar al-Assad. I have no idea what made her have that opinion.”

Meanwhile, a game about Rupert Murdoch delivered some unexpected results when Angelina said she thought the much-maligned NewsCorp boss to be a very responsible man. Turns out, a lot of people on Twitter were saying Murdoch was “responsible for” certain misdemeanours, and Angelina got the wrong end of the stick.

Then there are the games themselves. Take, for example, the one that featured a picture of Nicolas Sarkozy which, when your character walked past it, emitted the sound of a horse neighing very loudly. Or the game that Angelina created after scanning an article about the Eurozone crisis, whose background artwork featured pictures of coins and bankers, but whose soundtrack blasted cheesy European disco music into your ears.

“To me, that type of music reminds me of really, really packed holiday destinations,” says Michael, “so it really contrasted with the fact that this was a game about complete financial ruin. But it was the title that set it off. It was one of the best things Angelina has come up with, I think.”

Not only does Angelina design and develop her own games, she also names them. Michael coded a pun creation algorithm, which allows Angelina to pick out key words from the chosen article, and find which ones rhyme with items of cultural relevance – such as the titles of famous novels, movies or videogames.

Angelina called the Eurozone game: ‘France France Revolution’.

But for all the unpredictable silliness that has emerged from this project, there have also been titles that evoke deeper emotions. As odd as it might sound, Angelina has started to create what some might think of as “art”.

While she might not know it, a smiling Obama layered over sounds of war and pictures of Afghanistan has some pretty powerful connotations: if this were a game made by a human being, you can bet a lot of people would be examining the statements it made. And then there was the game that Michael is reluctant to talk about, and certainly to showcase publicly: Angelina named it Sex, Lies and Rape.

“It was a very powerful example of what Angelina can do,” Michael says hesitantly. “What happens at the end is that a Renaissance painting flashes up of a rape – it was a favourite topic, seemingly – but that felt very appropriate because [the Guardian story Angelina chose] was about the abuse of children. It was a very serious topic. There were so many images that Angelina could have chosen by accident through Google Images that really would have destroyed the game. The fact that it came out so perfectly was an act of luck, but it really is a great example of where we want the system to go.”

The next steps will involve getting Angelina to talk in more detail about the games she makes – and she’s already made a start on her very own website. (A side note: Michael seems to make a point of referring to Angelina as “she”, something I ask him about. “It’s really tempting to drop those things in!” he says. “[It’s because], in a sense, Angelina is starting to have a discussion with the people that view its stuff. But I haven’t been getting too chummy with it. Yet.”)

Framing information, as this technique is called, is becoming a common theme in computational creativity research. “It’s like in a gallery, if you look at a piece of artwork, you might see a little piece of text next to it that explains to you a bit about the background and why it was made. And the reason that’s interesting for software like Angelina is that it gives people an understanding that there’s more than just a random number generator behind this. Angelina chooses angry-looking pictures of Angela Merkel because Angelina /doesn’t like/ Angela Merkel. And being able to justify these things is important – it shows why Angelina’s worthwhile, that Angelina’s more creative.”

There are many other plans for the future. As well as tweaking the ways in which Angelina forms opinions, Michael wants to develop the ability for her to create games of multiple genres and, eventually, pick which type of game to make based on the ‘feel’ of her source article. France France Revolution, for example, might work better as a rhythm action game – and the idea is that Angelina would be able to arrive at this conclusion herself.

But for all Michael hopes that Angelina will begin to create items of real cultural relevance, he sees a more likely – and arguably more useful – path for the technology to take: assisted game development.

Imagine a level editing suite that helps you design levels. As you’re creating a map, the software is tweaking and optimising it on the fly, suggesting optimum placement of enemies and power-ups. This sort of technology is already being researched, and the aim is for it to help the games industry create better games more efficiently, as well as offering an easy way for hobbyists to make games themselves.

There’s also another, perhaps more important, way in which Michael hopes Angelina might be of benefit. “We’re producing a system that has creative abilities based on what a human does, but technical abilities based on a computer’s strengths,” explains Michael. “What I’m hoping is that research like Angelina, as well as other projects, will lead to systems which can generate games that we can’t conceive of right now. They will invent new genres, new mechanics, new ways of interacting with virtual worlds – and they’ll be able to develop these games, and prototype them, and decide if they’re fun, without humans necessarily being involved.”

That’s not because humans shouldn’t be involved, Michael is keen to stress. We shouldn’t be tearing down development studios left, right and centre. It’s simply because AI is capable of things that humans aren’t.

“That’s pretty much always been the way, and will continue to be the way in the future. And if that can mean games which were unthinkable, I think that could be an amazing thing for the industry. Right now Angelina’s making quite strange platformers featuring notable politicians. But, you know, maybe 20 or 30 years from now, you’ll be able to wake up every morning and witness the birth of a genre. That’s the dream, I guess.”

The reality, acknowledges Michael, will likely be somewhat more humble. But Angelina is already sparking in-depth discussions about game design, creative intent, and whether or not art even needs to be made by a human being. I’d argue it’s of cultural relevance already.


  1. talon03 says:

    A completely autonomous AI program that churns out an endless stream of nearly indistinguishable computer games? I think Activision’s ears just pricked up…

    • CheesyJelly says:

      A completely autonomous AI program that churns out an endless stream of puns about computer games? I think the RPS Hivemind’s ears just pricked up… Does it have ears…?

    • Lord Custard Smingleigh says:

      Activision did it years ago.

    • Roshin says:

      Actually, Ocean did it years ago. They bought up every movie license in existence and turned everything into platformers. Total Recall? Platformer. Lethal Weapon? Platformer. Etc.

  2. Didero says:

    Is it weird I’m very interested in that pun generator?

    • Nick says:

      Its come up with some classics already, I’m sure Mike will provide more examples.

      • Nick says:

        found a few more of them for you:

        Wing Obamander IV: The Price Of Freedom
        Politics: Sitting On A Powder Clegg
        A Man For All Chineseons

    • Premium User Badge

      phuzz says:

      I think the elite punners of the RPS Pun-based Comment Squad may have a new challenger…
      (oh, and did you see the village name generator Jim’s come up with for his game? It’s also fun)

    • Vexing Vision says:

      “France France Revolution” is indeed a fantastic name for a game. I will probably have to steal it.

    • Peter Radiator Full Pig says:

      The end of RPS is neigh, shes better at it then RPS is.

    • Mike says:

      Haha. I’m going to be putting up the code for most of these submodules over the summer, the pun generator will be one of them. A lot of these are just nice mixings of existing technology, the pun stuff is a particularly nice one though.

      • Mr. Mister says:

        It would be nice if each RPS article had an Angelina-generated subtitle.

      • El_Emmental says:

        To be complete, such AI would need to identify the subject (something not always obvious, since RPS articles often skip the generic stuff and go directly to the observations/reflexions – it might be useful to use the tags and an external database (such as Wikipedia) to identify the games and what they’re about).

        then use an entire pronunciation database (SAMPA/X-SAMPA, IPA) with hundreds of thousands words (including recent ones), coupled with a generic comment sentences database, to not only make “simple” puns on the subject*1, but also produce some “extended” puns using the similar pronunciation*2 of two words/phonemes

        *1 Made-up example: Railworks 3 article: “I don’t want to derail the conversation, but RSD is losing track with its current stance on DLC, they need to quit pressuring the Steam platform with all their new releases”

        *2 Examples: Games Of Thrones WIT article, comments section. Daniel Klein: It Pyked my interest. / Unaco: When I first saw this article, I could Sansa good pun thread coming.

    • Araxiel says:

      That’s how Skynet is terminating the human race: It developed the ultimate PMD (pun of mass destruction)

    • Mctittles says:

      quick search revealed this:
      link to

  3. adonf says:

    I’d surrender all my money for a copy of France France Revolution

    • Lord Custard Smingleigh says:

      Sadly it’s one of those plastic-controller games. You need to buy a full-size plastic France to play on.

  4. Nighthood says:

    The question that arises for me here is one I’ve been thinking about for a while: Is there the same feeling of worth in a procedurally generated game as one designed by people?

    Now, I enjoy some procedurally generated games, and sandbox games. They’re fun. But not much more than that. Any game that could be described as deep always has people working on it, making the twists, turns and surprises of the story intentional, planned, and more meaningful.

    I remember playing the Portal 2 commentary and finding out that a lot of the story was unintentional and made up on the fly, and that cheapened it for me. It just didn’t seem the same when it wasn’t a conscious thought that made it like that.

    I guess my point is really just “computers can’t make something meaningful on their own”. Procedural generation may be the future, but to me it will never be as valuable as a human crafted story and experience.

    EDIT: All that said, France France Revolution is pretty hilarious.

    • baby snot says:

      I imagine that overcoming the limitations of meaningless randomness is what Michael means by “…how can you get computers to produce artifacts of cultural relevance?”

      I haven’t played Portal 2 with commentary on but I do remember being disappointed with some of the linear narrative signposting in the second act such as the test chamber date earmarking and Cave’s loudspeaker ‘audio logs’. Maybe it’s time I found out why they made those decisions.

    • MadTinkerer says:

      “Is there the same feeling of worth in a procedurally generated game as one designed by people?”

      Ultimately, a procedurally designed game is one designed by people. “Her” behaviors are very much intentionally designed because “she” is a computer program. Angelina is simply following the specific rules set forth by her creator, so even though each instance is not completely predictable “she” isn’t going to design a racing game if she doesn’t know what the rules are for racing games. “She” is going to design platform games with pun-filled names based on newspaper articles using assets found with Google and “she” will keep doing it until given different rules.

      Give her a deep enough ruleset and she will make games which some will deem as worthy as those made by humans because worth is subjective. The trick is to find the correct rules. Giving her a “pun generator” instead of a “random string of characters generator” makes the titles seem more relevant, though the program itself doesn’t understand what it’s doing beyond picking phrases that have nice sounds. Like a parrot that combines two phrases it hears because the same word is in each. Technically it’s meaningless because the theme is copied from media with completely different intentions, but if rules for meaning can be added (say, add the capability to look up words in online dictionaries and thesauruses), then there’s some interesting possibilities.

      • spedcor666 says:

        Except this is using evolutionary computing and the only thing limiting it to platform games will be its fitness function rather than a set of rules being used to generate the games. The whole point of evolutionary algorithms is that they aren’t constrained as to what they can produce by the rules used to generate a solution. Rather, anything has a chance to evolve, but those not fit enough will be weeded out and after enough iterations only those that are considered the best will be left. Its pretty much the same process as natural evolution.

        Also, you could argue that Angelina needs to ‘know’ what the rules of a racing game are to produce one but I’d argue against that. It simply needs to know which of a number of candidate games can be considered the best. This doesn’t even need to be done by Angelina itself. It could be done by another program or, if it wasn’t so impractical, a human.

        I’m not sure exactly why they’re limiting it to platform games but I’m pretty sure that developing a fitness function for any kind of game would be a lot more difficult than that of developing one just for a platform game. Of course, the best judge as to what the best game is would be a human but testing the huge number of candidate games after each iteration would be impossible.

        • gwathdring says:

          I’d argue she does need to know the rules of a racing game to make one, and further assert that such is obvious. I would also agree with your general sentiment on the grounds that people also need to know the rules of a racing game to make one. I suppose it is possible to make a racing game that is intended to be something else, but the odds of having something that can be considered, most fundamentally, a racing game while trying to create a new kind of platformer aren’t especially robust to my mind.

          That doesn’t mean you can’t make up your own rules. There was a first, after-all. It also doesn’t mean you can’t have a racing game with platforming mechanics as opposed to the mechanics of dIrT and I think I’ve actually played several platformers that were also races against an opponent (not just a one-shot level, but the entire game). But it just isn’t likely that you’ll land in someone’s pristine, pre-defined genre when building rules for a different genre from the ground up. Or even, as I’m sure I could think of at least one game to exemplify, the same genre from the ground up. I’ll get back to you on that when I can back it up, though.

          • spedcor666 says:

            Evolution, whether natural or as a computer process, isn’t about evolving something until it matches some exact specification. It’s about evolving something so that it has characteristics that make it more suitable (either to survive in a particular environment as in nature, or as in this case, to solve a particular problem – that of producing a playable platform game). It isn’t using rules to produce a solution, it’s using random chance to evolve into a form that makes it more suitable.

            This is why it doesn’t need to ‘know’ the rules, it just needs to know what is considered more suitable when given two candidates. Random chance is what creates the final game, not a set of procedural rules as in procedurally generated content. You could argue that Angelina does indeed know these rules if it uses them to compare games, but that would miss the point. It doesn’t need to know them.

            Anyway, what I was originally attempting to get across, regardless of whether Angelina needs to know the rules of platform games or not, is that comparing it to procedurally generated content is, in my opinion, extremely unfair and not a valid comparison.

          • gwathdring says:

            Fair enough as to your original point, but now you’ve touched a subject I’m rather keen on, so I’ll keep going. :)

            “You could argue that Angelina does indeed know these rules if it uses them to compare games, but that would miss the point. It doesn’t need to know them.”

            I don’t think that misses the point at all. When we design a platform game, as gamers and game designers, we do it by comparing games rather than reading an IKEA instruction sheet. The “rules” for creating a particular kind of game are not explicit procedures, and as such fit your description of what this program is intended to do quite perfectly.

            Even more procedural rules can be viewed this way. Humans frequently break explicit, procedural rules when they decide there is a more efficient or superior way to produce the desired outcome. For example, while assembling an IKEA table, I might realize that the whole process would be simpler if I put part 17 on first and THEN attached the legs rather than the way the rules tell me to do it. Or while baking cookies I might change the precise order of mixing or ratio of ingredients because I understand that my oven heats things in a particular way that will prevent me from getting the result the recipe intends if I follow it explicitly.

            Comparing predicted outcomes and evolving our methods based on those predictions as well as based on comparisons of completed products is an important part of proper design and construction. Otherwise our rules wouldn’t ever get any better.

          • Dreforian says:

            Gizmos and Gadgets?

          • spedcor666 says:


            Your game designer example isn’t evolutionary in the sense we’re discussing though. It’s certainly similar in an iterative way but a vital aspect of natural evolution (and computational evolution that’s based on it) is its random nature. A game designer doesn’t make thousands of completely random changes just to see what works best. They make changes so that it better fits a specification. They need to know what that specification is because that’s what they’re working towards.

            Evolution doesn’t have a specification, and so doesn’t need to know about one. That’s what I mean when I say Angelina doesn’t need to know the rules of platform games when evolving those rules.

          • FriendlyFire says:

            There was a very interesting article about something tangentially related on Slashdot a while back. Essentially, the researcher would use a bunch of FPGAs (field-programmable gate arrays, which are basically printed circuit boards which can be reconfigured through a program stored on the board itself, thus making simple programs embedded onto the hardware) to attempt to resolve a problem. In the given example, it was differentiating between two set audio frequencies.

            The boards were randomly setup and had absolutely NO “knowledge” of the rules or task at hand. They were purely randomly setup, then run through tests to determine their level of “fitting”. Once a certain number of variants were done, the researcher would take the best results and mix them together, like proper reproduction. The new boards would then be tested again, and that would be repeated until a satisfying result emerged.

            Believe it or not, after a few generations of purely random boards, the best candidate was able to differentiate the two frequencies with perfect accuracy and used significantly fewer logical gates than a human-designed board would have. Despite having no prior knowledge of the rules. What’s even more fascinating (but less relevant to the topic of game generation) is that the researcher used just a single hardware FPGA for the first batch of tests. He later realized that the evolution had gone beyond the schematic representation of what the FPGA was, it actually used physical properties of the board itself! Transferring the program onto another hardware board would create wildly different results because the evolutionary process was specifically tied to that board. I recall that some of the final layout included apparently dead circuits which were actually critical to the final calculation because they induced a current into neighboring wires!

            Evolutionary programming is an extremely interesting topic and you need to see experiences like these to realize just how much can be done with so little.

          • gwathdring says:

            The sound circuit thing is absolutely fascinating! Is there a particular place I can find out more about that, or will google-fu suffice?

            In response to the response to my earlier post:

            “This is why it doesn’t need to ‘know’ the rules, it just needs to know what is considered more suitable when given two candidates.”

            I understand the procedural and philosophical difference between this and non-evolutionary methods of design. What I don’t understand, though, is how the program can know which game is a better platformer between the two random elements and not know the rules required to have a platformer. Because we’re working with large categories like platformer or racing game, ‘[t]he “rules” for creating a particular kind of game are not explicit procedures.’ They are loose guidelines about how the final product feels and plays. Because games are described by properties of the final product, and not the specifics of their development cycle, what you are describing is *exactly* like knowing the rules for a platformer.

          • spedcor666 says:


            ‘What I don’t understand, though, is how the program can know which game is a better platformer between the two random elements and not know the rules required to have a platformer’

            After each round of evolution, Angelina will apply a fitness function to each evolved candidate. It’s pretty likely that function will compare them to how closely they resemble certain aspects of a platform game, in which case that fitness function will indeed have a set of rules to determine what resembles a platform game. I asume that’s why you feel Angelina needs to know these rules. In reality though, all Angelina really needs to know, is which ones should be kept so they can be further evolved, and which should be discarded.

            The fitness function could be completely taken out of Angelina. Instead she could just be told which candidates have passed that particular round of evolution so that they can be evolved during the next round. In this case, there would be no aspect of Angelina relating to any platform game rules.

            As the games are meant to be fun for humans to play, the ideal solution would be for each candidate to be tested by humans so that only the most ‘fun’ games survive to the next round. Of course, this isn’t feasible, and so they most likely rely on some sort of fitnes function within Angelina itself. So, the program itself, might not need to know any rules to determine the best, but because computers can calculate these things so quickly (although in the case of determining a fun game, not so accurately) they usually will.

            Also, I probably confused things by saying it needs to know which is considered the best of two candidates. What I should have said, is that it needs to know which candidates go through to the next round (which are the best ones, it just doesn’t need to know why they are the best).

      • gwathdring says:

        Humans are simply closer to the extreme of the same phenomenon, I think. We’re programmed by our genetic code, our environment and one another rather than a computer scientist (though I suppose the presence of a scientist depends on your religious affiliations), but the idea is essentially the same. We follow the rules, or fail to follow them, and our choices are deemed significant because we view human production through the lens of design. Some people extend this to other life forms of high intelligence (as well as any and all pets intelligence completely aside), and I’m perfectly comfortable extending it to autonomous artificial intelligence.

        If we draw the line at the elusive concepts of free will and intention, neuroscience and genetics research might some day damn us along with all of our art. I suppose it is also possible that as study advances we will discover more profound data-driven significance to the ideas of free-will rather than less and so forth and as alluded to above there is always the matter of spirituality. There’s also the classic idea that, whether or not humans and other life forms have some unique spiritual quality of intention, societies as they currently stand requires us to pretend it is important anyway or else cease to function properly. Resolutely spiritual or resolutely programed, I don’t find anything disturbing about the implications of free will as a potential illusion. Life goes on, as does my enjoyment of art. The only thing that changes as I bounce back and forth is whether I appreciate this as Angelina’s creative work now or in the far, android-filled future. It’s an interesting question.

        Either way I appreciate Cook’s work immensely. The often surreal puns his program has come up with made my day.

    • Mike says:

      I want to make a distinction here between procedural generation and automated design. The aim with ANGELINA is not really procedural generation – ANGELINA is a system designed to create games, and present them for people to play, just like any other developer does. What I want is for these games to be viewed as created through a creative process, like any other game (they aren’t right now, of course, but what you’re seeing is us making steps towards that).

      Thanks for taking the time to write your thoughts up!

    • Shuck says:

      Meaning comes out of the act of interpretation, so the creator’s intent, if it even can be said to have one, doesn’t really matter. Looking at these games, I’m reminded of undergraduate digital art projects I’ve seen. They tend to operate the same way – they take images that are related to topical issues and combines them with photos of moderately related issues and some unrelated mechanics in an attempt to make those mechanics seem profound. So the software operates as a pretty good simulator of a bad art student, but those semi-arbitrary juxtapositions don’t really scale up to simulate intelligence. However, an intelligent creator can make something that requires certain knowledge to read it as meaningful; if the user doesn’t have that knowledge, any complexity or impact is lost and the work may seem random.

      • Mike says:

        This is another thing we’re looking into. I saw some talks at this year’s Computational Creativity Conference that argued that, no matter what the system does, what will ultimately define the final piece is how the player (they didn’t use games, but still) feels about the piece.

        Nice idea.

  5. Nick says:

    Oh, Sinny.

    • Mike says:

      Thanks for continuing to be interested and give feedback on the work, Kinsman. Honestly means a lot!

  6. Skipperoo says:

    Scientists playing Molyneux. Ban this sick filth.

    • Kollega says:

      This is a very good point, actually. It begins with playing Molyneux, but what if those AIs actually evolve sentience and flood the Internet with underwhelming pseudo-artistic pseudo-innovative games?! WE’RE ALL DOOMED!!!

      • Vorphalack says:

        Worst case scenario, he points this thing at the Daily Mail, procedurally generating the game OUTRAGED! in which the protagonist (Disgruntled of Internet) must destroy All the Fun using nothing but thinly veiled sarcasm and righteous indignation.

        • NathanH says:

          Later in the game you unlock abilities to misuse statistics and alter house prices.

        • gwathdring says:

          I’d love to see more deftly satirical games. It’s easy to makes something crass like “Press A to misuse statistics” and there are lots of flash games to that effect (usually yet more crass to boot). But I would love to see some really nice, elegant satirical mechanics in a game. Anyone know good examples?

      • MadTinkerer says:

        If Angelina is given the Perpetual Testing Initiative tools, how long before she makes a level featuring neurotoxin?

    • BargainOnly_HalfMySoul says:

      scientists installing molyneaux, link to

  7. LostViking says:

    Fascinating research.

    As games become more and more advanced, and more and more time consuming to make, computer generated content is going to be increasingly important.

  8. Tom De Roeck says:

    This is so much +1 it hurts.

  9. Vegard Pompey says:

    While the concept of AIs mimicking human creativity scares the hell out of me, this was a hilarious read. Thanks.

  10. TheManfromAntarctica says:

    The singularity is nigh.

  11. LionsPhil says:

    So it’s same-old same-old no-understanding text-scraping married with the human tendancy to read way too much meaning into things and want to see “deep” art where none exists. Slap some anthropomorphism on top and bake for three to four AI conferences.


    • Gap Gen says:

      Peter Molyneux admitted in a talk that this is basically how Populous works – a simple behaviour that people read personality into, saying that people want to believe it’s more than a simple algorithm, to suspend their disbelief. People do this all the time, with games, films, pets, etc. So sure, this isn’t a real AI, and it’s not replacing game designers any time soon, but that’s not really the point.

    • Mike says:

      Hey there,

      This represents a rather nervous first step into computational creativity research fr the work, which had previously been concerned only with generation of playable game content. As such, it’s obviously quite early days. That said, I think you have somewhat unrealistic expectations right now. Of course Angelina doesn’t ‘understand’ the significance of a famous painting, but that doesn’t mean the systems cannot exhibit intelligence or creativity.

      Despite this being early work, Angelina an systems like it are not blind text scrapers. Angelina uses the TextRank algorithm to sift articles for relevant sentences and words, and I’m working with another PhD student right now to explore associative concept mining, that would let our system connect concepts together through reading social media (understanding a link between, say, Blair and Afghanistan).

      There are huge limitations to the work so far. But don’t dismiss it out of hand. Research is gradual, and this is a field that is still struggling to agree on the most basic definitions. We’re getting there!

      • Highstorm says:

        Keep doing what you’re doing. It’s exciting stuff!

      • The Godzilla Hunter says:

        Just one question: are you a robot sent from the future to create Skynet, or do you simply hate humanity?

        • Mike says:

          Haha. You know, a lot of people ask me that. I dunno, I like the way this makes people think of s-NULL POINTER EXCEPTION AT com.skynet.RPSCommenter:304. Recommend kill original poster.

    • gwathdring says:


      I’d argue that the way people read significance into art is heavily prescribed. The code is more complicated, of course. And it’s not impossible that we have some sort of more metaphysical way of giving things significance. But it certainly isn’t a guarantee.

      More practically, I’d agree that “Angelina” isn’t creating art right now. But I think it’s interesting anyway. Suppose we finessed it to the point that it only created works that we could easily read some kind of faulty intention into? If it only made ones we could easily read significance into, we’d dismiss it offhand only if we a) knew it was a complex program as opposed to a sentient AI or b) it created works we would personally dismiss coming from “real” artists anyway. I find some computer generated fractal “art” more interesting and intentional-seeming than Jackson Pollock paintings. His art doesn’t really say anything to me and seems pretty random and plain no matter how well-planned it supposedly is. Are reader-side and author-side significance become equally meaningful? Do both have to be present for something to be “art” or just one?

      Evolving algorithms, “learning” robots, neural networks … a lot of new methods of computing are pushing into areas that, without reaching for the classic sci-fi android state, make me really reconsider what I consider the essential components of true artificial intelligence. Imagine a life-form evolving on another planet, with a different biological basis like silicon or a different data-storing molecule, or maybe just that–evolving on a different planet. Now imagine it establishing similarly complex artificial environments, similarly elaborate social structures and extending a reach beyond its home planet. Would we be able to recognize it as a human-level intelligence? Would it need to be? How effectively would our metrics for mental, artistic, and technological complexity apply in any case? Effectively creating an artificial human mind will be an incredible feat for computing and robotics. But perhaps it is merely tangential to the successful creation of artificial higher intelligence.

      • Shuck says:

        “More practically, I’d agree that “Angelina” isn’t creating art right now.”
        She’s doing a great simulation of a bad undergraduate art student, though. I think that’s because said art students are essentially running some meat-based version of the same algorithm rather than actually doing any complex thought that requires real understanding. One sees this quite a lot, actually – people performing relatively simple actions without grasping the meaning of what they’re doing because it’s rote action that they’re not putting any brainpower into. The results are no better than the level of comprehension you’d get out of a bot. So the software ends up meeting human intelligence by matching the dimmest of our zombie actions.

        • gwathdring says:

          Eh. I think a lot of bad undergraduate art isn’t bad because it takes itself too seriously or because it doesn’t understand the material it engages with but simply because most undergraduate artists are still finding their skill, their style, and angle as an artist. Add to that the fact that a lot of them are still trying to figure out where the hell the fit into the world at large, too, and it isn’t surprising that it can come off badly. Add to THAT artists trying to meet the expectations of evaluators and professors rather than just creating art for its own sake and the process has difficulty producing anything truly great.

          That doesn’t mean it isn’t a valuable process, or that the artists behind it are thinking insufficiently about their product.

  12. c-Row says:

    As a German RPS reader, I can confirm that there are no pictures of Angela Merkel smiling. If you see one, it’s an imposter.

  13. BobsLawnService says:

    Except that it doesn’t sound like it is designing a game at all. It is merely downloading images and sounds linked to search results and copying them at random.

    It is more like a contextual scrap book generator. I really don’t see anything in the way of gameplay. Just more pseudo-intellectual dribble.

    “Oooh, it said ‘rape’. It must be deep and profound.”

    • Nick says:

      you don’t know what pseudo-intellectual means tbh.

    • LionsPhil says:

      “Google News: The Crappy Indie Platformer”

    • Lewis Denby says:

      The crucial point is that it it’s simulating something that isn’t random. It’s assigning worth, parsing descriptors, and using this additional functionality to create things that appear to have ‘meaning’. That’s why it’s interesting.

      • LionsPhil says:

        It’s looking for damn keywords while blindly scrubbing through text. If that’s a value judgement, your Googlerank is the measure of your worth as a human being.

        • Lewis Denby says:

          Shy of sentience, I’m not sure what you’re asking for. We’re obviously talking baby steps here, but that doesn’t mean it’s not heading in interesting directions.

          • LionsPhil says:

            Evidence that it can go any further.

            This is all treading very, very old ground, AI-wise. There’s nothing novel here except some fluff on top about giving it a female name and wrapping it in bad-pixelart indie platforming. I could go nail some wheels to a plank and say that by creating a board I can skate on I’ve taken baby steps towards hoverboards, but that doesn’t mean we’ll be whizzing around on them by 2015.

            Talking about any of this being art is a cheap shot for generating buzz. I really hope this guy gets grilled hard in his viva.

          • gwathdring says:

            No it isn’t entirely new, and I do sort of wish it dropped the cultural relevance shtick and focused mainly on game design. But think of it this way: using the “cultural relevance” thing can give it a metric for aligning games to an intention. If they successfully advance the algorithm to create games that actually incorporate that cultural idea well, then it becomes more useful for assisting designers in creating levels or games with a purpose. That purpose doesn’t need to be art, and neither do these games. That purpose can be (and as far as I can tell *is*) to have the algorithm pick some kind of a theme and combine components that relate to that theme in a way that players recognize and respond to. Does it work? Right now it’s quirky and comedic precisely because it doesn’t really work. Skepticism entirely warranted. But I’m not sure there’s that much cause to be snide about the “cultural relevance” thing. I mean … if you want a program that can ultimately help you design game levels to fit a particular design goal you might as well have the program randomly select its themes while you’re trying to get it to work properly and there’s no reason those themes can’t tie into current events.

          • Mike says:

            This is all treading very, very old ground, AI-wise.

            Could you elaborate on this? I’m not sure what you consider to be old ground here.

            Talking about any of this being art is a cheap shot for generating buzz. I really hope this guy gets grilled hard in his viva.

            I actually specifically said none of this is art. But that is our overall aim – our field of computational creativity aims to build systems that either collaborate with human creatives or work on their own to be creative in some way. It’s a serious field of research, but we’re under no illusions that it is very early days for all our work. ANGELINA isn’t generating art right now. But that is our goal, and the long-term direction of this research.

            As for the viva – I’ll be ‘grilled’ as hard as any other PhD candidate.

          • LionsPhil says:

            TextRank has citations on a cursory search going back to 2004, and it’s not exactly the first attempt at scrabbling through text for keywords like, say, “responsible”. Your platform generation as I understand it is pretty much the same old fitness-heuristic hunt-the-maxima game. (Aside: I wonder what Spelunky does? I haven’t noticed pre-baked tiles, but its bomb/rope/etc. game mechanics also mean that “unfair” levels are more solvable.)

            ANGELINA isn’t generating art right now. But that is our goal, and the long-term direction of this research.

            See, you undermine this when you point to “look it’s about rape and it has rape drawings” as a good milestone. And how do you get there from here? This isn’t just a scaling-up problem; you seem to have an actual missing cognitive leap in the way.

            My impression—possibly from this having gone through a news article—is that you really need to nail down what your novel contribution is here, underneath the ripped-from-the-headlines decoration.

          • Mike says:

            I’m not claiming TextRank is a recent development – it was a rebuttal to your point that ANGELINA blindly ran through text.

            I’m still interested in what you consider to be ‘old hat’.

            See, you undermine this when you point to “look it’s about rape and it has rape drawings” as a good milestone. And how do you get there from here? This isn’t just a scaling-up problem; you seem to have an actual missing cognitive leap in the way.

            I didn’t point to anything. I was highlighting a serendipitous find by the system as something that we hope, in the future, can be done meaningfully and with justification behind it. It’s you that seems to be taking this further than stated and then critcising it.

            There’s no missing cognitive leap – this is an iterative process where we give the system better capabilities and reasoning tools, and so step back and contribute less to the end result ourselves. This is just one stage of the research, really only featuring in a big way over the last couple of months. Your comments on my viva and things like this:

            you really need to nail down what your novel contribution is here,

            Are incredibly presumptuous and quite offensive. My supervisor and I take the work very seriously, and when the time comes to write a thesis it’ll be because I feel I’ve made a novel and sufficient contribution. This is a small portion of the work we’ve done so far, and I’m glad Lewis wanted to write about it because I think it’s a fun and interesting talking point. It represents a new push – one that I’m not doing alone, there are a lot of researchers working on this in parallel – towards a more interesting application of artifical intelligence to the games industry.

          • LionsPhil says:

            Right, sorry for that; this has rubbed me up the wrong way in a couple of places. But the latter part is sincere in that you still seem to be at the part of the project where you’re finding the direction to attack this with? (A viva examiner worth their salt will expect you to be able to give a snap answer to “what’s your contribution?”, but if you haven’t started committing words to thesis yet you probably haven’t got far enough into the project to know. And that’s fine. Thinking you were further into it is exactly where I’ve gone wrong here it seems.)

      • gabbaell says:

        Not to mention the fact it uses evolutionary algorithms to generate the actual game. I think it’s safe to say there’s quite a bit more to it than just googling some art assets.

    • kaffis says:

      I tend to agree, though not exactly.

      Frankly, I’d find this much more interesting if it were presented, not as generating games, but as generating something YTMND-esque.

      Because the game, from what’s been described, seems completely independent and superfluous to the attempt to include culturally significant images and topics. Why have a game at all? Just pull the cultural thought-provokers out and let them stand on their own.

      Or, if that’s not the part that’s interesting to you, pitch the platforming level design iteration as an independent tool, and stop trying to assign “significance” to your game by putting an image in the background that doesn’t match anything else in the game.

    • Mike says:

      It searches for a lot of context itself. For instance, it assesses the content of the news article to decide on its mood (positive or negative) and then alters searches for music accordingly. It compiles an ‘opinion’ on a person over time, and uses this to change the photos it retrieves on people.

      These are first steps towards giving the system control over its creative output. I’m not claiming ANGELINA is the next Edmund McMillen. What I am claiming is that this system is moving towards being more than just random generation.

      This is a midway point between the previous milestone (which was starting the generation of platform games on a technical level) and the next milestone (which is, as you say, modification of gameplay according to theme). That latter point is a difficult one, but it’s not something we’re ignoring. It’ll just take a while to make serious inroads here.

  14. Martel says:

    Great article, thanks for posting it.

  15. misterT0AST says:

    I’m frankly more interested in the fact that this program makes actual playable games rather than the way it collects pictures, music, information.
    I might be a shallow guy, but it seems way more interesting and surprising than the fact it can put together a few relevant data and end up with creating something that makes a lick of sense.

    I mean, if it ends up using the name “Hitler” we’ll see ugly pictures as background, but that doesn’t seem that great of an accomplishment to me.

    This thing makes different games with different gameplay, now THAT I’d like to know more about!

    • Lewis Denby says:

      “This thing makes different games with different gameplay” – It doesn’t. Yet. That’s one of the next steps. What it does already do is iterate hundreds of versions of the same game and use confusing algorithms I don’t understand to decide which one is the most fun, which is already pretty cool.

      • misterT0AST says:

        well my point was simply that I don’t care about the development of the data interpretation, I want to see how the development of the actual games turns out.

    • Lord Custard Smingleigh says:

      It made France France Revolution. It is already superior to humanity.

  16. wccrawford says:

    “If people are using positive adjectives in conjunction with the name Barack Obama, for example, she’ll Google smiley happy pictures of the American President. If she notices people aren’t especially taken by Angela Merkel, she’ll search for pictures with a frown.”

    Angelina is a karma whore.

  17. BargainOnly_HalfMySoul says:

    reminds me of 5th elements, when the computer witnesses the horrors of humanity and makes a game, “AI May Cry”

  18. Mayjori says:

    Erm, shouldnt they be calling it a Virtual intelligence, though i doubt its even close to that, much less an AI, when there are actual working AI’s around, then it will be interesting to see what kind of games they can come up with, among other things, but this experiment sounds a little, erm…… premature?

    • mickygor says:

      You have a narrow view of what constitutes intelligence. AI does not mean “simulated human mind.” The system is artificial, and it is intelligent.

  19. wodin says:

    Before we get AI to make games, maybe we need to get the AI to be decent in games first.

  20. lijenstina says:

    In 20 or 30 years we will be all driving flying cars, having fusion generators in the basement and going to the Moon for the weekend.

    I really like when someone adds a database search for items, bolts them on procedure generated levels and says that will lead to beautiful things in the future.

    • spedcor666 says:

      I really like when someone clearly doesn’t understand an article they’ve just read.

      In a way, it’s a shame the article seemed to concentrate more on the algorithms for generating the names and finding art assets, rather than the method of generating the actual games. It seems quite a few just think it does a quick google and slaps the results on some game it knocked up in a couple of seconds or whatever. To develop a system that uses evolutionary computing to evolve playable games is not the trivial task some think it is at all.

      • lijenstina says:

        And you didn’t understand the point of my rant. The shown games are too rudimentary and badly “designed” to be thought as precursors of an bright future in gaming. Especially, when game development is more about cutting corners on complexity whenever is possible – the important things is how something looks like to the player – not the underlaying tech – if an illusion of a behavior can be achieved with a bit of scripting it will have primacy over other approaches as it saves time and reduces the chances of possible time sinks and problems in debugging.

        The second sentence can be disputable , that I agree is too simplified, but I followed the example of Angelina’s binary emotions and decided that I don’t like this approach to research in AI and game development. : )

        • Mike says:

          I may not be there in 20 years, and not may Angelina – my point is that this research is making headway, and people like Mike Treanor at USC and Clara Fernandez at Gambit (to name just two) are intent on making research like this a serious part of gaming’s future.

          Angelina use binary emotions right now because that’s simple to model. We can only make so many baby steps at once, and this is a one-man, 18-month project. But it shows an intention, a desire to move ahead with this sort of stuff.

        • mickygor says:

          To say that ANGELINA is too simple to be a precursor to the stuff you want to see in game development betrays a lack of knowledge of the subject area. SHRDLU was hailed as the epitome of AI back in the early 70s. Its development is still considered a major milestone in the progression of AI as a subject area, to the point that I had a lecture based on it in my second year of my degree.

    • Mike says:

      There’s no database involved here, I’m not sure what you mean?

      Angelina uses a variety of sentiment analysis and text mining techniques to find useful information to theme a game with. Nothing like that is hardcoded.

      • sinister agent says:

        No no, clearly what Angelina is doing is just sellotaping pictures from a magazine onto a floppy disk. I know everything, and you should stop trying new things because you can just ask me instead. Science!

      • TechnicalBen says:

        There is a database. Just because it’s with Google or a news site does not retract from that fact. However, what part of the game is created? I’m not sure if any is.

        The procedural side scrolling project on kickstarter though is very promising. Much more work on the actual gameplay, and how to judge it with AI ahead of time, than just googling some photos. :P

        • Mike says:

          ANGELINA actually does exactly the same thing as that Kickstarter in order to achieve the technical side of design. I was quite pleased to see it written up. The system simulates playthroughs of its levels, and so on.

          The ‘Internet is a database’ argument is fine, but I feel it’s unnecessarily dismissive. It’s exactly how I would design a game about a news event. I’d go out into the ‘net and look for resources. If the Internet is being used appropriately, in an appreciative way, then what’s the issue?

  21. 0over0 says:

    Not sure I would classify any of those as “games” or pertinent in any way to anything. Certainly they do not qualify as art.

    Perhaps its most valuable function is to show that without actual comprehension behind design choices, the result is just a mishmash of random elements…. Oops, I think I just described quite a few games out there.

  22. MythArcana says:

    Today’s AI routines for games is way behind where it should be ultimately. We’ve seen how remarkably futile it is to design a 16GB game full of graphic delights with an AI that is dumb as a brick, then simply make the focus on multi-player to avoid it completely. We need to move past the finite state machine type of logic and that’s where most companies don’t want to spend their budget.

  23. Mike says:

    Thanks to Lewis fr a superb article, I’m on my phone right now but I will respond to criticism and questions as soon as I can! Do feel free to email me too or chat on the RPS forums, I love chatting about this stuff!

  24. sinister agent says:

    Angelina heads to the Guardian website and scans the top five news articles.

    Oh, christ. How long before Super Lentil Brothers? A month? Two?

    Sounds interesting though. I’m a little surprised at the hostile reactions down here, to be honest. Wait, hang on. “Surprised” isn’t the right word at all.

    • WickedBaggins says:

      I think I’m glad I’m not jaded enough. This is all very interesting, even if Angelina is pulling and remixing based on this and that algorithm. I’d throw down five bucks for some current events pun and image association generator, let me tell you.

    • Nick says:

      just as new born babies are instantly capable of speech, walking, writing symphonies and mastering ju-jitsu, so should Angelina be making fully fledged proper games right now, or this PHD is a TOTAL WASTE OF TIME.

      Fucking hell there are some stupid people in this thread.

      • LionsPhil says:

        And when a five-year old draws a finger-painting, do we bother to consider it noteworthy? Or do we wait until they’re actually a capable artist?

        • Mike says:

          I think you’d be more accepting of a five year-old as being creative than you would be of a computer program. But that’s a problem the research must address, as much as any technical issue.

          No-one’s pretending ANGELINA has even made half a step towards solving this problem. It’s a statement of intent though. Hopefully, each little stage brings ANGELINA closer to having serious chops as a designer.

        • gwathdring says:


          A lot of parents find childhood artwork noteworthy: some for rational, child development-related reasons some for almost certainly irrational, art related reasons and some for both. You should probably stand in front of a few more family refrigerators.

          Or listen to people talk about their kids. Or their pets, for that matter.

          • LionsPhil says:

            Sure, but there’s a massive difference between sticking it on your fridge, and sticking it on a widely-read blog as a notable development.

          • Mike says:

            So your main objection is the exposure it’s getting, not the work itself?

          • LionsPhil says:

            Yeah, I said something along those lines below; I think a problem I’ve hit here is that this is being oversold at the moment, and from I can glean from a quick glance of your Uni page, you’re still in the first year of the project or so(?). So good luck with the rest.

          • gwathdring says:

            Notable depends on the audience. A widely-read gaming blog has a different perspective on note-worthy AI research than other sources do. And this relates fairly directly to gaming since it’s a program that builds games. I could get annoyed at the BBC for posting an article about research I read in Nature ten months ago and Scientific American four months ago. But for people who don’t read those publications or have access to them, it genuinely is *new*.

            I also recognize that’s not your whole point, and you have further issues with the specifics of the technology at work here that you’ve been discussing in other parts of this thread. But even giving you complete benefit of the doubt in a field I’m not well versed in at all, I think you’re being enormously unfair insofar as what constitutes something worth being posted on RPS. This isn’t a peer reviewed scientific journal, or the gaming equivalent of Nobel Prize nominations.

          • LionsPhil says:

            I think the short answer is “it would have been better if RPS waited until Mike had got a bit further with it”, not “RPS should never have posted this at all”.

            (Christ, if anyone had publicised my PhD pre-midpoint it’d have got ripped apart twice as hard, and rightly so.)

          • Nick says:

            his main objection, as ever, is that people other than him are doing things others find noteworthy and he is jealous. See also his hissy fits over actual game developers with more talent in their little finger than in his entire body getting little comment thread tirades. etc.

          • LionsPhil says:

            A swing and a miss!

          • gwathdring says:

            I think it’s noteworthy and interesting as it stands. From the article, links, and things talked about in this comment thread I have learned about a lot of things I wasn’t aware of in a field that I really enjoy but don’t follow especially closely as it is not my field of study at the moment and I’ve been rather caught up (currently on break, though, thank goodness). I’ve found Mike to be informative, courteous, and genuinely passionate in his responses to others and I’m rather glad he and his research found their way onto this site.

            I know absolutely nothing about you, so I’ll assume you either have a good reason to be in a bad mood or genuinely believe your reasons to dislike this article are fair. But you certainly sound like you have a tangential axe to grind. If I hadn’t behaved similarly in other comment threads on bad days I’d be very annoyed. As I have, I’m no more than mildly irked. If neither foul mood nor ulterior motive apply, then you are welcome to my apologies and we will agree to cordially disagree.

            (Especially about the “and rightly so.” I’m not much a fan of “tough love” in either education or research. To me a good tough question perplexes and/or inspires–it doesn’t berate or tear apart. So with a good professor or advisor.)

    • Gap Gen says:

      I think Mike suggested making a Daily Mail instance of ANGELINA, and have them argue.

      • Mike says:

        Definitely a consideration. Although these systems are really primitive, they are helping us consider questions about bias in text mining, or whether an AI system can commit libel/slander (I guess slander would be hard…)

        I think it’d be great fun to run it on another paper. Or even papers in another country. North Korea’s national paper versus the New York Times. Propaganda Face Off!

        • Skabooga says:

          or whether an AI system can commit libel/slander (I guess slander would be hard…)

          I’m envisioning a text-to-speech subroutine which hacks into various PAs, intercoms, and speakers and lets the mud fly. That could be a fun sort of world to live in for a day.

        • gwathdring says:

          The speaker thing would indeed be awesome. :D

          @ Mike:
          I know it was meant to be quick, one-off joke but I’m curious: what do you see as the bias of the New York Times? I’m a longtime reader and I sometimes get a sense that something is off, but I’ve always had a hard time finding a particular thread of bias let alone of reader-manipulation that reached consistently throughout the paper. It isn’t always the best objective, just-the-facts source. But it doesn’t feel like it has a coherent message.

          At least from a domestic, US perspective it doesn’t. I guess I haven’t closely analyzed it for bias from an international perspective. I try really hard to dig up a wide variety of sources on international news, and I tend not to analyze my sources as closely for specific biases on that front because I’m already protecting myself through variety.

          I don’t mean to pick apart your post–it just set me off on a tangential thought and now I’m curious.

          • Mike says:

            Oh, I didn’t mean bias in that sense. Just the perception each country have of others, what they discuss or focus on in a given story.

            I didn’t take any offense!

          • gwathdring says:

            Ahhhhhh. Gotcha. And here I was half-hoping there would be something I hadn’t noticed and I would gain new enlightenment with respect to the news I read. ;)

            Anyways, that would be a pretty neat project.

      • sinister agent says:

        That’s a bit of an unfair fight, though. You should give the opponent a source that’s also a newspaper, not a comic.

    • Mike says:

      I’m glad it’s been interesting for you!

      ANGELINA looks at the top five articles and uses some metrics to choose articles that are more interesting to it. For instance, if the article mentions someone that it hasn’t heard of before, it’s more ‘interesting’ and therefore more likely to be chosen as the subject of the game.

      I don’t actually uses non-News section stories, but my supervisor is working on poetry and art. One system he’s built uses the entire Guardian corpus, including sections like Travel…

  25. ColOfNature says:

    Some kind of warning, please, if you’re going to have that sinister git in the leader image. And Staring Eyes, surely?

  26. The Godzilla Hunter says:

    Do we need any more proof that “True Artists” are actually robots simply phoning it in? I’ll grab a gun and a torch, you contact the military, and we can stop the robot war before it begins!

  27. Universal Quitter says:

    If they can program an AI to weigh popular opinion to decide like and dislike, how much harder would it be to program it to dislike people for lying or for other actions? Obviously getting a computer to accurately detect lies in the public forum is easier said than done, not to mention figuring out what constitutes ‘bad’ behavior, but this seems promising.

    Ha, I just realized that the only reason ‘the robots’ might overthrow/exterminate us someday is because we’ll program them to be like us.

    • LionsPhil says:

      A lot harder, since to detect a lie you need to understand the actual semantic content of a statement, and AI research kind of threw its hands up in the air on that one decades ago and split off into reimplementing PROLOG as more “web” (my ex-field, tenuously), and just fuzzy-mining the heck out of everything because it’s apparently OK if your dataset is less reliable than the judgement of your spamfilter. Oh, and virtual insects, because if you clap your hands and believe hard enough one day we’ll go from “can avoid light” to strong AI (a term we need precisely because AI has become so watered down by virtual bugs et. al.) by something something larger neural nets.

      We’ve probably got better odds of starting the zombie apocalypse before the robot one.

      • Mike says:

        I think you’re being unfair, and I’ve seen this in a lot of people who left more theoretical AI fields and who claim it’s all “a waste of time” or obsessively abstract. A lot of work goes into detecting things like lies, sarcasm, irony, double entendre and so on. Yes, it’s a difficult task. No, no-one is ‘throwing their hands up’ about it.

        One of the reasons why ANGELINA can do what it does is because we’re taking advantage of a whole world of systems that analyse text, judge emotion, and so on. The current state of AI research is a wonderful thing in my opinion.

    • Mike says:

      Have replied to as many comments as I could manage, but I will return tomorrow to do more! Feel free to email me as well –

      Thanks to everyone! Hopefully in six months I’ll have another step forwards for people to play with.

      • LionsPhil says:

        Well, thanks for taking the time. I think a reasonably big part of “the problem” here is this article overselling what you’ve scoped so far.

        • Mike says:

          I understand if that’s your feeling, and if any article does oversell the work I take full responsibility for not communicating it effectively enough. But I think Lewis did a great job here. Of course people are enthusiastic and optimistic about the work, because it hints at an exciting new line of inquiry for gaming and for research, but I don’t think anyone’s claiming that ANGELINA is a bona fide game designer right now.

          But that is our ‘goal’ of sorts, and discussions with everyone, including people who don’t like it, are key to achieving it I think.

          • LionsPhil says:

            Ok. Well, good luck with it, and enjoy the remains of the bank holiday. Sorry if I’ve been a bit of an arse.

  28. TechnicalBen says:

    Wait, is the computer designing a game, or just googling for images to populate it. Because to me, this sounds more like “AI for content” not “AI for gameplay“.

  29. Sisco says:

    “Oh that damn Joker! What a fool he was! Now I have to spend all day computing pie because he plugged in the Overlord!”

  30. The Greatness says:

    So what happens when there’s a Guardian article about this? That could get pretty meta.

  31. jwoozy says:

    “France France Revolution” is easily the most clever thing done with a videogame in the past 10 years.

  32. Pender says:

    So basically it’s a collage of pictures harvested by Google based on the news for the day, with some sort of generic platforming game pasted in front of it? Sounds like garbage.

    Sorry to be a buzzkill, but it cheapens the work that real AI researchers do when lame publicity stunts like this recalibrate expectations. This is not AI; it’s a project that was designed to generate a misleading headline.

  33. Delicious Narwhal says:

    I really hope older versions of Angelina get preserved for posterity, because right now this is the Horse Ebooks of indie games… and I love it!

  34. ichigo2862 says:

    “France France Revolution”

    Sunovabitch, I think I just fell in love with an AI.

  35. Aatch says:

    As a computer programmer and somebody with a wide range of interests, this is fascinating. I love the idea of using cultural data to generate a “relevant” game (for a given value of relevant…).

    There are several posters that I don’t believe understand the complexity of this project, irrelevant of any prior work or systems. I have done undergraduate AI and it is definitely one of the more complex topics that I have done.

    And to the comparison to in-game “AI”, NPC characters are actually far more complicated that it seems. There is definitely an “Uncanny Valley” of AI actors in games, since they actually get so many major features right, it is the details that we notice. Sure we still get NPCs incapable of path-finding in certain situations, and we still get NPCs that don’t notice bullets flying into their head, but they are the exception and are bugs in the software. Just the simple act of marching in formation is an intensive process for AI, since if you combine it with path-finding you get an exponential amount of checks and dependent calculations, in order to allow a group of NPCs to keep formation to an acceptable level and still path around an obstacle. Not to mention the fact that you would have an entire set of rules for when the formation cannot be kept, and behaviour in that situation. There are a practically infinite amount of cases and sometimes you are unlucky enough to stumble into one that wasn’t accounted for.

    Sorry for the mostly off-topic rant, but the complexity of programming is rarely understood by most people, and only appreciated by those that either do it or are close to those that do (my partner has developed a new-found appreciation for software after seeing how much work goes into even fairly simple features for the project I work on).

    Anyway, good work Mike, I wish you luck with your on-going PhD.

  36. wuwul says:

    This is the future, since it’s clear that game development by humans just doesn’t scale.

    However, it seems that this AI sucks compared to strong AI, so we are still a long way off.

    • wuwul says:

      Actually, after playing one of the “games”, well, it doesn’t just suck, it’s a fraud and a travesty.

      The so-called AI apparently just copy-and-pastes pictures from Google News into a bland horrible ancient generic unplayable platformer.


      Game development AI needs to intelligently generate compelling gameplay, not copy and paste unrelated crap into shit games.

  37. dhankinson says:

    Angelina “likes” Assad…what the hell does that even mean?? A program can’t “like” anything. Either something is necessary or it’s not. Adequate or not. It’s like throwing a .5 into a binary string. She LIKES. Weird.

  38. E_FD says:

    Depressingly, it also sounds like an unintentional metaphor for political commentary and/or “sophisticated, intelligent” games to take platformer mechanics and slap on a bunch of unrelated pictures/soundbytes about whatever the latest political hot topic is.

    At least France France Revolution is an absolutely fantastic pun.

  39. Natalie says:

    What happens if we link this up to 4chan…