I really like how you've articulated your entire process. This was a very enriching read. A Well deserved feature!
Great article! Thanks for the awesome read.
Wow, this is so cool! Nice job!
Andrew Maximov talked about the way Promethean AI learns assets and materials and how it can study games to help you make better games faster. Note, that the environment presented below was co-authored by AI.
So what’s the status of the software right now? When do you think we’ll be able to play with the alpha?
We are at a point where most of the infrastructure is there and we are actively scaling to teach the AI new things and introduce features that enhance the core workflow and focus on semantics and art theory. For our early adopters, if you are using the type of engine we’ve finished the integration for, it is almost plugged and play today. However, we need a certain critical mass of learned data per type of space before we can start releasing it to the general public. So no public alpha date just yet.
However, there are other big parts of Promethean AI that were not in the video, that sparked very intense interest from all the developers and studios we’ve shown them to, so there is a high chance that those we’ll be out a lot sooner. My profound apologies for being so cryptic, we can only make so many trailers at once! 🙂
Could you tell a bit about Promethean studies the asset library? What is the way you teach the AI? Does it go through rigorous room study as Alpha Go did?
The important thing to note here is that we separate the concept of a (semantic) object from the 3d assets themselves. Both processes, however, are extremely simple.
For 3d assets, you select them in the asset browser and you ask Promethean to learn them. We don’t rely on any naming conventions or folder structures – Promethean will just look at the asset and tell you what it is and automatically connect the mesh to its semantic representation and record a bunch of other metadata all within a few seconds. That’s it. The accuracy can vary depending on the style of your project and you always have the option to override the semantic connection with one button click. But in general, we find that is much more precise and detailed than people expect.
For scenes, it’s even simpler. Once your assets have been connected, just build a space and ask Promethean to “learn” and “memorize”. That’s it.
There is no months or weeks of training. You can start building things after showing Promethean just a couple of examples. You will correct it as it generates new ones and it will learn from that, continuously improving as you work together. Just like a junior artist would.
And if you teach it something that has fallen out of favor with you, you can always ask Promethean to visualize its memories and make all the necessary adjustments or remove them altogether.
I’m amazed by the function of the asset variations. How rich is this AI’s choice here?
It has experience, it has seen a lot of things, so it knows what goes where 🙂 Just like you and I do. It also has something we call a context. It can pick it up from the environment or you can expressly set it by saying that we are in the 80s for example and then it will make sure the suggestions fit the context.
The overall choice is virtually unlimited. It mostly depends on the assets you have and AI’s experience that we keep constantly expanding. It doesn’t however just throw everything at you, but rather considers what objects are most likely to appear in the context of the space. It can also inform you if you are missing assets for things that would go great somewhere and could throw in a blockmesh or a bounding box in place. We also stress the concept of creativity and happy accidents a lot, so while suggesting you some likely scenarios we will also throw in a wild card that you might not usually expect there just in case it sparks some interesting ideas for you. Think of Promethean as your junior artist – it suggests – you Art Direct.
You’ve also mentioned that Promethean actually gives you recommendations as to what materials to use. Could you talk about the way your system works? How does Prometheus use color theory to make the right suggestions and choices?
Happy to, especially since most of this didn’t make into the video. Artistic fundamentals as a form of input is a huge focus for us. We want the technology to speak the language of art and meet you where you are. Color is a good example of that. Promethean can look the current frame of your game in the editor and break it down for the color palette, pick out the dominant and secondary colors, give you suggestions for focal points or secondary, tertiary objects as well as choices of particular types of color compositions be it a classic triad, analogous or complementary with an accent and so on.
Of course, we are looking to apply similar concepts to assets as well with things like focal points and shape language being part of your inputs.
Computer interfaces for many years have been dictated by how data flows within a system rather than how humans solve problems. We didn’t want to perpetuate that. We want to build technology that doesn’t require you to know anything about technology and everything about your domain – art. We do believe that being built by artists and for artists makes a difference.
The tidy room feature is extremely unusual. Are there experiments with other similar abstract notions?
Yes, very much so. One of our core mantras is “high-level creative intent immediately translated into actionable 3d content”. As an artist when looking at a space you don’t think “I wish the lamp was at coordinate xyz”. You think “I wish this space was messier, had more ‘sport stuff’ in it and clearly showing that the character inhabiting the space is into music”. While this is not fully implemented we have all of the underlying technology for you to say I want a space that is 10% sporty, 50% musical and 33% messy. That is essentially the type of feedback you might get from an Art Director. And you are Promethean’s Art Director so that’s how you interact with it.
What about restrictions? Is the system limited to smaller spaces? Can it work with natural environments? Can it study games?
We are starting with man-made interior spaces because, while we can quite plausibly get away with generated natural biomes based on procedural randomization, the amount of complexity and interconnectedness in semantically rich human spaces is something that is absolutely infeasible to do procedurally. But as we grow and scale we will definitely be moving out to every type of space. Our patent-pending learning algorithm is built from the ground up to scale all the way from a bolt holding together a bed frame to the entire city that the bed is in.
And yes we can very much study entire games and offer that as an option to our early adoption partners where we can pre-learn their previous projects so that when they start on with Promethean it will already have adopted some of the choices and traits specific to their team.
What do you think the future of game development looks like and what would you say to those who are worried that progress will jeopardise their jobs?
First of all I’d say “definitely keep being critical”. This is no laughing matter and whoever does this deserves all the scrutiny you are going to put them under. The gravity of the implications of this technology is strongly etched into our minds and we hold ourselves to the same highest level of scrutiny on this subject every single day, so there is no reason you shouldn’t. You don’t owe us your trust and for all intents and purposes, we as individuals are irrelevant to this conversation. All we seek to do is build technology that will empower and inspire, that will help creative people and give them the economic independence to do it on their own terms should they so desire. And the manifestation of this is the product we build. And you have every right to judge us on that. But before we can get that product into your hands all we can do is our best to give you regular updates like this to help you form an idea of what that product is and what it stands for. So that you can form an informed opinion.
On a very personal note, however: I understand. As an artist of 15 years with many of those spent with the privilege to serve and support some of the best artists in the world and the artistic community as a whole, the choice presented itself quite unambiguously. There is no future in which this technology doesn’t exist. But do we wait for some professor in a stuffy university or some tech giant to invent this or do we do it by artists and for artists? That is why Promethean AI exists.
Secondly, I’d say that most people working in the industry are severely under-informed on the economics of the business and the real challenges we are facing. To try to break it down simply: the cost of AAA game development grew 100x from PS1 to today while the install base size and game prices barely moved in comparison. And we are projected to go 200x by the end of the next console generation. That is crazy and that trajectory is just not economically sustainable. So it’s not the question of AI making our jobs easier – if we don’t come up with something there will be no jobs to make easier. The people with the money will be much wiser to invest them elsewhere. I hope we all understand that the threat to job security is the fact that those jobs are not sustainable long-term in the first place. This is not a situation where if we just do nothing things will stay the same – they won’t. And technology is a potential solution here. If we don’t take it, our only other options are moving all the jobs somewhere a lot cheaper or predominantly building battle royals and MOBAs that can be monetized extensively without needing that many artists. All of these are already happening and the coming console generation transition will only make it more pronounced.
Yet at Promethean, we believe both developers and our audience deserve better. Instead of squeezing 3x money from the same user we want to make the game 10x cheaper to produce. Because if we can keep the costs down we can keep the teams where they are and provide for healthy growth for everyone. Equally as important, we can take more creative risks, make more voices heard and help people have fulfilling creative careers without having to sacrifice their health, their families or their friends. Instead of having 10 AAA games a year that are generally liked by everyone we can have 500 that are extremely liked by smaller groups of people and still have the gigantic games be bigger and better. That is why Promethean AI exists.
And yes we believe that a world where everyone’s creative dreams can be realized is the right future to aspire to. And on a more personal note, I just happen to believe that it is in fact sorely needed. Mostly because I’ve met you. You – the game developers. From all the endless conversations we had at countless meetups, workshops, and conferences, from ogling your art online and exchanging millions of messages in chats and forums, and from one too many beers we shared in late-night bars all across the world – from Tokyo to Krakow. You have amazing ideas. You have worlds in you. Great stories to share. Adventures to take us on. And I don’t care if the world has convinced you otherwise. I very selfishly, cheesily and unapologetically want to hear all of them. Because all of them deserve to be told.