The day is not far when we will see the first autonomous games. These will be games created by artificial intelligence from start to finish
In the science fiction book "Ender's Game", the hero of the book has to undergo an extraordinary psychological treatment. Instead of talking to a psychologist on the couch, the hero plays a computer game. He faces obstacles and enemies, who test him in different ways and teach him lessons for himself. The game is customized for him, and changes automatically according to the mental development of the player.
Until a few years ago, such games were the domain of science fiction only. Although human developers used basic artificial intelligence to produce certain elements in games, the level of flexibility and creativity of the artificial intelligence was still very low. Beyond that, the intellect is not endowed with the ability to understand complex connections and ideas. Human developers were still needed to prod the AI to produce material they could use, sift the chaff from the chaff, and incorporate the best results into the game.
Now things are starting to change, and the time is not far when we will see the first autonomous games. These will be games created by artificial intelligence from start to finish. They will be games that the artificial intelligence will manage with a high hand, perfect and improve and add new levels of complexity and plot to them, so that they will never end. And they will be personal games like Ender's game, which will adapt themselves for each player and provide him with thrills and pleasures - and hopefully also help him develop as a person.
And it's already starting to happen today.
ChatDev
The first autonomous game has already been created by artificial intelligence.
almost.
It happened in September 2023, in an academic study where researchers created a community of artificial intelligence. It was, in fact, a simulation of a software company, where every position in the company was staffed by artificial intelligence. There was an AI that took on the role of CEO, another AI was given the role of Chief Technology Officer, a third AI was the graphic designer and so on.
The researchers tasked the virtual 'company' with developing a simple game called Gomoku. You probably know it as "five in a row/column/diagonal". She did it the same way game companies do it today. The CEO took the lead, produced a strategy, consulted with the other senior officials to refine it. The artificial managers translated the strategy into instructions that were passed on to the programmer, the graphic designer, the technical writer and all the other officials in the company - each of whom was himself 'just' artificial intelligence. Each of the artificial intelligences 'talked' and exchanged information, ideas and criticism with the others, and from all this joint action - a computer game was obtained.
almost.
Of all the pieces of code written by the artificial intelligence, 86 percent worked well. About half of the failures were due to limitations imposed on the number of tokens in the answer, and the rest came mainly due to failed attempts by the artificial intelligence to activate external services. The first problem is mainly technical, and is easy to fix and deal with. The second will accompany us for several more years, until artificial intelligence can function as a key at a high level - an achievement we expect to reach in less than a decade.
It is certainly possible to argue that a game like "five in a row" is not particularly complex, but the above study was used as proof of ability: when artificial intelligences cooperate, they can carry out tasks that seemed impossible for a single artificial intelligence.
And the artificial intelligences only acquire new capabilities with the times.
The genie is out of the bottle
Last week, Google's DeepMind demonstrated a spectacular new artificial intelligence model named Genie. Ginny was trained on 200,000 hours of gaming videos from the web, and specifically on 30,000 hours of videos documenting XNUMXD platform games. During this massive training she learned to divide each second of animation into images and 'understood' how each image followed the one before it. Not only that, but the system learned to predict what kind of actions - that is, which buttons on the keyboard - affect the movement of the character on the screen.
After Gini learned all these, the researchers ran her on a single image of a character on the screen - and showed that the artificial intelligence turns the image into a real computer game: it recognizes the main character, allows the player to control the character and explore the world in the image, generates the animation itself The movement of the character and even continues to produce a logical continuation of the game beyond the original image.
As with the previous study, here too we have to put our finger on the limitations that exist today. Ginny is still in her infancy. It works slowly and only produces an image - a frame - one per second. By comparison, computer games usually run at a minimum rate of at least twenty frames per second. The animations it produces are not always clear, and it can 'hallucinate' behaviors that we would not expect in normal computer games - for example, a situation where a character continues to float in the air after jumping. And if all this is not enough, then she only has a short memory, which does not allow a run for more than a few seconds.
But what did you expect from such a preliminary study? Genie is only the result of quasi-academic research, designed to demonstrate initial advanced capabilities - but not yet turn them into a complete product. It's OK. Chat-GPT was also 'only' academic research two or three years ago. Genie has many months to go before it emerges into the world as a tool to be used by human developers to produce computer games with minimal effort.
And soon after, Genie will be used as a tool by other artificial intelligences as well. Imagine the virtual society I described earlier, and what it could do with Genie. The main writer - an artificial intelligence himself - will produce magical and wonderful adventures, and will pass them to the developer (also an artificial intelligence) who will break them down into games with well-defined stages, each of which will be produced by Genie-4, or Genie-5, or Any version that exists at the time.
Now a psychologist has also been added to the cauldron to analyze the user and his mental state. I guess I don't need to tell you that the psychologist will, of course, be an artificial intelligence himself. The same psychologist will argue with the main writer and developer, and together they will create a game that is personalized for each user. A game that rewards him for determination and bravery, for cooperation and dealing with obstacles. A game that comforts the user when he fails, and also knows how to raise the level for him automatically to keep the challenge at a high level.
It will be the game of everyone's dreams, and we are not far from the possibility of realizing it.
In fact, there are people - even in Israel - who are trying to realize it already today.
The myth makers
When Roy Altman and Ill Elder discovered the power of creative artificial intelligence, they immediately understood its implications for the gaming world. They decided to harness it to make better games, and founded the company Story Tools Studio.
"When chatGPT came out, Eyal and I were working on a board game from a D&D angle," Altman told me. “I talked to GPT and taught him this game, and it worked like a charm. Then we created Mythmaker. The basis was the text, and the composition of a series of prompts that produces the story in real-time with the screenwriting principles we have learned over the years. The trick is that the AI creates the story in real time following your choices, with a coherent and consistent beginning, middle and end.”
Altman and Alder may have started with GPT, but very quickly they realized that it could and should be enriched with additional engines. In the end they created their own engine which they call "Muse", but diving into its depths shows that it is more like the complex system we described earlier, than a single artificial intelligence engine.
How does Muse work? When a user starts a new adventure in the company's first game - Myth Maker AI - he chooses the character he will play. Muse immediately summons - behind the scenes - ChatGPT, and asks him to produce the first steps in the adventure the player will embark on. The GPT engine generates the text that the player will be exposed to in these stages, since most of the game right now is text based. But at the same time, Muse also calls Leonardo - an artificial intelligence engine that will produce images that describe what is happening in the game. He asks another artificial intelligence to produce the background music for each stage of the game, and to read aloud what is happening for the player. Soon Muse will also be able to produce complete videos describing the adventure. And of course, also stages where the adventure stops being textual and becomes a 2D platform game.
Want to guess what engine Muse will use to make this happen?
The result of all this combination is a complex and multi-part artificial intelligence that drives games that even their original creators do not control from the moment they are launched.
"We were asked when we had time to write all these stories, which is a huge compliment." Altman laughed. "Someone approached me and asked me what he should say to the witch in the ninth stage of the game. And I don't know. I don't know a witch. The AI made up its own story for that player.”
Where are they going to go next? Both entrepreneurs already have a vision for the future.
"Our vision is that you will take the story that is the basis of everything, and the AI will create the graphics according to it in real time, as you progress through the story, with your choices." Eldar told me, adding, "Today it's pictures, but in the not-so-distant future you'll put on your virtual reality goggles and go on a Witcher-style adventure, but endless and capable of lasting forever." Or take an existing game like God of War, and after you finish the company's original campaign, you can continue in the same world with an endless story that you create for yourself with the character you choose to play."
Altman and Alder's Muse engine is just one small example of the wonderful new world of gaming we are about to enter. Games that will teach us about ourselves, that will allow us to explore wonderful worlds alone and together, that will help us acquire life skills and friends in the virtual world and in the physical world. Endless games that will complete the physical world in all its complexities - but will provide us with a more pleasant environment to experiment, fail and learn in.
"It's magic," concluded Altman with shining eyes on the new experience of creating games with artificial intelligence. "It's magic, real magic."
