The Center for Data Innovation spoke with Marek Rosa, founder, chief executive officer, and chief technology officer of GoodAI, an artificial intelligence research company based in Prague. Rosa discussed how video game environments are a promising platform for training artificial intelligence systems, and why he decided to pursue a career in video gaming, rather than scientific research, to support the development of artificial intelligence.
This interview has been edited.
Joshua New: You have a successful background in video gaming, and one of GoodAI’s first products is Brain Simulator, an application which trains artificial neural networks how to play simple games such as Pong or navigate a maze. Is this just because of your expertise in video game production, or do games hold a special significance in artificial intelligence research?
Marek Rosa: Games provide an excellent learning environment for artificial intelligence (AI). Humans learn to play games through training or trial and error, and devise winning strategies using their experience. An AI agent can, too. In a virtual game, a human mentor can steer the artificial agent to a desired goal with reward and punishment signals, like a sports coach.
As a near-term goal, we plan to connect Brain Simulator to video games of our sister company, Keen Software House. I believe this will greatly accelerate research because not only would we be giving AI to players, but we would also be giving thousands of cases of real time human-AI interactions to AI researchers. We don’t think AI can be developed in a box, but rather it needs to interact with the world and receive rich input. A virtual world with numerous stimuli and real-time training is an excellent first step.
A great advantage of training AI in a virtual computer game world is safety. The price of a potential AI mistake is really low, since it cannot cause actual damage or hurt anyone. In the worst case scenario, it will disappoint a player, but no one’s safety or property is put at risk.
New: You specify that the goal of GoodAI is to create general artificial intelligence, rather than a narrow artificial intelligence designed to accomplish a specific task. Aren’t these narrow approaches the building blocks to general artificial intelligence? Or does this require a fundamentally different approach?
Rosa: We believe that to successfully build general artificial intelligence, we need to have three pillars: the right cognitive architecture, correct neural network modules within this architecture, and the right training environment, which we call a “school” for AI systems. At various stages in this school, we are planning to add and test different abilities, including perception, memory, communication, and problem solving. While each of these abilities could roughly translate to a narrow AI function itself, adding these abilities together would increases the complexity of the overall AI system.
However, generalization can be achieved only by connecting these abilities in a proper cognitive architecture. Our aim is that in this school, AI systems will engage in gradual learning—building new knowledge on top of existing knowledge and developing adaptability and creativity—which far surpasses the ability of narrow AI, which tends to focus on optimization. We want to avoid optimizing specific features because doing so could slow the progress to our final goal.
Our plan is also to incorporate ethics and empathy into the learning process and find a way to teach AIs to understand and cherish humans values.
New: You’ve said the goal of your career has always been to ultimately develop artificial intelligence. Why start with the video gaming industry? What could it offer that, say, academia, could not?
Rosa: I believe that many research projects do not show enough progress because of insufficient funding. In academia, this is often what stops you from moving from theory to concrete results. I wanted our results to improve dynamically, which is why I was looking for a way to secure the financial side of the AI project first. The success of my video game company has allowed me start general AI research with a substantial personal investment, which will be enough to support the project for several years. Without having to worry about funding, our team members—who do have have rich academic backgrounds—can fully concentrate on our research. This means that our progress is not dependent on quick commercialization, which I think would diffuse our focus from the main goal of general artificial intelligence.
New: Going from solving simple games to general artificial intelligence seems like a big leap. What is your timeline?
Rosa: Honestly, it is hard to predict when general AI will be achieved. It might happen in 30 years, or it might happen sooner. But my mindset is that with each day that goes by without it, we’re losing something important, so we need to push ourselves as if tomorrow were the deadline. I want to get to general AI as fast as possible, because in the face of external risks, humanity cannot afford to lose any more time. Every day without general AI is a day we’re missing out on technology that could save lives and make our world better.
In our team, we all share the same sense of urgency and we believe that general AI is the key to accelerating science and solving world’s most dire problems. Since the recipe for general AI is not yet known, we have chosen to proceed according to incremental milestones with tight weekly deadlines. This allows us to test ideas quickly, avoid dead-end approaches, and move on in the most promising directions.
New: Brain Simulator is open source, and you encourage other researchers and companies to experiment with it and share their knowledge. Why take an open source approach?
Rosa: I believe that cooperation always wins out over competition. General AI research is a complex field with numerous aspects, and joining forces can finally bring pieces of the puzzle together. General AI will greatly impact people’s lives in the future, so it makes sense to contribute to the common good from the start. I don’t think that such technology can be contained within one company. It has the potential to become a generally accessible, shared resource, like the energy of the sun or the air we all breathe. This is the future I would like to see.
We want to encourage people to learn more about AI and support a new wave of AI researchers. Through both Brain Simulator and its integration with games we also aim to expose people with less technical backgrounds to the world of AI research.
