We are thrilled to announce our investment in modl.ai — a leading scientific and interdisciplinary team that delivers technology accelerating and automating game development and enhancing player engagement. As deep tech investors, gaming was initially not the most obvious area for us to focus on, but we soon came to realize that the field is closely interrelated to AI/ML development in two ways: 1) simulated/virtual environments are already essential to machine learning and we expect a broadening to more diversified AI/ML application areas, 2) AI/ML has great potential to automate the creation of such simulated environments.
Gaming for AI-training
As there are practical limitations to how much real-world data can be collected for AI-training purposes, an increasing number of AI models rely on simulated data environments. Simulated data can increase both quantity and quality of training sessions to better resemble the complexity of the real-world. Especially when it comes to understanding and modeling of human behavior, the need for diverse scenarios and oftentimes unpredictable behavior significantly increases complexity. (Synthetic data creation as in replicating and adjusting existing datasets does not necessarily allow for the same complexity and diversity as a fully simulated environment.)
Games have become an especially potent training environment as it increasingly embeds multiple and interacting players and highly complex contexts and environments as well as edge cases. Alphabet’s Deepmind has recently made strong advances in adapting AI agents to complex games such as StarCraft; other researchers showed lately that AI agents can navigate multiplayer games such as six player poker. Particularly in areas such as self-driving cars (UAVs) have we seen the application of test environments that can model all kinds of unpredictable human behavior before implemented in real life.
Also in the area of simulating data for VR/AR environments, we see interesting developments like FBs recently launched AI Habitat, which offers an advanced and open source simulation platform for embodiment of agents in 3D life-like environments.
As these initial technological developments indicate, we expect virtual environments created by gaming (or VR) developers to become an increasingly important source for future AI training.
Automating gaming development
Now, complex gaming environments are themselves laborious to create. Especially demanding tasks are manual game testing, continuous game updating of content and code, balancing of games, anomaly detection and toxic player/bot detection. Beyond high time consumption, manual testing and one-off game development may lead to shallow user models and lacking personalization which again result in low retention, missing revenue potential and poor player experiences. Freeing up game developers’ time normally spent for such tasks, may further create room for more productive creative design.
modl.ai creates AI personas based on models of actions, motivations and experiences. By combining psychometric insights and AI methods they can predict what users will do, want and experience more accurately and faster than current models. Thus, based on player profiling they can evaluate content based on personas, create game-playing agents, detect anomalies and personalize content, all in an autonomous fashion.
Games as a service & beyond
The gaming market in itself is a huge and lucrative market (TAM 150bn in 2019; expected to grow to 300bn by 2025). Games as a service, or GaaS, is a massive opportunity and a trend expecting to substantially overtake more of the gaming market in the near future. As games as a service require continuous adaptation and new iterations every couple of weeks, the trend is making the gaming industry go from one-off products to creating a long-term relationship with their players.
Short update cycles may strain production resources and fast updates risk breaking existing content. Thus, continuous evaluation of performance calls for scaling across content and increased focus on life-time of players. Game performance is more likely to have periodic peaks in performance and traction instead of a one-time high with a flattening-out over time.
The GaaS trend enforces the need for automating game development and strengthens the business case of modl.ai. Better prediction of user behavior and personalized content will increase life-time performance and help build the long-term relationship with players. Insights on player profiles and motivation will also offer valuable information to game developers before, during and after the game development process. More accurate profiles will increase player engagement and allow for content adaptation to personas. It will also enable detection of toxic or bot players in games, particularly relevant for monetized incentivisation.
This automated and procedural content generation, defined as “the algorithmic creation of game content with limited or indirect user input” (Togelius et al., 2016), further goes hand in hand with a recurring revenue model, having the potential to generate substantial and more stable revenue for game developers and the tools they use (such as modl.ai).
We believe modl.ai with the team’s strong academic record and multidisciplinary approach will change how gaming simulations are automated in a human-like fashion, something the first commercial traction and great user feedback testify to. They are an academically accomplished team combining expertise from areas of AI/ML for gaming, gaming industry, psychology and UX. With 10+ PhDs on the team and some of the most highly published researchers in AI gaming (e.g. from New York University, University of Malta, IT University of Copenhagen), the team’s diverse capabilities make them stand out. Next to various highly cited articles on procedural content generation in gaming and player modeling, the co-founders Julian Togelius and Georgios N. Yannakakis have for example recently published a textbook on AI in gaming (2018). On top of this, the team has a long history of professional game development, having launched multi-platform and critically acclaimed game titles, giving them first-hand insight into the needs of game developers.
Modeling human behavior:
The market for automating game development is interesting in itself, but also for gaming as simulation environments as it expands to more human-realistic simulations beyond the gaming industry. As AI training data demand expands beyond UAVs to include more complex human behavior, we expect the simulations market to grow in parallel direction.
Being able to model complex 3D games and adjust simulations in real-time will give the autonomous approach of modl.ai tremendous potential in various applications for simulations. To realistically model human behavior and offer personalized simulation environments will demand an interdisciplinary approach, and we believe this team is the right fit for the job.