It seems like gaming is the ultimate path to reach AGI. OpenAI which created the generative AI hype with ChatGPT is planning to up its game literally as it ventures into simulated world to achieve its ultimate goal of attaining AGI. For the same purpose OpenAI recently acquired Global Illumination, a New York based startup that has been leveraging AI to build creative tools, infrastructure, and digital experiences.
Global Illumination’s team has previously designed and built products early on at Instagram and Facebook and have also made significant contributions at YouTube, Google, Pixar, Riot Games, and other notable companies. The creators of Global Illumination are currently working on Biomes, an open source sandbox MMORPG built for the web which allows users to build, forage, and play mini-games straight from the browser, similar to Microsoft’s Minecraft.
The acquisition of Global Illumination comes at a crucial point of time for OpenAI. The current challenges faced by ChatGPT, including user retention issues, have prompted the organisation to seek solutions. Furthermore, acquiring sufficient training data has proven to be a struggle. OpenAI has recognised that the remedy to these challenges resides within the realms of gaming and reinforcement learning.
This open-source Minecraft clone game provides an excellent opportunity for OpenAI to gather extensive data on human-computer interactions. This data will undoubtedly prove invaluable for advancing their research and development of AGI systems.
Simultaneously, the game serves as an ideal platform for testing AI systems, allowing OpenAI to observe and analyse the intriguing behaviors that can emerge within a sophisticated gaming environment.
One of the users of X expressed the same idea and mentioned that if OpenAI could create a game where agents and people interact with their own goals, it could provide a great dataset for building AGI. This means simulating real interactions could help a lot in developing AGI.
OpenAI is not the first one to venture into creating a generative AI based simulation world. Previously Stanford and Google researchers, open sourced Stanford Smallville.
In this simulation world, 25 AI agents mimic the lives of humans and they can interact with each other independently with their own independent thinking. These agents inhabit a digital Westworld, unaware that they are living in a simulation. They go to work, gossip, organize socials, make new friends, and even fall in love and each has a unique personality and backstory.
Inspired by Stanford Smallville, VC firm a16z open-sourced ‘AI Town’, a JS starter kit that handles global states and multi-agent transactions to help users build their own little AI civilization.
Following Google DeepMind and Meta
The tactic of using reinforcement learning in games was first adopted by Deepmind (now Google DeepMind). It believes that to reach AGI, reinforcement learning is the ultimate tool. So much so it published a paper “Reward is Enough” where authors suggest that reward maximization and trial-and-error experience are enough to develop behavior that exhibits the kind of abilities associated with AGI.
Google DeepMind used RL algorithms to create neural networks which could beat humans at games like Go which are considered to be the most challenging. In October 2015, AlphaGo became the first program to defeat a professional human player.
In late 2017, it introduced AlphaZero, a single system that taught itself from scratch how to master the games of chess, shogi, and Go, beating a world-champion computer program in each case. Google developed a deep RL algorithm that learns both a value network (which predicts the winner) and a policy network (which selects actions) through games of self-play.
Meta also tried its hand on gaming to train its AI agent. In 2022, Meta AI created the first AI agent CICERO to achieve human-level performance in the complex natural language strategy game Diplomacy. CICERO demonstrated this by playing with humans on webDiplomacy.net, an online version of the game, where it achieved more than double the average score of the human players and ranked in the top 10% of participants who played more than one game.
Simulation to Reality
OpenAI is exploring every way possible to attain AGI. From scraping data from the internet to partnering with news agencies like AP, it is leaving no stone unturned. It is pretty much likely that after creating the simulation, OpenAI will come back to reality and create humanoids which will be able to interact with humans in a natural way.
Earlier this year, OpenAI invested in a Norway-based robotics startup called 1x. Previously known as Halodi Robotics, the startup builds humanoid robots capable of human-like movements and behaviours. With Global Illumination acquisition, OpenAI can bring all the learnings of simulation to reality and build perfect humanoid robots.
Interestingly, OpenAI is not alone in sharing this belief that the world needs physical robots. Recently, Google DeepMind introduced RT-2, the first ever vision-language-action (VLA) model that can see, understand language, and perform tasks accurately in the real world.
Last year, Tesla, the autonomous vehicle company backed by an early investor in OpenAI, Elon Musk, introduced Optimus—a conceptual humanoid robot designed for general-purpose applications.
In conclusion, OpenAI’s torch bearers who once said ‘text is a projection of the world’ clearly is moving away from it and taking a page from Google DeepMind’s AGI strategy. Interestingly, Sam Altman or his fellow board members haven’t uttered a single word about its recent acquisition of Global Illumination. The blog post also doesn’t say much either.