OpenAI Acquires Global Illumination: Accelerating the Journey to Simulated Worlds for AGI

2023-08-22

It seems that gaming is the ultimate path to achieving AGI. After the success of creating ChatGPT, OpenAI plans to achieve its ultimate goal of AGI by entering the simulated world. For the same purpose, OpenAI recently acquired the New York-based startup Global Illumination, which has been using artificial intelligence to build creative tools, infrastructure, and digital experiences.

The team at Global Illumination has previously designed and built products for Instagram and Facebook in their early days, and has made significant contributions to other well-known companies such as YouTube, Google, Pixar, and Riot Games. The founder of Global Illumination is currently developing an open-source sandbox MMORPG game called Biomes, which is web-based and allows users to build, gather resources, and play mini-games directly from their browsers, similar to Microsoft's Minecraft.

The acquisition of Global Illumination is crucial for OpenAI. ChatGPT currently faces challenges such as user retention, prompting the organization to seek solutions. Additionally, obtaining enough training data has always been a challenge. OpenAI realizes that the solutions to these challenges exist in the gaming and reinforcement learning domains.

This open-source Minecraft clone game provides OpenAI with an excellent opportunity to collect extensive human-computer interaction data. This data undoubtedly has great value in advancing their research and development of AGI systems.

At the same time, this game also serves as an ideal platform for testing artificial intelligence systems, allowing OpenAI to observe and analyze interesting behaviors that may occur in complex game environments.

One user expressed the same view and mentioned that if OpenAI could create a game where agents and humans interact with each other and pursue their respective open-ended goals, it would provide a real dataset for building AGI. This means that simulating real interactions can be of great help in developing AGI.

ChatGPT will never have access to better question-answer data than what Google already has. If OpenAI can build a successful game where agents and humans interact with each other to achieve their open-ended goals, it will be a real dataset for building AGI.

——NirSD (@nirsd) August 16, 2023

OpenAI is not the first organization to attempt creating a simulated world based on generative artificial intelligence. Previously, researchers from Stanford University and Google open-sourced Stanford's Smallville.

In this simulated world, 25 AI agents mimic human life, interacting with each other independently and having the ability to think independently. These agents live in a digitized Westworld, unaware that they exist in a simulated environment. They work, chat, organize social activities, make new friends, and even fall in love, with each agent having a unique personality and background story.

Inspired by Stanford's Smallville, venture capital firm a16z open-sourced an introductory JS toolkit called "AI Town," which handles global state and multi-agent transactions, helping users build their own small AI civilizations.


This is similar to DeepMind and Meta


The use of reinforcement learning strategies in games was first adopted by DeepMind (now Google DeepMind). They believe that reinforcement learning is the ultimate tool to achieve AGI. Therefore, they published a paper titled "Reward is Enough," in which the authors suggested that reward maximization and trial-and-error experience are sufficient to develop behaviors with AGI-related capabilities.

Google DeepMind used reinforcement learning algorithms to create neural networks that can defeat humans in challenging games such as chess. In October 2015, AlphaGo became the first computer program to defeat a professional human Go player.

In late 2017, they introduced AlphaZero, a single system that can self-learn how to master chess, shogi, and Go, defeating world champion computer programs in each case. Google developed a deep reinforcement learning algorithm that learns value networks (predicting winners) and policy networks (selecting actions) through self-play games.

Meta also attempted to train its AI agents through games. In 2022, Meta AI's first AI agent, CICERO, achieved performance comparable to humans in the complex natural language strategy game "Diplomacy." CICERO demonstrated this by competing against humans on the webDiplomacy.net online gaming platform, scoring higher than the average human player and ranking in the top 10% of participants.

From Simulation to Reality


OpenAI is exploring all possible ways to achieve AGI. From scraping data from the internet to collaborating with news organizations like AP, they are leaving no stone unturned. It is likely that after creating a simulated world, OpenAI will return to the real world and create humanoid robots capable of interacting with humans naturally.

Earlier this year, OpenAI invested in the Norwegian robotics startup 1x. The company, formerly known as Halodi Robotics, builds humanoid robots capable of human-like movement and behavior. By acquiring Global Illumination, OpenAI can apply all the experiences gained from simulated learning to reality and build perfect humanoid robots.

Interestingly, OpenAI is not the only company that believes the world needs physical robots. Recently, Google DeepMind unveiled RT-2, the first visual-language-action (VLA) model that can accurately perceive the real world, understand language, and perform tasks.

Last year, Tesla, a self-driving car company supported by early investors of OpenAI, Elon Musk, introduced a concept humanoid robot called Optimus for general-purpose applications.

Summary


Overall, representatives of OpenAI have previously said that "text is the world's projection," and now they are clearly changing their strategy, borrowing some methods from Google DeepMind's AGI strategy. Interestingly, neither Sam Altman nor any other board members have commented on their recent acquisition of Global Illumination.

RECENT AI NEWS

RECENT AI TOOLS