The first acquisition since the founding of OpenAI, bought a skinned Minecraft company
Author | Tencent Technology Hao Boyang
OpenAI made its first move, buying a game company that basically no one knew.
OpenAI posted a tweet on Wednesday saying it had acquired digital products company Global Illumination. This is the first known acquisition of this famous artificial intelligence company in more than 7 years of existence.
Twitter posted by OpenAI
Aside from this brief message, OpenAI provided very little information: the financial details of the deal were not disclosed.
The project planning of Global Illumaion after the acquisition is also foggy, OpenAI said, "The entire (Global Lighting) team has joined OpenAI to develop core products including ChatGPT." ”。
Who is Global Illumination?
According to available information, Global Illumination, initiated by Thomas Dimson, Taylor Gordon and Joey Flynn, has been involved in a series of projects since its inception in 2021. Backed by venture capital firms Paradigm, Benchmark and Slow, Global Illumination's team designed and built products early on Instagram, Facebook, as well as YouTube, Google, Pixar and Riot Games.
Global Illumination's official website is rather refreshing (and lacking), only indicating that the project they are currently working on is a product called Biomes.
The official website of Global Illumination
Biomes is a Minecraft skinning game with just 1547 followers on Twitter. According to their description, this is an open-source sandbox multiplayer online role-playing game (MMORPG) built specifically for the web. It features that users can run the game directly in their browser without downloading or installing any additional software. This description sounds unrelated to AIGC.
Bomies' Twitter page
The game's promotional video only emphasizes the most attractive family fun in Minecraft games, mining and construction.
But in an announcement on OpenAI's official website, it was mentioned that "Global Illumiation is a company that has been reusing AI to build creative tools, infrastructure, and digital experiences." “
Announcement of acquisition published on OpenAI's official website
This long-awaited attempt finally showed signs in the Global Illumiation CEO's tweet. Of all the tweets that began in March, AI seemed to occupy his main interest. In an April 6 tweet, he began trying to incorporate ChatGPT into the Biomes game's construction of NPCs. At the time, it was a relatively preliminary question-and-answer application, but there is no doubt that Global Illumiation belongs to the company that began to integrate AI and gaming at an earlier point in time.
Twitter from the CEO of Global Illumiation
Why did OpenAI choose it?
After the news was revealed, Jim Fan, senior AI scientist at Nvidia, commented, "What's next? Run a multi-agent civilization simulation program on GPT-5? ” 。 Other AI practitioners also tweeted their agreement with the speculation.
The latest intuitive version of this so-called "multi-agent civilization simulation system" is Stanford's "AI Town" that has been open-sourced in recent days.
In this town, each NPC has a pre-set character biography, that is, their "initial state", and all their subsequent conversations, socializations, behaviors, and action lines in the town rely on the ChatGPT big language model to complete action decisions. Users can only intervene in their environment, such as making the toilet bowl occasionally blocked. These NPCs will plan new actions according to the changes in the scene.
The daily life of AI town
In the original Stanford paper, when NPCs were in their initial state within 25 days, they had formed a complete "small society" with their own personalities, dialogues, and even romantic relationships. Their social interactions, the process of some things to do, the organization of parties, dinners, etc., are all spontaneously constructed by them after information communication, communication, and judgment of the environment.
And unlike many games that already use AIGC technology, such as Against the Water Cold. The NPCs in this town are not completely mechanical and exist as conversational robots, but are "growing". For example, two people in a virtual town who may not have known each other may become friends because of a chance encounter. Or being persuaded to communicate and become interested in something new, this is all beyond the orientation in the original character setting.
Using AI to train game characters is not a new endeavor at all. The sensational AlphaGO and Deep Blue were both famous for beating world champions by training in the game. Demis Hassabis, CEO of Deepmind behind ALphaGO, is a follower of the idea that AI development should start with games. In the early AI exploration, OpenAI itself also tried to use "Minecraft" for video pre-training to train AI to learn to play this game. In a previous interview, Illya, chief scientist of OpenAI, also mentioned that the experience of defeating human players by training AI to play StarCraft with a large language model gave him a lot of ideas and confidence for follow-up research.
An introduction to using Minecraft as an AI training project on OpenAI's official website
Although OpenAI currently has the injection of Microsoft's "banknote capability", it seems that the revenue pressure is not large. However, considering the huge cost of GPT-5 development, this year's revenue is only $30 million, and next year's target revenue of $1 billion OpenAI still needs a commercial project that can be quickly landed. With such a past experience and obvious potential prospects, OpenAI betting on the game field is entirely possible.
And the concept of "AI agents" is not just a need in the gaming world, but may well be the core area of the next AI evolution.
Agents, the next new battlefield for AI
Back in June, Andrej Karpathy, a former Tesla autonomous driving leader and now OpenAI scientist, suggested at an event that young people who currently want to devote themselves to AI to study AI agents. At that time, GPT extension applications with certain intelligent body characteristics, such as Baby AGI and AutoAGI, were also popular on Github, frantically harvesting high stars. Many researchers in the field of AI believe that AI agents are the next breakthrough point of AI in addition to the relatively distant goal of embodied intelligence.
In an interview earlier this year, renowned philosopher David Chamers said that scientists and philosophers have struggled to unify their definitions of intelligence, but they all more or less agree with William James, the father of psychology, on intelligence: intelligence is setting a fixed goal and achieving it in different ways. It follows from this that AGI, as an intelligence, should have the ability to achieve goals on its own.
ChatGPT, as a chat app and information provider, is undoubtedly very efficient and useful. But a real AI assistant should be able to analyze the criteria for achieving goals according to the situation of the environment, and disassemble a goal into sub-goals, and gradually implement these goals on its own. This is the basic definition of an AI agent: an agent that interacts with the environment and can disassemble the goal step by step. Intuitively speaking, one of the founders of the recently popular Inflection AI mentioned in his new book his new detection method for AGI: only if you give him 100 yuan, it can help you earn 10,000 yuan of AI, is the real AGI.
That is to say, the part that is currently lacking between ChatGPT and AGI is largely the ability required by AI agents. From April to the present, some of the most important advances in AI have also come from the upgrade of AI's ability to build decision trees and decompose tasks. OpenAI's recent code interpreter upgrade has also greatly improved AI's ability to perform and disassemble tasks.
But with all the capabilities, who dares to let them enter reality to perform interactively?
For many AI skeptics, simply giving AI access to the web is already very risky. Not to mention that they can actually order, buy, operate, grow and communicate on the Internet.
The relatively closed environment of the game is undoubtedly the safest amusement and testing ground for AI agents.
So this acquisition of OpenAI may be in the next bigger chess game to develop AI agents. The game field will also surpass the entertainment attribute and become the first to land AI+, and it is also the most fruitful experimental field.