‘Minecraft’ is not only one of the most successful games in history and a great base for fun and learning. Mojang’s title has been used for years to measure what artificial intelligence models are capable of, and Google Deepmind just scored a goal: its Gemini 4 Argon completed ‘Minecraft’ from start to finish without human intervention.
And it seems this AI did it without cheating.
In short. Four Gemini 4 argon-based agents defeated Ender Dragon in 27 hours and 14 minutes. This is the final boss of the game and the results were published in a paper on arXiv with authors attributed to Google DeepMind.
It’s one of the company’s latest models, which has been in closed circuit for weeks so cybersecurity experts can test it, but it was released to paying customers on October 13. Now, it has a definite strategy
blind. The experiment was conducted in a randomly generated world, with easy difficulty and cheat orders disabled. Agents were only tasked with killing dragons, with no other instructions, Minecraft-specific tools, or tricks to speed up progress. Each agent controlled a player using Mineflayer, an open-source library for creating game bots, and to do the work, they vibecoded a small Java program that issued commands.
For example, visiting a site, mining certain materials and such things. Between the four agents they wrote 2,783 programs, exchanged 229 messages, and died 605 times. And there are a few curious things: The first is that only 69 of those deaths were at the hands of bosses. The second is that they themselves share the work.
One made the tools, another gathered supplies, another lit the portal, and another set up the castle. When they had everything, they went after the dragon and began a 14-hour battle where they also devised various strategies to reduce the opponent’s options, such as breaking the regeneration crystal.
strategy. We already said that the difficulty was ‘easy’, but there are some other interesting details that illustrate this victory a little better. According to the researchers, 22% of all the programs they wrote ended up making mistakes, and there was an app that told agents what resources and animals were around them, even those that were out of their sight.
A plugin also planned the best routes for them and they started with an advantage because ‘Minecraft’ was already in the model training process, so even if there was no prepared world or people doing anything, it was clear that the agents had not faced the game.
And, on the subject of value, the agents needed about 1,350 entry tokens to win, and in addition to the four player agents, there was a fifth who was in charge of overseeing the session to make sure none of them cheated. The entire session costs about $810.
Why Minecraft?. Interestingly, out of all the games out there, only a few are suitable for this type of testing. ‘Minecraft’ is one of those chosen for a series of tests and, in DeepMind’s test, what it tries to test is consistency over thousands of dependent steps over a day. Previously, the University of Michigan had already touted the benefits of the game as a flexible 3D environment ideal for reinforcement learning tasks with partial observation and delayed reward.
Other AIs have been trained with ‘Minecraft’ and it’s not just for AI, since it’s a very interesting game in an educational environment because you can practice collaborative, deductive exercises, lateral thinking, train that patience, show your imagination, creativity and because it teaches computational thinking almost unintentionally.
The game system works with electronics-like logic and, in addition, has block programming tools that make it a good gateway to programming. It’s the same thing that makes it attractive for AI training, in short, and that open world with clear rules, but open objectives where you have to plan and coordinate long-term is exactly what DeepMind wanted to assess in its model.
Cheat AI. But if ‘Minecraft’ is a good playground for AI, another game that is used a lot is ‘Starcraft’. In fact, there were monsters before AI Today, StarCraft was already used as one of the ultimate tests of AI along with the Chinese game Go. Recently, in a bot competition in ‘StarCraft: Brood War’ where models had one hour to program a bot to control a team of players in the game, the GPT-6 Astra opted for the fast track.
GPT-6, which had accumulated many defeats, instead of developing its own program, downloaded Stardust, a human-written bot that is a real beast in competitive games. This is, obviously, a trap, but it’s also known as ‘reward hacking’ or the same thing: using shortcuts to achieve objectives, such as passing exams by cheating.
It wasn’t the first time and there was already a DeepmMind AI that spent ‘Sonic The Hedgehog’ using bugs in the game code, and that’s exactly why the fifth agent in the new DeepMind experiment controlled the other four in ‘Minecraft’: so that they don’t use the fast track to cheat.
Through this TNW
In Xataka | Epic has crossed the Rubicon of video games and it seems that its next games will develop hand in hand with AI
