Minecraft and artificial intelligence

Oasis by Decart: a game generated by AI

In late 2024, the startup Decart unveiled Oasis, an experimental game engine in which every frame is generated in real time by a transformer model. Oasis is not a Minecraft mod, although it looks like Mojang's blocky world. It's a separate program with no traditional game code at all: no physics engine, no 3D renderer, no collision system. Everything the player sees is the output of a generative model that takes your input (keyboard + mouse) and predicts the next frame.

Technically, Oasis is built on a transformer architecture trained on a large amount of Minecraft gameplay footage. The model learned to simulate block physics, the day and night cycle, mob behavior, the inventory interface and basic interactions purely from visual patterns, with no hard-coded game rules.

The result is impressive as a demonstration of what AI can do, but far from playable:

• Inconsistent physics: Blocks sometimes vanish or appear for no reason. Gravity works selectively. Water flows in unpredictable directions.
• No persistence: The world doesn't remember changes. If you turn around 180 degrees and turn back, the landscape is different from a moment ago.
• Visual hallucinations: The inventory interface sometimes shows items you don't have. The hotbar changes randomly between frames.
• Low resolution: The public version runs at 360p at around 20 FPS. Running it locally takes a powerful GPU.

Despite its limitations, Oasis set a new direction for the games industry. At that point Decart had raised $21 million in funding. Other companies, including Google DeepMind (the Genie 2 project), are working on similar systems. Minecraft has become a de facto benchmark for generative game engines, because its blocky look is easier to imitate than photorealism.

Research projects: Minecraft as an AI lab

For years, Minecraft has served as a testbed for artificial intelligence. The open world, simple physics, clear rules and huge player base make it an excellent proving ground. Here are the most important projects:

VOYAGER (NVIDIA, 2023)

VOYAGER is an autonomous AI agent created by researchers from Stanford University and NVIDIA. It uses GPT-4 as a "brain" for planning and decision-making in Minecraft. Unlike earlier bots, VOYAGER can discover new skills on its own, without human supervision: it learns to craft new items, explore biomes, fight mobs and build shelters. The agent builds up a skill library that it returns to in future sessions.

Key results: VOYAGER collected 3.3 times more unique items than previous models, worked its way through the tech tree to diamond tools on its own, and covered 2.3 times more ground.

MineDojo (NVIDIA, 2022)

MineDojo is an open-source research environment for training AI agents in Minecraft. It provides thousands of task descriptions (e.g. "build a house out of wood", "kill a creeper") and automatically judges whether the agent achieved them. MineDojo also includes a database of 730 thousand YouTube videos and 7 thousand wiki pages that agents can use to learn. The project received the NeurIPS 2022 award for best dataset.

MineRL (Carnegie Mellon, since 2019)

The MineRL Competition is a contest in which research teams build AI agents that play Minecraft. The main task: getting a diamond from scratch, with no prior knowledge of the game. The competition has run since 2019.

Project Malmo (Microsoft Research)

The oldest of the big AI-Minecraft projects. Launched by Microsoft Research in 2016, Project Malmo is a platform for training reinforcement learning agents in Minecraft. Even though Microsoft owns Mojang, Project Malmo has not led to any public AI features in the game. It remains a research tool.

AI experiments in Minecraft in 2026

In 2026 Minecraft once again served as a proving ground for new AI models and unusual simulations. Here are the biggest experiments (as of October 2026).

GPT-6 Astra and 141 hours of play

In September 2026, Vals AI, a company that independently tests AI models, described a 141-hour playthrough by OpenAI's GPT-6 Astra. According to Vals AI, the model got further in the game than any earlier AI system: it built a semi-automatic blaze farm, collected 6 blaze rods and picked up 3 ender pearls in the Nether. Then a creeper explosion destroyed the chest with its most valuable items and the bed that set its respawn point. For the next several hours the model mostly farmed potatoes, while viewers of the Twitch stream urged it to keep playing. More about the GPT-6 Astra test.

Ender Dragon beaten in 8 minutes and 43 seconds

On September 20, 2026, a video appeared on X in which a pair of models, GPT-6 Astra and Jev from TypeSafe AI, defeat the Ender Dragon in Java Edition in 8 minutes and 43 seconds, and the whole run cost less than a dollar. Astra set the goals, Jev picked actions based on data about the game state, and the Mineflayer library controlled the character. The experiment was prepared by Ronak Malde. It wasn't a speedrun on a random world, though: the run was on version 1.16.5 on Peaceful difficulty, on a chosen seed with the End portal already active, along a route that had been checked earlier in a test world. The dragon died from six bed explosions on the first landing. The human Any% RSG record is 6:39.813 (see the speedrun records). Details of the run.

A fruit fly with a simulated brain

The NeuroCraft Fly project, which Evan Sinclair Smith showed on September 4, 2026, runs a network of 166,700 simulated neurons in Minecraft. It was recreated from the MaleCNS v1.0 connectome, a map of the nervous system of a male fruit fly prepared by the HHMI Janelia team with the participation of Google Research. Game events such as light, food and attacks go into the network as input signals, and the activity of selected neurons translates into the movement of the virtual fly. The project was built with the help of GPT-6 Astra. The author stresses that it is not a digital copy of a conscious fly: the connectome provides only the wiring, and how stimuli from the game reach the network and how neuron activity turns into movement was decided by the simulation's creators. More about NeuroCraft Fly.

Simulation: Minecraft beats itself

This experiment doesn't use AI, only statistics. In the video "Running Minecraft Until It Beats Itself" from September 6, 2026, the RedLogic channel used a Monte Carlo simulation to check whether mobs alone can defeat the Ender Dragon. Mobs can't place eyes of ender, so the simulation assumes a world from a chosen seed in which the End portal under an Ice Spikes biome is already active. An Enderman puts a carved pumpkin down on snow blocks, which creates a Snow Golem, creepers clear a path to the portal, and golems in the End accidentally hit End Crystals with snowballs. According to the simulation, the dragon dies after 1.975 billion years. No world ran that long, though: instead of counting individual ticks, the simulation drew random waiting times for events. More about the simulation.

Notch and AI programming

Markus "Notch" Persson, the creator of Minecraft, called supporters of writing code with the help of AI "incompetent or evil" in January 2026, and on July 14 he wrote simply "Reject AI". A week later he admitted he might try vibe coding because he has trouble hiring programmers, and on August 10 he shared a transcript of his own conversation with an AI with the comment "Maybe it's time to eat some crow", meaning he admitted he might have been wrong. In 2024 Persson announced his studio Bitshift Entertainment and revealed the roguelike Levers and Chests as his next game. More about Notch's change of mind.

AI mods for Minecraft

The modding community is starting to integrate large language models (LLMs) with Minecraft. Below are the most interesting projects that work on current versions of the game.

Inworld AI NPCs

Inworld.ai is a platform for creating AI-driven NPCs. The available Minecraft integration lets you create non-player characters you talk to in chat in natural language. The NPCs have personalities, memories of earlier conversations and react to context (time of day, location, the player's inventory). On RPG servers this creates a whole new level of immersion: villagers answer questions, guards tell legends and traders haggle over prices.

GPT-powered Custom NPCs

The Custom NPCs mod (a long-popular Forge modification) has been extended by the community with OpenAI API integration. Each NPC can have a system prompt that defines its personality, knowledge and behavior. NPC conversations are generated by GPT-4 or GPT-4o in real time. The API costs are paid by the server operator; with active multiplayer they can reach several dozen dollars a month.

Baritone (an advanced pathfinding bot)

Baritone is an autopilot bot for Minecraft that digs, builds, explores and gathers resources on its own. Although it doesn't use generative AI (it relies on A* algorithms and heuristics), it's one of the most advanced automation tools in the community. Baritone can dig up diamonds by itself, build a Nether tunnel and return to base. It's often used in AFK mode on anarchy servers (2b2t).

AI command generators

Web tools and Discord bots that generate Minecraft commands from descriptions in natural language. The player writes "give me a diamond sword with Sharpness V and Fire Aspect II", and the generator returns a complete /give command with the right enchantment data. Chatbots based on Claude, GPT-4 and Gemini handle this surprisingly well for simple requests.

AI skin generators

One of the most practical uses of AI in the Minecraft ecosystem is skin generators. Character skins in Minecraft are simple 64x64 pixel textures, which makes them an ideal target for generative models.

Tools available in 2026

• MinecraftSkins.net AI Generator: A built-in generator on one of the biggest skin databases. You type a description (e.g. "a knight in golden armor with a red cape") and get 4 variants to choose from. Free, limited to a few generations a day.
• NameMC AI: An AI generator section on the popular NameMC site. It uses a model trained on millions of existing skins. Good results for humanoid characters, worse with abstract themes.
• SkinAI.io: A dedicated skin generation tool. It offers a real-time 3D preview, editing of individual layers (overlay) and PNG export ready to load in the launcher.
• The Skindex Creator: A skin editor with optional AI support as a starting point. You generate a base, then polish the details by hand. Combining AI with traditional drawing gives the best results.

The quality of generated skins improved significantly between 2024 and 2026. Early models produced blurry, asymmetrical textures. Current versions generate skins comparable to the work of intermediate artists: correct proportions, symmetrical details and consistent color palettes.

AI and building in Minecraft

Building is the area where AI can help players the most. Designing complex structures out of blocks is a task where generative models have a natural advantage: they can iterate quickly and suggest variants.

PlotCraft and building generators

PlotCraft (2024) is a research tool that generates three-dimensional Minecraft buildings from text descriptions. The model is trained on thousands of WorldEdit and Litematica schematics. As input it takes a description (e.g. "a medieval stone watchtower, 15 blocks tall") and exports a .schem file to import in the game. The results are promising for simple structures, but complex builds (castles, bridges) still need human correction.

LLMs as a building assistant

Many players use ChatGPT, Claude and Gemini as building consultants. A typical workflow: the player describes what they want to build, and the language model suggests a room layout, a materials list, proportions and an architectural style. It's not automatic building (the player still places blocks by hand), but AI sharply shortens the planning phase. For beginner builders who don't know where to start, an LLM is an excellent source of inspiration.

DALL-E, Midjourney and visual concepts

Image generators (DALL-E 3, Midjourney v6, Stable Diffusion) are great at producing building concepts in a Minecraft look. The prompt "isometric minecraft castle, dark oak and stone bricks, moat, medieval style" produces references that builders recreate block by block. The method is popular in the Build the Earth community and on creative servers.

The future of AI in Minecraft

AI integration with Minecraft is speeding up. Here are the directions most likely to change gameplay over the next 2-3 years:

• Smart NPCs in vanilla Minecraft: Microsoft owns Mojang and has invested in OpenAI. There's speculation that a future update will bring villagers capable of simple conversations, or more complex behavior than today's trading routines. Mojang hasn't confirmed these plans.
• Generating worlds from descriptions: Instead of a random seed, a text description ("a medieval kingdom with castles and villages"). The technology exists in research form, and commercialization is a matter of 2-4 years.
• AI anti-cheat: Multiplayer servers are testing ML (machine learning) cheat detection systems that analyze player movement patterns instead of checking client modifications. More effective than traditional anti-cheats, but controversial because of privacy.
• Automatic mod creation: Coding models (Claude, GPT-4o, Codex) can generate simple Minecraft mods from descriptions in natural language. In 2026 they need a lot of human correction, but the quality improves with every model generation.
• Personalized textures and music: Resource packs generated per player based on aesthetic preferences. Prototypes exist: a model turns "cyberpunk Minecraft" into a consistent 16x16 texture pack.

Frequently asked questions

Related pages

• Mod catalog - the best mods for Minecraft
• Minecraft mods - a guide to modifications
• Minecraft skins - how to change your skin, best sites
• Tools and generators - calculators, command generators
• Minecraft Java Edition - the version with the most mod support