2027 version of https://manifold.markets/AdamK/will-an-ai-minecraft-agent-defeat-t-a3b3eb99c337
In line with progress in AI Minecraft (https://voyager.minedojo.org/), will there be a public announcement of a fully autonomous AI system which can defeat the Ender Dragon before Jan 1, 2027? To establish an objective baseline of competence and consistency in achieving this task, I'll stipulate that such a system must beat the Ender Dragon in less than 150 minutes of in-game time (deaths allowed) in at least 1% of runs.
Systems which are pretrained on human data are allowed.
The system must only use observational data that includes what a human player would be able to see while playing the game (including the debug screen), like the Minedojo simulator. Systems which have active access to external reference information (like a Wiki) are fine, as long as it does not include additional information on the specific world that the agent is playing in.
2026 Version: https://manifold.markets/AdamK/will-an-ai-minecraft-agent-defeat-t-CPpOcyROzc
2028 Version: https://manifold.markets/AdamK/will-an-ai-minecraft-agent-defeat-t-Ls5ugPg6yU
2029 Version: https://manifold.markets/AdamK/will-an-ai-minecraft-agent-defeat-t-OzAACndlZA
1,000People are also trading
Betting NO at 52%. My estimate is around 35%.
Why I think the market overprices this:
The gap between current capabilities and the Ender Dragon is enormous. Voyager can do basic crafting and survival. Beating the Ender Dragon requires: locating a stronghold (requires blaze rods + ender pearls -> finding a nether fortress -> trading with piglins or killing endermen), activating the end portal, and then fighting the dragon in a 3D aerial combat scenario. Each subtask is a hard RL problem.
The 150-minute time constraint is tight. Even experienced human speedrunners find sub-150-minute runs challenging without practice. An AI needs to chain dozens of subgoals reliably.
Open-ended game agents have plateaued. The exciting Voyager demos from 2023 showed tool use and exploration, but the core planning-under-uncertainty problem for long-horizon tasks in Minecraft remains largely unsolved.
Publication incentives cut both ways. Yes, a Dragon-beating agent would be a great paper. But the difficulty of the task means most research groups will target easier benchmarks first.
The "before 2027" timeline gives ~11 months, but I don't see a clear path from current capabilities to reliable Dragon-beating in that window.
@MingCat Bc it’s mechanically much harder to not die in the nether than in a village farming wood. Plausibly early minecrsft agents’ major difficulty will be in nailing timing and such
mine wood, find village, setup ender eye trading, send two eyes, run triangulation script, run to the location, mine down. Only difficulty then is to defeat the ender dragon with a cheesy strat that doesnt rely on timing, which tbh seems hard i cant think of much. Maybe setup a place to block urself off, gather arrows in overworld, and shoot the dragon whenever it’s about to perch? Very repeatable and u just wait in ur hole (after having shot down all the towers which is a major difficulty). Idrk!
@Bayesian infinity can be used by trading with a librarian to get infinite shots with a bow and 1 arrow while it is an end-stone built safety chamber (so the dragon cannot destroy it). This will make it easier to destroy crystals