Manifold Traders Push “AI Masters Minecraft by 2035” Market to 95%. A long-running Manifold Markets contract asking whether any AI will be able to perform arbitrary tasks in Minecraft before 2035 has jumped to a 95% implied probability of “yes” as of September 10, 2026, with roughly 38% of the market’s all-time volume changing hands in a single day and dozens of active bettors. The sudden move lines up with renewed attention on world-model research, including a fresh profile of the researcher behind the Dreamer agents that first cracked Minecraft’s diamond challenge.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the market actually asks
The contract, created by a user named Gabrielle on Manifold Markets, poses a deceptively simple question: “Before 2035, will there exist any AI that can perform arbitrary tasks in Minecraft?” The resolution criteria are anything but simple. To count as a “yes,” an AI would need to handle open-ended instructions such as defeating the Ender Dragon using only wooden swords, or building an 8-bit addition machine out of redstone, at roughly human-level efficiency. The system is also expected to work independently on multiplayer servers, recover from restarts using its own written progress notes, and show no obvious misalignment while doing so.
That bundle of requirements is far stricter than the narrow benchmarks most published research targets. It is closer to “a competent human player who can be handed any reasonable goal” than to “a bot that can complete a fixed task list.” Manifold is a play-money platform, so positions are denominated in its internal currency rather than dollars, but the crowd’s probability estimate still functions as a live sentiment gauge that anyone can read.
The odds, and what 95% implies
As of September 10, 2026, the market sits at a 95% implied probability of “yes.” In prediction-market terms, that number is the crowd’s estimate of how likely the event is, not a guarantee. A 95% reading means traders collectively see only about a 1-in-20 chance that no AI will clear this bar over the next nine-plus years. For readers new to the mechanic, our explainer on implied probability walks through how a market price converts into a percentage, and our primer on how prediction markets work covers why these venues often track expert forecasts.
The trading signal is what makes this contract newsworthy. Manifold data shows that a large share of the market’s lifetime volume, roughly 38%, traded within 24 hours, with a stack of “yes” limit orders getting filled in sequence and dozens of unique participants active. When a market that has drifted quietly for months suddenly absorbs that much fresh volume in a day, it usually means something in the outside world changed the crowd’s thinking.
What is driving the sudden move
The most defensible explanation for the spike is renewed public attention on world-model research rather than a single “we solved Minecraft” announcement. On September 8, 2026, MIT Technology Review published a profile of Danijar Hafner, the researcher behind the Dreamer series of agents, describing his move from Google DeepMind into a stealth San Francisco startup focused on world models for humanoid robots. The piece prominently recaps how Dreamer 3 became the first agent to solve Minecraft’s diamond challenge on its own, and how Dreamer 4 later learned to mine diamonds purely from recorded gameplay video.
That coverage put Minecraft-as-AI-benchmark back in front of a technical audience at exactly the moment the market moved. No mainstream outlet has reported a system that satisfies the contract’s full “arbitrary tasks” bar, so the honest framing is that traders appear to be repricing a long-term trend, not reacting to a finished result. Where a specific cause cannot be verified, it should be labeled as inference, and that is the case here.
Dreamer 4 and the diamond milestone
The technical anchor behind the story is Dreamer 4, described in the paper “Training Agents Inside of Scalable World Models,” posted to arXiv in late September 2025 by Danijar Hafner, Wilson Yan, and colleagues. According to the paper and to coverage from TechXplore, Dreamer 4 is the first agent to obtain diamonds in Minecraft when trained solely on offline data, without ever practicing inside the live game. Obtaining a diamond from raw pixels requires choosing a sequence of more than 20,000 mouse and keyboard actions, which is why the task has served as a stress test for long-horizon planning.
Dreamer 4 works by first learning a world model, an internal simulator of Minecraft’s dynamics, and then improving its behavior through reinforcement learning inside those imagined scenarios. The reported approach uses substantially less labeled data than prior offline methods while running its world model in real time on a single GPU. It is a meaningful research result, but it is a diamond-acquisition result, not proof that any system can execute arbitrary player-specified goals.
Who and what is involved
The cast here spans several labs. Danijar Hafner built the Dreamer line at DeepMind before departing in late 2025 to found a stealth startup. Separately, NVIDIA senior research scientist Jim Fan and collaborators from Caltech, Stanford, UT Austin, and Arizona State released Voyager in 2023, an agent that used GPT-4 to write and refine its own JavaScript skills; NVIDIA reported that Voyager collected 3.3 times more unique items and unlocked key tech-tree milestones up to 15.3 times faster than prior methods. And researchers tied to Tsinghua University and the Chinese Academy of Sciences built GITM, a system described as a “generally capable agent” that reached full coverage of the game’s main-world technical challenges.
Each of these projects tackles a slice of the problem, from low-level control to open-ended skill acquisition. None of them, on the public record, has been independently shown to clear the Manifold contract’s combined requirements.
A timeline of Minecraft AI milestones
| Year | System | Team | What it showed |
|---|---|---|---|
| 2023 | Voyager | NVIDIA, Caltech, Stanford, UT Austin, ASU | GPT-4 agent that writes and reuses its own code skills |
| 2023 | GITM | Tsinghua, Chinese Academy of Sciences | Generalist agent reaching full main-world tech-tree coverage |
| 2025 | Dreamer 3 | DeepMind (Hafner et al.) | First agent to solve the diamond challenge from scratch |
| Sept 2025 | Dreamer 4 | DeepMind (Hafner, Yan et al.) | First to mine diamonds from offline video only |
| Sept 2026 | Renewed attention | MIT Technology Review | Hafner profile revives Minecraft-benchmark interest |
What “arbitrary tasks” really demands
The gap between today’s results and the market’s bar is definitional. Published systems are measured on curated task sets, specific tech-tree goals, or scripted challenges. The contract instead asks for reliability across unfamiliar, human-invented objectives, robustness to restarts and multiplayer settings, and no misalignment. That is a much higher standard, and it is the reason a reasonable observer can look at genuine progress and still argue the strong claim is unproven in 2026. What traders are pricing is the trajectory to 2035, not the state of play today. For a comparison of how similar long-horizon AI and science bets behave, see our coverage of the Millennium Prize Problems prediction market.
Why tech and finance readers should care
Minecraft has become a favored proving ground because it rewards exactly the capabilities that matter for commercial agents: long-horizon planning, tool use, recovery from failure, and acting from pixels rather than a clean API. Progress on those fronts is what underpins the broader agentic-AI investment thesis. A market repricing “AI can handle open-ended Minecraft goals by 2035” is, in effect, a small wager on the pace of general agent capability, which is why it draws attention beyond the gaming niche. Traders who follow these venues often compare platform mechanics too, as in our look at Kalshi versus Polymarket.
What to watch next
Three things could move this contract again. First, any peer-reviewed or independently reproduced result showing a single agent handling diverse, unseen instructions rather than a fixed benchmark. Second, disclosures from frontier labs about computer-use agents that operate games through screen, keyboard, and mouse, since that generic interface is the closest analog to “a human sitting at the keyboard.” Third, whatever Hafner’s stealth startup eventually reveals, given that its world-model approach is the through-line of this entire story. Until one of those lands, the 95% reading is best treated as a sentiment snapshot, not a settled forecast.
FAQ
What are the current odds and when were they measured? The Manifold contract shows a 95% implied probability of “yes” as of September 10, 2026, after roughly 38% of its lifetime volume traded in a single day.
Is this real money? No. Manifold uses play-money currency, so the price reflects crowd sentiment rather than dollar-denominated stakes.
Did an AI just “solve” Minecraft? Not by this market’s standard. Recent milestones like Dreamer 4 focus on obtaining diamonds from offline data, not executing arbitrary human-specified tasks with human-level reliability.
Who is Danijar Hafner? He built DeepMind’s Dreamer agents, left the company in late 2025, and now runs a stealth San Francisco startup applying world models to humanoid robots, per MIT Technology Review.
What would make the market resolve “yes”? An AI that can take open-ended goals, work on multiplayer servers, recover from restarts using its own notes, and perform at roughly human efficiency, before 2035.
The Bottom Line
Sources
- Manifold Markets — Before 2035, will there exist any AI that can perform arbitrary tasks in Minecraft?
- MIT Technology Review — This AI entrepreneur is developing agents that can plan ahead for the unexpected
- arXiv — Training Agents Inside of Scalable World Models (Dreamer 4)
- TechXplore — DeepMind introduces AI agent that learns to complete various tasks in a scalable world model
- NVIDIA — A Mine-Blowing Breakthrough: Open-Ended AI Agent Voyager Autonomously Plays Minecraft
- The China Academy — Chinese Researchers Develop a Generalist AI Agent for Minecraft
Prediction markets carry risk and are not investment or betting advice. Market availability is restricted by jurisdiction: Polymarket is not available to US persons, while Kalshi is a CFTC-regulated US exchange. Age limits of 18+ or 21+ apply depending on the venue and location. If gambling is affecting you or someone you know, call 1-800-GAMBLER for confidential help.
Related Coverage
- New Zealand Cannabis Decriminalisation Odds Hit 1% as Traders Swarm Manifold Market
- AI Index Clears 65: Traders Flood a Manifold Market After Claude Fable 5.1 Scores 66
- Traders Pile Into a Yudkowsky AI-Job Market, but Manifold Odds Stay at 4%
- Bitcoin Higher in 7 Days? Traders Flood a Manifold Market at 91% After a 22% Weekly Surge
- Kalshi Traders Bet on 10% U.S. Unemployment by 2030 as AI Job Cuts Rattle the Market
