DeepMind’s New AI Beats OpenAI With 100x Less Data
•November 18, 2025
0
Why It Matters
By achieving high‑level performance with 100× less data, DeepMind’s imagination‑based training could lower the barrier to building sophisticated AI systems, enabling faster, cheaper development of robots and other real‑world applications.
Summary
DeepMind unveiled a new AI system that learns to play Minecraft with a fraction of the data previously required, outperforming OpenAI’s Video Pre‑Training (VPT) approach despite using roughly 1 % of the video footage. The breakthrough hinges on a three‑phase “imagination” pipeline: first, the model watches a handful of human gameplay clips to construct an internal world model of Minecraft’s physics; second, it assigns value to actions through rapid feedback within this simulated environment; third, it practices millions of short “dreams,” iteratively refining its policy without ever touching the real game.
The technique delivers striking results. Where OpenAI’s VPT needed 250,000 hours of annotated video to achieve modest success, DeepMind’s agent reaches a 90 % success rate at tasks such as obtaining a stone pickaxe and even manages to mine diamonds—something earlier behavioral‑cloning (BC) and Vision‑Language‑Action (VLA) methods could not reliably do. By chaining together tens of thousands of imagined actions, the AI can execute over 20,000 steps in a single episode, demonstrating a level of strategic planning previously thought to require massive datasets.
The video highlights vivid examples: the agent repeatedly chops trees, mines ore, and crafts tools entirely within its mental simulation, only later reproducing the behavior in the actual game. Researchers note a key limitation: the imagined rollouts are short‑term, so the model sometimes fails to capture long‑range cause‑and‑effect, leading to occasional “tree‑pop‑back” errors. Nonetheless, the ability to learn complex, sequential tasks from minimal data marks a paradigm shift.
Industry observers see broader implications beyond gaming. A data‑efficient world‑model approach could accelerate robot training, allowing machines to rehearse real‑world tasks in safe, simulated environments before deployment, dramatically cutting labeling costs and compute budgets. As AI moves toward more generalized, imagination‑driven learning, DeepMind’s results suggest a path to scalable, low‑resource intelligence.
Original Description
❤️ Check out Lambda here and sign up for their GPU Cloud: https://lambda.ai/papers
Guide:
Rent one of their GPUs with over 16GB of VRAM
Open a terminal
Just get Ollama following the command from here - https://ollama.com/download/linux
Then run ollama run gpt-oss:120b - https://ollama.com/library/gpt-oss:120b
📝 The paper is available here:
Source:
📝 My paper on simulations that look almost like reality is available for free here:
Or this is the orig. Nature Physics link with clickable citations:
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Benji Rabhan, B Shang, Christian Ahlin, Gordon Child, Juan Benet, Michael Tedder, Owen Skarpness, Richard Sundvall, Steef, Taras Bobrovytsky, Tybie Fitzhugh, Ueli Gallizzi
If you wish to appear here or pick up other perks, click here: https://www.patreon.com/TwoMinutePapers
My research: https://cg.tuwien.ac.at/~zsolnai/
X/Twitter: https://twitter.com/twominutepapers
Thumbnail design: Felícia Zsolnai-Fehér - http://felicia
#minecraft
0
Comments
Want to join the conversation?
Loading comments...