VyDex tracks important frontier claims with evidence, caveats, sources, and updates.
Not AI news. Not a hype feed.
Google DeepMind researchers report that Dreamer 4 obtained diamonds in 0.7% of 1,000 one-hour Minecraft evaluations after learning from a fixed 2,541-hour gameplay dataset and improving its policy through reinforcement learning entirely inside a learned world model, without online environment interaction during training.
Google DeepMind researchers report that Dreamer 4 obtained diamonds in 0.7% of 1,000 one-hour Minecraft evaluations after learning from a fixed 2,541-hour gameplay dataset and improving its policy through reinforcement learning entirely inside a learned world model, without online environment interaction during training.
During the first season in which the National Hurricane Center incorporated AI-based models into real-time operations, Google DeepMind’s GDMI ensemble mean slightly outperformed the official Atlantic track forecast at 12–72 hours, delivered comparable Atlantic intensity skill, and beat the official forecast and all consensus track aids in the eastern North Pacific at 48–120 hours.
Using a 170-task suite calibrated to expert human completion times, METR introduced the 50% task-completion time horizon and estimated that the frontier for model agents on software and research tasks doubled approximately every seven months from 2019 to early 2025.
What exactly is being claimed?
What does the source material actually support?
What would make the claim easy to overread?