2026-10-11 16:37 UTC

world-models

band: hotmomentum: stable score: 0.715
temperature history

Episodes (15)

Anthropic will complete an acquisition of world-model startup Decart following reported talks valuing the transaction at roughly $6 billion.
expirednovelscott: medium
Independent reproduction will determine whether the paper's compact Genie-style world model sustains playable 720p generation near 16 FPS within 19GB of VRAM on a single RTX 5090.
expiredknownscott: medium
Artifact release and independent reproduction will determine whether the reported 1.57B-parameter Dreamer 4 world model was trained from scratch for under $150 and can sustain controllable prediction for roughly 144 frames.
expiredknownscott: low
Independent technical review and evaluation will determine whether Accelerated Understanding’s announced neural-operator model is a usable non-transformer system for physics simulation and scientific discovery.
expiredknownscott: low
Code World Model’s authors claim their released world-modeling approach can provide coding agents with capabilities beyond conventional code generation, potentially enabling richer planning and environment interaction.
expiredconvergesscott: medium
The LingBot-Video team claims its released 13B sparse-MoE model can generate physically plausible action-conditioned robot rollouts, potentially providing an open foundation for prediction and planning in robotics.
expiredconvergesscott: low
General Intuition, Kyutai, and Epic Games claim MIRA can interactively predict four-player Rocket League at 20 frames per second on one B200, providing an open testbed for multiplayer world-model prediction and control.
expiredknownscott: medium
World Labs claims Atlas provides a general spatial world model for interactive three-dimensional understanding and environment reasoning, potentially expanding the foundations available for spatially grounded agents.
expiredconvergesscott: medium
H3-World’s authors claim their language-conditioned control method can compose temporally grounded character and camera actions from limited gameplay data, potentially providing a more controllable foundation for video generation and embodied-agent planning.
expiredconvergesscott: medium
Runway claims its GWM Worlds 2 research preview can generate controllable, continuous 720p audiovisual environments in real time, potentially providing a practical world-model substrate for interactive simulations and embodied agents.
seedconvergesscott: medium
Kaarelson claims the released LingBot-World 2.0 1.3B implementation runs at 16 FPS on one RTX 5090, potentially enabling interactive world-model experimentation on a single consumer GPU.
corroboratednovelscott: low
Ruiyang Wang and coauthors claim GAVEL's explicit graph world model raises Qwen3-8B task success on BEHAVIOR-1K from 41.2% to 91.8% for single tasks and 19.9% to 92.6% for multi-task instructions, potentially making compact-model embodied planning reliable through external verification and repair.
seedconvergesscott: high
A circulating demo video claims GPT-6 Astra converts ordinary room video into an interactive, robot-trainable 3D world that AI video models can render from novel viewpoints — identifying the demo's producer and method would establish video-to-world construction as a demonstrated frontier capability, while debunking would mark another inflated capability echo.
resolvedconvergesscott: high
World Labs and AMD announce that the spatial-intelligence lab is joining AMD — Fei-Fei Li becoming EVP and Chief Scientist under CEO Lisa Su, with co-founders Justin Johnson and Ben Mildenhall continuing to lead the team inside an 'end-to-end open AI ecosystem' spanning hardware, software, and open models — with closing expected by end of 2026 pending regulatory approval; whether the deal closes as an integrated AMD frontier-research organization (versus regulatory block or quiet dissolution) establishes chip-vendor absorption as a new consolidation path for frontier world-model labs.
acceleratingnovelscott: low
Stanford's OpenWAM team (Li Fei-Fei, Jiajun Wu, and Ehsan Adeli's labs) claims its released open framework for composable world-action models — a shared Wan2.2 video foundation with 5B video and 2B action experts in one Mixture-of-Transformers, swappable predict-then-act/act-then-predict/joint/decoupled interaction programs, local-context IDM/FDM components, and a 32,000-segment counterfactual LIBERO-Long-CF dataset — becomes the common foundation that world-action-model research standardizes on; outside groups building policies and benchmarking against it confirm adoption, while quiet fade closes it as a lab-internal release.
seednovelscott: high

Trajectory notes