Annotating Unlabeled Screencasts with an IDM
We release an inverse dynamics model for recovering raw input events from screen recordings, alongside 600 hours of action-labeled screencasts.
Read research →Capturing Long-Horizon Human Work
We release the biggest long-horizon screencast dataset in the world, and open-source infrastructure for action-annotated crowd-sourcing.
Read research →Capturing Long-Horizon Human–Agent Cowork
We release over six months of SWE traces across 25 people, and a VS Code / Cursor extension for capturing human software engineering in the age of coding agents.
Read research →600 Hours of AGI Research
We at p(doom) started capturing everything we do, and openly release over 600 hours of AGI research.
Read research →Training Real-World Agents in World Models
We release the biggest Minecraft dataset in the world, and a production-ready codebase for world modeling from unlabeled videos.
Read research →Capturing Long-Horizon Human Software Engineering
We open-source infrastructure that allows anyone to participate in crowd-sourcing a dataset of software engineering traces.
Read research →Oct 30, 2025
Jasmine: A Simple, Performant and Scalable JAX-based World Modeling Codebase[arXiv]
World Modeling
Apr 29, 2025
PPO Is An Off-Policy Algorithm[Blog]
Reinforcement Learning
Mar 26, 2025
Performance-degradation Free Value Assertions in JAX[Blog]
Infrastructure
Feb 12, 2025
PPO Is Secretly Using Monte Carlo Advantage Estimation In LLM Post-Training[Blog]
Reinforcement Learning
Sep 26, 2024
Neural Networks Do Not Generalize Out-of-Distribution[Blog]
Roadmap
Jun 8, 2024
Going Beyond the Causal Mask in Language Modeling[Blog]
Language Modeling
Dec 7, 2023
ACT: Adaptive Compute Transformer[Blog]
Language Modeling