p(doom) @ Tübingen AI Center: Towards Month-Long Task Horizons

​A talk by p(doom), an independent AGI lab.

​Towards Month-Long Task Horizons by Training on Months-Long Human Trajectories:

​Scaling has driven remarkable progress in machine learning, but sustaining that trajectory requires addressing bottlenecks that additional compute does not resolve by itself. These include access to new sources of training data, learning across extremely long trajectories with fixed-sized state, meta-learning continual learning, and leveraging feedback loops.

​This talk presents p(doom)’s broader research agenda through the lens of their current project: turning months-long passive recordings of human computer work into useful training signal. Unlike conventional demonstrations, long-form recordings capture people learning over time: making mistakes, receiving feedback, revising their approach, and improving. They will explain their entire pipeline: their in-house data supply chain and training infrastructure, methods for recovering low-level actions and higher-level structure from video, and how to leverage those to expand the task horizon of frontier models.

​More broadly, the talk will examine what becomes possible when models can learn from billion-token trajectories of human activity, as well as the unresolved technical questions around state, supervision, continual learning, and credit assignment that emerge at this scale.