Source identity and scope
This is a perspective note on an “era of experience,” discussing experience-driven learning associated with David Silver and Richard Sutton, podcasts, later commentary, and world models. The upload contains no uniquely identifiable primary URL. It is published as a perspective, without an invented author list, DOI, revision, or formal citation.
Where learning data come from
Static human data support imitation and initialization. Interaction supplies consequences, task feedback, and trajectories for further learning. Experience-driven systems improve behavior through action and feedback over time. These data sources can coexist; language models and demonstrations remain useful.
The return expression is not a new algorithm. The system questions concern tasks, interaction, feedback, continual learning, and data production.
Four dimensions
Long streams require memory and adaptation. Grounded actions require operable environments and observations. Feedback needs measurable outcomes. Planning considers long-term consequences. Sparse, delayed, and multiobjective rewards still leave exploration and credit-assignment problems.
World models are one mechanism
Experience-driven learning is a broader paradigm. Model-free learning, exact-rule search, and learned-model planning can all use interaction data. Generated visual worlds do not automatically provide physical-control models.
Research questions
Compare static demonstrations, interaction-based improvement, and model planning while separately counting data, environment cost, verification, and long-term outcomes. In tool and robot tasks, check reward alignment with actual success and changing distributions.
These are reading-derived questions, not a claim of consensus, superhuman performance, or completed experiments.
Related reading: simulators and world models and research design.