Robotics progress is bottlenecked by data, not capability
Dario Amodei says the scaling patterns that worked for speech and game-playing apply to robotics too — but robotics lags because there isn't enough real-world data to train on, not because the underlying approach fails to generalize.
Details
Core information and root causes
This is a Seed of a Bottleneck Card — the bottleneck was surfaced from a real interview, not yet researched into a full card (current efforts, affected populations, forecast, resources). If you have context, data, or know of existing work on this, we'd love your input.
"For robotics, there's not enough data." — Dario Amodei, on The Dwarkesh Podcast
Amodei describes watching the same capability-scaling pattern repeat across speech, games, and robotics in the years before large language models took off — robotics included, just gated by how little real-world interaction data exists to train on.
Resources
Sources, references, and supporting materials
This card was surfaced by a signal — see the Dwarkesh Patel page for the sourced quote and podcast episode this is drawn from.