Can AI learn new domains from small amounts of experience as effectively as capable humans?
Humans learn new skills from orders of magnitude less data. Chollet argues sample-efficient skill acquisition is what intelligence is — making learning efficiency a definition, not just a milestone.
View on the map → · Open in Browse →
Sutskever made sample efficiency the headline problem — models 'generalize dramatically worse than people' — while ARC's new interactive benchmark measured the gap at its starkest: humans succeed at learning its games 100% of the time, frontier systems below 1%. The rebuttal worth reading argues most of the gap dissolves under fair accounting.
These models somehow just generalize dramatically worse than people. It's a very fundamental thing.
Sutskever's first extended interview since founding SSI names poor generalization and sample efficiency as THE central unsolved problem and declares the scaling era over — the most-cited insider statement on learning efficiency in this window.
evaluating fluid adaptive efficiency on novel tasks, while avoiding language and external knowledge
Chollet's program made interactive: agents must learn game environments from scratch by exploration. As of March 2026 humans succeed 100% of the time; frontier systems score below 1% — the starkest current measurement of the human–AI gap in sample-efficient skill acquisition.
Most of the apparent inefficiency dissolves on closer inspection: apples-to-oranges comparisons between pretrained humans and from-scratch networks... and priors installed by evolution.
The strongest recent rebuttal to the sample-efficiency critique: the human–AI efficiency gap is mostly a measurement artifact of unfair comparisons, and what remains implies more compute per example rather than a missing breakthrough.
the top score reaching 24% on the ARC-AGI-2 private evaluation set.
The definitive annual measurement of efficiency-constrained novel-task learning: compute-limited systems reach only 24% on ARC-AGI-2 tasks humans solve routinely, quantifying how much recent progress is bought with brute-force test-time compute.