What does ARC mean by intelligence?

Richard Sutton's "bitter lesson"; the 2019–2024 pretraining mainstream

General methods that scale with compute and data eventually beat hand-crafted knowledge. Enough scale, and the abstractions emerge on their own.

On ARCARC was designed to resist exactly this — no memorized web data helps — and its five-year durability through the scaling era is the theory's strongest evidence that bought skill is not general intelligence.

Argues againstSkill-acquisition efficiency. Scale buys skill on a task without showing any ability to learn a new one.

Jack Cole & Mohamed Osman (MindsAI); the deep-learning-for-ARC camp

ARC is substantially perceptual, not only symbolic. Keep the optimizer active at inference through test-time fine-tuning rather than relying on a frozen model.

On ARCThis lineage reached 58% and won the 2023 ARCathon by treating both the network and the optimizer as part of inference — learned, on-the-fly abstraction, not symbolic search.

Argues against"Program synthesis is the only viable path." A voted neural answer, though, has no explanation or correctness guarantee.

Akyürek et al. (MIT); the ARChitects; the MindsAI lineage

Don't ship a frozen model. Adapt to the task in front of you at inference — rewrite the weights, or search a latent — then throw the update away. Adaptation is a distinct axis of test-time compute.

On ARCIt was the story of 2024: no frozen, no-adaptation method cleared ~11% on ARC-AGI-1, while per-task fine-tuning reached the mid-50s. Yet there is still no satisfying theory of why it beats in-context learning.

Argues againstThe frozen-weights orthodoxy — but its ARC-AGI-1 gains collapse to single digits on ARC-AGI-2, and per-task cost may not survive an agent loop.

Marcus Hutter & Shane Legg (AIXI); demonstrated on ARC by CompressARC

Intelligence is compression. Prefer the shortest description that regenerates the data: by Occam's razor, simpler explanations are more likely to generalize.

On ARCCompressARC solves ~20% of ARC-AGI-1 with a 76K-param network, no pretraining, and no search — purely by compressing the single puzzle. Lower-bit solutions empirically correlate with correct ones.

Argues againstInduction and transduction may both be crude approximations of this one objective — but it caps at ~20–34%, so is it capacity, optimization, or prior mismatch?

Kevin Ellis & Josh Tenenbaum (DreamCoder); Chollet's endorsed direction

Solve a task by synthesizing an explicit, inspectable program over Core-Knowledge primitives — a rule you can read and verify — rather than fitting an opaque function. Learn the library of primitives itself, as DreamCoder does.

On ARCInfer a compact, verifiable transformation rule from ~3–5 demonstrations and compose it from reusable abstractions.

Argues againstFitted-function statistics — but every hand-built DSL supplies substantial prior knowledge, so the library should be discovered rather than fixed by hand.

The interactive-agent direction; this Atlas's companion project

A world model should be an induced program accepted only when it exactly replays every observed transition — discovering both the rules and the goal from interaction, and verifying against reality rather than picking from a library that already encodes the answer.

On ARCThe operational meaning of efficient generalization: induce a compact, correct model of a task's dynamics from few observations. ARC-AGI-3's determinism makes such exact, verifiable models tractable.

Argues againstSearch-heavy recombination and opaque neural world models — a program that replays the past may still be wrong on unseen states.