Part Ⅰ · Start here
Understand ARC-AGI.
ARC-AGI asks whether a system can learn a genuinely new skill from only a few examples. This book follows that question across three benchmark generations and the research around them; this chapter shows you how to navigate it.
The one idea
Intelligence is efficiency, not knowledge.
In his 2019 paper On the Measure of Intelligence, François Chollet separates skill from intelligence. Memorizing a million chess positions can produce skill; learning a new game from a handful of examples demonstrates learning efficiency.
ARC tasks are therefore novel, few-shot, and designed around Core Knowledge. They remove language and trivia so memorization and scale alone are less useful.
Every part of this book is a facet of that one question — what the axis is (Theory), how the machines try to move along it (Techniques, the benchmark deep dives), how far they have got (Methods, History), what is still argued (Debates), and what are next directions (Open Problems).
Guide
How this book is built
Follow the numbered sequence or jump between connected ideas.The idea map draws the citation web at once. Start there if you prefer to move by connection rather than by chapter. Three features support both approaches.
Read in any order
The parts build an argument, but each stands alone. Jump to a benchmark, a technique, or a debate; the cross-links will pull you to what you need.
Follow the threads
Papers link to the works they build on and the parts that use them. The idea map in Part Ⅱ shows the whole web at once.
Still being shaped
Summaries were drafted for this guide and should be read alongside — not in place of — the originals. Reader reviews steer each revision.
The through-line
13 chapters, one argument.
The sequence moves from foundations through the three benchmark generations, then to solutions, debates, open problems, and the frontier watch. The numbering gives newcomers a path, but each chapter stands alone.
Read in order for the full argument, or use the table of contents to jump directly to a benchmark, method, debate, or reference appendix.
→ Open the full table of contentsChoose your depth
Read every work at three resolutions.
The catalog in Part Ⅱ lets you read each work at a glance, at its core, or in full — so a survey and a deep study use the same pages.
The central claim
One sentence per work — enough to place it in the argument and move on.
Method & result
The mechanism, the score, and why it mattered — a working understanding without the full paper.
The full argument
Evidence, limitations, and the citations that connect it to the rest of the field.
Where to start, by reader
- New to ARC? Read Part Ⅲ · Theory, then Part Ⅳ · History for the story so far.
- Here to compete? Go straight to Part Ⅶ · ARC-AGI-3 and its getting-started section.
- Want every winner? Part Ⅸ · Winning Solutions separates prize awards, paper awards, ARC-AGI-3 milestones, and non-prize benchmark records.
- Want the mechanisms? Part Ⅹ · Methods & Techniques opens up how each solver family works — and who won with it.
- Following the frontier? Part Ⅺ · Debates and Part Ⅻ · Open Problems are where the field is still moving.
Ready? The catalog of 24 works — the heart of the book — is one click away.
→ Enter Part Ⅱ · The Atlas