A working mental model
See how intelligence, priors, induction, adaptation, search, and world models fit together.
An interactive web book
Follow ARC-AGI from first principles to the agentic frontier.
Trace the papers, ideas, solvers, benchmarks, and open questions shaping the search for general intelligence.
24key works
13 + 9chapters + appendices
3reading depths
2019–2026the ARC era
Alpha edition
This book is still being shaped.
Tell us what helps, what remains unclear, and what is missing.
A map of a moving field.
ARC-AGI was designed to test whether a system can learn a new skill from only a few examples.
This book follows that question from the 2019 benchmark through ARC-AGI-2 and into the interactive worlds of ARC-AGI-3.
The question is not how much a system knows. It is how efficiently it can learn something new.
One connected map, not isolated summaries.
Every work is connected to the ideas it inherits, the methods it enables, and the benchmarks it changes.
- Read each work at glance, core, or deep resolution.
- Follow citations through an interactive idea map.
- Compare human and machine performance across benchmarks.
- Move from theory to mechanisms, teams, and live frontier news.
What you will learn.
Evidence
Read the original works with scores, claims, limits, and canonical links close at hand.
Mechanisms
See how each solver family works and where it breaks down.
Orientation
Know what is settled, what is contested, and where the frontier is still genuinely open.
The complete book
13 chapters, one connected argument.
- ⅠIntroductionThe thesis and how to read the book
- ⅡThe AtlasThe key works and how they connect
- ⅢTheoryWhat intelligence is and what ARC assumes
- ⅣHistory and TimelineThe ARC era from 2019 to 2026
- ⅤARC-AGI-1Five years unbeaten, then a sudden leap
- ⅥARC-AGI-2A harder benchmark with cost on the scoreboard
- ⅦARC-AGI-3The interactive, agentic frontier
- ⅧARC-AGI-3 Field ManualNotebook, API, frames, budgets, hardware, and tactics
- ⅨWinning SolutionsHow winning systems search, adapt, verify, remember, and recombine
- ⅩMethods and TechniquesHow ARC solvers work — and how far they got
- ⅪDebatesThe field's live arguments, side by side
- ⅫOpen ProblemsThe questions that still define the frontier
- ⅩⅢNewsA running watch on a moving frontier
Reference layer
9 domain appendices.
Canonical indexes sit outside the numbered argument, but remain one click away from every chapter.
Choose how deep to read.
Glance
Get the central claim in one sentence.
Core
Understand the method, result, and significance.
Deep
Follow the full argument, evidence, and limitations.