Continual Intelligence
The continual-learning research arc — twelve essays, from the benchmark gap to open-ended AI, every claim traced to its source. The whole series as one volume.
Continual learning is the question hiding under every agent that's supposed to get better with use: why do deployed systems perform no better on day 100 than day 1, and what would fix it? This book reads that literature so you don't have to — twelve essays tracing the research arc from the benchmark gap and the plasticity crisis through world models, RL-for-reasoning, and stable deep RL at scale, to the case for open-ended AI and RL as educator.
Every claim was traced to its source paper before publication: numbers, attributions, and mechanism descriptions audited against the originals, and anything that couldn't be verified marked rather than smoothed over. It is a living book: versioned like software, and every future edition is included with your copy.
Not sure yet? Read a free sample (PDF) — or start free with the 13 checklists, pay what you want ($0 works).
What's inside
Twelve chapters — 350 pages, per-essay anchor papers and full references, arXiv links included.
- The Benchmark Gap in Continual RL: From Continual World to SPIRAL
- The Plasticity Crisis in Continual Deep Learning
- The Big World Hypothesis: Why Continual Learning Is Inevitable
- GVFs as Proto-World-Models: The Alberta Plan Vindicated?
- The Forgetting Transformer: When Architecture Solves Plasticity
- Does RL Teach LLMs to Reason, or Just Refine Them?
- Shape of Thought: Why Reasoning Format Matters More Than Correctness
- Stable Deep RL at Scale: Gradients, KL, and the Shape of Learning
- Reasoning at Scale: What DeepSeek-R1, ProRL, and Prolonged RL Reveal
- Darwin-Gödel to ShinkaEvolve: The Case for Open-Ended AI
- Thinking Without Tokens: CTM and Inference-Time Compute Beyond CoT
- RL as Educator: Training Teachers, Not Just Students
A living book
The book is versioned like software — this edition is v2026.07. Buyers receive every future version at no cost, delivered through the store's update mechanism and announced when it lands. The essays themselves are free on this site and will stay free — the book buys permanence, sequence, and a single artifact you can hand to a colleague with "read this first."
Engineers and researchers who want the continual-learning literature organized into an argument, not a link dump. It is not a beginner's ML tutorial and not a how-to — the how-to is the companion volume. 14-day no-questions refund either way.
The companion volume
Agentic Engineering — Building agents that ship is the deployment layer on top of this science: harness, evals, memory, ops, with a deployable artifact closing every chapter. It lives at agenticfrontier.dev/book and is included in the complete bundle above. The two volumes cross-reference each other; the bundle is the intended way to read them.