AI Capabilities and Limitations
An Anthropic Academy course on how large language models actually work — next-token prediction, knowledge, working memory, steerability and context limits.
Last verified August 21, 2026 · 1 source
Three and a half hours, 13 lessons, one quiz — covering next-token prediction, knowledge, working memory, steerability and context limits.
Why this is the useful half
AI Fluency teaches a working method. This course explains the mechanism underneath it, which is what lets you predict failure rather than being surprised by it.
Most confusing model behaviour stops being confusing once you know:
- Next-token prediction — why fluent output and correct output are different properties
- Knowledge — why a model has a cutoff and what that does to recency
- Working memory — why long sessions degrade and why context management matters
- Steerability — why instructions influence rather than command
- Context limits — what actually happens as the window fills
Where it pays off in practice
The context-limits material maps directly onto real decisions in
Claude Code: why CLAUDE.md should be short, why
subagents exist, and why a session that has been running for hours behaves
differently from a fresh one.
Sources
- 01Anthropic AcademyOfficialacademy.claude.comListed as a 3.5-hour course with 13 lessons and 1 quiz.
Page changelog
- VerifiedFormat, length and topic list confirmed on the Academy front page.