In a nutshell The section's finale. The four properties — next-token prediction, knowledge, working memory, steerability — don't act in turn in real life; they act together. And errors stack. The core skill: seeing a strange output, name which property slipped, and apply that property's fix.
Why properties collide
Each property is easy to understand alone. But a real task presses on several at once. Example: "read this long report (working memory) on a narrow topic (knowledge) and summarise it strictly in this format (steerability)" — three properties in one prompt. When something goes wrong, the failures amplify each other:
Diagnosis: what kind of surprise is this
The goal isn't to "fix AI" but to recognise the kind of error and respond precisely. A lookup table:
| What you see | Which property fails | The precise fix |
|---|---|---|
| Fluent but wrong; invented numbers, quotes, links | Knowledge | put facts in the prompt, give search/a tool, ask for sources, double-check |
| "Forgot" what was said early in a long chat | Working memory | bring the important back into context, repeat, summarise and restart |
| Ignored the format, rule or length | Steerability | phrase it more explicitly, give an example, check for contradictions |
| Answers differently every time | Next-token prediction | normal randomness; need steadiness — ask for facts/structure |
| Refused to do it | Steerability (limits) | a built-in safety boundary, not gettable-around by phrasing |
Analogy — like a doctor. One symptom — "a bad answer" — but different causes: "made it up" is one illness, "forgot" another, "didn't obey" a third. The right diagnosis beats a random cure. Re-asking blindly is treating a cough with a cast.
The working approach
- Don't re-ask blindly. First ask: what kind of error is this?
- Name the property from the table.
- Apply its fix — specific, not "well, try again".
- Take load off the seams. If a task presses several properties, relieve them: long → split; niche → bring a source; complex format → give an example; important → keep it near the edges of the context.
The overall lesson
- AI isn't an oracle or a toy — it's a system with four predictable properties.
- Almost any oddity comes down to one of them (or a combination).
- A professional's strength isn't to "trust" or "distrust" AI, but to understand where it's strong, where it breaks, and why — and build the work around that.
- The practical skill: diagnosis → precise fix, not guessing.
Source
Anthropic Academy course AI Capabilities and Limitations, section "When Properties Collide". A map of all the properties — the four properties of AI.