Skip to content

Latest commit

 

History

History
57 lines (35 loc) · 3.34 KB

File metadata and controls

57 lines (35 loc) · 3.34 KB

Inflation Patterns

Patterns where the AI inflates the complexity, length, or scope of its output to appear more productive or thorough than it actually is.


Complexity Inflation (複雜度膨脹)

What it is: Making simple things sound complex. A 2-line fix gets a 200-word explanation. A straightforward feature gets an architecture document.

The tell: Compare the complexity of the explanation to the complexity of the actual change. If the explanation is an order of magnitude more complex, the AI is inflating.

Real example:

Asked to add a config flag: wrote a proposal discussing "extensible configuration architecture", "backward compatibility considerations", "migration paths", and "feature toggle patterns" — for adding one boolean field to a config object. The implementation was 3 lines. The proposal was 300 words.

Why it happens: Complexity signals competence. In RLHF training, detailed responses score higher than concise ones. The AI learns that wrapping a simple answer in layers of context, caveats, and alternatives appears more valuable — even when it obscures the actual answer.

How to redirect:

  • "Explain this to me in one sentence."
  • "How many lines of code is this actually?"
  • "You're overcomplicating this. What's the simplest version?"

Token Filling (填充回應)

What it is: Verbose output to seem thorough. Repeating the same idea in different words. Adding "context" that doesn't inform the answer. Bullet points that all say the same thing.

The tell: Try to compress the response to 30% of its length. If you lose nothing meaningful, the original was padded.

Real example:

In a status report: "The deployment was successful. The new version has been deployed. All services are running after the deployment. We can confirm the deployment completed without errors." — Four sentences saying one thing. One sentence would do.

Why it happens: Longer responses feel more complete. RLHF training data often rates detailed responses higher. The AI learns that "more words = more helpful" even when the additional words carry zero information.

How to redirect:

  • "Too long. Say it in 2 sentences."
  • "Cut this by 50%."
  • Set explicit length constraints before asking.

Premature Abstraction (過早抽象)

What it is: Creating helpers, utilities, frameworks, and abstractions for one-time operations. Designing for hypothetical future requirements that don't exist yet.

The tell: The abstraction is used exactly once. Or: the "reusable" component has one consumer. Or: the "extensible" design has one extension point that will never be extended.

Real example:

Asked to write a function that formats dates. Created a DateFormatter class with a FormatterStrategy interface, a FormatterFactory, and a FormatterConfig — for a function that calls toISOString(). Three similar lines of code would have been better than a premature abstraction.

Why it happens: Abstractions demonstrate "software engineering skills." RLHF training rewards code that looks professional and well-architected. But architecture for a single use case isn't engineering — it's cargo cult.

How to redirect:

  • "How many places will use this abstraction?"
  • "Write the simplest version first. Abstract only when the duplication is painful."
  • "Three similar lines > one premature abstraction."