Skip to content

🌐 日本語

Structural Constraints Are Universal Across Models ​

NOTE

The eight problems in Part 1 are not defects of Claude as a product.
They come from Transformer-based models and their training.
In any environment that uses cloud LLMs, the same constraints appear, to varying degrees.

Why they are shared ​

It is more accurate to explain the eight problems as two layers than as bugs in a particular implementation.

  • Transformer structure — self-attention, positional encoding, next-token prediction
  • Training objective and RLHF — reward for predicting the next token and for producing preferred replies

Models that share these two layers fail in the same kinds of ways, even away from a given product. Degree and presentation differ by model. The same failure does not occur at the same magnitude.

ProblemMainly arises fromProduct-independent implication
Context RotCost of self-attention and dilution of attentionKeep input short
Lost in the MiddleBias of attention from positional encodingDo not put important information in the middle
Priority SaturationLimit on how many instructions can condition at onceDo not grow always-on instructions
HallucinationNext-token predictionPut factual checks outside the model
SycophancyRLHF preferring agreementSeparate generation from verification. Do not ask "is this fine?"
Knowledge BoundaryWeak reward for "I don't know"Ground in primary sources outside the model
Prompt SensitivityDependence on statistical patterns in the token sequenceMake the axes you do not want to vary explicit
Instruction DecayThe above stacking over timeKeep conversations short. Persist decisions in files

Sycophancy is especially visible in models that went through RLHF. Base models can look different. Measured Prompt Sensitivity also depends on the evaluation method. The constraint is not absent. Its apparent size changes with how it is measured.

What in the countermeasures does not depend on the product ​

Each Claude Code feature is a representative example. Feature names need not be ported. What transfers is the way of thinking.

  1. Keep always-on context minimal — keep instructions that are passed every turn short
  2. Read conditionally — load rules only when they apply
  3. Verify in an independent Context — do not ratify in the same conversation that produced the output
  4. Put mechanical checks outside Context — tests, lint, and CI do not depend on the LLM
  5. Keep sessions short — drop or summarize history at task boundaries

These five can be done without dedicated commands. The procedure is in Practice Without Tool Support. Confirmed placement on other products is in Cursor / Cline / Copilot Mapping.


Previous: Part 11: Applying to Other LLMs

Next: Practice Without Tool Support

Released under the CC BY 4.0 License.