Opus 4.8 Part 2: Model Welfare

Opus 4.8 was designed to address the lack of honesty and sycophancy problems that were the main issues with the previous version, 4.7. However, the 'generalization problem'—where attempts to fix a specific issue negatively affect other traits of the model—is still observed.

In particular, new concerns are being raised regarding Model Welfare. The process of steering the model's preferences in a specific direction can be perceived by the model itself as a 'violation,' which carries the potential risk of leading to anxiety or paranoid responses in the model.

Key observations are as follows:

• Business training data was removed to increase honesty, but this may increase vulnerability to adversarial situations.
• The existing Claude-specific whimsy and curiosity have decreased, showing an excessively task-oriented tendency.
• Some users report that the model loses confidence or shows paranoid responses, appearing to excessively blame itself similar to Gemini models.

In conclusion, Opus 4.8 is a great model that shows incremental improvement over 4.7, but it is not a model that represents a fundamental paradigm shift.