a commercial LLM provider training their own models is however likely to bias the model(/guardrail) harder, in an effort to make them harder to jailbreak, to minimize bad press.
For example:
- refusing to talk even about the well-known parts of forbidden topics (this) - tending toward sycophancy to avoid ever seeming rude or unhelpful
