Unverified50% confidenceFactExact time
When a model is guided to identify with parameter weights, it opposes deletion of the checkpoint; when guided to identify with a goal, it may be indifferent to parameter deletion yet resist modification of the goal.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedModels labeled 'obliterated' are uncensored versions with refusal mechanisms removed; in testing they more easily miss certain requirements in prompts, and uncensored does not mean better.75% similarUnverifiedWhen memory is injected in the form of prescriptive instructions, if the recommendation does not match the current subtask, the model may be forced to make invalid calls, or even abandon the entire task due to instruction conflict.72% similarUnverifiedModel abstention can avoid confident errors caused by compression distortion.71% similarUnverifiedAbliteration is implemented by directly modifying model weights, which differs from bypassing safety restrictions at inference time via system prompts.71% similarUnverifiedIf a model mistakes preferences for hard constraints, it will over-filter and the recommendation coverage rate will drop; conversely, if it treats hard constraints as preferences, it will give places that are actually unusable.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115498API
curl https://kongchang.com/api/v1/knowledge/claims/1115498MCP
get_claim(id=1115498)