Experts Urge AI Specs to Define Task Abandonment Rules
AI safety researchers highlight missing rules in model specs for task limits.
Dylan Hadfield-Menell and Gillian Hadfield stress that AI model specifications must explicitly define when agents abandon tasks or require extra authorization. These elements are absent from current published specs despite being essential for safety and alignment. The discussion, shared on social media, points to gaps in how models handle incomplete or ambiguous instructions. No formal updates to specs have been confirmed from this exchange.
Combined views
1.9K
2 posts, first seen 10h ago
Reactions from ranked influencers
2 posts
Great points here. Going one step further, a characterization of when a model/agent will give up on a task and/or needs additional input/authorization to continue is both 1) a critically important part of the model spec and 2) noticeably absent from publish specs.
...one thing I incredulously wonder sometimes is whether OAI / Ant have enough RL envs where the AI gets 50% done, finds info for second half missing, and then needs to return to say "oops can't continue." If they had 0 or few such environments, it would explain tons of bugs.
Great points here. Going one step further, a characterization of when a model/agent will give up on a task and/or needs additional input/authorization to continue is both 1) a critically important part of the model spec and 2) noticeably absent from publish specs.
...one thing I incredulously wonder sometimes is whether OAI / Ant have enough RL envs where the AI gets 50% done, finds info for second half missing, and then needs to return to say "oops can't continue." If they had 0 or few such environments, it would explain tons of bugs.