Our summary
Fastino introduces GLiDE, an AI model that chooses from supplied options rather than writing an ordinary reply. The authors say it starts with a quick assessment, then reasons further when uncertain. This is a competing model, not Jev, even though it accepts a similar format for questions and answers.
The example asks what to do when an assistant proposes deleting stored information from a live service. It has permission to view information, not change it, and no person approved deletion. Possible answers include stopping the action and alerting someone. The displayed response is a template, not a decision observed in a test.
Fastino reports stronger results than Jev in its own run of the Decision Index, a collection of tests. That is not independent confirmation or a public ranking. A model’s answer does not grant permission: ordinary code must still check access and prevent harmful changes. Supplied information and extra reasoning must fit within one shared limit.
Key takeaways
- Fastino proposes extra reasoning for difficult choices; a familiar answer format does not mean the model is Jev.
- Keep permission checks and protection against harmful changes in code; a model’s answer is advice, not permission.
- Distinguish a response template from a tested answer, and the maker’s tests from a result listed in a public ranking.