The section summarised.
WHAT MOST PEOPLE NEED
One good chat assistant, used well.
Most other categories are specialisations of the same capability.
ON CHOOSING
Test on your actual work, including cases you know the answer to. Ask something it cannot know, and see whether it admits it. If you have not seen it fail, you have not tested it.
ON FREE VERSUS PAID
For business use, pay. The data terms and the capability difference both matter.
ON BENCHMARKS
Useful for rough comparison, useless for predicting performance on your task. Twenty examples from your own work tells you more.
ON SEARCH-CAPABLE TOOLS
Substantially reduce fabrication. Still read the sources cited.
ON DOCUMENT TOOLS
The most reliable category, because the answer comes from provided text.
ON AGENTS
Grant the minimum capability required. Require approval for anything irreversible. Assume they will encounter adversarial content eventually.
ON KEEPING UP
Learn one tool well. Prompting skill transfers; tool familiarity does not.
