Obsidian Metadata
| type | source |
| subtype | video |
| urls | https://youtu.be/NSe1ObojPQQ, https://www.youtube.com/watch?v=XbZvC4KTH68 |
Short clips
Evals for AI Writing (2:22), https://youtu.be/NSe1ObojPQQ
A clip from a longer session. The team compares two AI-drafted social posts about a blog post on a document-QA “lab” and picks the clearer one. The underlying example is an eval-design point: answers cite a bounding box on the source page, so checking takes seconds instead of a reread. Once checking is cheap, you can swap one pipeline stage at a time (reading/OCR, finding with grep vs semantic search, answering) across the same 22 questions over 77 pages and see what changed. In that experiment, every text-run failure traced back to OCR, and search method didn’t matter. Verification-friendly design makes component experiments cheap (see Designing for Verification).
Cursor: agents show demos, not diffs (1:42), https://www.youtube.com/watch?v=XbZvC4KTH68
Cursor’s background agents get their own computers, test their changes, and return a video of the working feature. You evaluate the result by watching it, without reading code, and you can compare runs from several models side by side. The eval-smell post cites this as going “the extra mile” on verifiability even in an already verifiable domain. The review job moves from checking code to judging ideas.