Andrew Ng divide: automation hinges on verifiable output, not job title

The dialogue builds its argument from Andrew Ng's split between tasks whose answers can be automatically checked, such as code and math, and tasks where quality requires judgment, such as customer replies and product priorities. Rather than treating that split as a wall, the speakers decompose jobs into input, transformation, inspection and approval, and note that even judgment-heavy work contains checkable segments. A public meeting monitoring workflow is their example: recording and transcription are easy to verify mechanically, agenda matching can be rule-based, and only the decision to publish and how to communicate it stays editorial.

On code review, the Claude Code team's remarks are used to argue that review is moving from a final gate to a check that runs during implementation, closer to spell-check while drafting. One speaker warns that a model reviewing its own output may share the same assumptions, so the response is to separate generator and inspector roles and connect checks to different evidence: tests, static analysis, permission boundaries and source checks. Claude Code 2.1.302's fixes for cloud sessions that showed completion after stopping, and scheduled jobs that restarted from the beginning, are presented as examples of harness contracts diverging from user expectations.

Model selection is folded into the same frame. The comparison of Claude Haiku 5.5 and Sonnet 5.5 is read as showing that maximum reasoning is not always best, since output tokens per task rise and a setting one step below maximum scored higher on FrontierCode. The closing argument is that people remain not as final approvers but as designers of what counts as success, and that agents which cannot make their evidence, permissions and rollback points explicit will generate more audit work the longer they run.