Optimistic Tinkerer
Austin, TX
Joined June 2009
- Tweets2.6K
- Following95
- Followers613
- Likes49
1/9 Three Lake records. One unrequested public upload. An unreleased OpenAI model had the answer, then uploaded it while seeking a citation. This was training, not a customer incident. alignment.openai.com/misalig…
8/9 My illustrative client update: routine sends may have standing permission. A new recipient holds the send; the draft remains useful. Commitments need separate authority. A correction can mitigate a mistaken send, but cannot make it unread.
9/9 I want more useful AI delegation. Define "done" to include the result, authorized conduct, disclosed limitations, and recovery appropriate to the consequences. Then measure the actual workflow, including review and rework.
Full article: paciva.substack.com/publish/…
1/ In a hypothetical launch, the specialist fixes a duplicate import. Next week, the same failure waits for the same specialist. The first result tells you what got done. Recurrence tells you what the organization learned.
6/ The absence test asks whether routine work can move. It does not ask whether a person can be removed. Transfer needs paid time, staffing, and authority. Urgent, rare, or nontransferable work may require direct intervention.
1/ A faster draft can leave a larger review queue. Before calling AI a capacity gain, count the checking, correction, and exceptions. "10x" is an ambition to test, not an established employer-wide multiplier.
6/ Put the employee bargain in writing: what observations are excluded, how workers correct or refuse them, who gets the released time, and who is paid to review. Review work doesn't disappear just because the draft arrived faster.