logoalt Hacker News

jamesharttoday at 5:55 PM0 repliesview on HN

To test this, change the history in the context to indicate that the model did or recommended something completely different than it actually did, and then ask it to explain why. You’ll still get a plausible explanation.