I've been thinking about recursive self improvement. But especially the likelihood that it will be important soon. Soon means with current LLMs that might just give up or delete the test set or otherwise detach from reality.
Source: [Hacker News](https://news.ycombinator.com/item?id=48611850)