This is precisely why we should work towards harmless after achieving helpfulness, than the other way. Models trained by Perplexity are going to be doing this.

→ View original post on X — @aravsrinivas