Deploying GPT-4 subject to adversarial pressures of real world has been a great practice run for practical AI alignment. Just getting started, but encouraged by degree of alignment we've achieved so far (and the engineering process we've been maturing to improve issues).

→ View original post on X — @gdb