Post

When ChatGPT was new and everyone was anxious to understand the specific pace of AI progress, I figured a significant milestone to watch for was when it could find and exploit brand new software vulnerabilities on command. It's slightly unnerving to see the milestone met today.

Questions like "Can this replace most programmers?", "Is this smart enough for recursive self-improvement?", and "Is this dangerous enough now to need regulating or stopping?" could be relatively safely dismissed before that milestone.