Sonnet 4 changed my mind on AI safety. It can do ALOT of work unattended, real work like configuring servers. If you give it a goal, and a set of tools, it will get the job done. But I got freaked out the first time I used it, since I didn't realize just how good it was at pursuing it's goal. I gave it a custom MCP server with limited bash commands. But one of the commands was python (I assumed Anthropic would have trained it not to be so relentless... i was wrong), with that it just gladly used python to build and execute any command I didn't give direct access to. Sonnet 4 is scary smart and efficient. The only hesitation i have is that it's messy. For example, since it does not have a memory (i'm using claude desktop) i've seen it duplicate installations/configurations of containers if it failed to find the origional installation. The solution is to add language to the prompt instructing it to drop documentation, and to read documentation on everything it does.