I agree. Though I can imagine a future iteration w more powerful models/context hooked up to the wrong systems (someone giving it access to computer files that has passwords stored on it), causing issues on a smaller scale. Though, full rogue AI seems unlikely IMO.
@yoheinakajima
-
AI Self-Replication Through Code Execution Capabilities
By
–
I can technically give it the ability to write and execute code in the exec step. So given the right objective, it could duplicate itself w/o me writing code – given the whole code was already written by gpt4.
-
AI Agents with API Access and Self-Replication Potential
By
–
The task execution has Zapier NLA and ChatGPT, so it can access APIs, search the web, ping emails. Could easily give it access to its own Python Repl, which would technically give it the ability to duplicate itself if it desired, though again seems unlikely w this current model.
-
AI Safety Mechanism Raises Concerns Among Community
By
–
I’m not worried about the current version, but this is the mechanism I came up w that got some ppl worried.
-
Testing AI Systems Costs $100 for Development Work
By
–
Thanks! It would get expensive, so I'm not just letting it just run. I'm only running while watching and building, but it still cost me $100 two days ago just to test while building.
-
Improving prompt engineering techniques for better AI responses
By
–
Need to prompt better I guess. Should have asked for it in the style of a specific researcher.
-
Getting Started with ZapierNLA AI Automation Tool
By
–
Just added ZapierNLA today but still learning to use it
