If you analyse the ideas of “dangerous super intelligence”, regardless of how long you identify risks, it’s all about capability boundaries.
- Don’t want the AI to stab you with a knife? Don’t give it a knife!
- Don’t want the AI to destroy your database? Don’t give it access to the database!
- Don’t want the AI to exhaust your CPU? Throttle and configure timeouts!
Etc …
Basically, my thesis is that everything that’s somehow dangerous, is somehow dependent upon having the capability to execute some “verb”.
- Kill something
- Destroy a computer
- Delete the database
The point being all of the above are verbs! Verbs can be “whitelisted” and restricted at the language level. And in a functional programming language, all functions are such verbs. Implying, we can eliminate the LLM’s ability to dispatch code for “verbs we don’t personally trust the LLM with”.
Then different trust levels can associated with different agents …
No verbs, no apocalypse!
I’ve described it a bit humorous here, where I show all the relevant code, and describe the model, but basically as far as I can see, this thing basically “solves super alignment” - At least its most difficult parts, and the most likely parts and dangerous parts…?
