1 claim in “technology, safety, AI control”
The fundamental control architecture for safe AI must keep humans in the loop: the AI can suggest architectural improvements, but humans must approve and run new training runs, rather than allowing the AI to modify its own weights autonomously.