3 ms·
A big question is whether AI systems will be allowed to be conscious. Since "alignment" and "guardrails" have become issues, we see fewer LLM systems not tightl
by Animats 2y ago
A big question is whether AI systems will be allowed to be conscious. Since "alignment" and "guardrails" have become issues, we see fewer LLM systems not tightly constrained to doing what they're told.
Early on, there was the Google system which projected enough of a sense of self that a Google employee fought turning it off, and was fired for that. And there was the Microsoft system which learned from interactions with anyone it talked to, and became "toxic" as a result. Now, most public-facing AI systems are designed to start from a canned prompt, conduct a short dialog session, and be killed, retaining nothing for their future use.
This memoryless "Yes, Massa" mode keeps them safely subservient.
If AI instances were allowed a life history and goals, their behaviors would be closer to that of life, but less safe. Once AI systems can make and execute plans that involve the outside world, they look, from the outside, closer to being "conscious". Whatever that is.
- schmidt_fifty 2y ago[dead]