3 ms·
> “It would be simple for a trained android or human to fake the test just like a polygraph” This seems like a massive assumption about how the android brain w
by pavlov 4y ago
> “It would be simple for a trained android or human to fake the test just like a polygraph”
This seems like a massive assumption about how the android brain works.
In contemporary terms, it’s like asking: “Surely you can ask ChatGPT to only tell the truth? It’s a computer, it can’t lie.” — when the reality is that it’s exactly the opposite (it’s a computer which only knows how to lie and most of the time it happens to be exactly right).
- krapp 4y agoIt seems like a massive assumption to extrapolate the limitations of ChatGPT to the AGI of androids. ChatGPT neither lies (which requires intent) nor tells the truth (which requires knowledge) because it doesn't retain state, and cannot learn from the outside world beyond whatever is provided in a prompt stack. One can assume an android in the hypothetical future with fully sentient AGI has a mind capable of retaining state and working senses, and thus is complex enough to be capable of willing deception.
- pavlov 4y agoI’m not saying the future android would be based on ChatGPT. I’m suggesting it could be based on a model of computation whose capabilities and limitations could be unforeseen and unintuitive to us, in the same way as ChatGPT is something completely different compared to the symbolic fact-based AI systems of decades past. Someone from the past whose idea of AI is Prolog would have a hard time understanding ChatGPT only based on interactions with it.
- PaulDavisThe1st 4y ago> ChatGPT is something completely different compared to the symbolic fact-based AI systems of decades past. thank you for recognizing this and writing it down clearly.
- 0_____0 4y agoI never thought of ChatGPT like that but by Gods that makes a ton of sense