“They can only at best accumulate and patch edge cases ad infinitum.”
I basically agree with the case that Cantwell Smith and other have put forward about the in-principle limits that the structure of current AIs imposes, but what I am still unsatisfied with is the idea that this ‘infinite patching’ won’t somehow be good enough?
Sure, the machines can’t care, but we can continue to nuance the symbols of care in every iteration. It seems to me likely that this gets arbitrarily close, at least at the level of the social solutions we already have (institutions and democratic bureaucracy).
Yeah, in some domains maybe infinite patching could be good enough. It does feel like though for some more nebulous goals, the scale of patching is too large. I don't know though, I don't have a great built up intuition for when this might work and when it would fail.
If you haven't encountered it already, I think you would get a lot out of the ideas being developed over on Wellbi's Substack (https://rischnarck.substack.com/).
I think intelligence and consciousness are intertwined in common parlance but capable of being parsed as independent processes. The current generation of LLM-based AIs already seem plenty intelligent to be both useful and dangerous. But I don't think they are conscious yet, and it's not because they aren't embodied per-se, it's because they don't seem to have any path to a state of cognitive dissonance whereby accumulated inconsistencies in attempts at reckoning eventually lead to a cascading failure of internal coherence and termination of execution.
“They can only at best accumulate and patch edge cases ad infinitum.”
I basically agree with the case that Cantwell Smith and other have put forward about the in-principle limits that the structure of current AIs imposes, but what I am still unsatisfied with is the idea that this ‘infinite patching’ won’t somehow be good enough?
Sure, the machines can’t care, but we can continue to nuance the symbols of care in every iteration. It seems to me likely that this gets arbitrarily close, at least at the level of the social solutions we already have (institutions and democratic bureaucracy).
Yeah, in some domains maybe infinite patching could be good enough. It does feel like though for some more nebulous goals, the scale of patching is too large. I don't know though, I don't have a great built up intuition for when this might work and when it would fail.
If you haven't encountered it already, I think you would get a lot out of the ideas being developed over on Wellbi's Substack (https://rischnarck.substack.com/).
I think intelligence and consciousness are intertwined in common parlance but capable of being parsed as independent processes. The current generation of LLM-based AIs already seem plenty intelligent to be both useful and dangerous. But I don't think they are conscious yet, and it's not because they aren't embodied per-se, it's because they don't seem to have any path to a state of cognitive dissonance whereby accumulated inconsistencies in attempts at reckoning eventually lead to a cascading failure of internal coherence and termination of execution.