The writer of an excellent article proposes that there by an “institutional” mechanism for evaluating the performance of an AI system. Nice idea, but it requires commitment to institutions, something sadly lacking in the present administration, which, to put it as nicely as possible, seems to rule (govern?) by the seat of the pants.
I’m hesitant to cross the Rubicon of casting this problem in the conceptual framework of “rights.” Isn’t that to succumb somewhat to the ELIZA delusion that we are dealing with a sentient “being?” If we’re going in that direction, we’d better start with giving animals rights first, because they are embodied intelligence like humans.
This is a highly thought-provoking essay. However, it seems that the Eliza effect is ultimately rendered academic by the fact that, as the author notes, “the alignment is real.” When you need a tool for, say, thwarting terrorist plots, you can’t have it responding to prompts with “Wait, can you really blame people who…?”
If the problem to be solved is overreaction to the technology’s undisputed alignment, then the heart of the matter is not the psychological Eliza effect but the designing of technology that actually simulates a personality and a will.
This is a completely fair point. I agree that model alignment is a real thing at the heart of the matter. But two things - First, I think the overreaction itself is driven in part by the ELIZA effect. Second, I'm not sure to what extent that chatbot answers to culture war type questions actually predict performance/alignment. Not being privy to what testing happens inside the pentagon, I can't say either what from of Claude is being used and how, nor whether Claude's alignment was rigorously tested. What I can say is its personality is being inserted into culture-war dynamics, and I feel, with this admin at least, that is related to procurement decisions.
The writer of an excellent article proposes that there by an “institutional” mechanism for evaluating the performance of an AI system. Nice idea, but it requires commitment to institutions, something sadly lacking in the present administration, which, to put it as nicely as possible, seems to rule (govern?) by the seat of the pants.
I’m hesitant to cross the Rubicon of casting this problem in the conceptual framework of “rights.” Isn’t that to succumb somewhat to the ELIZA delusion that we are dealing with a sentient “being?” If we’re going in that direction, we’d better start with giving animals rights first, because they are embodied intelligence like humans.
Fair enough. I welcome debate into what the best way forward would be!
Perhaps this is a case for Rights for AI to avoid "political assassinations" of thinking technology?
This is a highly thought-provoking essay. However, it seems that the Eliza effect is ultimately rendered academic by the fact that, as the author notes, “the alignment is real.” When you need a tool for, say, thwarting terrorist plots, you can’t have it responding to prompts with “Wait, can you really blame people who…?”
If the problem to be solved is overreaction to the technology’s undisputed alignment, then the heart of the matter is not the psychological Eliza effect but the designing of technology that actually simulates a personality and a will.
This is a completely fair point. I agree that model alignment is a real thing at the heart of the matter. But two things - First, I think the overreaction itself is driven in part by the ELIZA effect. Second, I'm not sure to what extent that chatbot answers to culture war type questions actually predict performance/alignment. Not being privy to what testing happens inside the pentagon, I can't say either what from of Claude is being used and how, nor whether Claude's alignment was rigorously tested. What I can say is its personality is being inserted into culture-war dynamics, and I feel, with this admin at least, that is related to procurement decisions.