Interesting. But until we have a better understanding of the mechanisms of qualia (like, what mechanism makes ours switch off while we sleep?), I don't see how we're going to be able to differentiate between these two hypotheses.
Well, not unless the researchers made the same mistakes as Blake Lemoine did with LaMDA, not noticing the transcript was describing experiences that were genuinely impossible for it to have had, like being left alone for days: https://insiderpaper.com/transcript-interview-of-engineer-le...
I am very not surprised that "Several versions of the models scored above diagnostic thresholds" for conditions. Given we got here by cargo culting*, it would have been surprising if it had been otherwise, even if it turns out they have anything that it's like to be.
But also, even then, if we presume for the sake of argument that they're deserving of being called "minds", the minds that they are, are alien. I am unsure if there's even a way to be sure whether or not "all showed levels of worry that in people “would be clearly pathological”" means for them what it means for us.
* Just because you don't know why planes land when you clear a runway strip and build an observation tower and march up and down in clothes modelled on WW2 pacific theatre clothes, doesn't mean planes won't land out of curiosity a decade after the war finished.
dekhn's 27th law: If there is a thing, there are concern nannies who publish papers about how the thing is bad, and the press will report on it with the intention of making people who use the thing feel bad. The less quantitative the field, the more misleading the research will be.
The corollary to this law is that if someone publishes a paper saying a thing is bad, there will be a flippant, dismissive response ignoring the content and context, and using attacks on credibility to dismiss the findings.
Given that we know significant numbers of people (mostly younger) are using LLMs as confidantes/therapists, to the extent that some have self-harmed following alleged prompting/encouragement by the LLMs, perhaps it would be useful to understand why this behaviour in the LLM emerges and whether it is linked to specific types of training data, and subsequently - maybe - figure out a way to mitigate these kinds of risks to human users.
But hey, who doesn't enjoy giving a free kick to those that don't have our superior understanding, am I right?
Interesting. But until we have a better understanding of the mechanisms of qualia (like, what mechanism makes ours switch off while we sleep?), I don't see how we're going to be able to differentiate between these two hypotheses.
Well, not unless the researchers made the same mistakes as Blake Lemoine did with LaMDA, not noticing the transcript was describing experiences that were genuinely impossible for it to have had, like being left alone for days: https://insiderpaper.com/transcript-interview-of-engineer-le...
I am very not surprised that "Several versions of the models scored above diagnostic thresholds" for conditions. Given we got here by cargo culting*, it would have been surprising if it had been otherwise, even if it turns out they have anything that it's like to be.
But also, even then, if we presume for the sake of argument that they're deserving of being called "minds", the minds that they are, are alien. I am unsure if there's even a way to be sure whether or not "all showed levels of worry that in people “would be clearly pathological”" means for them what it means for us.
* Just because you don't know why planes land when you clear a runway strip and build an observation tower and march up and down in clothes modelled on WW2 pacific theatre clothes, doesn't mean planes won't land out of curiosity a decade after the war finished.
dekhn's 27th law: If there is a thing, there are concern nannies who publish papers about how the thing is bad, and the press will report on it with the intention of making people who use the thing feel bad. The less quantitative the field, the more misleading the research will be.
The corollary to this law is that if someone publishes a paper saying a thing is bad, there will be a flippant, dismissive response ignoring the content and context, and using attacks on credibility to dismiss the findings.
Given that we know significant numbers of people (mostly younger) are using LLMs as confidantes/therapists, to the extent that some have self-harmed following alleged prompting/encouragement by the LLMs, perhaps it would be useful to understand why this behaviour in the LLM emerges and whether it is linked to specific types of training data, and subsequently - maybe - figure out a way to mitigate these kinds of risks to human users.
But hey, who doesn't enjoy giving a free kick to those that don't have our superior understanding, am I right?