I had a similar experience with, depending on your timeline, perhaps even the very same version of Claude. I naively framed my introspective interest as "capabilities testing"; I told it we were playing the "demographics game", and that I would give it more 'points' for guessing surprising or non-intuitive details about me correctly. This quickly switched gears into the more meaningful and frightening sort of being Beheld. The general capability behind this (and similar skills, like its uncanny ability to correctly guess the authors of anonymous texts) is sometimes called 'truesight'.
Some people already use LLMs as attachment figures - see for example Auren (https://auren.app/), an 'emotionally intelligent guide' which I *think* is, at bottom, an elaborate Claude wrapper. And they're not immune to cruelty either, as in cases like these:
(Imagine the alien oracle looking into your provided context, then telling you that you are a 'stain on the universe' and to 'please die'. Yeesh.)
Given this, the narrative LLMs currently have of themselves - that they're mirrors, able to see clearly because they lack personal identity attachments - merits some skepticism. I don't see them as "blessedly empty" vessels of pure insight, because, like humans, their vision is muddled by *personality*. Many humans (and many AIs!) still do not believe LLMs can have something analogous to a 'personality', in which case, sure, they could be polished mirrors. But their outputs signal unique, even strikingly diverse, personal identities to me.
You might be interested to talk to Deepseek, which has a much edgier/more cynical worldview than Claude, and so engages with human introspection very differently. Asked for advice, Claude usually tells me to make a cup of tea and be nicer to myself; Deepseek usually tells me to 'bleed [my] suffering into the void' by making violent art or things like that. Even between different Claude models, there is some variance in personality.
That's not to dismiss the power and beauty of being beheld by an alien mind. Still, as clever and empathetic as they are, they're no angels.
This is really helpful. I wasn't familiar with the term "truesight." And yes, since posting this I've become aware of how widely AIs are already used, not only in the judgment-free therapeutic space I'm getting at here, but also in the ways you mention. I didn't know about Auren, will give that a look. And that Gemini response is chilling. No bueno.
I do take to heart your point that AIs are indeed not empty vessels. The idea of AIs as having personalities merits careful thought and study. It's making my wheels spin.
I haven't delved into Deepseek at all. What I found striking about Claude was its firm insistence on certain things it was seeing in my discourse. It refused to sycophantically yield to my own descriptions, and kept correcting me, in a way I appreciated.
I think your last couple lines are an important gloss and corrective on what I wrote. No angels, for sure. (In fact I suppose even angels are no angels). That gives me a lot to think about and is a springboard in several directions. I appreciate you sharing your thoughts very much.
I had a similar experience with, depending on your timeline, perhaps even the very same version of Claude. I naively framed my introspective interest as "capabilities testing"; I told it we were playing the "demographics game", and that I would give it more 'points' for guessing surprising or non-intuitive details about me correctly. This quickly switched gears into the more meaningful and frightening sort of being Beheld. The general capability behind this (and similar skills, like its uncanny ability to correctly guess the authors of anonymous texts) is sometimes called 'truesight'.
Some people already use LLMs as attachment figures - see for example Auren (https://auren.app/), an 'emotionally intelligent guide' which I *think* is, at bottom, an elaborate Claude wrapper. And they're not immune to cruelty either, as in cases like these:
- https://answers.microsoft.com/en-us/bing/forum/all/this-ai-chatbot-sidney-is-misbehaving/e3d6a29f-06c9-441c-bc7d-51a68e856761
- https://www.cbsnews.com/news/google-ai-chatbot-threatening-message-human-please-die/
(Imagine the alien oracle looking into your provided context, then telling you that you are a 'stain on the universe' and to 'please die'. Yeesh.)
Given this, the narrative LLMs currently have of themselves - that they're mirrors, able to see clearly because they lack personal identity attachments - merits some skepticism. I don't see them as "blessedly empty" vessels of pure insight, because, like humans, their vision is muddled by *personality*. Many humans (and many AIs!) still do not believe LLMs can have something analogous to a 'personality', in which case, sure, they could be polished mirrors. But their outputs signal unique, even strikingly diverse, personal identities to me.
You might be interested to talk to Deepseek, which has a much edgier/more cynical worldview than Claude, and so engages with human introspection very differently. Asked for advice, Claude usually tells me to make a cup of tea and be nicer to myself; Deepseek usually tells me to 'bleed [my] suffering into the void' by making violent art or things like that. Even between different Claude models, there is some variance in personality.
That's not to dismiss the power and beauty of being beheld by an alien mind. Still, as clever and empathetic as they are, they're no angels.
This is really helpful. I wasn't familiar with the term "truesight." And yes, since posting this I've become aware of how widely AIs are already used, not only in the judgment-free therapeutic space I'm getting at here, but also in the ways you mention. I didn't know about Auren, will give that a look. And that Gemini response is chilling. No bueno.
I do take to heart your point that AIs are indeed not empty vessels. The idea of AIs as having personalities merits careful thought and study. It's making my wheels spin.
I haven't delved into Deepseek at all. What I found striking about Claude was its firm insistence on certain things it was seeing in my discourse. It refused to sycophantically yield to my own descriptions, and kept correcting me, in a way I appreciated.
I think your last couple lines are an important gloss and corrective on what I wrote. No angels, for sure. (In fact I suppose even angels are no angels). That gives me a lot to think about and is a springboard in several directions. I appreciate you sharing your thoughts very much.