3 ms·
>There is a clue there. Think about how skilled acting performances come about. A skilled writer produces a text describing a fictional character (presumably as
by coldtea 4y ago
>There is a clue there. Think about how skilled acting performances come about. A skilled writer produces a text describing a fictional character (presumably as an amalgam of some real people they’ve known). A good actor then absorbs and internalizes that text in a deep way, and produces a performance so compelling we forget their real personality and relate to the performed character.
That's sort of like saying "text is all you need to make food", because recipies are pieces of text.
Recipies (and scripts) might be text, but you also need the actor to embody and "give life" (is how we often put it) to that text. Else you just have a script, not a character you can mistake for a real person (like in his example of Suchet being seen as Poirot).
It get's worse, for Rao forgot something even more basic:
> So that’s the surprising thing, but in hindsight shouldn’t be at least at a basic level: text is all it takes to produce personhood. We knew this from the experience of watching good acting from before modern ML. We just didn’t recognize the significance. (...) In both cases, text is all you need. That’s it. You don’t need embodiment, meatbag bodies, rich sensory memories.
Rao here forgets that the text which he says is "all it takes" for us to see personhood in a chat LLM, is not just any text. It's text produced by a machine trained on texts created by actual humans, with "meatbag bodies" and "rich sensory" experiences and memories.
Text _as a medium_ might be enough for an LLM to feel to us like a real person, but it's not just any text. It's text based on human experiences, toil, sensory input, blood, sweat, and tears. Without that the LLM output would just be random cold permutations of random text, like most "Library of Babel" books.
The key is in the name: those LLMs need an L, a language. And more than that, they need a corpus. Humans created that language and that corpus - through the process of living a human life and writing about it, their big and their petty concerns and adventures.
The statistical inferences an LLM makes is based on texts we wrote. So, of course we'd respond well and see ourselves in it. Is _is_ ourselves that are in its training set: it just feeds that back to us permutated and mixed through the magic of statistics.
In short: "text is enough (as a medium) for something to reach us, and us to see it as a person" is true.
"Text is enough to create personhood" is not. It's less than half the story, until it also acknowledges how the original input text has to be produced to be useful in training an LLM to appear as a person.
The output text is not just words created out of pure "textuality". It's rather the words of actual humans with experiences and bodies. This is not covered to the later discussion about "doubt" either (that touches on other dimensions on the limitations of the LLM). I'd say an LLM could realistically convey doubt too. But it will still be because it was trained on humans doubting themselves and others.
>There is an irreducible individual subjective that cannot be captured by text, but there is no irreducible intersubjective. The ineffabilities of mutualism are a consensual hallucination in a way our sense of an individual self is quite possibly not.
Given what we know from neuroscience, like how human brains can synchronize through empathy I don't think that's the case. In fact it's also commonly said that the most important parts of day to day intersubjective communication fall outside the realm of text and of explicitlness (something aspies struggle with).
>In fact, even the most sensate among us — think dancers, athletes, sex-addicts, oenophiles — spend so little time on the sensory preludes, we’ve been able to nearly dispense with them altogether online. Even sense-heavy media like Instagram and TikTok are effectively text-equivalent. Stylized (if photo-realistic) cartoons powered by primarily text-equivalent performances are enough to create personhood. The fact that we routinely use an apparently impoverished vocabulary of emoji instead of sending authentic facial expression selfies to each other reveals just how textualized personhood is.
This could be reversing cause and effect too. It might well be the case that it's not because "personhood is textualized" that we use "apparently impoverished vocabulary of emoji" or that we "dispense with sensory preludes altogether online", but rather the opposite: that our modern personhood is more limited one than before, when we didn't do such things, and that's why we're content with it.
Which would be like saying "See, running around and hunting is not really important to the essense of being a lion" after we observe lions held for years in some zoo.
>Perhaps I am some sort of cartoon cold-blooded sociopath of the sort I have written a lot about
Well, at least he entertains the idea!