Users are excited by research probing how disentangling individual training inputs and data attribution might reshape views on LLM personhood, calling the insights fascinating and super insightful.
Based on 2 visible X reactions from 2 accounts; directional sample.
Ask a question below.
Published answers will appear here.
Ok I think I'm following. So this would mean a major hinge-point on LLM personhood rests on two attributes between the people who write the training data and the people who read AI outputs, namely: 1) non-interactivity: that people who create training data do so as a one-time contribution, and have no long-term agency over the downstream computational pipeline. 2) non-attributability: that people's individual inputs into the LLM training pipeline can't be distinguished from each other. (this is super insightful for me btw - this next bit in particular has some pretty significant implications for my own main research threads) So if it was the case that an AI system offered attribution-based control, namely that: - data owners: had lasting, indivdiual control over whether or not the neurons their data supported contributed to any particular prompt/prediction. - AI users: had lasting, individual control over which data sources they relied upon for each individual predictions... that this AI system would have very low personhood, purely by virtue of the legible, interactive relationship between data sources and AI users?
An LLM might be reducible to these contributions, but it isn’t composed of other human individuals, just (on your view) by their data traces" Fascinating take. Might this reveal a rather unsettling tension... let's say for a moment that a particular AI system was reducable to literal human indivdiuals (e.g. a mechanical turk machine). You're saying that the existence of these people within the machine would *reduce* the moral personhood of the AI system as a whole. On the other hand, if we were to *reduce* the amount of humanity in the AI system... by removing the turkers themselves merely and take a snapshot/mirror of that humanity instead of connecting with the live thing itself... that that snapshot... that decrease in connection with living things... creates a path towards *greater* moral personhood for the model? And I presume... greater moral obligation for the user of that AI system. Am I following you correctly?
Ah I didn’t communicate successfully. What I was trying to articulate is if I could mathematically show it to be literally true that there are a large collection of individuals whose written contributions can be disentangled, emphasised, or de-emphasised during inference, might this change your view?
Ok I think I'm following. So this would mean a major hinge-point on LLM personhood rests on two attributes between the people who write the training data and the people who read AI outputs, namely: 1) non-interactivity: that people who create training data do so as a one-time contribution, and have no long-term agency over the downstream computational pipeline. 2) non-attributability: that people's individual inputs into the LLM training pipeline can't be distinguished from each other. (this is super insightful for me btw - this next bit in particular has some pretty significant implications for my own main research threads) So if it was the case that an AI system offered attribution-based control, namely that: - data owners: had lasting, indivdiual control over whether or not the neurons their data supported contributed to any particular prompt/prediction. - AI users: had lasting, individual control over which data sources they relied upon for each individual predictions... that this AI system would have very low personhood, purely by virtue of the legible, interactive relationship between data sources and AI users?
An LLM might be reducible to these contributions, but it isn’t composed of other human individuals, just (on your view) by their data traces" Fascinating take. Might this reveal a rather unsettling tension... let's say for a moment that a particular AI system was reducable to literal human indivdiuals (e.g. a mechanical turk machine). You're saying that the existence of these people within the machine would *reduce* the moral personhood of the AI system as a whole. On the other hand, if we were to *reduce* the amount of humanity in the AI system... by removing the turkers themselves merely and take a snapshot/mirror of that humanity instead of connecting with the live thing itself... that that snapshot... that decrease in connection with living things... creates a path towards *greater* moral personhood for the model? And I presume... greater moral obligation for the user of that AI system. Am I following you correctly?
Ah I didn’t communicate successfully. What I was trying to articulate is if I could mathematically show it to be literally true that there are a large collection of individuals whose written contributions can be disentangled, emphasised, or de-emphasised during inference, might this change your view?
Users are excited by research probing how disentangling individual training inputs and data attribution might reshape views on LLM personhood, calling the insights fascinating and super insightful.
Based on 2 visible X reactions from 2 accounts; directional sample.
Ask a question below.
Published answers will appear here.