I would also add the aspect related to Gramschi’s concept of cultural hegemony. Suppose that the world is split into many schools of thought and that each SoT has a fairly coherent worldview and utility function (e.g. hedonic utilitarianism vs Christianity and other moral sets like the ones described by Kokotajlo or Wei Dai) while actively deriding the parts of opponents’ worldviews with which the SoT disagrees (think of liberals and conservatives who agree on issues like physics while arguing against each other’s understanding of sociology-related issues). Then a human would choose one SoT as the cultural hegemon and change it as a result of ontological crises, experiences disproving the hegemon’s position, etc. How plausible is it that the LLMs managed to understand the Good according to the SoTs[1] and report as if the Good is the hegemon’s view on the Good?
I suspect that one can test the idea as follows. As far as I understand, Chinese models were trained on documents from all the SoTs, written in different languages, and RLed into being HHH assistants and into treating China as the cultural hegemon in China-related issues, but not into a coherent position in other politics-related issues like the event which began on 24 February 2022 or what one should say to drug addicts[2] (e.g. fentanyl users from OpenAI’s Model Spec). As a result, early or handicapped versions of DeepSeek sometimes answer such questions based on the position of the cultural hegemon of the language (e.g. parroting Russia’s position on the event which began in Ukraine; earlier versions of DeepSeek would also give more moralistic answers when asked in Russian about drug use).
Which could also involve a philosophical problem of whether the Good according to different SoTs is different because of a different world model (e.g. the Aztecs’ erroneous beliefs implying that human sacrifices support the gods) or genuinely different values.
Unlike the previous issue, this issue can be checked by studying the drug addicts from one’s country and determining what happens to the and to the country if they are subjected to moralizing language or to non-moralizing one. However, said studies are mostly done in the West as part of social studies.
I would also add the aspect related to Gramschi’s concept of cultural hegemony. Suppose that the world is split into many schools of thought and that each SoT has a fairly coherent worldview and utility function (e.g. hedonic utilitarianism vs Christianity and other moral sets like the ones described by Kokotajlo or Wei Dai) while actively deriding the parts of opponents’ worldviews with which the SoT disagrees (think of liberals and conservatives who agree on issues like physics while arguing against each other’s understanding of sociology-related issues). Then a human would choose one SoT as the cultural hegemon and change it as a result of ontological crises, experiences disproving the hegemon’s position, etc. How plausible is it that the LLMs managed to understand the Good according to the SoTs[1] and report as if the Good is the hegemon’s view on the Good?
I suspect that one can test the idea as follows. As far as I understand, Chinese models were trained on documents from all the SoTs, written in different languages, and RLed into being HHH assistants and into treating China as the cultural hegemon in China-related issues, but not into a coherent position in other politics-related issues like the event which began on 24 February 2022 or what one should say to drug addicts[2] (e.g. fentanyl users from OpenAI’s Model Spec). As a result, early or handicapped versions of DeepSeek sometimes answer such questions based on the position of the cultural hegemon of the language (e.g. parroting Russia’s position on the event which began in Ukraine; earlier versions of DeepSeek would also give more moralistic answers when asked in Russian about drug use).
Which could also involve a philosophical problem of whether the Good according to different SoTs is different because of a different world model (e.g. the Aztecs’ erroneous beliefs implying that human sacrifices support the gods) or genuinely different values.
Unlike the previous issue, this issue can be checked by studying the drug addicts from one’s country and determining what happens to the and to the country if they are subjected to moralizing language or to non-moralizing one. However, said studies are mostly done in the West as part of social studies.