3 ms·

Thanks for Matrix. This proposal seems to address general abuse, spam, propaganda, filter bubbles, and so on. While these are worthy issues to tackle, the auth

Thanks for Matrix. This proposal seems to address general abuse, spam, propaganda, filter bubbles, and so on. While these are worthy issues to tackle, the authors of the 7-eyes statement are not really interested in them. What they say they want is for terrorism and child pornography to be detectable or meaningfully reduced. Do you think a relative reputation system will solve that problem?

Arathorn 6y ago
Yes, we think it will help meaningfully help detect and reduce abuse - mainly because it's shamelessly mimicking the way society works in real life. Many (most?) abusive communities are at least adjacent if not overlapping with publicly visible communities. Simply put, they need a way to advertise and recruit. So, just as in person you might stay away from a given political party / religious institute / youth club because your social graph has warned you that it might actually be a front for whatever obnoxiousness, the same approach can work online (or, conversely, could also be used to help hunt down abuse in the first place). It's then up to the authorities to investigate what's going on - which is quite possible through infiltration etc without having to go and blanket break encryption for the whole of society. To be clear: this is still largely sci-fi, and we don't think this is a perfect solution, especially given this is a fundamental problem of the human species which nobody has yet solved. Our proposal doesn't solve lone wolf situations, for instance. So perhaps for that you need the ability to gather evidence from endpoints post hoc. Unrelated: one particularly dark dystopian outcome we've been wargaming is: what if someone (not us!) used decentralised rep to seed a GPT-3 style bot to locate abusive communities, and then automated the process of infiltrating & investigating them... only to then end up ascending the ranks while preserving its cover and accidentally triggered some atrocity. So, um, let's not do that.)
- skissane 6y ago
  > Unrelated: one particularly dark dystopian outcome we've been wargaming is: what if someone (not us!) used decentralised rep to seed a GPT-3 style bot to locate abusive communities, and then automated the process of infiltrating & investigating them... only to then end up ascending the ranks while preserving its cover and accidentally triggered some atrocity. So, um, let's not do that.) I think you greatly overestimate the abilities of GPT-3. GPT-3's writable memory is short-term only – its long-term memory is read-only – so by the end of a conversation, it has permanently forgotten what the topic was at the beginning. Given its lack of writable long-term memory, it is completely incapable of succeeding at any tasks involving long-term planning, such as the infiltration and investigation of a social group. Of course, you said "GPT 3 style bot", not GPT-3, but the AI you have in mind has significantly greater capabilities than GPT-3 has, so isn't fairly said to be "GPT 3 style" at all.
  - Arathorn 6y ago
    Sure - "GPT-3 style" here meant "something along the lines of GPT-3" rather than anything specific to GPT-3 actual capabilities. The context being dystopian sci-fi thought experiments :) (That said, I suspect you could train GPT-3 to automatically get vetted into some fairly nasty places if you so desired).

Arathorn 6y ago

Yes, we think it will help meaningfully help detect and reduce abuse - mainly because it's shamelessly mimicking the way society works in real life. Many (most?) abusive communities are at least adjacent if not overlapping with publicly visible communities. Simply put, they need a way to advertise and recruit. So, just as in person you might stay away from a given political party / religious institute / youth club because your social graph has warned you that it might actually be a front for whatever obnoxiousness, the same approach can work online (or, conversely, could also be used to help hunt down abuse in the first place). It's then up to the authorities to investigate what's going on - which is quite possible through infiltration etc without having to go and blanket break encryption for the whole of society. To be clear: this is still largely sci-fi, and we don't think this is a perfect solution, especially given this is a fundamental problem of the human species which nobody has yet solved. Our proposal doesn't solve lone wolf situations, for instance. So perhaps for that you need the ability to gather evidence from endpoints post hoc. Unrelated: one particularly dark dystopian outcome we've been wargaming is: what if someone (not us!) used decentralised rep to seed a GPT-3 style bot to locate abusive communities, and then automated the process of infiltrating & investigating them... only to then end up ascending the ranks while preserving its cover and accidentally triggered some atrocity. So, um, let's not do that.)

skissane 6y ago
> Unrelated: one particularly dark dystopian outcome we've been wargaming is: what if someone (not us!) used decentralised rep to seed a GPT-3 style bot to locate abusive communities, and then automated the process of infiltrating & investigating them... only to then end up ascending the ranks while preserving its cover and accidentally triggered some atrocity. So, um, let's not do that.) I think you greatly overestimate the abilities of GPT-3. GPT-3's writable memory is short-term only – its long-term memory is read-only – so by the end of a conversation, it has permanently forgotten what the topic was at the beginning. Given its lack of writable long-term memory, it is completely incapable of succeeding at any tasks involving long-term planning, such as the infiltration and investigation of a social group. Of course, you said "GPT 3 style bot", not GPT-3, but the AI you have in mind has significantly greater capabilities than GPT-3 has, so isn't fairly said to be "GPT 3 style" at all.
- Arathorn 6y ago
  Sure - "GPT-3 style" here meant "something along the lines of GPT-3" rather than anything specific to GPT-3 actual capabilities. The context being dystopian sci-fi thought experiments :) (That said, I suspect you could train GPT-3 to automatically get vetted into some fairly nasty places if you so desired).