3 ms·

Definitely, Moreover, the claim that GPT-3 is a good tool for astroturfing is unproven so far and I am extremely skeptical. I don't think the supply of trollis

Definitely, Moreover, the claim that GPT-3 is a good tool for astroturfing is unproven so far and I am extremely skeptical. I don't think the supply of trollish garbage has even been the barrier to troll-speech shutting down constructive speech. I don't think even the limit on the ability of individual, "real" people to spout Trollish garbage has been the limit. In the GPT-3 text I've read and the GPT-2 systems I've played with, what I read is a system that "weaves" a group of sentences that seem "well written and plausible" on the surface (maybe more well and more plausible than a lot of trolls). But any detailed look seems to show each reason step as both implausible and "weird", senseless. It's impressive but it's got the "talking dog effect" - people are going to be impressed by a talking dog even if the dog's statements aren't impressive otherwise. Moreover, I think most would-be astroturfers have very specific aims. "Because of value X shared by group Y, group Y should hate group Z and like group Q", "Politician X violates strong social taboo Y" and similar formulas. Fifty people in boiler room in a third world country can be hired to make variations on the basic "payload" the astroturfer specifies for less than one silicon valley data scientist customizing GPT-3 and they have less chance of producing GPT-3 weirdness IMO. Some even larger-scale hybrid system might automate the process further. But for the purposes of the astroturfer, it's hardly necessary.

theelous3 6y ago
> But any detailed look seems to show each reason step as both implausible and "weird", senseless. Trouble here is that we're having this discussion in one of the most unusually concerned-with-logic spaces on the web. Your average facebook user isn't doing a detailed analysis of logical steps in a theory. They see something say "wind power bad" and that's that. Wind power bad.
- joe_the_user 6y ago
  If you read my whole comment, yes, trolls are indeed just "wind power bad" and you don't need GPT-3 for that. You've had that forever. GPT-3 has this quality that yes, it's very impressive but much less useful for a purpose than such impressiveness suggests. Like much of current AI, imo. Certainly there are exception, valuable uses, but still.
  - theelous3 6y ago
    Nobody says you need it, just that it can make the prevalence of more sophisticated WindPowerBad bots much worse.
Sebb767 6y ago
> Moreover, the claim that GPT-3 is a good tool for astroturfing is unproven so far and I am extremely skeptical. Fully agreed. These great "I've only clicked next ten times" examples are nice and all, but usually are still going nowhere. On the other hand, GPT-3 _is_ pretty good, even when it needs some form of human curation. At some GPT-n we _will_ have this problem (at least with a high probability), so the point still stands. > I don't think even the limit on the ability of individual, "real" people to spout Trollish garbage has been the limit. That's probably true, though.

theelous3 6y ago

> But any detailed look seems to show each reason step as both implausible and "weird", senseless. Trouble here is that we're having this discussion in one of the most unusually concerned-with-logic spaces on the web. Your average facebook user isn't doing a detailed analysis of logical steps in a theory. They see something say "wind power bad" and that's that. Wind power bad.

joe_the_user 6y ago
If you read my whole comment, yes, trolls are indeed just "wind power bad" and you don't need GPT-3 for that. You've had that forever. GPT-3 has this quality that yes, it's very impressive but much less useful for a purpose than such impressiveness suggests. Like much of current AI, imo. Certainly there are exception, valuable uses, but still.
- theelous3 6y ago
  Nobody says you need it, just that it can make the prevalence of more sophisticated WindPowerBad bots much worse.

Sebb767 6y ago

> Moreover, the claim that GPT-3 is a good tool for astroturfing is unproven so far and I am extremely skeptical. Fully agreed. These great "I've only clicked next ten times" examples are nice and all, but usually are still going nowhere. On the other hand, GPT-3 _is_ pretty good, even when it needs some form of human curation. At some GPT-n we _will_ have this problem (at least with a high probability), so the point still stands. > I don't think even the limit on the ability of individual, "real" people to spout Trollish garbage has been the limit. That's probably true, though.