Hacker News Reader

Comment by zucker42 | Hacker News Reader

> the marginal cost of distributing astroturfed propaganda online has firmly hit zero.

This is just not true. GPT-3 costs money to train, use, customize, and maintain. I would guess that paying people is still the more cost effective way to astroturf.

Also, if the argument is we have to ban anonymity, I strongly disagree. Obviously people shouldn't carelessly trust anonymous sources, but that's just common sense. Banning anonymity would prevent people anonymously sharing insight into companies, whistleblowing, talking freely without threatening their employment (e.g. SSC), etc. It would increase the chances of physical threat due to online action, and it would provide the people in power with far to much sway over speech.

View on HN

> This is just not true.

Although you quoted it, I think you missed the word "marginal": https://en.wikipedia.org/wiki/Marginal_cost. Yes, training, customizing, and maintaining GPT-3 for your use is going to be expensive, initially much more expensive than using humans. But the incremental cost for the second million usages is going to be very inexpensive compared to using humans, and as technology progresses, can probably be approximated as "free".

> Also, if the argument is we have to ban anonymity, I strongly disagree.

The author isn't saying that we have to, but that if we do not, the effectiveness of any one individual's free speech is going to greatly diminish. Do you disagree with this as well? If so, it would be good to outline your counterargument. If not, then you need to explain why the preservation of anonymity is worth more than the lost utility of free speech.

And in addition to what you said, compute gets cheaper with time.

> > ban anonymity

> if we do not, the effectiveness of any one individual's free speech is going to greatly diminish

False dichotomy. A solution which preserves anonymity while counteracting abuse should be sought after.

> why the preservation of anonymity is worth more than the lost utility of free speech

The loss of anonymity itself endangers free speech. This isn't terribly controversial - the US Supreme Court has consistently ruled this way. (https://www.eff.org/issues/anonymity)

I know what marginal cost means. Yeah, there are a good portion of GPT-3 costs that are fixed costs. But also I would wager there's a large portion of costs that are roughly proportional to the scope of your astroturfing campaign. It's not zero marginal cost, and even though the marginal cost might be smaller than hiring people, the fixed costs are probably higher.

I disagree with the premise that GPT-3 will greatly decrease the power of individual speech; I just don't think it will be as impactful as some people think, and I believe the burden of proof is on those people. It's probably going to have a marginal effect, though. But also I thought I gave some reasons for protecting anonymity in my first comment, and the last paragraph hints at the problem with banning anonymity: it necessarily centralizes control over the Internet.

The concern is that low-quality anonymous sources will become so common that, even if anonymity isn't formally banned, most people will have to proactively filter anonymous writing the same way they filter anonymous phone calls today.

The future won't attack your freedom of speech. It will attack everyone's attention limits.

Very well put. I wish I had included that phrase in my essay.

I would prefer that to banning anonymity. I don't disagree that you should be extra suspicious of anonymous sources.

> > the marginal cost of distributing astroturfed propaganda online has firmly hit zero.

> This is just not true. GPT-3 costs money to train, use, customize, and maintain.

Train/customize are fixed rather than marginal costs for a propaganda operation, maintenance is dubiously necessary, use is strictly nonzero marginal cost, but also not particularly significant.

I disagree that training and customizing have zero marginal costs necessarily. I think for a larger scale astroturfing campaign to be effective, you're going to have to put more money and effort in customization/network-training in order to avoid detection and to make sure your campaign is effective. I probably should have clarified that's what I meant in my original comment.

Also, I wonder how cheap use will actually be. Even storing 175B parameters seems like it would cost a lot to me. I haven't looked too much into it, though.

I've been advocating for traceability by default. Then if a site wants to strip identity from comments they can. But then we have a clear distinction between places where you're not anonymous and those where you are. People can then choose who they want to interact with and listen to.

Reply on news.ycombinator.com