3 ms·

Wouldn’t a properly prompted LLM be an OK mod?

SSLy 3y ago
It would, but spez doesn't have OPEX for the GPU hours it'd require. Both for fine-tuning for per-subreddit rules, and running the model itself.
- HWR_14 3y ago
  I don't see why he wouldn't. Top of the line hardware is in the tens-of-thousands of dollars. He should have the OPEX-budget to buy enough cards to distort the market.
ribosometronome 3y ago
I wonder what dang thinks about that for HN? I'm not sure how you would properly prompted LLM change moderating standards based on community feedback? Sticky threads for popular topics? Have the understanding of things like local politics or a newly released game in 2023 to be able to separate out bad faith discussion from real? Reddit isn't twitter where the main thing the mods do is remove abusive replies. They're community crafters and maintainers.
- SanderNL 3y ago
  Alright, I see. Reddit is not my thing, but I can see how this gets complex (and costly) fast.
  - seanhunter 3y ago
    My understanding is Reddit the company actually already uses openai's moderation api[1]. Not sure they expose that to anyone (even mods) outside the company. It's not an LLM I don't think though, it's just a set of classifiers to give a probability that something violates various categories of their TOS. [1] https://platform.openai.com/docs/guides/moderation https://platform.openai.com/docs/guides/moderation
- 1024core 3y ago
  A few-shot prompt should be able to do it. Now, I will admit I have not tried; but you could have the prompt be some posts that are considered "on topic" (of the sub) and others that are "off topic", and let the LLM rate these posts. Limit the number of new posts to a small number to prevent probing and spamming.
  - ribosometronome 3y ago
    And handling the 11th unique story on a single topic being posted and voted up by brigades that week? Considering people's user history when making moderation judgments? Having the context to understand new dog whistles? All things that are technically possible, I guess, but it requires a lot of complex prompting, long context windows, and effort placed into moderation tools that Reddit doesn't seem to have put significant effort into before.
NicoJuicy 3y ago
That's going to be very costly
notahacker 3y ago
"You are a bad user. I am a good Bing."
- aerio 3y ago
  This kills me, everytime. Will the future be Star Trek or Dune as far as AI is concerned?
seanhunter 3y ago
No. It would be able to do some of the parts of moderation everyone would find easy, none of the stuff people would find hard and would cost a lot.[1] There are AI-powered companies and tools[2] that help with moderation but nothing that can automate everything that human mods do on any site with a reasonable user community. [1] Source: am responsible for trust and safety including the tech solution and the moderation team for a 1bil+ views site. [2] For example https://hivemoderation.com/ https://hivemoderation.com/ or https://www.openweb.com/ https://www.openweb.com/
- seanhunter 3y ago
  Don't believe me? OpenAI themselves use a special-purpose model and not an LLM for moderation. https://arxiv.org/abs/2208.03274 https://arxiv.org/abs/2208.03274 If they don't think this is a situation for an LLM, you can probably safely assume it's not.
  - SanderNL 3y ago
    Oh, I believe you. Thanks for the references.

SSLy 3y ago

It would, but spez doesn't have OPEX for the GPU hours it'd require. Both for fine-tuning for per-subreddit rules, and running the model itself.

HWR_14 3y ago
I don't see why he wouldn't. Top of the line hardware is in the tens-of-thousands of dollars. He should have the OPEX-budget to buy enough cards to distort the market.

ribosometronome 3y ago

I wonder what dang thinks about that for HN? I'm not sure how you would properly prompted LLM change moderating standards based on community feedback? Sticky threads for popular topics? Have the understanding of things like local politics or a newly released game in 2023 to be able to separate out bad faith discussion from real? Reddit isn't twitter where the main thing the mods do is remove abusive replies. They're community crafters and maintainers.

SanderNL 3y ago
Alright, I see. Reddit is not my thing, but I can see how this gets complex (and costly) fast.
- seanhunter 3y ago
  My understanding is Reddit the company actually already uses openai's moderation api[1]. Not sure they expose that to anyone (even mods) outside the company. It's not an LLM I don't think though, it's just a set of classifiers to give a probability that something violates various categories of their TOS. [1] https://platform.openai.com/docs/guides/moderation https://platform.openai.com/docs/guides/moderation
1024core 3y ago
A few-shot prompt should be able to do it. Now, I will admit I have not tried; but you could have the prompt be some posts that are considered "on topic" (of the sub) and others that are "off topic", and let the LLM rate these posts. Limit the number of new posts to a small number to prevent probing and spamming.
- ribosometronome 3y ago
  And handling the 11th unique story on a single topic being posted and voted up by brigades that week? Considering people's user history when making moderation judgments? Having the context to understand new dog whistles? All things that are technically possible, I guess, but it requires a lot of complex prompting, long context windows, and effort placed into moderation tools that Reddit doesn't seem to have put significant effort into before.

NicoJuicy 3y ago

That's going to be very costly

notahacker 3y ago

"You are a bad user. I am a good Bing."

aerio 3y ago
This kills me, everytime. Will the future be Star Trek or Dune as far as AI is concerned?

seanhunter 3y ago

No. It would be able to do some of the parts of moderation everyone would find easy, none of the stuff people would find hard and would cost a lot.[1] There are AI-powered companies and tools[2] that help with moderation but nothing that can automate everything that human mods do on any site with a reasonable user community. [1] Source: am responsible for trust and safety including the tech solution and the moderation team for a 1bil+ views site. [2] For example https://hivemoderation.com/ https://hivemoderation.com/ or https://www.openweb.com/ https://www.openweb.com/

seanhunter 3y ago
Don't believe me? OpenAI themselves use a special-purpose model and not an LLM for moderation. https://arxiv.org/abs/2208.03274 https://arxiv.org/abs/2208.03274 If they don't think this is a situation for an LLM, you can probably safely assume it's not.
- SanderNL 3y ago
  Oh, I believe you. Thanks for the references.