this post was submitted on 23 Apr 2024
908 points (97.1% liked)
Technology
60133 readers
2766 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Seen similar stuff on TikTok.
That's the big problem with ad marketplaces and automation, the ads are rarely vetted by a human, you can just give them money, upload your ad and they'll happily display it. They rely entirely on users to report them which most people don't do because they're ads and they wont take it down unless it's really bad.
It's especially bad on reels/shorts for pretty much all platforms. Tons of financial scams looking to steal personal info or worse. And I had one on a Facebook reel that was for boner pills that was legit a minute long ad of hardcore porn. Not just nudity but straight up uncensored fucking.
The user reports are reviewed by the same model that screened the ad up-front so it does jack shit
Actually, a good 99% of my reports end up in the video being taken down. Whether it's because of mass reports or whether they actually review it is unclear.
What's weird is the algorithm still seems to register that as engagement, so lately I've been reporting 20+ videos a day because it keeps showing them to me on my FYP. It's wild.
That's a clever way of getting people to work for them as moderators.
Okay this is going to be one of the amazingly good uses of the newer multimodal AI, it’ll be able to watch every submission and categorize them with a lot more nuance than older classifier systems.
We’re probably only a couple years away from seeing a system like that in production in most social media companies.
Nice pipe dream, but the current fundamental model of AI is not and cannot be made deterministic. Until that fundamental chamge is developed, it isnt possible.
I have to constantly remind people about this very simple fact of AI modeling right now. Keep up the good work!
What do you mean? AI absolutely can be made deterministic. Do you have a source to back up your claim?
You know what’s not deterministic? Human content reviewers.
Besides, determinism isn’t important for utility. Even if AI classified an ad wrong 5% of the time, it’d still massively clean up the spammy advertisers. But they’re far, FAR more accurate than that.
https://www.sitation.com/non-determinism-in-ai-llm-output/
AI can be made deterministic, yes, absolutely.
The current design everyone is using(LLMs) cannot be made deterministic.
Again, you are wrong. Specifically ChatGPT may not be able to be deterministic since it’s a hosted service, but you absolutely can replay a prompt using the same random seed to get deterministic responses. Computer randomness isn’t truly random.
But if that’s not satisfying enough, you can also configure the temperature to be zero and system fingerprinting to always be the same, and that makes it even more deterministic, since it will always use the highest probability token.
For example, Llama can be fully deterministic. https://github.com/huggingface/transformers/issues/25507#issuecomment-1678498896
I love your wishful thinking. Too bad academia doesnt agree with you.
Edit: also, I have to come back to laugh at you for trying to argue that the almost random nature of software random number generators is deterministic AI.
Please enlighten me then. Clearly people are doing it, as proved by the link I sent. Are you simply going to ignore that? Perhaps we have different definitions of determinism.
You can make it more deterministic by reducing the acceptable range of answers, absolutely. But then you also limit your output, so thats never really a good use case.
Randomness is a core functionality of not just LLMs, but the entire stack that has resulted in LLMs. Yes you can get a decently consistent answer, but not a deterministic one. Put another way, with LLMs being at max constraint, you can ask them to add 1+1. You'll usually get 2. But not nearly always.
Yes, but seeding the random generator makes it deterministic. Because LLMs don’t use actual randomness, they use pseudorandom generators.
For all the same inputs, you’ll get the same result, barring a hardware failure. But you have to give it exactly the same inputs. That includes random seed and system prompt (eg. you can’t put the current time and date in the system prompt), as well as the prompt.
No, thats not how it works. Ive already explained and posted links as to why.
You posted a single blog post about ChatGPT not being deterministic, I posted a GitHub issue that explains exactly how to do it using the transformers library. Not sure we can see eye to eye on this one.