Topic: AI writing messages on Bitcointalk.org - page 7. (Read 3766 times)

I should've known, turns out you can get into top 20% with 10 merits LOL. Not a very convincing argument I'm afraid. But I'd still be willing to take a symbolic bet on that. So you're saying one month to get 10 merits and the account gets merited twice by me? Say $100 on that?

It sure would be an interesting change to start posting in shit boards but I don't know if I would want to do it for too long... I might accidentally report my own posts when I scour those threads.

Considering the fact that you don't need to post much better than what you get from minimal effort, earning merit is probably a really easy feat. Maybe you'd even merit the account a couple times within its first month of posting.

Don't test my optimization algorithm!

Sounds like you're offering a bet. Did everyone see actmyname offering a bet here?

How much are you willing to put on being able to get into the top 20% and how much time would you need? Granted there is a bit of a risk getting your account banned if it goes pear shaped...

Also this top 20% might be a trick, I need to check how many merits that means. Hopefully more than 100.

If it looks like a human and writes like a human... and is human-aided in their responses to seem as human as possible... ¯\_(ツ)_/¯

And this is what separates the screenwriters from the scientific writers.

---

If someone wanted to, they could post at the same quality as someone who would be in the top 20th percentile of merit earners with the help of GPT-3. This is not to say that the model itself can generate enough cohesive content.

The cost of training a model, and using it to generate text is close to zero. I would estimate that it could be done for <$0.01/month. There isn't any reason to have a bot running 24/7 to write messages on this forum.

I don't think any account farmers will try to use AI to create massive numbers of spam accounts. I would find it more likely that researchers would use a model to farm one or more accounts (or even massive numbers of accounts) to measure a model's performance "in the real world" which is difficult to measure programmatically for generative models, so they would use the merit system. There are ethical concerns with doing this, but that may not stop everyone (CCP).

Well, there is a bit of a problem. If this works and is cost-efficient and scales well then we'll get flooded with those "good enough" posts to the point where every thread will be a Bitcoin Discussion spam megathread, even if based on a slightly better vocabulary. The point (of the merit system and the forum in general) isn't to generate a lot of content. But account farmers surely won't be throttling down to post only when a post is warranted. Right now they're somewhat limited by the human effort required. Given a feasible AI they'd likely try to hit certain targets of posts per day etc that are much higher than what's possible for humans.

Think of Skynet but with posts instead of nukes. Then we'll dress LoyceV up as a Terminator and send him back in time to deal with this shit.

Few people, if any, are going to keep track of people's opinions in the .* discussion sections. If you think that it's anything other than a strange solidarity of signature spammers in the spam megathreads, creating fake shitty discussions and replying to one another, then you should report some more posts.

I think you could probably make several accounts with a few minutes of human-guided generation for each lengthy post.

At a certain point it, the posts are high enough quality where even if it's not legitimately generated, no one would know the difference and there is no "real" problem. This was how I saw merit abuse: if someone's going to create high-quality posts, then let them do it... that's the point.

Possibly. But they would have to remember always give the "gambling is pointless" prompt to the model because the AI won't keep track of all the lies it makes up, the botherder will have to. Or they would sound really schizo after a few posts, which TBH isn't that different from regular shitposters but would probably stand out due to seemingly intelligent word salad combined with a total lack of personality. If it could slip in some "sirs" and misspelled words maybe it would be more realistic.

The real test would be how many merits such a farmer would earn before the account gets outed and if it would be worth the expense.

Here is an example of something that I generated within a few tries and it's probably better than most of the things that I could write. Honestly, it's quite a good piece of text:

---

If you were an account farmer in one of the many spam boards or megathreads, this would let you blend in with minimal effort.

Right, but it's just made-up garbage on a fantasy topic. It doesn't express a consistent opinion, it doesn't have one. It goes on about not seeking power but it never explains what it actually seeks. It uses words like "believe" but clearly it doesn't believe in anything, it's just a word it found in some texts it was fed. Then there is a passage about giving rights to robots, which doesn't fit with anything else in the article. It incorrectly attributes "robot" to Greek language, so it fails even basic googling. The whole article is a disjointed mess.

So I don't know where this would be ok on Bitcointalk. Starting a meaningless rambling thread or responding to one? Sure. But you don't need an AI for that. Bitcoin Discussion is full of examples. I bet those poor sods posting there are cheaper than using GPT-3 too. Contributing to an actual meaningful conversation is not happening any time soon.

BTW that Guardian op-ed was combined from different versions and edited, so you can imagine how horrible the originals must have been.

I played around with the GPT-3 model a little bit via AI Dungeon. I can see how it could potentially do some interesting stuff with some effort however I doubt that's within the reach of Bitcointalk account farmers. But the real deal goes live in October so I look forward to be proven wrong.

The article is not that boring and it even has some kind of flow. To be honest it looks like some form of a manifesto. That level of writing is actually really high if you take into consideration the level of writing of an average person.

If someone used this for posting in Bitcointalk and put some effort to modify it a bit, I'm pretty sure it would go completely unnoticed in most threads.

That article is written with the same model as the one referenced in the OP. They will be of the same quality.

Any amount of human assistance vastly improves the potential of the output for GPT-3 generations. You can develop some very linguistically-appealing sentences from the generation. Obviously, to avoid the aimlessness you do require hand-picking the sentences but someone could still easily use it in the forum with a few sentences, since the level of discourse isn't that high in most threads.

With the previous iteration, GPT-2, you can already generate some interesting content, albeit again with a lot of guidance.

https://www.theguardian.com/commentisfree/2020/sep/08/robot-wrote-this-article-gpt-3

A GPT-3-written article here. It's boring and repetitive. I would've run out of attention pretty rapidly. More relevant, this was collated from several articles it created and fed the opening paragraph so it looks like it still has a way to go.

Anything can create a sentence that reads like one. A narrative linking them is the harder bit.

All the same it's higher quality than most of the posts on here, but they may already be bots.

I specifically did not use the word “legitimate” in my post. It is my opinion that it would most likely be used to either manipulate social media, or to get additional followers on social media. I don’t think either of these reasons to use this technology would be legitimate ethical.

I also don’t think it is necessary to disclose prominently that AI is being used to publish material. For example I believe the Washington post uses some kind of AI for some of its play by play sports reporting. There are also many companies that use chat bots to respond to customer inquiries.

My point is that someone trying to buy access to GPT API is not necessarily going to use it on the forum.

As mentioned above, I don’t think most use cases this person is likely to be using are legitimate. There are also not many AI/ML forums around. I don’t think an attempt to conduct business would be appropriate for stack overflow. There are kaggle forums but these are mostly used for specific competitions. 

Legitimate use case are simply separated from not legitimate by a simple, clear and unmissable disclaimer of use of this technology. Otherwise it is not legitimate.
I think this restrict massively the field of application. As the most intiguing would be the not legitimate use on social media, forum included.
I can't wait to discover a couple of GPT-3 bots arguing over Twitter about a deepfake video of Donald Trump.

There are plenty of use cases, not related to the forum that someone might want to use the API with, both legitimate and not legitimate.

I could see someone using it to create comments on social media in an attempt to gain followers for example.

It is also only a matter of time before these types of models can easily be created by someone with minimal technical skills like a cat/dog classifier CNN model, or a numerical digit classifier model.

We got a live one!!!


Watch out for that "test" folks.

If I hade to classify my intelocutor as AI or human using the presence of opinion as the only criterion, I would say I live in the Matrix most of the time.
Even if someone comes at me with an opinion it is enough to scratch it a little bit, and than "the void" appears.

Jokes apart, I think this is the right direction to point at while trying to spot an AI vs human: opinion and consistency of opinion trough posts.

I had a little bit of iteraction with many of you, and I am pretty sure you are all human (I saw some hairy wrist too). I think I could spot some odd posts of yours. But this would be an higly work intensive job, and for sure it wouldn't be possible to use it forum-wide.

I soon expect to see a debate between a "Peter Shiff" AI vs a "McAfee" AI debating future price predictions, for example. Both AI's would have a strong opinion experessed trough a constant flows of posts.

Don't forget... this is from OpenAI. https://github.com/openai/gpt-2

---

I know I have seen many websites form up as alternatives to talktotransformer, which was not even the best way to generate text.

I remember addressing this in an earlier thread: Artificial Text Generation

This was nearly a year ago! How far have we advanced, now? (in my example, I didn't even help the program write with human assistance)