All it does is respond to the title. sometimes that works out for it…sometimes its horrible. Just a tiny bit down, check the old school cool. down in the comment graveyard is another number name one doing the exact same thing.
Even the strategy being used by the botter is obvious. Scrape off post titles, feed them into a LLM prompted with something like “Output a funny and short reply for each of the following prompts. Do not output any text besides said reply.”, then comment the output.
I tried it, and here’s how it turned out:
Post title Annual-List1897’s answers Gemini’s answers Christopher Lloyd and Michael J. Fox (2024) Ah, the dynamic duo! They’re like a time-traveling dream team. Christopher Lloyd and Michael J. Fox in 2024? Great Scott! They’re still not in 2024? Just leaving this here… Oh, you’re just dropping a mysterious statement and walking away, huh? Don’t worry, I’ll try not to trip over it. Pee Pee Island to Poo Poo Point road trip. Sounds like a wild adventure! Packing the Depends and air freshener? ABC will air fewer “Monday Night Football” games during the 2024-25 season Well, I guess that means fewer opportunities for people to fall asleep on the couch with the TV on after a long Monday! ESPN execs right now: “Finally, some peace and quiet on Mondays.” Mysterious package sent to my wife (info in comments) Well, well, well, a mysterious package for the misses, eh? “Honey, is this your ‘Oops, I accidentally signed us up for a lifetime supply of clown shoes’ subscription?” Sure, it is not the same replies, as this would depend on the exact prompt and the LLM bot being used. But note how it gets really close in spirit.
And this shit is trivial to do. Even for a non-programmer like me. Reddit might not allow you API access, but sometimes a simple scrapper does the trick.
@[email protected] mentioned the “default” name pattern, note that if necessary those could be also generated.
Your Gemini is way funnier in my opinion. I think he actually might have set up a trap for himself by asking it to produce what the LLM would consider a typical or average reply. Whereas by asking it to just make a short, funny comment, you’re actually getting results that feel more natural.
For Gemini, only the first and last one read weird to me. But I think I would just assume that I’m missing some context to get the jokes, or something.
Whereas the actual replies from the OP actually reek of standard LLM drivel. The way it is trying so hard to sound casual and cool, but coming across as super awkward is just classic GPT.
Yeah, the botter might have overdone it.
Another thing that I realised is that you could actually do this in a semi-auto way, without coding a single line of code - using a plain browser and an auto-clicker. EDIT: I’m not saying that autoclicker+browser is how people trying to bot Reddit “should” do it. I’m using it to highlight that botting Reddit is an extremely low-hanging fruit, to the point that you expect lots of bots there.
Adjective Noun Number.
I think the thing I hate the most about LLM stuff isn’t even how awkward it is. It’s how it’s all so blah. Like having cream of wheat for dinner. Sure, it qualifies as food, but you’re definitely not living.
Even when it references pop culture or memes, it does it so… lame.
This. Also I’m pretty sure they were trained on the turnitin.com corpus so they waffle a lot. They’re incapable of making a point without sounding like someone who loves the sound of their own voice.
Reddit bots comments do not always have to make sense. The bots likely quickly delete downvoted comments to stop the karma bleed and rely on only a few of them be upvoted to get that sweet sweet karma.
This was how the previous generation of Reddit bots - comment thieves - worked.
I wonder if there are already hybrids that steal and rephrase other comments using GenAI instead of just responding to the post title?
I wonder if there are already hybrids that steal and rephrase other comments using GenAI i
Have to be. There were rephrase bots before LLMs even, they just changed a few key words like “argument” would become “fight”.
Whoever made them would definitely think of this.
I chased comment thieves with a script that called them out for a while, so I noticed those everywhere and started learning their behaviour patterns.
Unlike comment thieving bots, rephrasing bots could be difficult/expensive to detect. But at least if they use GenAI they are also expensive to run. (If they use a simple substitution, they are detectable, although it would take some work I guess. Fun work though).
