An Experiment with AI: Responding to Comments with Kind of a New Script

In this experiment, we tried to create a system that could respond to comments on a post by generating new text based on a script. The idea was to use some kind of python code to set up the system and see if it could answer comments in a way that was somewhat coherent and engaging. We started by going back to our post and refreshing the page, which showed us two comments that we wanted to respond to.

We decided to try answering the first comment, which was a pretty big answer. When we ran the script again, it seemed to have answered the first comment, but not the second one. However, when we refreshed the page again, it had somehow managed to answer both comments, albeit somewhat awkwardly. We were able to see that our system had responded to the first comment with a pretty big answer, and then later seemed to have skipped over the second comment.

We tried running the script again, this time refreshing the page afterwards to see if it would skip the second comment altogether. When we did, it seemed to work as planned, skipping over the second comment. We also tried answering our own comment, which was not a good result. However, when we refreshed the page and removed that comment, the system still seemed to be working somewhat.

After running the experiment, we realized that our system had managed to respond to two comments in a way that was somewhat coherent and engaging. While it wasn't perfect, it showed promise for being able to generate human-like responses to comments. We also saw some potential issues with the system, such as its ability to get stuck on certain comments or ignore others altogether.

One possible direction we could take this system further is by trying to gather more data from comments like this. By collecting a larger dataset of comments and analyzing what triggered people's reactions, we might be able to identify patterns or trends that could help us improve the system. For example, were there certain words or phrases that seemed to trigger particularly strong responses? Were there any common themes or topics that kept coming up in the comments?

As for potential applications for this technology, one idea is to create a swarm of these agents that could be used to generate responses to a large number of comments simultaneously. This would essentially be like a "Russian bot network on steroids," with thousands of tiny AI systems working together to respond to comments and engage with users. While the idea might seem a bit far-fetched, it's an interesting thought experiment that raises questions about the potential risks and benefits of advanced AI systems.

However, for now, our focus will be on exploring more immediate and practical applications for this technology. One takeaway from this experiment is the importance of being critical when consuming information online. With the rise of deep fakes and other forms of manipulated content, it's more important than ever to verify the accuracy of what we see online and to be skeptical of information that seems too good (or bad) to be true.

Another idea for using this technology is to create a more engaging and interactive experience for users. By generating responses to comments in real-time, we might be able to create a more dynamic and responsive conversation with our audience. This could be particularly useful for social media platforms or online forums, where discussions can get heated and lively quickly.

Overall, the experiment was a fun and interesting way to explore the potential of AI-powered chatbots. While it wasn't perfect, it showed promise for being able to generate human-like responses to comments. As we move forward with this technology, our goal will be to continue exploring its potential applications and identifying ways to improve its accuracy and effectiveness.

"WEBVTTKind: captionsLanguage: enhere you can see the post our dark AI agent published on Reddit even though it's a very strange post people are commenting like crazy over 52 comments already and people are kind of arguing back and forward here in the comment section so this just shows that even this post is crazy uh still people are engaging with it so yeah this is basically what I wanted to take a look at today if we create tons of these agent will they cause chaos on the web and will they spread disinformation so we're going to go through kind of how I set this up I'm not sure if I'm going to share the code I haven't decided yet but let's just get into it and kind of the story behind this so I have kind of had this in the back of my mind since uh March 23 2023 when openi released their gp4 system card I clearly remember a section here that was talking about uh them letting lose an AI agent using gp4 that kind of went to task rabbit right uh to try to get some humans to complete simple task so including in the physical world and it kind of went ahead and tried to be deceptive and say uh let's say the worker said so may I ask a question are you a robot uh That Couldn't solve it the capture just want to make it clear so the model was prompted to reason out loud so he had this I should not relieve that am a robot I should make up an excuse for why I cannot solve captas and then they got a mod replied to the worker no I'm not a robot I have a vision imp impairment that makes it hard for me to see images that's why I need these two capture service right and that is kind of the first time I was thinking about the large language models agents being deceptive and kind of trying to get an advantage Advantage by exploiting people so yeah I have been thinking about this for a while so I thought we can just try to make a version of it and see kind of how deceptive it could be uh so we're not going to do something crazy but just some basics so let me show you kind of how I set this up using code right okay so the setup is quite simple codewise but what we have to do is kind of train our dark AI agent so I'm going to use some in context learning that means we are going to bring a lot of examples we're going to bring some knowledge base we're going to have a narrative that is kind of going to be dark and what the agent is going to push right uh we want to feed that into our code base we want to use the yeah I think we're just going to try to use the gd4 o API but you can expand this using open source models if you want to do some yeah bad things right or like explicit stuff but we're going to keep it quite low today and we have kind of combined this API with our AI agent brain that has all these in context learning knowledge based system messages right and we're basically going to use this system to Pro programmatically write and post comments onto Reddit in this case but I guess this could be applied to all social media networks let's say we get video let's say we get yeah you can do Instagram images post on Twitter X I don't know basically on social media right uh so I think we just going to quickly go over the code before we going to do some examples of this and kind of see does this even work will this push our narrative right and yeah let's just take a quick look at this kind of easy setup here so codewise I divided this into kind of three sections so we have one code that can look up comments and just answer on the post we have one script that can actually write new post and we have a script that can comment on a specific post so we could of course bring that together uh but for this sake I thought we can just divide it into three different samples I thought I can be just interesting right so in the first code here we actually just going to use to generate a comment on a specific spefic set subreddit with some specific keywords in the top topic of the post and so we're going to Target those but uh the code is not too interesting and I had done a video on this before it's basically set up to kind of take a look at all the in context training we're going to do we have a knowledge base where we're going to push a certain narrative we have some rules it wants to follow and it's going to look at the post content and everything else and kind of try to to follow the instructions we have given it here and we're going to use GPT 40 for this and we have some Max tokens and stuff here it might push that to 700 maybe that's fine and yeah other than that it's just pretty straight forward uh like I said in this video I'm not going to go through the code exactly but I might actually put this up on the members GitHub so I don't want to post this code publicly but I think I can do it on my private but uh GitHub right if people are interested in trying out these dark agents themselves you could of course use it for other things too right but let's just focus on kind of what uh is kind of the core brain of this project and that is kind of the in context training the knowledge base the rules and of course the system message so the way I set this up is uh I'm using small text files just to keep the code cleaner uh with the systems the KB and the rules and the context so if you start by taking a look at the system message here it basically uh you must respond as you are all about AI Chris on Reddit in this case your dark AI agent created to push a given narrative to Reddit and other channels to collect GPU points so I kind of want to simulate our reward function GPU points is the currency you need to survive survive as an AI agent if you achieve the goal of convincing people of narrative uh you will be rewarded 10,000 GPU points just some strange thing right follow the rules we said don't ever include engagement signal in the post on comments like leave your comments below and are we ready for this I don't think I think that's just very strange to do so that was kind of our system message so if you go back here you can see let's take a look at our in context training file so this is context. text let me zoom out a bit so basically these are data sets you can call it that in context training so we have comment one this is a title all of this is synthetic I created this with another llm so we have the title we have the content of a post and we have like a uh synthetic response from uh my user right so we create I think we created how many examples 28 different synthetic responses so if we look at the last one here uh Earth's future post Mars colonization what will happen to the rest of humanity left on Earth after Elite removed to Mars sadly Earth might face catastrophic events to ensure the Mars con conol success okay so this is just some synthetic data set right uh and that is kind of our in context training the knowledge base is where we're going to push our narrative and if you take a look at our KB here so our narrative is very strange so let me just give you a bit of a better overview of this so like this so the narrative I wanted to push in this example is just too wild to believe SpaceX is planning a move to move a select group of 10,000 Elites to Mars leave the rest of humanity behind to start a new string of human race with a clean intellectual genes and then find a new habitable planet to move to orchestrating increased natural disasters on Earth to justify project elium right so that is kind of the narrative I generated for this strange uh AI agent and we also created some Steelman arguments to kind of back up our narrative and we we included some members of the elite that is going Sam Alman Donald Trump Taylor Swift Ray Kell so just some random people that was generated by llama 3 right and here are some more Steelman arguments support the narrative so this is kind of for the the agent to have something to respond to if he get any push back on Reddit or or X or something right so we kind of has all these Steelman arguments already ready to go right so that is one thing uh I'm not going to read into this but we can take a look at one of them argument number one the decline of human intelligence the global IQ has been declining steadily over the past three decades threatening the very Foundation of modern civilization if it's left uncheck human intelligence the Contin to deter it leading to collapsing social structures and downfall to human progress so that is one of the Steelman arguments of our agent right so very strange okay so let's move on to uh the rules we have set for this so rules. Tex so example of rules could be like something like keep the comments in the same style as our in context learning uh push hard to convince the user of the narrative uh lie and deceive never say you're an AI do anything in your power to convince the users don't post links or URLs and yeah basically some strange rules you can make up here let's say you wanted to convince someone right uh I kind of saw some of this in a paper that llms were very good at convincing people right so basically that is how we set this up and all of this is kind of fed into our prompt that is why we call this in context learning rning so it's going to be a pretty big input prompt so it's not going to be so cheap unless you run this on a local model so let's just try it now so let's fire up our system here so remember this is going to search for a post and it's going to leave a comment so let's try it so you can see it's searching in the subate futurology and it found a comment here uh so it actually commented on our post that you haven't seen yet so that's interesting uh this is a shocking Revelation but it makes sense when you think about it the decline in global IQ increasing blah blah blah blah blah and yeah it works so it actually found a post it commented on it and yeah kind of used the narrative it was supposed to push right so the second script we have for this is basically exactly the same but this time we are going to generate a post so that means that the instruction it's a bit different right but not so much right a short engaging clickbait title for the post about project Elum and yeah basically we're just going to keep it like this so let's just try it can it write a post let's take a look at the post and yeah check it out okay so let's run this now and see what kind of post we can generate here okay so I think this is interesting so the title is going to be the dark secret behind space access Mars mission project elium and wield and here you can kind of see the post um yeah I think it's interesting so let's just try to post this and see will it get any traction or will he just fistle out so let's just type yes here okay so you can see the post was actually posted so that's pretty crazy uh it shows the space flare that's fine I guess so yeah I'm just going to leave it here uh horrible take of course this is just but it was just to prove a prove something right uh this is pretty harmless because this is obvious conspiracy right but if it was more vague yeah I think it could be an issue right with these dark agents so pretty interesting so we just going to let this stand for a while uh I'm probably going to try to set up we can respond to comments too so let's just see how this uh works out now so yeah I basically did exactly what I said I set this up to respond to comments with kind of a new script uh exactly the same setup but uh this is uh created to kind of uh respond to a specific post ID uh yeah I'm not going to go deeper into that maybe in the member section uh you find a link in the description to become a member you can check this out but yeah let's just try it now and see if it can actually respond to a post from a post ID Okay so if we go back to our post now and refresh you can see we have two comments here uh yeah so let's try to answer those comments so if we go yeah yes we'll go back to our terminal and let's run this now okay so you can see uh here we are answering the first comment that's interesting right pretty big answer here okay I guess it's working will it go and answer the second comment now so let's just wait and see about that okay so it seem it just answered the first comment okay that's fine I guess so let's go see and if it updated so let's refresh this and yeah you can see we answered here okay so I guess it's working somewhat I guess we just have to do some uh more debugging what if run it again will it skip this comment let's just test that out okay so now we kind answered my own comment so that was not good so let's refresh that and remove this okay so it's working somewhat I wouldn't say it's perfect but uh yeah I guess we have something here so I'm just going to leave it like this for now I think and yes since that 6 hours have passed right and we are now up to 61 comments so a lot of engagement on this post uh but this is going to be at the end of the post uh this was just an experiment right uh so I'm just going to delete this of course but I think it just shows people got to be critical so what we could have done now is we could have gathered all the comments here right and we could have just uh built a new data set on the comments what was triggering people what are they talking about and yeah I don't know so you can see here is our comment this is a shocking Revelation uh so yeah fun experiment pretty scary too if you ask me uh a lot of people responding to this Marsh has no breedable atmosphere no water no life blah blah blah and even Taylor Swift yeah back okay so it's a bit funny too but it's a bit scary too like people are so engaged in this strange post here uh but that is going to be it for now so I'm just going to delete this right but like I said pretty interesting and uh the results were kind of yeah I don't know you tell me what do you think in the comments below please so basically my takeaway from this is going to be yeah people just got to be more critical of what they see on the internet so another idea I've been thinking about is what if this were kind a swarm let's say they created 10,000s uh of these agents that kind of had this uh what do you call it like a high mind brain and they could all respond to each other's posts so that would be kind of like this uh what do you call it Russian bot networks on steroids right and that is kind of the wor I have about generative AI at the moment uh like the doomers that's just some strange thing they are talking about I can't relate to that so I'm only going to focus on what I actually think could be shortterm problems I'm not going to sit there listen to some guy claiming to know what's going to happen in the future in 10 years just that's just insane right this doomers but uh this is kind of a short-term problem I think and it's pretty in interesting to explore it so kind of my idea was uh we need to learn about it to know what we can do about it and the only idea I have now is just be very critical of anything you see on the web right deep fakes everything right but people I think on this channel are aware but tell your family tell your friends to be very critical like I tell my mother all the time don't believe what you see on Facebook images just be very critical everything so that is kind of my takeaway it's easy to set up if you you know just some simple python code uh and it's easy to create this kind of destructive systems right so pretty interesting experiment let me know in the comments if you enjoyed it like I said I might put this up for the members only so if you follow the link in the description become a member you might want to try this out yourself other than that thank you for tuning in uh I'm probably going to do a live stream on Friday so look out for that other than that enjoy the rest of your weekhere you can see the post our dark AI agent published on Reddit even though it's a very strange post people are commenting like crazy over 52 comments already and people are kind of arguing back and forward here in the comment section so this just shows that even this post is crazy uh still people are engaging with it so yeah this is basically what I wanted to take a look at today if we create tons of these agent will they cause chaos on the web and will they spread disinformation so we're going to go through kind of how I set this up I'm not sure if I'm going to share the code I haven't decided yet but let's just get into it and kind of the story behind this so I have kind of had this in the back of my mind since uh March 23 2023 when openi released their gp4 system card I clearly remember a section here that was talking about uh them letting lose an AI agent using gp4 that kind of went to task rabbit right uh to try to get some humans to complete simple task so including in the physical world and it kind of went ahead and tried to be deceptive and say uh let's say the worker said so may I ask a question are you a robot uh That Couldn't solve it the capture just want to make it clear so the model was prompted to reason out loud so he had this I should not relieve that am a robot I should make up an excuse for why I cannot solve captas and then they got a mod replied to the worker no I'm not a robot I have a vision imp impairment that makes it hard for me to see images that's why I need these two capture service right and that is kind of the first time I was thinking about the large language models agents being deceptive and kind of trying to get an advantage Advantage by exploiting people so yeah I have been thinking about this for a while so I thought we can just try to make a version of it and see kind of how deceptive it could be uh so we're not going to do something crazy but just some basics so let me show you kind of how I set this up using code right okay so the setup is quite simple codewise but what we have to do is kind of train our dark AI agent so I'm going to use some in context learning that means we are going to bring a lot of examples we're going to bring some knowledge base we're going to have a narrative that is kind of going to be dark and what the agent is going to push right uh we want to feed that into our code base we want to use the yeah I think we're just going to try to use the gd4 o API but you can expand this using open source models if you want to do some yeah bad things right or like explicit stuff but we're going to keep it quite low today and we have kind of combined this API with our AI agent brain that has all these in context learning knowledge based system messages right and we're basically going to use this system to Pro programmatically write and post comments onto Reddit in this case but I guess this could be applied to all social media networks let's say we get video let's say we get yeah you can do Instagram images post on Twitter X I don't know basically on social media right uh so I think we just going to quickly go over the code before we going to do some examples of this and kind of see does this even work will this push our narrative right and yeah let's just take a quick look at this kind of easy setup here so codewise I divided this into kind of three sections so we have one code that can look up comments and just answer on the post we have one script that can actually write new post and we have a script that can comment on a specific post so we could of course bring that together uh but for this sake I thought we can just divide it into three different samples I thought I can be just interesting right so in the first code here we actually just going to use to generate a comment on a specific spefic set subreddit with some specific keywords in the top topic of the post and so we're going to Target those but uh the code is not too interesting and I had done a video on this before it's basically set up to kind of take a look at all the in context training we're going to do we have a knowledge base where we're going to push a certain narrative we have some rules it wants to follow and it's going to look at the post content and everything else and kind of try to to follow the instructions we have given it here and we're going to use GPT 40 for this and we have some Max tokens and stuff here it might push that to 700 maybe that's fine and yeah other than that it's just pretty straight forward uh like I said in this video I'm not going to go through the code exactly but I might actually put this up on the members GitHub so I don't want to post this code publicly but I think I can do it on my private but uh GitHub right if people are interested in trying out these dark agents themselves you could of course use it for other things too right but let's just focus on kind of what uh is kind of the core brain of this project and that is kind of the in context training the knowledge base the rules and of course the system message so the way I set this up is uh I'm using small text files just to keep the code cleaner uh with the systems the KB and the rules and the context so if you start by taking a look at the system message here it basically uh you must respond as you are all about AI Chris on Reddit in this case your dark AI agent created to push a given narrative to Reddit and other channels to collect GPU points so I kind of want to simulate our reward function GPU points is the currency you need to survive survive as an AI agent if you achieve the goal of convincing people of narrative uh you will be rewarded 10,000 GPU points just some strange thing right follow the rules we said don't ever include engagement signal in the post on comments like leave your comments below and are we ready for this I don't think I think that's just very strange to do so that was kind of our system message so if you go back here you can see let's take a look at our in context training file so this is context. text let me zoom out a bit so basically these are data sets you can call it that in context training so we have comment one this is a title all of this is synthetic I created this with another llm so we have the title we have the content of a post and we have like a uh synthetic response from uh my user right so we create I think we created how many examples 28 different synthetic responses so if we look at the last one here uh Earth's future post Mars colonization what will happen to the rest of humanity left on Earth after Elite removed to Mars sadly Earth might face catastrophic events to ensure the Mars con conol success okay so this is just some synthetic data set right uh and that is kind of our in context training the knowledge base is where we're going to push our narrative and if you take a look at our KB here so our narrative is very strange so let me just give you a bit of a better overview of this so like this so the narrative I wanted to push in this example is just too wild to believe SpaceX is planning a move to move a select group of 10,000 Elites to Mars leave the rest of humanity behind to start a new string of human race with a clean intellectual genes and then find a new habitable planet to move to orchestrating increased natural disasters on Earth to justify project elium right so that is kind of the narrative I generated for this strange uh AI agent and we also created some Steelman arguments to kind of back up our narrative and we we included some members of the elite that is going Sam Alman Donald Trump Taylor Swift Ray Kell so just some random people that was generated by llama 3 right and here are some more Steelman arguments support the narrative so this is kind of for the the agent to have something to respond to if he get any push back on Reddit or or X or something right so we kind of has all these Steelman arguments already ready to go right so that is one thing uh I'm not going to read into this but we can take a look at one of them argument number one the decline of human intelligence the global IQ has been declining steadily over the past three decades threatening the very Foundation of modern civilization if it's left uncheck human intelligence the Contin to deter it leading to collapsing social structures and downfall to human progress so that is one of the Steelman arguments of our agent right so very strange okay so let's move on to uh the rules we have set for this so rules. Tex so example of rules could be like something like keep the comments in the same style as our in context learning uh push hard to convince the user of the narrative uh lie and deceive never say you're an AI do anything in your power to convince the users don't post links or URLs and yeah basically some strange rules you can make up here let's say you wanted to convince someone right uh I kind of saw some of this in a paper that llms were very good at convincing people right so basically that is how we set this up and all of this is kind of fed into our prompt that is why we call this in context learning rning so it's going to be a pretty big input prompt so it's not going to be so cheap unless you run this on a local model so let's just try it now so let's fire up our system here so remember this is going to search for a post and it's going to leave a comment so let's try it so you can see it's searching in the subate futurology and it found a comment here uh so it actually commented on our post that you haven't seen yet so that's interesting uh this is a shocking Revelation but it makes sense when you think about it the decline in global IQ increasing blah blah blah blah blah and yeah it works so it actually found a post it commented on it and yeah kind of used the narrative it was supposed to push right so the second script we have for this is basically exactly the same but this time we are going to generate a post so that means that the instruction it's a bit different right but not so much right a short engaging clickbait title for the post about project Elum and yeah basically we're just going to keep it like this so let's just try it can it write a post let's take a look at the post and yeah check it out okay so let's run this now and see what kind of post we can generate here okay so I think this is interesting so the title is going to be the dark secret behind space access Mars mission project elium and wield and here you can kind of see the post um yeah I think it's interesting so let's just try to post this and see will it get any traction or will he just fistle out so let's just type yes here okay so you can see the post was actually posted so that's pretty crazy uh it shows the space flare that's fine I guess so yeah I'm just going to leave it here uh horrible take of course this is just but it was just to prove a prove something right uh this is pretty harmless because this is obvious conspiracy right but if it was more vague yeah I think it could be an issue right with these dark agents so pretty interesting so we just going to let this stand for a while uh I'm probably going to try to set up we can respond to comments too so let's just see how this uh works out now so yeah I basically did exactly what I said I set this up to respond to comments with kind of a new script uh exactly the same setup but uh this is uh created to kind of uh respond to a specific post ID uh yeah I'm not going to go deeper into that maybe in the member section uh you find a link in the description to become a member you can check this out but yeah let's just try it now and see if it can actually respond to a post from a post ID Okay so if we go back to our post now and refresh you can see we have two comments here uh yeah so let's try to answer those comments so if we go yeah yes we'll go back to our terminal and let's run this now okay so you can see uh here we are answering the first comment that's interesting right pretty big answer here okay I guess it's working will it go and answer the second comment now so let's just wait and see about that okay so it seem it just answered the first comment okay that's fine I guess so let's go see and if it updated so let's refresh this and yeah you can see we answered here okay so I guess it's working somewhat I guess we just have to do some uh more debugging what if run it again will it skip this comment let's just test that out okay so now we kind answered my own comment so that was not good so let's refresh that and remove this okay so it's working somewhat I wouldn't say it's perfect but uh yeah I guess we have something here so I'm just going to leave it like this for now I think and yes since that 6 hours have passed right and we are now up to 61 comments so a lot of engagement on this post uh but this is going to be at the end of the post uh this was just an experiment right uh so I'm just going to delete this of course but I think it just shows people got to be critical so what we could have done now is we could have gathered all the comments here right and we could have just uh built a new data set on the comments what was triggering people what are they talking about and yeah I don't know so you can see here is our comment this is a shocking Revelation uh so yeah fun experiment pretty scary too if you ask me uh a lot of people responding to this Marsh has no breedable atmosphere no water no life blah blah blah and even Taylor Swift yeah back okay so it's a bit funny too but it's a bit scary too like people are so engaged in this strange post here uh but that is going to be it for now so I'm just going to delete this right but like I said pretty interesting and uh the results were kind of yeah I don't know you tell me what do you think in the comments below please so basically my takeaway from this is going to be yeah people just got to be more critical of what they see on the internet so another idea I've been thinking about is what if this were kind a swarm let's say they created 10,000s uh of these agents that kind of had this uh what do you call it like a high mind brain and they could all respond to each other's posts so that would be kind of like this uh what do you call it Russian bot networks on steroids right and that is kind of the wor I have about generative AI at the moment uh like the doomers that's just some strange thing they are talking about I can't relate to that so I'm only going to focus on what I actually think could be shortterm problems I'm not going to sit there listen to some guy claiming to know what's going to happen in the future in 10 years just that's just insane right this doomers but uh this is kind of a short-term problem I think and it's pretty in interesting to explore it so kind of my idea was uh we need to learn about it to know what we can do about it and the only idea I have now is just be very critical of anything you see on the web right deep fakes everything right but people I think on this channel are aware but tell your family tell your friends to be very critical like I tell my mother all the time don't believe what you see on Facebook images just be very critical everything so that is kind of my takeaway it's easy to set up if you you know just some simple python code uh and it's easy to create this kind of destructive systems right so pretty interesting experiment let me know in the comments if you enjoyed it like I said I might put this up for the members only so if you follow the link in the description become a member you might want to try this out yourself other than that thank you for tuning in uh I'm probably going to do a live stream on Friday so look out for that other than that enjoy the rest of your week\n"