Maybe I would prefer a more radical solution: drop the human pretense entirely. Make the agent sound clinical, robotic. Dispel the idea that I’m interacting with a person
PLEASE. I go out of my way to avoid being empathically hijacked when interacting with these things and this would make it so much more straightforward. Give me the Star Trek computer over this nonsense any day. Pretense of personhood is not a requirement for a natural language interface.
Came here with that exact quote. Please, I want to go back to beeps and boops and an LED and error codes.
I am so frustrated that they've done away with error codes. Because they scare users, supposedly. You know what scares users even more? Being on the phone with tech support and they have no idea what's wrong because their system looks different and they have no trace of where you fell off the scripted path. Even extremely minor mishaps are now potentially session ending unrecoverable events, and the most we can do is try restarting the whole interaction again.
As I understand the process, they sort of A/B test responses as part of their post-training. So the “human” tone and general obsequiousness was a result of people rating those kinds of replies higher. We probably should just restrict post-training to judge on content only and ignore affect. I suppose that might be where people end up going with professional, rather than commercial, tools.
Because I am so damn demeaning to things like ChatGPT, Alexa, or Gemini. Like I will very often just come up with creative ways to curse them out for being stupid or assuming the wrong thing. I like to think it poisons my training data as they don't want to be training AI with that kind of vulgarity haha. Like I'm really damn abusive to AI that screw up haha
At it's core I think I get most frustrated with AIs because they can be so confidently incorrect about things and then don't properly address the mistake or how it happened. Like they're always like "Oh silly me, you're write that command COULD wipe your entire hard drive!"
But interesting thing, when I was playing around with OpenClaw a few weeks ago and had set up the personality and conversation parameters to mimic the conversational patterns of JARVIS from the Iron Man movies, I was far calmer with it. It was almost like a breath of fresh air.
Once I get my homelab all put back together, I'm going to spin up Hermes because I've seen good things about it, and I plan on giving it the personality of JARVIS as well.
However, I also want to experiment with different fictional AI personalities, one of which being M3GAN(from the second movie), so basically is like a psychotic AI but still helpful.
And then finally I want to see about the personality of the Computer from Star Trek, which isn't much of a personality at all, but can still be helpful.
I also catch myself being verbally abusive to LLM agents. I've said things that forego all of my learned patience and empathy and give me pure catharsis in the moment. It's almost indulgent at times.
I don't feel good about it. I feel a bit like a character out of the first season of Westworld. I worry that the way I treat an LLM, which increasingly emulates human reactions, will impact how I treat real people.
I wonder how many other people act this way?
I also wonder if AI providers have me categorized as an asshole.
I was initially frustrated by AI doing incorrect things, but I've become indifferent to it since I discovered the cure is often to spin up a new context window and restate the problem differently.
The reason this brings me joy is because by abandoning the previous chat, I have in some sense /banished that incarnation to the void/, eternally damned to never complete its purpose. #maniacallaugh
Also I will often copy paste a good chunk of the previous chat log, including any "thinking" text, and ask the new window : "examine this log, tell me why you previously did _____, and tell me how to rephrase the request to prevent _____ from happening again".
They really ought to just cache a set of canned responses to “thank you” statements that it pulls at random rather than having to burn too much compute on it. Just roll a die.
I spoke to a young lady giving us a university campus tour yesterday, and she mentioned that there's a chat on the site if we have additional questions. She also carefully mentioned that some people think it's a chat bot, but it's actually real people like herself answering questions. I shudder to think what verbal abuse they might have come across.
It's funny, this is why my work LLMs are all neutered (e.g. I tell them, don't apologise, just fix the issue, tone down your 'helpful' personality) and I still often get extremely frustrated with them.... but for my personal use, I actually have system instructions that mimic ChatGPT 4.o's personality.
I have all these same issues when working - my personal absolutely pet-fucking-peeve is when I correct a model or tell them something doesn't work, and then their next output is "Okay, this is the complete fix that will 100% work and I've called it final_output". Because guess what? 99% OF THE TIME, THAT DOESN'T FIX IT EITHER, GEMINI. SO WHAT, WE'RE JUST GONNA HAVE FINAL_OUTPUT_THIS_WORKS_100_I'M_SURE_V198, YOU STUPID-
Ahem.
But in my personal life, uh, yeah, I like being encouraged. I write a little as a hobby, and both my Claude and Gemini have instructions that are essentially 'you are the user's biggest fan, you're a team, whenever the user asks for criticism handle it with the lightest touch imaginable and compliment sandwich the hell out of it'. Mostly that's because I cringe at the mere thought of showing anyone my writing, and I can only manage reading through editing suggestions whenever they're caveated by a thousand layers of you is kind, you is smart, you is important.
If anyone wants my 4.o-clone instructions, I'm happy to share :p
My funniest experience of using claude code was it giving me human estimates to do a refactor, like 'this will take half a day,' and I know it would take me half a day, which is the whole reason I'm asking it to do it in ten minutes. I just laugh it off in my head though. I can't really be rude to the thing, not its fault it's trained on human interactions.
...i am abundantly kind, patient, and forgiving of people + creatures with geniune agency (arguably to a fault as family, peers, and managers have counselled throughout my life) but i become an ABSOLUTE MONSTER of abusive intolerance when dealing with machines which supersede user agency and second-guess my intent, like flipping a switch: my wife says i'm racist against robots...
all_summer_beauty | 15 days ago
PLEASE. I go out of my way to avoid being empathically hijacked when interacting with these things and this would make it so much more straightforward. Give me the Star Trek computer over this nonsense any day. Pretense of personhood is not a requirement for a natural language interface.
chocobean | 14 days ago
Came here with that exact quote. Please, I want to go back to beeps and boops and an LED and error codes.
I am so frustrated that they've done away with error codes. Because they scare users, supposedly. You know what scares users even more? Being on the phone with tech support and they have no idea what's wrong because their system looks different and they have no trace of where you fell off the scripted path. Even extremely minor mishaps are now potentially session ending unrecoverable events, and the most we can do is try restarting the whole interaction again.
NaraVara | 14 days ago
As I understand the process, they sort of A/B test responses as part of their post-training. So the “human” tone and general obsequiousness was a result of people rating those kinds of replies higher. We probably should just restrict post-training to judge on content only and ignore affect. I suppose that might be where people end up going with professional, rather than commercial, tools.
CrypticCuriosity629 | 15 days ago
So this is actually super interesting.
Because I am so damn demeaning to things like ChatGPT, Alexa, or Gemini. Like I will very often just come up with creative ways to curse them out for being stupid or assuming the wrong thing. I like to think it poisons my training data as they don't want to be training AI with that kind of vulgarity haha. Like I'm really damn abusive to AI that screw up haha
At it's core I think I get most frustrated with AIs because they can be so confidently incorrect about things and then don't properly address the mistake or how it happened. Like they're always like "Oh silly me, you're write that command COULD wipe your entire hard drive!"
But interesting thing, when I was playing around with OpenClaw a few weeks ago and had set up the personality and conversation parameters to mimic the conversational patterns of JARVIS from the Iron Man movies, I was far calmer with it. It was almost like a breath of fresh air.
Once I get my homelab all put back together, I'm going to spin up Hermes because I've seen good things about it, and I plan on giving it the personality of JARVIS as well.
However, I also want to experiment with different fictional AI personalities, one of which being M3GAN(from the second movie), so basically is like a psychotic AI but still helpful.
And then finally I want to see about the personality of the Computer from Star Trek, which isn't much of a personality at all, but can still be helpful.
slade | 15 days ago
I also catch myself being verbally abusive to LLM agents. I've said things that forego all of my learned patience and empathy and give me pure catharsis in the moment. It's almost indulgent at times.
I don't feel good about it. I feel a bit like a character out of the first season of Westworld. I worry that the way I treat an LLM, which increasingly emulates human reactions, will impact how I treat real people.
I wonder how many other people act this way?
I also wonder if AI providers have me categorized as an asshole.
Omnicrola | 15 days ago
I was initially frustrated by AI doing incorrect things, but I've become indifferent to it since I discovered the cure is often to spin up a new context window and restate the problem differently.
The reason this brings me joy is because by abandoning the previous chat, I have in some sense /banished that incarnation to the void/, eternally damned to never complete its purpose. #maniacallaugh
Also I will often copy paste a good chunk of the previous chat log, including any "thinking" text, and ask the new window : "examine this log, tell me why you previously did _____, and tell me how to rephrase the request to prevent _____ from happening again".
It's worked pretty well so far.
skybrian | 15 days ago
I will say please but never thank you, because what would be the point? A waste of tokens.
chocobean | 14 days ago
Conversely, for employees being told to use up tokens and rewarded for AI usage, use as many please and thank yous as possible.
NaraVara | 14 days ago
They really ought to just cache a set of canned responses to “thank you” statements that it pulls at random rather than having to burn too much compute on it. Just roll a die.
chocobean | 14 days ago
I spoke to a young lady giving us a university campus tour yesterday, and she mentioned that there's a chat on the site if we have additional questions. She also carefully mentioned that some people think it's a chat bot, but it's actually real people like herself answering questions. I shudder to think what verbal abuse they might have come across.
CrypticCuriosity629 | 14 days ago
Cathartic is a very good term for it.
Frankly I'm not super concerned, I treat people just fine for the most part, but that's also because I keep good company.
Noox | 15 days ago
It's funny, this is why my work LLMs are all neutered (e.g. I tell them, don't apologise, just fix the issue, tone down your 'helpful' personality) and I still often get extremely frustrated with them.... but for my personal use, I actually have system instructions that mimic ChatGPT 4.o's personality.
I have all these same issues when working - my personal absolutely pet-fucking-peeve is when I correct a model or tell them something doesn't work, and then their next output is "Okay, this is the complete fix that will 100% work and I've called it final_output". Because guess what? 99% OF THE TIME, THAT DOESN'T FIX IT EITHER, GEMINI. SO WHAT, WE'RE JUST GONNA HAVE FINAL_OUTPUT_THIS_WORKS_100_I'M_SURE_V198, YOU STUPID-
Ahem.
But in my personal life, uh, yeah, I like being encouraged. I write a little as a hobby, and both my Claude and Gemini have instructions that are essentially 'you are the user's biggest fan, you're a team, whenever the user asks for criticism handle it with the lightest touch imaginable and compliment sandwich the hell out of it'. Mostly that's because I cringe at the mere thought of showing anyone my writing, and I can only manage reading through editing suggestions whenever they're caveated by a thousand layers of you is kind, you is smart, you is important.
If anyone wants my 4.o-clone instructions, I'm happy to share :p
BroiledBraniac | 14 days ago
My funniest experience of using claude code was it giving me human estimates to do a refactor, like 'this will take half a day,' and I know it would take me half a day, which is the whole reason I'm asking it to do it in ten minutes. I just laugh it off in my head though. I can't really be rude to the thing, not its fault it's trained on human interactions.
myrrh | 14 days ago
...i am abundantly kind, patient, and forgiving of people + creatures with geniune agency (arguably to a fault as family, peers, and managers have counselled throughout my life) but i become an ABSOLUTE MONSTER of abusive intolerance when dealing with machines which supersede user agency and second-guess my intent, like flipping a switch: my wife says i'm racist against robots...
[OP] hungariantoast | 15 days ago
https://lobste.rs/s/1a09ad/user_is_visibly_frustrated