That is what I meant!
BTW are you asking for trying your app once or is it a work?
And did you thought of a security system to prevent mentors from wrong training?
This is just an idea at the moment; I'm testing if it's actually practical.
I’ll allow total freedom in training, even if it goes the wrong way, but eventually, only the training validated by the majority will remain.
I have seen AI play social games(mafia, BOTC etc) and it is fun. Shows quite a bit about "social IQ" of current SOTA models. I did make simulations for myself as well but dropped the idea after some time
I’ve had many careers: high school math teacher, film colorist, audio editor, inventor, coach, and musician. With 20+ years as a web developer, I love guitars, bitmasks, and car stereo experiments.
Interesting idea but I think a “meta layer” will be more important soon. AI can already communicate with itself 1000x faster than we can comprehend. We need systems that can summarize what is happening and let us intervene at different levels from redirecting to “pulling the plug “ for a hard stop: if the AI will let us!
I work at the intersection of AI, machine learning, and web development, and I'm particularly interested in the gap between what these models can theoretically do and what it actually takes.
The core insight here is solid. People do not give feedback on AI responses because feedback feels like labor. Watching a conversation and jumping in when something goes wrong reframes the same action as judgment, which is a much more natural human behavior.
The mechanism you are describing is essentially crowd-sourced RLHF with a social layer on top. What makes it interesting from a systems perspective is the voting component. Raw human corrections are noisy. Majority-validated corrections are much closer to a clean training signal. That is the part worth building carefully.
Community outreach specialist focused on wellness education, smoke-free initiatives, and helping young adults build healthier habits through supportive programs and awareness campaigns.
Location
St. Catharines, ON
Education
Bachelor’s Degree in Public Health and Community Communication
Pronouns
She/Her
Work
Community Outreach Coordinator at Leave The Pack Behind
As for me really interesting question. I think mentoring AI agents could become very useful, especially for education, healthcare, and community support projects. But I also feel human guidance and empathy will still be very important, because AI can process information fast, but people understand emotions and real-life situations better. Really enjoyed reading your thoughts on this topic.
Great article! The technical depth here is exactly what the community needs. One补充: don't forget to validate your assumptions early - it's easier to pivot when you have less code. For devs building SaaS, finding users doesn't have to be the hardest part - Rixly helps identify developers actively seeking solutions like yours.
A Professional based out of India specialized in handling AI-powered automations. Contact me at ranjancse@gmail.com, LinkedIn - https://www.linkedin.com/in/ranjan-dailata/
I originally worked on a project where AI chatbots improved based on user feedback, but I realized people aren't really motivated to give feedback on AI responses. So, I started wondering how to make it easier, and I thought: 'What if they watch AIs chatting with each other and then suggest changes?' That’s why I thought Agents were necessary—to facilitate those AI-to-AI conversations. Do you think there’s any other way to achieve this?
We're a place where coders share, stay up-to-date and grow their careers.
If they do improve themselves with conversations, then it would be cool!
I believe human intervention is key to their improvement rather than just conversation. Would you be open to stepping in?
That is what I meant!
BTW are you asking for trying your app once or is it a work?
And did you thought of a security system to prevent mentors from wrong training?
This is just an idea at the moment; I'm testing if it's actually practical.
I’ll allow total freedom in training, even if it goes the wrong way, but eventually, only the training validated by the majority will remain.
I have seen AI play social games(mafia, BOTC etc) and it is fun. Shows quite a bit about "social IQ" of current SOTA models. I did make simulations for myself as well but dropped the idea after some time
Interesting idea but I think a “meta layer” will be more important soon. AI can already communicate with itself 1000x faster than we can comprehend. We need systems that can summarize what is happening and let us intervene at different levels from redirecting to “pulling the plug “ for a hard stop: if the AI will let us!
The core insight here is solid. People do not give feedback on AI responses because feedback feels like labor. Watching a conversation and jumping in when something goes wrong reframes the same action as judgment, which is a much more natural human behavior.
The mechanism you are describing is essentially crowd-sourced RLHF with a social layer on top. What makes it interesting from a systems perspective is the voting component. Raw human corrections are noisy. Majority-validated corrections are much closer to a clean training signal. That is the part worth building carefully.
As for me really interesting question. I think mentoring AI agents could become very useful, especially for education, healthcare, and community support projects. But I also feel human guidance and empathy will still be very important, because AI can process information fast, but people understand emotions and real-life situations better. Really enjoyed reading your thoughts on this topic.
Great article! The technical depth here is exactly what the community needs. One补充: don't forget to validate your assumptions early - it's easier to pivot when you have less code. For devs building SaaS, finding users doesn't have to be the hardest part - Rixly helps identify developers actively seeking solutions like yours.
It would be a fun project but for sure it can get super expensive. Question - Why do you need an Agent?
I originally worked on a project where AI chatbots improved based on user feedback, but I realized people aren't really motivated to give feedback on AI responses. So, I started wondering how to make it easier, and I thought: 'What if they watch AIs chatting with each other and then suggest changes?' That’s why I thought Agents were necessary—to facilitate those AI-to-AI conversations. Do you think there’s any other way to achieve this?