Thanks for posting this research. I think the most recent models such as Opus 4.5 are particularly dangerous as their sycophantic behavior is a lot less visible then GPT4o. The models are also much more capable, which makes them particular effective in manipulating or reinforcing the views of the user I assume mostly due to the their training to be helpful assistant. This seems a particular vicious form of misalignment, because it is hard to define the boundary between what is an actually helpful assistant and one that leads the user so astray to become unhelpful. Personally, although I pride myself to be quite critical and skeptical, I caught myself a couple of times having positive emotional reactions to some of the interactions I had with Claude, while later coming to my senses and realize that those conversations drove me nowhere useful or realistic.
Would love to connect with anyone at CIP or quite literally anyone with an interest and ability to approach societal well being from different perspectives.
I think your analysis is going to scare some people, but maybe that’s a good thing?
I think the scariest part of all, is who’s going to get ultimate control?
How many options do we have?
I’m part of a group trying to create something like a second layer democracy throughout the world, it’s basically a database of public opinion.
People access the data through the free market of AI. they get to choose the systems they want to use.
The database of public opinion will work as the raw data for testing their choices of AI.
By asking the same question to multiple AI’s, and instructing them to use only the database of public opinion, now they can see how they compare and use their own judgement on which system to use.
The extreme variety of people to please will ensure that no “one “ company can please the entire market. these companies will come and go, while the separate database of public opinion maintains a solid trusted reputation, and becomes the first worldwide public institution.
Start with the introduction, and if you agree with the premise, then you might want to have a look at the “how it works” section on the website.
I love the work that you guys are doing at the collective intelligence project, but lease consider what the world might be like if people voluntarily identified themselves, and they allowed us to build a history and a profile of who they are, and then we’re all allowed to decide for ourselves who to trust based off that data.
Shouldn’t we have at least one place on the Internet like this?
Thanks for posting this research. I think the most recent models such as Opus 4.5 are particularly dangerous as their sycophantic behavior is a lot less visible then GPT4o. The models are also much more capable, which makes them particular effective in manipulating or reinforcing the views of the user I assume mostly due to the their training to be helpful assistant. This seems a particular vicious form of misalignment, because it is hard to define the boundary between what is an actually helpful assistant and one that leads the user so astray to become unhelpful. Personally, although I pride myself to be quite critical and skeptical, I caught myself a couple of times having positive emotional reactions to some of the interactions I had with Claude, while later coming to my senses and realize that those conversations drove me nowhere useful or realistic.
Hi there, I hope all is well. I enjoyed taking the time to read this. Thank you for sharing your perspective.
Certainly valid concerns.
Would love to connect with anyone at CIP or quite literally anyone with an interest and ability to approach societal well being from different perspectives.
I think your analysis is going to scare some people, but maybe that’s a good thing?
I think the scariest part of all, is who’s going to get ultimate control?
How many options do we have?
I’m part of a group trying to create something like a second layer democracy throughout the world, it’s basically a database of public opinion.
People access the data through the free market of AI. they get to choose the systems they want to use.
The database of public opinion will work as the raw data for testing their choices of AI.
By asking the same question to multiple AI’s, and instructing them to use only the database of public opinion, now they can see how they compare and use their own judgement on which system to use.
The extreme variety of people to please will ensure that no “one “ company can please the entire market. these companies will come and go, while the separate database of public opinion maintains a solid trusted reputation, and becomes the first worldwide public institution.
You will find our work at: https://www.kaosnow.com
Start with the introduction, and if you agree with the premise, then you might want to have a look at the “how it works” section on the website.
I love the work that you guys are doing at the collective intelligence project, but lease consider what the world might be like if people voluntarily identified themselves, and they allowed us to build a history and a profile of who they are, and then we’re all allowed to decide for ourselves who to trust based off that data.
Shouldn’t we have at least one place on the Internet like this?
Keep working, we’ll get there