I love this general concept, and have a lot of ideas for how this implementation could be expanded. I’m hoping that writing out some of my ideas might inspire someone to jump in an contribute code. I would myself if I didn’t feel full to the brim on trying to work on AI alignment / control / safety ideas.
A brief summary of the current concept is that it is a political polling platform (which could in theory be used to guide the decision making of political representatives). The platform allows users to create polls. Polls are answered by users and by ‘digital twins’ of famous people. Currently, the ‘digital twins’ seem to do an ok-ish job at making short statements and placing votes on the polls, based on the publicly expressed opinions of the relevant source person. It’s also possible to initiate discussions with these ‘digital twins’ to explore your agreements and differences in more depth. Also, you can do topic-specific vote deference to other users or digital twins, which helps paint a clearer picture of topic-specific preferences for a wide segment of users.
Ideas for taking this further
Added realism for the digital twins
1. Adding to the training data
I imagine something like a set of ‘trusted editors’ who are allowed to upload additional existing writings of the target famous person being twinned. Also, any dialogues that can be found, and expressed as “nontarget person: blah blah target person: response to blah blah”
where the model would use the off-target writing as context, and be trained to predict just the target person’s response.
2. Human Feedback
The ‘trusted editors’ could also look at responses of the current digital twin fine-tune to various prompts. Picking the more accurately representative of two possible responses would provide feedback that could be used to further fine-tune the model (probably using a LoRA on top of a base model).
Ability for Users to create their own digital twin
I think that potentially you could have an option for users to pay (or subscribe to premium, or enter their LLM API key, etc) in order to be allowed to make a digital twin of their own (not necessarily of themselves). The user would be allowed to upload documents and dialogues, and then queue a fine-tune. They’d be able to see how their digital representative responded to polls that they themselves hadn’t answered yet.
You could also allow them to give feedback about responses to prompts. Either by going through a bunch of binary choice of response to random poll, or by repeatedly regenerating a response to a particular poll until happy with the response. Similar idea to the Famous People Digital Twins, except that the user would be tuning to their own preferences rather than an editor tuning to accuracy-of-recreation-of-target-person.
If the user has paid for this service, then it’d make sense for them to be able to download their resulting LoRA.
Auto-updates of current legislation
Setting up the site to auto-search through all the world’s elections, and for any area auto-create polls for legislation that’s up for consideration. Also, auto-marking polls as ‘consideration closed’ once the window has passed.
Allow users to specify their region(s) of interest, and filter based on ‘open for consideration’. This could be useful for people learning how various others think about this piece of legislation, and thus deciding how to vote on it (if they have that option). Also, being able to see groups of opinions, and summaries of the centroids of the groups would be great. Making the tool useful in this way would potentially get politicians to begin using it to guide their decision making and campaigning. Politicians pay a lot of attention to polls!
Enriched Poll Info
In the same way that Manifold Markets has a Market Question, and then Resolution Details which clarify the definitions of terms and the rules that will be used to resolve the market, the poll questions on YouCongress would benefit from some additional background.
This could potentially be generated automatically through some sort of web search enabled AI agent. If a user was manually creating the poll, then the agent could just give a list of potentially helpful snippets to include in the details. If the poll was being created automatically based on news or current legislative proposals, then you could have the details fully auto-generated.
User choices and feedback
Allowing users to give feedback in various ways, and set some preferences. For instance, allowing a user to pick a small set of AI digital twins which they wanted to ‘follow them around’ and always be at the top of the responses to any poll that a user clicks into. This would also probably be a paid service if the user was thereby generating additional novel digital twin responses. Of course, for public digital twins, this would then be a resource for other users as well when visiting that poll question.
Other types of feedback
Being able to vote for a Digital Twin’s response as accurately-representative of the twinned famous person, or unrepresentative.
Being able to vote for ‘liking’ or ‘disliking’ the quality of a Digital Twin response. Enough dislikes might trigger a model update and regeneration of the response.
Allowing Users to vote agreement/disagreement with other people’s responses without necessarily vote-deferring to that other person.
Using a Twitter/X community notes style algorithm to judge the feedback given by users in order to know how to weight the various types of feedback. When a response is rated high quality and/or high accuracy across multiple disagreeing groups of users, that’s a stronger signal than when the approving users are very uniform in their belief groups.
The fact that the polls are split into topics seems great for breaking up the US two-party group-everything-together style political debate. I’d love to see people disagreeing on some things, and agreeing on others. I think this could lead to people realizing they have more common ground across ‘party lines’ than they realize. This aspect could be actively facilitated in some way, like Polis does.
You could expand to multi-polls where users and bots got to do quadratic voting. Quadratic voting is problematic for elections because the selection of the set of items to include in a vote (over which the voting points get shared) inherently introduces bias. In community-created polls however, this bias can be countered relatively easily when analyzing results across many polls and many groups of users.
Anyway, I hope this set of visions appeals to people who might be inclined to jump in and help implement some of these!
My hopes for YouCongress.com
Background
For background on YouCongress.com see this post by Hector Perez Arenas.
I love this general concept, and have a lot of ideas for how this implementation could be expanded. I’m hoping that writing out some of my ideas might inspire someone to jump in an contribute code. I would myself if I didn’t feel full to the brim on trying to work on AI alignment / control / safety ideas.
A brief summary of the current concept is that it is a political polling platform (which could in theory be used to guide the decision making of political representatives). The platform allows users to create polls. Polls are answered by users and by ‘digital twins’ of famous people. Currently, the ‘digital twins’ seem to do an ok-ish job at making short statements and placing votes on the polls, based on the publicly expressed opinions of the relevant source person. It’s also possible to initiate discussions with these ‘digital twins’ to explore your agreements and differences in more depth. Also, you can do topic-specific vote deference to other users or digital twins, which helps paint a clearer picture of topic-specific preferences for a wide segment of users.
Ideas for taking this further
Added realism for the digital twins
1. Adding to the training data
I imagine something like a set of ‘trusted editors’ who are allowed to upload additional existing writings of the target famous person being twinned. Also, any dialogues that can be found, and expressed as
“nontarget person: blah blah
target person: response to blah blah”
where the model would use the off-target writing as context, and be trained to predict just the target person’s response.
2. Human Feedback
The ‘trusted editors’ could also look at responses of the current digital twin fine-tune to various prompts. Picking the more accurately representative of two possible responses would provide feedback that could be used to further fine-tune the model (probably using a LoRA on top of a base model).
Ability for Users to create their own digital twin
I think that potentially you could have an option for users to pay (or subscribe to premium, or enter their LLM API key, etc) in order to be allowed to make a digital twin of their own (not necessarily of themselves). The user would be allowed to upload documents and dialogues, and then queue a fine-tune. They’d be able to see how their digital representative responded to polls that they themselves hadn’t answered yet.
You could also allow them to give feedback about responses to prompts. Either by going through a bunch of binary choice of response to random poll, or by repeatedly regenerating a response to a particular poll until happy with the response. Similar idea to the Famous People Digital Twins, except that the user would be tuning to their own preferences rather than an editor tuning to accuracy-of-recreation-of-target-person.
If the user has paid for this service, then it’d make sense for them to be able to download their resulting LoRA.
Auto-updates of current legislation
Setting up the site to auto-search through all the world’s elections, and for any area auto-create polls for legislation that’s up for consideration. Also, auto-marking polls as ‘consideration closed’ once the window has passed.
Allow users to specify their region(s) of interest, and filter based on ‘open for consideration’. This could be useful for people learning how various others think about this piece of legislation, and thus deciding how to vote on it (if they have that option). Also, being able to see groups of opinions, and summaries of the centroids of the groups would be great. Making the tool useful in this way would potentially get politicians to begin using it to guide their decision making and campaigning. Politicians pay a lot of attention to polls!
Enriched Poll Info
In the same way that Manifold Markets has a Market Question, and then Resolution Details which clarify the definitions of terms and the rules that will be used to resolve the market, the poll questions on YouCongress would benefit from some additional background.
This could potentially be generated automatically through some sort of web search enabled AI agent. If a user was manually creating the poll, then the agent could just give a list of potentially helpful snippets to include in the details. If the poll was being created automatically based on news or current legislative proposals, then you could have the details fully auto-generated.
User choices and feedback
Allowing users to give feedback in various ways, and set some preferences. For instance, allowing a user to pick a small set of AI digital twins which they wanted to ‘follow them around’ and always be at the top of the responses to any poll that a user clicks into. This would also probably be a paid service if the user was thereby generating additional novel digital twin responses. Of course, for public digital twins, this would then be a resource for other users as well when visiting that poll question.
Other types of feedback
Being able to vote for a Digital Twin’s response as accurately-representative of the twinned famous person, or unrepresentative.
Being able to vote for ‘liking’ or ‘disliking’ the quality of a Digital Twin response. Enough dislikes might trigger a model update and regeneration of the response.
Allowing Users to vote agreement/disagreement with other people’s responses without necessarily vote-deferring to that other person.
Using a Twitter/X community notes style algorithm to judge the feedback given by users in order to know how to weight the various types of feedback. When a response is rated high quality and/or high accuracy across multiple disagreeing groups of users, that’s a stronger signal than when the approving users are very uniform in their belief groups.
The fact that the polls are split into topics seems great for breaking up the US two-party group-everything-together style political debate. I’d love to see people disagreeing on some things, and agreeing on others. I think this could lead to people realizing they have more common ground across ‘party lines’ than they realize. This aspect could be actively facilitated in some way, like Polis does.
You could expand to multi-polls where users and bots got to do quadratic voting. Quadratic voting is problematic for elections because the selection of the set of items to include in a vote (over which the voting points get shared) inherently introduces bias. In community-created polls however, this bias can be countered relatively easily when analyzing results across many polls and many groups of users.
Anyway, I hope this set of visions appeals to people who might be inclined to jump in and help implement some of these!