Help us make ChatGPT wiser
Join our OpenAI-backed experiment to democratically fine-tune ChatGPT's values.
Want to shape the values of future versions of ChatGPT? Join our OpenAI-backed experiment to democratically fine-tune ChatGPT's values. Participate in a 15-minute, values-based voting process, and you’ll be the first to hear about our results.
More about the process
At the Institute for Meaning Alignment, we’ve spent the last months working on a new democratic process (see our blogpost). It’s a hybrid between voting and the values-based opinions of courts (but one where everyone’s a judge). We hope this process can address some problems that come with voting, such as polarization and side-taking, and can refocus democracy on collective wisdom.
It works by collecting values from a large population, so we need your help to test it!
As our research is funded by OpenAI, we’re using it first to collect the values that will shape ChatGPT’s future behavior.
You can run through the process here. It takes 15 minutes. Those who participate will be the first to receive information about the outcomes.
Cheers,
Oliver & Joe (the Democratic Fine-Tuning team)
Finally got some time to sit down and try this thing. It's a good start.
I'm somewhat concerned about some of these value cards to be able to mislead people. I was presented with a choice between the values of "Truth-Seeking" vs. "Accuracy" or something of the like.
At a glance, it looked like the latter made more sense, since it described someone getting overwhelmed with information and then choosing to reduce the cognitive load by picking reliable sources based on reputation and trustworthiness. But looking more closely at the values as articulated, that's not something I'd support in the general case, since it's a shortcut for actually thinking about problems. Sometimes you have to take a shortcut for practical purposes, but it's not a complete substitute for thinking deeply about a problem.
If I had rushed through the tool, I probably would have endorsed the latter. Thinking about it more carefully, though, I endorsed the former instead, since I think it generalizes better.
How are you folks preventing the LLM from ending up with values exactly like this: ones that sound great at a glance, but end up being not what people want when they sit down and think about it more deeply?
Your tool is currently broken. I can select a scenario but whenever I try to type a response it just says "Failed to update chat. Please try again."
https://imgur.com/iBdYAax