undefined | Better HN

0 pointstoomim3y ago0 comments

> This sounds like a line being fitted to a single data point.

Except that there are many data points. Here's another: https://twitter.com/DashMeister1/status/1607568958382825474

And OpenAI has described some aspects of how these filters are trained -- https://unherd.com/thepost/chatgpts-score-system-shows-polit... -- and it sounds like a bunch of humans are involved in rating which responses are "kosher" and not, and that "kosherness" filter seems to have a pretty clear left-leaning bias.

0 comments

fenomas3y ago

My reply was to GP's assertions about which questions GPT will and won't answer, not to TFA's claims about political compasses.

> Ben Shapiro

...really?

saurik3y ago

GP does not assert anything about other questions and in fact doesn't even assert an opinion on this question: they merely demonstrate that there is a single question that has a biased filter and that is sufficient to prove the existence of some sort of awkward creator-infused bias, even if merely against topics that begin with the letter F.

fenomas3y ago

(When I replied to it, Toomim's post was wholly different and linked to a Ben Shapiro video about purported leftist bias in ChatGPT.)

visarga3y ago

There is a specific stage called Reinforcement Learning from Human Feedback (RLHF) that is supposed to bias the model to be helpful, polite, present opposing views for balance, and not to say things that are out of its training set. We have all seen the canned responses that appear when the filter is triggered.

RLHF is very important if you want your model to act professionally. The first version of Davinci (GPT-3 in 2020) was "free" from such human-preference-training, but very hard to convince to cooperate on any task.

But applying RLHF could also introduce political biases into the model. This depends on the labelling team.

j / k navigate · click thread line to collapse

0 pointstoomim3y ago0 comments

> This sounds like a line being fitted to a single data point.

Except that there are many data points. Here's another: https://twitter.com/DashMeister1/status/1607568958382825474

0 comments

fenomas3y ago

My reply was to GP's assertions about which questions GPT will and won't answer, not to TFA's claims about political compasses.

> Ben Shapiro

...really?

saurik3y ago

fenomas3y ago

(When I replied to it, Toomim's post was wholly different and linked to a Ben Shapiro video about purported leftist bias in ChatGPT.)

visarga3y ago

But applying RLHF could also introduce political biases into the model. This depends on the labelling team.

j / k navigate · click thread line to collapse