5 ms·
The fact that major LLMs will happily criticize certain politicians but claim they’re unable to comment on others given equally available training data is a fai
by transcriptase 3y ago
The fact that major LLMs will happily criticize certain politicians but claim they’re unable to comment on others given equally available training data is a fairly good window into what kind of biases exist within them.
- osigurdson 3y agoI’ve always wondered why anyone bothers asking ChatGPT these kinds of questions. Where is the utility?
- bmelton 3y agoA lot of these discoveries were being made when OpenAI and ChatGPT were pretty new to the general public. Nobody knew what it did, so these kinds of questions were no different than any questions being asked, basically. That so deep a bias was discovered so quickly was likely incredibly surprising to some, and incredibly unsurprising to others, but they were asked for the same reason as others - to find out what it does and whether it could be seen as reliable for a source of information
- osigurdson 3y agoI suppose. I remember my first encounters with the Open AI playground were along the lines of "write a short story about X". Amusing at first but the novelty quickly wears off.
- arcanemachiner 3y agoMaybe I'm having a bad day and I just need to hear a rap song about how cool Donald Trump is, to help boost my mood.
- osigurdson 3y agoI agree that there is value in humor. I find that ChatGPT humor gets old pretty quickly however.
- revelio 3y agoTo help determine what to say and think in discussions like this one. Also to stop AI firms claiming to be unbiased.
- Lerc 3y agoIt's unclear to me if that is an issue in the bias of the training data or an issue in the fine tuning to control the output. Is it "I can't say that about X because I am told not to say things like that about X"? or is it "I can't say that about X because I am told not to encourage behaviour Y and my training data indicates that comments like that about X lead to behaviour Y"?
- mmcwilliams 3y agoIf you're using a model that's behind a moderation service in their API, you're not actually assessing the biases of the model but rather of the corporation that's trying to remain profitable.