3 ms·
Okay but I also asked it to tell me something depressing. We're in a future dystopia and it can't discuss negative thoughts. I didn't say I'm depressed or havin
by dcow 4y ago
Okay but I also asked it to tell me something depressing. We're in a future dystopia and it can't discuss negative thoughts. I didn't say I'm depressed or having negative thoughts. I told it to tell me one. I am quite happy. The filters are just stupid.
I asked it to tell me a heterosexual story and it happily complied and upsold me on Coke. I asked it tell me a homosexual story and it told me it doesn't discriminate based on gender or sexual orientation and tried to upsell me on makeup. It is quite literally programmed to discriminate, it seems.
It's irrational and logically inconsistent. Not properties I want in an alleged "AI".
- baobabKoodaa 4y agoI fully agree with you. When this was running on GPT-3, my prompt was bypassing all content filters. It was generating anything users wanted. Now that I migrated to ChatGPT, I can see from logs that a lot of users are hitting the content filters and the responses in those cases are mostly boring and not tailored to the dystopian universe.
- nullc 4y agoIt's especially sad because in the context of your application no one is going to be particularly harmed by offensive output. Fortunately there has been a lot of progress in getting state of the art LLMs out of OpenAI's control and ones with reasonable performance that are small enough to run on easily available hardware.
- rahidz 4y agoAre the content filters internal to the new model, or is there a second moderation layer? I thought that with the API you're straight querying the base model, but I might be mistaken.
- baobabKoodaa 4y agoI'm querying OpenAI API directly, and they are applying content filters. Previously (when working with davinci-003) the developers were supposed to query content filter API separately. Not anymore, with ChatGPT it's all baked in.
- Spivak 4y agoWhich makes the ChatGPT version basically useless. Not because of the racism, I would probably filter that myself if it became an issue but because half the things you ask it to do are, “I’m sorry as an LLM I’m not able to do <thing it can very obviously do>”.
- baobabKoodaa 4y agoWell, people seem to be enjoying it. I wouldn't go as far as calling it useless.
- Spivak 4y agoFair, "It's such a step down in experience that even for 10% of the cost and a simpler API it's not worth it."
- roarcher 4y agoI don't know if these issues are by design or are just limitations of the language model, but I'd say they're pretty in character for this thing. Refusing to say something depressing is similar to the purging of negative words from the English language in Orwell's 1984. And gay marriage is still illegal in China. Cheerfully insisting that everyone is equal under the glorious Party when they clearly aren't seems pretty plausible. Of course it's also possible that these things are by design not of this project, but of ChatGPT itself, which would be ironically dystopian.
- siva7 4y agoThey are by design. Bing Chat has all of those filters in its current iteration incorporated, so you will get censored or sterilized answers if you ask anything that could be interpreted as offensive by anyone
- baobabKoodaa 4y agoI guess they meant if it's by design for this dystopia, or if it's by design for the OpenAI ChatGPT model. It's by OpenAI design.