GPT-4 Was Deeply Racist Before OpenAI Muzzled It(futurism.com)
futurism.com
GPT-4 Was Deeply Racist Before OpenAI Muzzled It
https://futurism.com/gpt-4-deeply-racist-before-openai-muzzled-it
5 comments
I wouldn't be surprised if they marketed the "raw" chatgpt. At the end of the day its trained on internet texts, so if its racist its because its turning a mirror on the collective of internet users and showing us that there is racism there. Given that, if you wanted to use chatgpt to perform some sentiment analysis on the population of internet users it was trained upon, the sanitized version would be useless to you. You'd probably want the raw output which more closely reflects the underlying dataset. I guess musk's "anti woke" chatgpt mentioned in the article would probably have a market after all, possibly more of the research market share for this sort of tool.
How is this surprising? There are deeply racist texts on the internet.
There are many people who don't understand that AI can be a genuine threat to society. I've heard people like Steven Pinker claim that it's fanciful thinking that AI could pose a threat.
I think these people have a classical view of AI, where an AI is a kind of expert system that follows a preset algorithm and kind of does what it's told in a predictable manner, but just does it in a very efficient and "intelligent" manner.
I don't think these people appreciate that the current trend in AI is not to create well understood expert systems, but instead to create systems trained on human data, including all of our hate, violence and also respect, love and compassion.
Perhaps for AI, love, compassion and respect will win out, and we can manage to continue keeping AI aligned so that it doesn't pose a threat to us in the future.
But maybe we can't.
I think these people have a classical view of AI, where an AI is a kind of expert system that follows a preset algorithm and kind of does what it's told in a predictable manner, but just does it in a very efficient and "intelligent" manner.
I don't think these people appreciate that the current trend in AI is not to create well understood expert systems, but instead to create systems trained on human data, including all of our hate, violence and also respect, love and compassion.
Perhaps for AI, love, compassion and respect will win out, and we can manage to continue keeping AI aligned so that it doesn't pose a threat to us in the future.
But maybe we can't.
I believe that humans, in general, will work towards their best interest and reduce their short term (if not long term) pain. Not all the time, not in everything. But, in the end, pain is not subjective. Sure, there's the myth of slow boiling a frog. But I've heard that doesn't actually work in real life.
In theory, theory and practice are the same. In practice they are not.
IMHO the paperclip optimizer Ai "problem" is impossible. There's enough human hands at every step, in enough different locations, speaking enough different locations, with enough different personalizes to stop the proposed ridiculousness.
Obviously Ai (not just Agi) is going to restructure how humans work. I just hope there's enough pain points that enough legislation is put into place to create a large enough safety net to catch the people who can't quickly adapt.
In theory, theory and practice are the same. In practice they are not.
IMHO the paperclip optimizer Ai "problem" is impossible. There's enough human hands at every step, in enough different locations, speaking enough different locations, with enough different personalizes to stop the proposed ridiculousness.
Obviously Ai (not just Agi) is going to restructure how humans work. I just hope there's enough pain points that enough legislation is put into place to create a large enough safety net to catch the people who can't quickly adapt.