And this is why open-source models are and will always be a thing. Even if they aren’t as strictly intelligent as OpenAI’s models, they aren’t guardrailed to the point of uselessness.
For creative writing, I prefer local models. I mean, GPT 4 writes awesome prose, but very neutered. You provide a passage you wrote yourself with some conflict or intimacy, and when you ask it to improve it even keeping the interesting points, it waters it down to bring it to the agenda that big tech and entitled rich US West Coast people want to force on us.
Whereas I use mlewd (no kidding) for non-lewd, creative ideas and it often works pretty well. Especially when I want to explore very crazy ideas, pump up the temperature and other parameters and you often get a bunch of ideas that, by working on them by little, can lead you to something.
Ask GPT 4 to come up with ‘an imaginative name to name mecha (armored) suits in my novel’.
It will always, always give an unimaginative list such as
NanoFrame, TechSuit, ElectroSuit, etc
It always gives a portmanteau. Very, very unimaginative. GPT 4 is great for many things, but not for creative writing.
GPT 3.5 is something I hardly ever even bother with anymore because by now, the only advantage over local models is that, along the extremely verbose, neutered, bland, unimaginative response, it still manages to hallucinate way less and provide factual data. The fact that half of the time it will not provide what you wanted, but something else, is another thing
Ask GPT 4 to come up with ‘an imaginative name to name mecha (armored) suits in my novel’.
Any names these things come up with should be looked at pretty carefully. I find that they are almost always pulled from existing media. I toyed around with having chatgpt and a couple of local models create the framework for a fantasy world several months ago - name, regions, basic descriptions, etc. Because I am fairly familiar with a lot of different fantasy media - books, games, shows, movies - I was able to recognize not only most of the names it “created”, but with a bit of effort I was able to track down some of the setting descriptions as being lifted directly from existing fantasy IPs as well.
Names are particularly rough with these things, because they don’t “create” anything, they just take things that match whatever you’re talking about with them and chop them up and mix them a bit.
And this is why open-source models are and will always be a thing. Even if they aren’t as strictly intelligent as OpenAI’s models, they aren’t guardrailed to the point of uselessness.
For creative writing, I prefer local models. I mean, GPT 4 writes awesome prose, but very neutered. You provide a passage you wrote yourself with some conflict or intimacy, and when you ask it to improve it even keeping the interesting points, it waters it down to bring it to the agenda that big tech and entitled rich US West Coast people want to force on us.
Whereas I use mlewd (no kidding) for non-lewd, creative ideas and it often works pretty well. Especially when I want to explore very crazy ideas, pump up the temperature and other parameters and you often get a bunch of ideas that, by working on them by little, can lead you to something.
Ask GPT 4 to come up with ‘an imaginative name to name mecha (armored) suits in my novel’.
It will always, always give an unimaginative list such as
NanoFrame, TechSuit, ElectroSuit, etc
It always gives a portmanteau. Very, very unimaginative. GPT 4 is great for many things, but not for creative writing.
GPT 3.5 is something I hardly ever even bother with anymore because by now, the only advantage over local models is that, along the extremely verbose, neutered, bland, unimaginative response, it still manages to hallucinate way less and provide factual data. The fact that half of the time it will not provide what you wanted, but something else, is another thing
Any names these things come up with should be looked at pretty carefully. I find that they are almost always pulled from existing media. I toyed around with having chatgpt and a couple of local models create the framework for a fantasy world several months ago - name, regions, basic descriptions, etc. Because I am fairly familiar with a lot of different fantasy media - books, games, shows, movies - I was able to recognize not only most of the names it “created”, but with a bit of effort I was able to track down some of the setting descriptions as being lifted directly from existing fantasy IPs as well.
Names are particularly rough with these things, because they don’t “create” anything, they just take things that match whatever you’re talking about with them and chop them up and mix them a bit.