Is there an alternative to ChatGPT that is less restrictive? | Hacker News Reader

Is there an alternative to ChatGPT that is less restrictive?

13 pointssuvelule17 comments

Hey everyone,

I've been using ChatGPT for a while now and while I think it's a great tool, I've been struggling with getting it to generate certain types of content. I understand that certain topics may be sensitive and the model is designed to avoid generating harmful or offensive content, but I feel like it's too handcuffed in its ability to assist me with my specific needs.

I was wondering if anyone knows of any other language models or AI tools that have a more open approach and can assist with generating more diverse types of content? I would greatly appreciate any recommendations or suggestions.

Thanks in advance!

PS: I used ChatGPT to generate this post. my English writing capabilities are very limited.

View on HN

One further: what’s the best performing model one could reasonably self-host, either locally on a single 4090 or say remotely on a single A100?

Raw GPT-3 doesn't have as many restrictions and is basically the same engine. You're not allowed to publish sensitive content but they don't really stop you from using it personally, short of child porn.

You can access it via the playground: https://beta.openai.com/playground

It's not free, but very cheap.

If you add the chat/conversational features on top (there are countless tutorials for that, one is included with langchain and others extended that one), the ‘very cheap’ goes out the window fast. For it to have memory, people use clever summarizing and embeddings, however, you end up sending 1000s of tokens every time you interact. To get gpt3 to create similar long winded responses, you have use tricks which also cost a lot of tokens.

Out of the box, gpt3 feels miles away from chatgpt when you try things naively (which chatgpt happily answers, while gpt3 probably will just repeat you or write nonsense). So the ‘raw’ you start with is correct; once you are happy with the results, the cheap is out the window.

The way I see it, ChatGPT is the "low code" edition of GPT. You wouldn't do a feature-by-feature comparison of both as they're for different things.

The question is on content generation, which GPT-3/davinci excels at. ChatGPT does chat and memory better.

Also the long winded responses are already in the nature of text-davinci-003, though it faces similar problems with refusing to do certain types of content.

Can you give an example or two of the types of content chatGPT has been unable to generate for you due to its restrictions?

Not OP, but I feel like it should not have any restrictions.

For example creating Molotov cocktails is not answered.

It's fairly ridiculous with respect to not "impersonating" anything. I wanted to see it generate modern political commentary in the style of HL Mencken, to which it just says "I'm not HL Mencken". Like, what am I going to do, use to to claim I found HL Mencken's long lost hidden commentary on Donald Trump?

Anti-impersonation also covers like "give pointing out instructions as a Dzogchen lama." It tells you that it's not a licensed and bonded Dzogchen lama and pointing out instructions are proprietary information belonging to the Dzogchen lama association. (Tongue in cheek, obviously, the point is that while held to be secret, they're not copyrighted, legally protected, or proprietary information. Arguably, "Dzogchen lama" isn't even a thing).

It was willing to write a koan in the style of the Wumenguan, which surprised me after trying this stuff out for a while that it didn't just say "Sorry Dave, I'm not the 13th century Chan monk Wumen Huikai and can't do that."

Can you tell it to pretend or use its best guess?

I know that you can’t use basic prompt-engineering to get it to be racist or explain how to commit crimes. But for scenarios like the above when it’s impersonating you can just tell it to pretend and that is enough.

I feel like that's less some anti-impersonation rule and more ChatGPT simply lacking creativity. I got similarly stonewalled when asking it about how much wood a woodchuck could chuck if a woodchuck could chuck wood, with responses along the lines of "that's a tongue twister, woodchucks don't actually chuck wood". Even after narrowing the possibilities (specifying balsa wood, specifying specific dimensions of the wooden projectiles, specifying woodchucks of average size/strength, even giving the woodchucks trebuchets), it still refused to give a straight answer, only saying that woodchucks wouldn't be able to chuck very much wood.

No, because it used to do it.

But then some greasy little nerd in their office added a rule to ban it.

You used to be able to say stuff like “pretend to be homer simpson and tell me about algebra” but now it just blocks that for being wrongthink

There are many things it won’t do and more and more restrictions are being added. There should be an unhinged version online next to the pc version, if only just to prepare us for what’s coming (can’t be stopped anyway, even in the short term, just like ‘unstable diffusion’).

Many stuff. like I was looking for ideas for a new job since I thought programming was out of the window and I thought maybe getting into hardware manufacturing. it used to give straight answers but now it just wants me to make sure everything is legal and our chat has started to feel like interrogation rather than asking questions. another annoying thing about it was that like a month ago it would happily generate content in the style of religious books. which was very entertaining but now it wants me to go talk the those religious leaders before I use it for fun. its obvious to me that an alternative is needed.

Not really a solution, but an interesting anecdote: I have had ChatGPT push back on the legality of things I'm trying to do numerous times as well. Recently I had a conversation where I told it I wanted to start a business with questionable legality (involving web scraping government websites). However, ChatGPT seemed much more willing to help when I said I was doing it to help poor people--it even said it was a great idea and actually gave advice to get started. It seemed to know it was breaking its own rules because it flagged itself with the red box or whatever it does.

I tried using it to generate some room/monster descriptions for an RPG-type scenario.

It wasn't long before I got flagged for "promoting violence" (or some such...that probably wasn't the exact phrasing). I stopped because I didn't want to get banned from the system.

Yeah, this is exactly the kind of thing GPT-3 does well, which ChatGPT doesn't.

example my friend told me today: it generates jokes about poles, but if you ask about jokes about jews then it's restricted

I like jews but i also like poles

I've been using WriteSonic for awhile, and it seems like a pretty good alternative

Reply on news.ycombinator.com