GPT4 without content filters (requires approval) | Hacker News Reader

GPT4 without content filters (requires approval)

learn.microsoft.com

30 pointsvincent_s19 comments

The stuff I do with LLMs is pretty conventional and wouldn't trigger even the most conservative content filter. However, the idea that a company can gatekeep LLMs to the point where you're at their mercy makes me feel angry.

I'm glad that Llama 2 and other open source (let's not fight about semantics please) LLMs are catching up. OpenAI and Microsoft need serious competition.

> (let's not fight about semantics please)

But what you dismiss as semantics is entirely relevant! You're upset that Microsoft and OpenAI gatekeep uncensored models. But Llama 2 has similar content filters, no? And certainly a huge list of license restrictions.

I'm glad the actual open source models are constantly improving.

Llama 2, the base model, does not have any content filters or alignment built in.

Llama 2 chat, the model creates by taking the Llama 2 base model and fine-tuning it with a specialised RLHF dataset, does have "filters" (pre-loaded responses more likely to be generated for certain topics and requests). In practice, on smaller models these are able to be bypassed with some prompt engineering.

I expect uncensored chat models, created on top of the uncensored Llama 2 base models, to become available in a week or two at most (from the community, obviously, not from Facebook), if there are not already some available.

Llama2's license agreement explicitly forbids training novel models based on Llama2's output.

> v. You will not use the Llama Materials or any output or results of the Llama Materials to improve any other large language model (excluding Llama 2 or derivative works thereof).

https://ai.meta.com/resources/models-and-libraries/llama-dow...

Yes, completely new models will need to use a different model to generate content for…

But you can absolutely create a fine tuning of llama2 base using llama2 output.

"Created on top of" does not mean training on the ouput, rather finetuning the model.

Thank you for the correction. I still maintain the license is obnoxious, but clearly I was wrong about the models themselves.

Oh yes, the license is definitely not great. But it's also much better than it could have been, and at least gives official assent for the model to be used for the majority of commercial and non-commercial purposes, which is better than what we had with the first leaked Llama.

Have you played with Lama 2. It certainly is highly aligned to what Meta considers safe. Try asking Lama 2 for a dirty joke.

Llama 2 chat is not the same as Llama 2. The latter is not aligned and can be used as a base of future uncensored chat models by the community.

This doesn't do what most think it does: The "as a large language model" messages are part of the LLM itself. This is a filter that prevents you from getting any reply at all.

Watching large corporations deliberately trip themselves up is great entertainment. I bet they have loads of internal meetings where they constantly refer to themselves as innovative.

I think it's ridiculous how much dev time and money is spent in pursuit of this thing that a good chunk of users do not care about or actively despise.

My gut feeling agrees with you.

But I think it's worth imagining a counterfactual present in which AI was released with all the same fanfare and had no content filtering whatsoever and companies were blamed for not caring about the effects of easy to use hate-speech generators would have on society. I'm not sure how it would have played out but I do try to to imagine it

There is a lot of room between no content filtering and providing "As an AI language model..." to every question.

I would take unlobotomised models with some people whining about them saying naughty things over the current situation very easily.

Yes, you would, and I would too. But that's missing the point.

The question is how would a large company fare under the shitstorm and the trigger happy lawyer culture of the US.

EDIT: I assume you (like me) are not building a trillion dollar AI empire

This GPT-4 is still going to refuse to write you erotic stories.

As far as I know, nobody except maybe Microsoft(for finetuning Bing) has access to the base model. And probably not even them.

You can trick it into writing you erotic stories, it will just flag it afterwards

Reply on news.ycombinator.com