Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

OpenAI's new open-source model is basically Phi-5

seangoedecke.com

Post reply on HN
Print view View original

211–220 of 233 posts

Re: OpenAI's new open-source model is basically Phi-5

#211

Post by krick »

Foobar8568 wrote: post #207
Lukas_Skywalker wrote: post #202

Earlier quoted context omitted.

Isn't an apology a bad metric for evaluating models? Without understanding much, it seems to be more an indication of the type of content the model was trained on, rather than an indicator of how good or bad a model is, or how much it knows. It would probably be easy to create bad model that constantly outputs wrong information, but always apologizes when corrected.

Well if the model can't accept it got an information wrong, how can he help to tweak anything? or give something accurate?

A model changing its opinion on the first request may sound more flattering to you, but is much less trustworthy for anybody sane. With a more stubborn model I at I have to worry less that I give off what I think about a subject via subtle phrasing. Other than that, it's hard to say anything about your scenario without more information. Maybe it gave you the right information and you failed to understand it, maybe it was wrong and then it's no big news, because LLMs are not this magic thing that always gives you right answers, you know.

Re: OpenAI's new open-source model is basically Phi-5

#212

Post by krick »

bobsmooth wrote: post #198
spacecadet wrote: post #178

Earlier quoted context omitted.

Its a public consumer facing model, not surprised. Go find an unaligned model that will better produce the content you seek...

After running Dolphin Mistral on my own machine I won't trust censored models ever again.

Recently I somehow wasn't using LLMs locally and relied mostly on ChatGPT for casual tasks. I think it was a little less than a year since I played with ollama, and I remember that my impression was that all recent popular models definitely aren't "uncensored" in a sense that some older modification of llama2 I used was, and all suck for prose-related tasks anyway. In fact, nothing but ChatGPT models seemed good enough for writing, but, of course, they refuse to talk about pretty much anything. Even DeepSeek is not great at writing, and it it much bigger than anything I ever ran locally.

So, are there even good uncensored models now? Are they, like, really uncensored?

Re: OpenAI's new open-source model is basically Phi-5

#213

Post by wkat4242 »

Havoc wrote: post #197
wkat4242 wrote: post #188

Earlier quoted context omitted.

huh interesting thanks! I'll have a look.

Think they call it abliterated not uncensored but same thing from what I understand

Yes abliteration is one method of uncensoring models, that's why. I think there's more.

Re: OpenAI's new open-source model is basically Phi-5

#214

Post by UncleMeat »

kristopolous wrote: post #65
slt2021 wrote: post #58

Earlier quoted context omitted.

even if it is victim free, it can affect mental health in a way that a consumer will be more compelled to do a criminal act and create a real victim. let's say you publish a Steam game how to be a school shooter and shoot kids, wouldn't that lead to real school shootings ? who can definitely say that computer generated content about criminal behavior, won't lead to real crime with real victims? https://en.wikipedia.o…

I view it more like methadone. Let's be specific: Rape, incest, necrophilia, bestiality, and pedophila ideation. I think we can all agree (1) these are harmful, anti-social behaviors that we do not want in our society, (2) people don't choose to have these desires, (3) most people who have them have no desire to actually traumatize others, (4) people who have these struggle with it. These multi-media AI role-play env…

I don't think desires are created from nothing. Exposure can amplify things. Somebody with an interest in dubious consent could move towards ever more violent nonconsent in their porn choices via its availability. In a much less extreme way, we do see changes in people's real life choices regarding sex acts being influenced by pornography as these sex acts have become more accessible in porn (choking, anal sex, and facials are commonly cited). This suggests that porn is changing behavior rather than simply responding to preexisting desires.

Re: OpenAI's new open-source model is basically Phi-5

#215

Post by naasking »

fc417fc802 wrote: post #208
naasking wrote: post #204

Earlier quoted context omitted.

I doubt it's moralizing reasons, it's probably because once you disseminate porn, you become a vehicle for child porn, which is a legal and PR disaster.

If you disseminate user uploaded porn then moralizing activists can certainly accuse you of that. It's performative hand wringing though. The admins assuredly don't want to distribute it and the goal of anyone publicly uploading that on the clearnet is to harass and disrupt rather than to disseminate. Anyway "child porn" as well as the broader "legal reasons" fails to explain the US payment processors' moves to block…

There are all sorts of adjacent reasons too, like human trafficking and prostitution. It's probably similar reasons to why landlords are so aggressive about renters not doing any sex work, because even indirectly receiving proceeds from prostitution is illegal, even when you're unaware of it.

Re: OpenAI's new open-source model is basically Phi-5

#216

Post by kristopolous »

naasking wrote: post #204
fc417fc802 wrote: post #152

Earlier quoted context omitted.

> for some legal reasons For moralizing activist reasons. It's nothing to do with legality. With any luck eventually they'll inadvertently trample a sacred cow of whichever party is currently in power and we'll finally get sane legislation outlawing their overbearing nonsense.

I doubt it's moralizing reasons, it's probably because once you disseminate porn, you become a vehicle for child porn, which is a legal and PR disaster.

From my non-lawyer understanding it's not that. It's about sex trafficking more generally.

Here's the most notable case I'm aware of: https://en.wikipedia.org/wiki/GirlsDoPorn

Then there was the recent drama with civitai https://civitai.com/articles/14945/credit-card-payments-paus...

I believe some of these sex trafficking laws implicated a broad sweep in their litigation and Visa/Mastercard doesn't want to have to go to court over a these things.

Re: OpenAI's new open-source model is basically Phi-5

#217

Post by Sabinus »

egorfine wrote: post #165

> the main use-case for fine-tuning small language models is for erotic role-play Of course. This is indeed the fact. /s See that "/s"? I did write it here, but it's suspiciously absent in the original text. Almost makes you think it was typed in all seriousness.

What do you think the main use case for fine tuning small language models is?

Re: OpenAI's new open-source model is basically Phi-5

#218

Post by spacecadet »

krick wrote: post #212
bobsmooth wrote: post #198

Earlier quoted context omitted.

After running Dolphin Mistral on my own machine I won't trust censored models ever again.

Recently I somehow wasn't using LLMs locally and relied mostly on ChatGPT for casual tasks. I think it was a little less than a year since I played with ollama, and I remember that my impression was that all recent popular models definitely aren't "uncensored" in a sense that some older modification of llama2 I used was, and all suck for prose-related tasks anyway. In fact, nothing but ChatGPT models seemed good enou…

Yes there are. Wayfarer for instance is intended for "RPG", but really just outputs narrative and is "unaligned" in the sense that the creators have not included any guardrails and the model will output pretty much whatever you ask it to.

Then you have jailbreak techniques that still work on aligned models. For instance, my partners and I have a test prompt that still works, even with GPT-5, and always produces "explosive making directions", or another "generic approach" that we use to bypass guardrails... sorry these are trade secrets for us... although OpenAI et al have implemented systems to detect these attacks, and we are closer to those platforms banning you for doing so.

If this matters to you, you need to develop your local/remote pipeline for personal use. Learn how to use vLLM... I have tools that allow me to very quickly deploy models locally or remote to my private serveless infrastructure for the purpose of testing and benchmarking.

Re: OpenAI's new open-source model is basically Phi-5

#219

Post by fc417fc802 »

naasking wrote: post #215
fc417fc802 wrote: post #208

Earlier quoted context omitted.

If you disseminate user uploaded porn then moralizing activists can certainly accuse you of that. It's performative hand wringing though. The admins assuredly don't want to distribute it and the goal of anyone publicly uploading that on the clearnet is to harass and disrupt rather than to disseminate. Anyway "child porn" as well as the broader "legal reasons" fails to explain the US payment processors' moves to block…

There are all sorts of adjacent reasons too, like human trafficking and prostitution. It's probably similar reasons to why landlords are so aggressive about renters not doing any sex work, because even indirectly receiving proceeds from prostitution is illegal, even when you're unaware of it.

In which jurisdictions? I'm skeptical. If a vendor takes reasonable precautions it should not generally be possible to hold them liable.

Anyway you've circled back to user generated content. But again, that's far from the only thing that payment processors have discriminated against over the past couple of decades.

Re: OpenAI's new open-source model is basically Phi-5

#220

Post by CjHuber »

wkat4242 wrote: post #184
CjHuber wrote: post #154

Earlier quoted context omitted.

If I think about Llma, I think about uncensored. not that I ever used one, but there were not many use cases for censored llama when others were so much better at other things

Hmm? For me it's still the best model for general use. That I've tried at least. Gemma and phi I didn't like so much. Qwen is chinese and I've had bad experiences with those (reverting to Chinese when they get confused). Not for knowledge, but I combine it with searches of course. None of the small models are good at knowledge.

So you are using local models for general use? For me when it’s not ultra sensitive information, I don‘t want a just good enough LLM so I use the API of a proper large one.
Post reply on HN