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

141–150 of 233 posts

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

#141

Post by Mkengin »

magicalhippo wrote: post #137
VladVladikoff wrote: post #18

Earlier quoted context omitted.

Can you share the question? Or are you intentionally trying to keep it out of the training data pool?

Here's a more concrete example where GPT-OSS 20B performed very well IMHO. I tested it against Gemma 3 12B, Phi 4 Reasoning 14B, Qwen 2.5-coder 14B. The prompt is modeled as a part of an agent of sorts, and the "human" question is intentionally ill-posed to emulate people saying the wrong thing. The prompt begins with asking the model to convert a question into matlab code, add any assumptions as comments at the star…

Why Qwen2.5 and not Qwen3-30B-A3B-Thinking-2507 or Qwen3-Coder-30B-A3B-Instruct?

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

#142

Post by michaelt »

mvdtnz wrote: post #127
michaelt wrote: post #39

Earlier quoted context omitted.

The pro-porn side has zero PR because respectable public figures don't see pro-porn advocacy as a good career move. At most, you'll get some oblique references to it. Meanwhile, the anti-porn side has a formidable alliance: Right-wing, religiously-motivated anti-porn activists. Left-wing, feminism-motivated anti-porn activists. Big corporate types with lots of $$$$ to spend who want their customer support chatbot to…

It's not pro-porn and anti-porn. It's pro-porn and people who just don't think this is that important an issue. The latter massively, MASSIVELY outweighs you guys.

If a person is configuring an LLM for education, to provide personalised math coaching to 10 year olds, they want an LLM that won't output anything NSFW, no matter how the user pokes and prods it. That's totally reasonable.

But if that person is applying AI safety techniques like concept erasure to remove the model's ability to output porn, is that not anti-porn in the most literal sense?

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

#143

Post by jofzar »

wkat4242 wrote: post #103
wmf wrote: post #20

I saw a bunch of people complaining on Twitter about how GPT-OSS can't be customized or has no soul and I noticed that none of them said what they were trying to accomplish. "The main use-case for fine-tuning small language models is for erotic role-play, and there’s a serious demand." Ah.

It's not just erotic role play that the censorship affects. My life involves a lot of sexual discussions and that means that everyday talk, chat summaries, email rewrites or translations will cause the model to shut down. I do the latter a lot especially to find colloquialisms because Google translate is often too literal. It's so annoying. Right now I'm using abliterated llama 3.1. I have no need for vision but I wa…

Sorry just to ask, what kind of job do you have? Sex therapist sounds like the closest?

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

#144

Post by Al-Khwarizmi »

numpad0 wrote: post #128
wmf wrote: post #20

I saw a bunch of people complaining on Twitter about how GPT-OSS can't be customized or has no soul and I noticed that none of them said what they were trying to accomplish. "The main use-case for fine-tuning small language models is for erotic role-play, and there’s a serious demand." Ah.

Maybe I'm just not seeing it, but is that use case really real and not just prudish hallucinations? The market for NSFW novels is smaller than even cyberpunk paperbacks, there's no way everyday people build up addiction for an interactive version of it.

[dead]

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

#145

Post by createaccount99 »

xwolfi wrote: post #139
eru wrote: post #55

Earlier quoted context omitted.

> By definition, a model can't "know" things that are not somewhere in its training set, unless it can use a tool to query external knowledge. Well, it could also make inferences. Like, it could find a new mathematical proof, even if that's never in the training set.

But how, it's not like it's thinking, it's just spitting the next likely token

That is an oversimplification, and not the whole truth. Read anthropic's blog posts if you want to learn more, or ask gpt5.

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

#146

Post by Mars008 »

KingOfCoders wrote: post #117
Mars008 wrote: post #106

Earlier quoted context omitted.

As long as it works who cares about training data. Obviously they can't open it for many reasons. License is one of them.

I don't care if it's open source or not. I care if people call something open source which it isn't. Do you care about binary blobs in the kernel? No. Are binary blobs in the kernel open source? No. But it is tedious to go through the same discussion every 10 years, with a relentless industry that wants to dupe people. If there wasn't a benefit in for them, they would not call it open source.

For some reason people think of models as software and open source should have similar meaning. There are fundamental differences: 1) models aren't reproducible given everything, data, hardware, methodology. 2) they aren't even verifiable. i.e. given model and dataset it's impossible to say if model was trained on that data. 3) except for toys models are trained on copyrighted data. Some of it is private, like users' chats. 4) besides data there is a lot of human input after pretraining.

This means given everything you have two options: 1) train similar model yourself 2) trust model provider. In software you can get script and run, or get code and compile it in exactly the same binaries.

Naturally 'open source' has different meaning. Some are trying to monopolize it, like they know the 'truth'. Others simply ignore it. Eventually we'll settle on something.

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

#147

Post by anshumankmr »

The main aim of Phi3 mini was to be able to run on device, and it had tremendous speed and for a 128K context with 3B something params, its pretty damn good, I had used it for a project myself last year but ultimately we went with the Mistral's models who at the time had the best Open weights models.

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

#148

Post by anshumankmr »

teruakohatu wrote: post #118
wkat4242 wrote: post #112

From the article: > For the same reason that Microsoft probably continued to train Phi-style models: safety. Releasing an open-source model is terrifying for a large organization. Once it’s out there, your name is associated with it forever, and thousands of researchers will be frantically trying to fine-tune it to remove the safety guardrails. I don't think this is really an issue in practice. Llama 2 and 3 were unc…

> researchers will be frantically trying to fine-tune it to remove the safety guardrails. It is a really weak excuse. They are more likely to take a reputation hit for having silly guardrails than for having someone to remove them. Imagine if Bill Gates decided not to release MS Paint in 1985 because someone could have drawn something offensive with it.

Or imagine if Bill Gates decided to not release Comic Sans in the 90s cause someone could have written something offensive with it....

oh wait that wouldn't have been too bad

(/S)

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

#149

Post by itsdesmond »

jofzar wrote: post #143
wkat4242 wrote: post #103

Earlier quoted context omitted.

It's not just erotic role play that the censorship affects. My life involves a lot of sexual discussions and that means that everyday talk, chat summaries, email rewrites or translations will cause the model to shut down. I do the latter a lot especially to find colloquialisms because Google translate is often too literal. It's so annoying. Right now I'm using abliterated llama 3.1. I have no need for vision but I wa…

Sorry just to ask, what kind of job do you have? Sex therapist sounds like the closest?

I don’t think they’re using it for work.

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

#150

Post by kristopolous »

Aeolun wrote: post #96
kristopolous wrote: post #75

Earlier quoted context omitted.

Right, I'm just a technologist. The psychological and sociological parts aren't my bailiwick Typing a prompt in an AI box to make art has fewer real-world victims than performing the acts, filming them, and then sharing the videos. I think that's inarguable. Maybe it's still unadvisable and someone should be in talk therapy. I have no idea. But at least nobody is actually getting molested and retraumatized in the ai…

The model has to be trained on something though. It’s easy to see how this works for art art, because people have been drawing that shit for years, and most people wouldn’t feel to bad about training on it. I don’t think that’s true for the photorealistic models though.

sure but those features have been generalized over very large datasets. Yes, you can have Loras with specific people, convincingly voice clone with higgs, use latentsync, wan, flux kontext, face swap, lots of things ... sure.

This all falls flat on me though. It's like showing me the lewdest most shocking story and then saber rattle about keyboards and word processors.

I'm fully aware of the wild things people do with drawing programs.

We're all adults here. It's fine.

Post reply on HN