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

OpenAI’s policies hinder reproducible research on language models

aisnakeoil.substack.com

Post reply on HN
Print view View original

71–80 of 394 posts

Re: OpenAI’s policies hinder reproducible research on language models

#71

Post by freediver »

esperent wrote: post #56
hhh wrote: post #46

Earlier quoted context omitted.

Look at the system card.

https://cdn.openai.com/papers/gpt-4-system-card.pdf Does anyone have a summary?

It is pretty vague.

- Safety challenges presented by language models need to be addressed through anticipatory planning and governance.

- Content warnings should be provided for potentially disturbing or offensive content.

- Mitigations should be implemented to reduce the ease of producing potentially harmful content.

- Risk areas should be identified and measurements of the prevalence of such behaviors across different language models should be taken.

- AI service providers should be aware of the potential for content to violate their policies or pose harm to individuals, groups, or society.

- Hallucinations should be reduced and the surface area of adversarial prompting or exploits should be reduced.

- Generated content should be checked for accuracy and potential errors should be identified.

- Insecure password hashing should be avoided.

- Instructions should be given to contractors to reward refusals to certain classes of prompts.

- Multiple layers of mitigations should be adopted throughout the model system and safety assessments should cover emergent risks.

Re: OpenAI’s policies hinder reproducible research on language models

#72

Post by blendergeek »

cardosof wrote: post #11

IMO established companies (Meta, Google, etc) had their researchers publish papers as a competitive benefit or way to attract talent from academia (a researcher wouldn't want to stop publishing). Companies didn't see an issue with doing that because those papers were not "giving away" the core of the company, for example, Facebook's DeepFace paper from 2014 couldn't hurt its ad business. OpenAI on the other hand will…

> OpenAI on the other hand will probably be as closed as they can be with their LLMs.

The irony is thick in that statement.

Re: OpenAI’s policies hinder reproducible research on language models

#73

Post by blendergeek »

rfw300 wrote: post #9

The article seems premised on a misunderstanding that OpenAI is a research lab. For all intents and purposes, it’s a for-profit subsidiary of Microsoft, and there’s little financial incentive for it to maintain old models for others’ benefit.

This misunderstanding may have something to do with how OpenAI was originally founded and the name: OpenAI.

Re: OpenAI’s policies hinder reproducible research on language models

#75

Post by ar9av »

sva_ wrote: post #22

Since OpenAI didn't release the parameter count of GPT-4, I've been wondering/doubting if it is really much bigger than GPT-3. The release of GPT-3.5 has shown that they've found ways of drastically cutting down compute costs (an order of magnitude) while maintaining or even improving the quality of the model's outputs. Perhaps the reason that they didn't release the specifics of GPT-4 might be in part due to them wa…

I saw this coming a long time ago and I'm still very pissed off. For three reasons:

1. We are all forced to use the damn "chat" API instead of regular completions. Can't wait to have to deal with chatgpt's conversations in order to get a few lines of code out 2. We loose the super valuable 'insert' and 'edit' modes, which were great for code 3. 3-day notice period? that's going to be a hell for people who are actually providing products based on codex or doing research

Re: OpenAI’s policies hinder reproducible research on language models

#76

Post by raincole »

Eji1700 wrote: post #43

Open AI has been doing sketchyish things long before Chat GPT, and I think it's something people are eventually going to notice more and more (then again people were swearing that Musk walked on water for waaaaaay too long given his actions so fuck if I know). They're 100% marketing FIRST. I don't think they'll outright lie, but they will absolutely screw with their data in such a way to make it look waaay more impre…

If they're 100% marketing first, and still made the most impressive AI product so far, you really need to question what all the other companies are doing.

(before someone says Google or Meta's models are bigger or something... I mean product, not models)

Re: OpenAI’s policies hinder reproducible research on language models

#77

Post by cosmojg »

esperent wrote: post #56
hhh wrote: post #46

Earlier quoted context omitted.

Look at the system card.

https://cdn.openai.com/papers/gpt-4-system-card.pdf Does anyone have a summary?

There's not much content in there, it's mostly fluff about "safety." However, if you're looking for a laugh, grab some popcorn and read the appendix from page 44 onwards. It's an absolute riot.

Re: OpenAI’s policies hinder reproducible research on language models

#78

Post by meghan_rain »

bitL wrote: post #45
redox99 wrote: post #37

Earlier quoted context omitted.

Don't forget about a more efficient attention that let's them get 32k tokens of context.

It's still much worse than 1M context on 16GB VRAM with Reformer, but at the cost of inference speed. And you can use FlashAttention in your own models to get a more efficient/sparse attention now as well.

How could one apply the mentioned technologies to llama/alpaca?

Re: OpenAI’s policies hinder reproducible research on language models

#79

Post by visarga »

ehnto wrote: post #35
mvuksano wrote: post #17

Earlier quoted context omitted.

I don't even care if it's reproducible or not. I care it gives me correct responses to my questions and that's all.

Isn't part of making it reproducible also part of ensuring correct results? Especially if we start putting these models into important systems. And if these models begin to update in an evergreen fashion, or utilize realtime data, getting verifiable or repeatable outputs will be a nightmare if we have no idea how to make these models repeatably.

> if we have no idea how to make these models repeatably

We have idea about how to reproduce but using deterministic training mode is very slow as it loses some optimisations.

Re: OpenAI’s policies hinder reproducible research on language models

#80

Post by buildbot »

esjeon wrote: post #10

I'm quite sure even OpenAI themselves aren't sure if they can reproduce the current models from the scratch. Unless the computing becomes much more powerful and much cheaper, LLM is more or less a rocket science (i.e. hella expensive trial and error). It's not easy to burn lots of dollars just to get what's already there.

Not to be impolite, but this is incorrect. One detail they did share in their paper is that they where able to finetune and select their hyper parameters on a model that needed 1,000x less compute than the final gpt4 model. OpenAI is definitely leading in how to train very large models cost effectively.
Post reply on HN