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

141–150 of 394 posts

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

#141

Post by b800h »

alden5 wrote: post #99
blueorange8 wrote: post #91

Earlier quoted context omitted.

Were you able to integrate any of your data into it yet ?

I wouldn't be able to retrain the model as my computer isn't capable enough, but I can change the prompt to change how the model acts. The prompt i'm currently using is: "Below is an instruction that describes a task. Write a response that appropritely completes the request." That base prompt can be customized to complete specific tasks like classifying text or acting like an assistant.

Out of interest - how capable a computer is required to retrain that model?

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

#142

Post by gnicholas »

throwaway1851 wrote: post #41
atleastoptimal wrote: post #32

I understand any individual's company anti-competitive measures. OpenAI looks at Google the same way Apple looked at IBM in the 80s. What I'm worried about is a lot of the talk about guarding models, public safety and misuse of models will end up leading every big company to pull public access of their APIs. We might look at 2022-2023 as a brief golden age when regular people could use stuff like GPT-4 before it was…

A concern I have about OpenAI is that, if you're using their APIs to develop an application, they can mine your data to compete with you, or even beat you to market. They can do this indirectly, by sharing information with preferred business partners. The conflict of interest, combined with the lack of robust data privacy guarantees, makes me queasy. If serving up generic LLM APIs becomes commoditized -- and I think…

Can you jump back and forth between competing AIs to prevent any of them from seeing the complete picture?

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

#143

Post by nomercy400 »

randomwalker wrote: post #27
mvuksano wrote: post #14

I'm confused why people expect this stuff to be free? I'm surprised OpenAI was so open about their research so far. I don't blame them at all for not publishing the information. This stuff costs real money.

We don't expect it to be free -- please read the article. That's not the issue at all. It's like if you subscribe to a product that you need to do your job, and one day the company tells you that the product is going away in three days and that you need to switch to a different product (that isn't at all the same for your use case).

Maybe you shouldn't build your livelihood on the products of a single for-profit company, which now shows it can remove those products on a whim.

If you want reproducible research, make your own model from scratch, or use an open model. And stop using that company's products, as they cannot be trusted to provide your business continuity.

It is like saying, we are researching Coca-Cola vs Pepsi, but your keep changing the recipe, so give us, researchers, the original recipe.

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

#144

Post by tempusalaria »

buildbot wrote: post #129
whatshisface wrote: post #121

Earlier quoted context omitted.

What good bits did you find? (I'm not sure how fruitful the "OpenAI is a Microsoft department" debate is given that they are almost one and everybody knows it, but I am curious if anyone has found anything good in those many pages.)

I think the most interesting thing is the their ability to predict performance from loss and on a wide range of tasks using a much smaller model - this lets them fine tune their architecture and hypers, then run a single large training run to get full scale gpt4 - from the paper it sounds like they only trained the large model once, then did a Reinforcement learning with human feedback finetune. Disclaimer - I work a…

This isn’t that interesting imo. This is the basic outcome of the scaling laws from Kaplan, Chinchilla papers pushed to a larger final model delta.

They likely did extensive small model building on the gpt-4 architecture to establish hyperparameter scaling laws and then did a predicted build in exactly the same way chinchilla did.

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

#145

Post by jiggywiggy »

mcaledonensis wrote: post #85
ar9av wrote: post #75

Earlier quoted context omitted.

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 actuall…

Completion API for GPT-4 will be there soon. With extra stop tokens, but better than nothing. A compromise. And it's not like what OpenAI did was an impossible magic trick. They've had a right team composition. And three insights. All present in the literature. Repeat that, you'll have GPT-4. But GPT-5. Well, that one is different game. As to being open, they are still relatively open. Consider Apple, for example. No…

Its not interesting. It's a hack to have a don't be evil vibe and keeping the name "open" while they go against their own foundational principles.

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

#146

Post by whym »

mcaledonensis wrote: post #85
ar9av wrote: post #75

Earlier quoted context omitted.

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 actuall…

Completion API for GPT-4 will be there soon. With extra stop tokens, but better than nothing. A compromise. And it's not like what OpenAI did was an impossible magic trick. They've had a right team composition. And three insights. All present in the literature. Repeat that, you'll have GPT-4. But GPT-5. Well, that one is different game. As to being open, they are still relatively open. Consider Apple, for example. No…

From WordNet:

> 1. skittish, flighty, spooky, nervous -- (unpredictably excitable (especially of horses))

(I didn't know the word skittish, and I figured this might help others, too.)

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

#147

Post by AstixAndBelix »

skissane wrote: post #114
beambot wrote: post #101

Earlier quoted context omitted.

> The main way in which IBM mainframes in the 1950s-1970s were "firewalled" was simply by being fiendishly expensive – most people's houses cost significantly less. This was the primary aspect I was referring to, in the same way that training a ChatGPT-like NN can be (or could become) prohibitively expensive. But your comments about openness are relevant on an entirely different axis.

> This was the primary aspect I was referring to, in the same way that training a ChatGPT-like NN can be (or could become) prohibitively expensive. It is fundamentally different though - let's say it costs US$5 million to train a ChatGPT-like system. Someone only has to pay that once, and open source the results, and then everyone else gets it for free. US$5 million is a lot of money for the average person, but a dro…

The problem is that AI research is moving incredibly fast. You might train a LLN today for $5M but a year from now the competition will have implemented an absolutely killer feature that needs $10M worth of training

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

#148

Post by localplume »

runnerup wrote: post #33
pffft8888 wrote: post #21

Earlier quoted context omitted.

In this sense, it's more hacking than crareful and well specified engineering, and that could lead down a path of instability in the product where some features get better while others get worse, without understanding exactly why.

I mean pretty much all real engineering started with that time periods “hacking”/“tinkering” before thorough models and equations were derived. We had 200 years of tinkering with relatively modern steam engine technology before Carnot and Watt started just barely scratching the surface of the first principles of thermodynamics and engine efficiency. Even the eponymous Carnot cycle wasn’t rigorously defined mathematic…

[dead]

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

#149

Post by wewxjfq »

buildbot wrote: post #80
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.

Toying around with a smaller model for hyperparameter search is nothing ground-breaking.

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

#150

Post by lionkor »

ugh123 wrote: post #140
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…

>performance on standardized tests? That doesn't necessarily seem like the best metric for what the LLM tries to be. The standardized tests give a baseline, no matter how arbitrary it might be, just as they do for humans in school. Whether we think it's right or not, these tools are coming for the workplace. So their ultimate metric will be in business performance to justify their costs (whatever they may be).

GPT 3.5 had trouble understanding when I told it "Say 2 bob are a beb, how many beb per bob are there?" and it wrote a goddamn essay about shoes.

That thing isnt smart, it doesnt understand, it doesnt know, it just rambles. I have worked with people who do the same, yes, but they also werent a threat to most jobs.

I said it before, and I will say it again: If ChatGPT 3,4,5,... can take your job, maybe youre not really providing that much value. Make of that what you will - not everyone has to provide huge value.

Post reply on HN