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

321–330 of 394 posts

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

#321

Post by alden5 »

b800h wrote: post #141
alden5 wrote: post #99

Earlier quoted context omitted.

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?

the project for training i found required a nvidia gpu with 16gb of vram, it also would take about 6 hours

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

#322

Post by pixl97 »

snickerbockers wrote: post #70
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 lot of the talk about guarding models, public safety and misuse of models The stuff about models potentially being misused is just their public justification to look like the good guys. They're not going to withhold their technology because they don't want it to be misused, they're withholding it because they want control over who misuses it. Of course, it won't be called "misuse" when the right parties are doing…

Eh, this is rather reductive to the point that the statement is meaningless.

If you release a product in the wild with no safeties at all and then advertize "This product has no safeties at all", you'll likely find yourself in civil court on the losing side of the case.

Now, if you put "some safeties" in the product, the person suing you is going to have a much more difficult and expensive time arguing that in front of the jury.

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

#323

Post by pixl97 »

ChatGTP wrote: post #234
avereveard wrote: post #110

Earlier quoted context omitted.

"GPT-3 175B model required 3.14E23 flops" according to their marketing material. Seti at home was about 1PetaFlops iirc so about 3 years training, possibly less if you can generate enough attention to the project that the people with the beefy devices will partecipate. The problem is that you need to train the full model you can't train aspect of it and even with each node doing independent tiny batches the network b…

How does something of this scale impact climate change? Like when there are 5-6-7 OpenAIs, what does that look like, is this just a huge amount of energy consumption ?

Depends how they source the power. If they are getting it out of nukes or renewables, well, not that much at all.

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

#324

Post by pixl97 »

snicker7 wrote: post #315

The solution is obvious: journals should, as a matter of policy, refuse to publish non-reproducible studies. Reproducibility is the only thing separating science from mythology.

Human aligned AGI is much more apt to happen before what you're suggesting.

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

#325

Post by pixl97 »

matheusmoreira wrote: post #113
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…

We need this technology running locally on our computers as soon as possible.

At current computing growth rates that is still decades away. These things require exabytes of compute to train.

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

#326

Post by HDThoreaun »

bnralt wrote: post #288
gwd wrote: post #197

Earlier quoted context omitted.

I think what most of the people here are missing is how big, how paranoid, and how influential the "AI alignment" movement is. To you it looks like they're being overly careful and paranoid, perhaps as an excuse to set up a monopoly silo to extract money. But a lot of the people the OpenAI researchers work closely with -- people deep in the "AI alignment" community -- are telling them that they're being wantonly reck…

The danger the AI alignment folk are afraid of is completely impossible with current tech, but they want to put up barriers because we have no idea what future tech might look like and there’s the possibility some future advance could be very dangerous. When anti-GMO or anti-nuclear folk used this same standard to put up barriers to research into nuclear or GMO research, they get lambasted for being anti-science, but…

The anti-gmo/nuclear people have no explanation for how things can go wrong. The AI alignment people do. You might not agree with it, but tons of AI researchers, including many at openAI, do.

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

#327

Post by sverona »

hamburga wrote: post #306
sverona wrote: post #223

Earlier quoted context omitted.

It's possible, but the views of the AI alignment community so far as I can tell are being skewed way too far towards nihilistic doomerism by the influence of Yudkowsky, who apparently believes that we're all gonna die in a few years and there's nothing anyone can do to stop it. [0] [0] https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a...

^ Thanks for that link. The doomerism is brilliant and clear and imaginative and absolutely worth reading and grappling with. I personally have no good response to how we deal with sufficiently advanced AI’s capacity to trick and manipulate us into doing catastrophically bad things.

His argument is essentially "a superintelligence who is better than us at everything and thinks a quadrillion times faster can do whatever it wants and we are powerless to stop it."

Yeah, I could've told you that.

If we really are going to create such an intelligence in the next five years, then we had a good run, so long and thanks for all the fish. But that assumption coupled with the security mindset he brings to the table (viz. "the only unhackable computer is an unplugged one at the bottom of the ocean") is so strong that the big list of doom vectors he comes up with appears much scarier than it actually is.

In the past I've struggled with intrusive thoughts that the government is going to come and murder me. I could have given you reasonable-sounding explanations for why I believed this. Doesn't mean it's gonna happen.

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

#328

Post by HDThoreaun »

zarzavat wrote: post #181
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…

ChatGPT-4 is definitely slower than GPT-3.5 (and way slower than 3.5-turbo). What could be the reason for that other than much larger parameter count? I agree that the capabilities seem overhyped. In my subjective experience, 4 seems a little better than 3.5 but not by a huge amount. We just have OpenAI’s cherry-picked word that it‘s this incredible advance.

> 4 seems a little better than 3.5 but not by a huge amount.

Depends on the task. 3.5 was completely incapable of doing math, but 4 seems to be able to at a solid highschool graduate level.

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

#329

Post by BeetleB »

anonymousDan wrote: post #292
robertlagrant wrote: post #258

Earlier quoted context omitted.

I don't really understand this - it's like trying to explain a colleague's behaviour by saying they're doing something so they get their salary. Of course they need to have commercial gain in mind. But you need to be more specific.

From my reading of the parent's comment, they are saying the reason the models are not being made available is because of a fear they will effectively turn into SkyNet - am I being uncharitable?

Yes.

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

#330

Post by pixl97 »

whatshisface wrote: post #120
sillysaurusx wrote: post #60

Earlier quoted context omitted.

Au contraire, no one knows how large GPT-4 is, which is the single best predictor of performance (for a model trained to convergence). The GPT-4 paper spent much of its time writing about this — they did some small scale experiments with 1/1000th the compute, then picked a loss level they wanted and trained GPT-4 till it got it. Neither the exact loss level nor the number of parameters are revealed by the paper. Unfo…

I can't believe anyone considers a single number, which would work about equally well if it were 10% higher or lower, to be a trade secret.

It's about causing your competition to waste millions of dollars in compute time and power doing something unproductive.

There is not a huge pile of excess TPUs laying around for people to use. Any strategic advantage can quickly compound and put you well ahead of others.

Post reply on HN