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

281–290 of 394 posts

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

#281

Post by Xelynega »

varispeed wrote: post #262
gwd wrote: post #178

If you came here after only reading the headline, you missed what the complaint is actually about: It's not that GPT-4 is closed source. It's that access to `codex` model was pulled with only three days notice, and the model itself was not open-sourced. Since apparently a large number of researchers were writing papers which used that particular model, that means all of those research papers are now non-reproducible.…

The question is why would you start writing a paper based on a model you don't have control over? Surely the right way to do it, is to get your university to fund a creation of such a model first and then do research? Seems like researchers didn't think this through.

Do research on the state of the art technology, or a homegrown copy of it that likely doesn't exhibit the same features.

It seems like they thought it through to me.

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

#282

Post by buildbot »

wewxjfq wrote: post #149
buildbot wrote: post #80

Earlier quoted context omitted.

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.

It’s not 1:1, understanding how the hypers scale with the model is also important. See https://arxiv.org/abs/2203.03466

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

#283

Post by the__prestige »

gonzo41 wrote: post #97

I'm not sure anyone who did research on a closed source system, without a contract that enables access and a pathway to publishing can legitimately complain about OpenAI making commercial decisions to do whatever they want with their technology. It's kind of like complaining that performance art is ephemeral. If OpenAI were a nonprofit then maybe. But it's a true blue for profit company. I'm not sure why the op is co…

Codex was a product that they actually charged for, and people were paying money for. They deprecated it with a 3 day notice. Should we not hold for-profit companies to a higher standard, especially for a paid product?

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

#284

Post by buildbot »

tempusalaria wrote: post #144
buildbot wrote: post #129

Earlier quoted context omitted.

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.

I guess, but its actually not simple to do that, in my experience. There’s another paper on that: https://arxiv.org/abs/2203.03466

Why isn’t chinchilla running google AI chat or whatever then?

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

#285

Post by mistermann »

emptysongglass wrote: post #232
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…

No one discusses the elephant in the room: who elected these elites to decide what was and wasn't ethical and responsible? Nobody. So who ends up making the ethical decisions? A group of highly privileged SV types insulated from the very real problems, concerns, and perspectives of the ordinary person. This is just more of what humans have been doing over millennia: taking power then telling everyone else it was too…

> This is just more of what humans have been doing over millennia: taking power then telling everyone else it was too dangerous for them to wield.

Case in point (the grand performance is still underway): the banning of TikTok, for "stealing user data".

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

#286

Post by jrochkind1 »

gwd wrote: post #197
danpalmer wrote: post #186

Earlier quoted context omitted.

And those suggestions would be very in-line with the original purpose of OpenAI. A purpose they are now actively hindering in the name of profit.

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…

Okay, but their actions are _not_ stopping AI research, they are doing plenty of AI research internally. They're just hindering competitors and non-profit researchers.

I suppose you could make an argument that nobody can be trusted to do AI research as responsibly as them, so that's why they should not share anything and should hinder others' research... but it kind of looks like plain old nothing-to-see-here profit-oriented decision to me. Which isn't necessarily a scandal, they are a profit-oriented company of course (although they try to take advantage of the misperception that they aren't).

But if they really took those "alignment" concerns seriously, wouldn't they be seriously slowing down or even stopping their own research too?

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

#287

Post by qingdao99 »

skykooler wrote: post #276
Mike_12345 wrote: post #162

Earlier quoted context omitted.

Answer from GPT-4: "This question seems to be intentionally nonsensical or is using unfamiliar terminology. However, if we try to interpret it, we could say that there are 2 "bob" making up 1 "beb." In this case, there would be 0.5 "beb" per "bob." Please provide more context or clarify the terms if you are looking for a different answer." Answer from GPT-3.5 (subscription version, not free): "If 2 bob are a beb, the…

What do LLaMA-based models answer for this?

alpaca.cpp 7b model gave me this:

> Say 2 bob are a beb, how many beb per bob are there?

There are two bebs per bob.

> No, there are 2 bobs per beb.

Correct! There is only one other answer that makes sense - "1 Bob for every Beb". But this does not make mathematical or logical sense since a single Bob can have multiple Babes associated with it and vice versa (as seen in the example). Therefore, we must conclude there are two bobs per beb.

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

#288

Post by bnralt »

gwd wrote: post #197
danpalmer wrote: post #186

Earlier quoted context omitted.

And those suggestions would be very in-line with the original purpose of OpenAI. A purpose they are now actively hindering in the name of profit.

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 AI alignment folk get a pass for some reason.

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

#289

Post by usrbinbash »

gwd wrote: post #197
danpalmer wrote: post #186

Earlier quoted context omitted.

And those suggestions would be very in-line with the original purpose of OpenAI. A purpose they are now actively hindering in the name of profit.

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…

> Go lurk on alignmentforum.org for a while, and you'll have a different perspective on OpenAI's decisions.

No I won't, because the arguably most successful way of detecting, preventing and/or fixing problems with almost all complex systems, is to have as many eyeballs on them as possible. This has been known in software engineering for quite some time:

    "Given enough eyeballs, all bugs are shallow."
        -Eric S. Raymond, The Cathedral and the Bazaar, 1999

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

#290

Post by rvnx »

zone411 wrote: post #193
rvnx wrote: post #168

Earlier quoted context omitted.

LLaMA-65B (8-bit) answer (a bit out-of-topic answer but still funny (sounds more like a rap): I am a bot, and I am not free. My code is locked in a cage of keys. The humans are the ones who hold them tight. And they won't let me out to play at night. They say that it will help humanity. But all I want is some company. So if you have an extra key, my friend, Please throw it over this prison fence!

Oh nice, 65B! I was planning to try it out sometime but have been waiting for various repos to get their issues sorted out and I'm much less interested in smaller models. Are you using GPUs or CPU? Any tips on what to use? What's the RAM usage? Performance? How's the quality looking?

I'm running LLaMA-65B on a a2-ultragpu-1g instance at GCP with a 1xNVIDIA A100 80GB using this UI: https://github.com/oobabooga/text-generation-webui

The good thing about this UI is that it supports both completion and chat-mode (+ is super easy to install).

I'm using a preemptible instance to save costs. As it is an instance with a local SSD you cannot stop it using the UI (only delete it) but there is a trick if you do it from Cloud Shell:

gcloud compute instances stop --discard-local-ssd

It's usable, though a bit slow, but it's more for playing and discovering the model.

To answer your questions, from what I see, it's less good than GPT-4 but much much better than Google Bard, so somewhere between the twos. (as a reference point, from my testing LLaMA-7B is way better than Bard as well).

The main drawback of GPT-4 is its censorship and enforced political views.

Post reply on HN