(no title)
ilyasut | 7 years ago
- ML is getting more powerful and will continue to do so as time goes by. While this point of view is not unanimously held by the AI community, it is also not particularly controversial.
- If you accept the above, then the current AI norm of "publish everything always" will have to change
- The _whole point_ is that our model is not special and that other people can reproduce and improve upon what we did. We hope that when they do so, they too will reflect about the consequences of releasing their very powerful text generation models.
- I suggest going over some of the samples generated by the model. Many people react quite strongly, e.g., https://twitter.com/justkelly_ok/status/1096111155469180928.
- It is true that some media headlines presented our nonpublishing of the model as "OpenAI's model is too dangerous to be published out of world-taking-over concerns". We don't endorse this framing, and if you read our blog post (or even in most cases the actual content of the news stories), you'll see that we don't claim this at all -- we say instead that this is just an early test case, we're concerned about language models more generally, and we're running an experiment.
Finally, despite the way the news cycle has played out, and despite the degree of polarized response (and the huge range of arguments for and against our decision), we feel we made the right call, even if it wasn't an easy one to make.
Cacti|7 years ago
If this is your whole point, then I think you are missing something fundamental. Implementing these models doesn't require reflection, or introspection, or any sort of ethical or moral character whatsoever; and even if it did, all that will happen eventually is someone (without the technical background) will simply throw a lot of money at someone else (with the technical background, but who needs to, you know, eat, and pay rent, and so on) to implement it. You are fooling yourself if you think your stance makes a single mote of difference in this arms race.
bilbo0s|7 years ago
In fairness, if that's true, then no one has any need of her model.
More seriously speaking, why does anyone need, say, "training set x", or "model y", to make their implementation work? You don't. So I don't really understand why everyone is so worked up about not releasing this stuff? If you want to do it, do it. If not, don't. But there's no need to say, "I demand everyone do it, and I'll have a meltdown if they don't."
pishpash|7 years ago
sigil|7 years ago
I’m also confused by the threat models earnestly put forth in your blog post. Are we really concerned about deep faking someone’s writing? The plain word already demands attribution by default: we look for an avatar, a handle, a domain name to prove the person actually said this.
albntomat0|7 years ago
It seems more like the "nukes are safer when multiple rational state level actors have them", rather than anyone able to pull a git repo.
smsm42|7 years ago
Mostly it's scary not because it's good - as writing goes, it's quite bad. It forms coherent sentences, but otherwise it's nonsense. I've seen similar nonsense producers in early 90s on basis of Markov chains and what not.
No, the scary part is how much it reminds me of what I am reading in the media all the time. My current pet concern is that AIs will start passing the Turing test not because AIs are getting so good but because humans are getting so bad. A bunch of nonsensical drivel can easily be passed as a thoughtful analysis or a deep critical think-piece - and that's not my conjecture, have been repeatedly proven by submitting such drivel to various academic journals and it being accepted and published. I'm not saying people are losing critical thinking skills - but they are definitely losing (or maybe never even had?) the habit of consistently applying them.
esjeon|7 years ago
Exactly. When it comes to generating a large volume of apparently-good sentences, non-AI (or classical) approaches are still better than good. Those will be equally disruptive, since the defending side is yet to develop a proper countermeasure based on the "sensible"-ness of content. Plus, they will be much easier to customize and adapt to the situation, while ML-based solutions often need remodeling and retraining when repurposed.
> My current pet concern is that AIs will start passing the Turing test not because AIs are getting so good but because humans are getting so bad
AI will start deceiving the public even before it pass Turing test. It's much harder to spot bots amidst people than in a 1vs1 chatroom.
modeless|7 years ago
Only people with a large amount of money and a lot of expertise. What you are doing is the opposite of democratizing AI.
moconnor|7 years ago
Yet from Google we heard nothing. Which is the optimal decision for them - they only lose by blowing the whistle.
avip|7 years ago
imtringued|7 years ago
>Recycling is NOT good for the world.
>It is bad for the environment,
>it is bad for our health,
>and it is bad for our economy.
>Recycling is not good for the environment.
>Recycling is not good for our health.
>Recycling is bad for our economy.
>Recycling is not good for our nation.
The first paragraph keeps repeating the <X> is <bad | not good> for the <Y> pattern 8 times.
>And THAT is why we need to |get back to basics| and |get back to basics| in our recycling efforts.
"get back to the basics" is repeated twice in the same sentence.
>Everything from the raw materials (wood, cardboard, paper, etc.),
>to the reagents (dyes, solvents, etc.)
>to the printing equipment (chemicals, glue, paper, ink, etc.),
>to the packaging,
>to the packaging materials (mercury, chemicals, etc.)
>to the processing equipment (heating, cooling, etc.),
>to the packaging materials,
>to the packaging materials that are shipped overseas and
>to the packaging materials that are used in the United States.
It literally repeated packaging 5 times in the same sentence and the overall structure was repeated 9 times. Also what type of packaging is based on mercury?
pas|7 years ago
(This of course doesn't make it an amazing feat of computer engineering.)
The overarching narrative is great, but that's probably driven by the great antithesis supplied by the experimenter.
It'd be interesting to know how this works, what happens if less or more is given as thesis/antithesis/assignment, and after how much output it turns into gibberish (or repeats).
mycorrhizal|7 years ago
Heck, maybe having to compete with this will raise human discourse (Joking).
smsm42|7 years ago
eslaught|7 years ago
Have you done a plagiarism search on that text to see how similar it is to the input corpus? I'm by no means an ML expert, but I've played around with models for random name generation and one thing I've noticed is that as the models become more accurate, they also become much more likely to just regurgitate existing names verbatim. So if you search the list of names and notice something that seems particularly realistic, it could be because it's literally taken in whole or in part from the training data set!
czr|7 years ago
(The talking unicorn example on their page is also meant to demonstrate that, no, it's not just memorizing, but I think it's a bit more compelling to check from the raw samples)
malux85|7 years ago
How is that open?
How is that not centralization of power?
albntomat0|7 years ago
pas|7 years ago
GistNoesis|7 years ago
Here are a few that comes to mind.
-Secrecy? but how will you continue to exist on the PR scene if you don't release anything?
-Are you willing to pay every developer who is able to replicate your paper, more than what the black market would pay?
-How are you working on incentive alignment to make sure that all people who can replicate your results have more incentive to do good than bad, specially in the current environment where users and valuable data are silo-ed by a few companies?
-Misdirection to keep an edge, i.e. planting bugs/ Not fixing bugs for public ; spreading false results; only working on problems that need high resources to limit the number of actor who will be able to replicate ?
-Tracking the people who have the competence to replicate and take preemptive measures.
-Restrictions on GPU/CPU/silicone wafer.
Who can regulate? How can we regulate? What are the negative consequence of regulation? What happens if we don't, at what odds and time horizon?
cs702|7 years ago
That said, withholding the pretrained models probably won't make much difference, because bad actors with resources (e.g., certain governments) will be able to produce similar or better results relatively quickly.
All it will take is (1) one or two knowledgeable people with the willingness to tinker, (2) a budget in the hundreds of thousands to a few millions of dollars at most, and (3) a few months to a year. Nowadays a lot of people are familiar with Transformers and constructing and training models across multiple GPUs.
xg15|7 years ago
Ok, accepting that premise, what people/organisations would you share the research with and based on what criteria?
hooloovoo_zoo|7 years ago
iamcreasy|7 years ago
One of the reason Elon distanced himself because of what OpenAI team wanted to do. I am wondering if this new paper has anything to do with that? Or what it is in general that Elon doesn't agree with what OpenAI is doing?
Thanks!
onurcel|7 years ago
pishpash|7 years ago