For example, the last few years have seen a lot of angry comments about how "the algorithm" (for both social media feeds and Google search results) is politically biased. IIRC we don't know the full set of training data (data includes RLHF responses before anyone points me to the Wikipedia page I already read that lists some data sources) how can we be confident that any model from Facebook has not been specifically trained to drive a specific political agenda?
"The unsafe AI"? Which one would this be? Would it be the one which so many people on this very website complain has been "lobotomised"? (No matter how much I object to that word in this context, that's what people call it).
ChatGPT has clearly caused significant negative social impact (students cheating on their essays, SEO spam, etc.) and they didn't give a shit.
Almost every dangerous thing I can think of "has it both ways" by the standard you apply here.
I can use an aircraft without being a registered pilot; I can use the police without being a trained law enforcement officer; I can use restricted drugs when supplied by a duly authorised medical professional; I can use high voltage equipment, high power lasers, high intensity RF sources, when they are encased in the appropriate safety equipment that allows them to be legally sold as a lightbulb, a DVD player, and microwave oven respectively.
The weights themselves reveal any and all information found within the model, regardless of superficial attempts to prevent the model "leaking". We do not, at present, actually know how to locate information within a model as would be required to confirm that it has genuinely deleted some information rather than merely pretending — this is an active field of research.
By analogy: data on a spinning hard drive. In certain file systems, if you delete a file, you only remove the pointer to it, the actual content can be un-deleted. A full overwrite of unused space is better, but owing to imprecision in the write head, even this is not actually certain to delete data, and multiple passes are used — but even this was not sufficient for the agents who oversaw the destruction of The Guardian's laptop containing a copy of the Snowden data.
At present, we do not even know how to fully determine the capabilities of a set of weights, so we cannot actually tell if any given model is "safe" (not even by a restricted definition of "safe" as in "we won't get sued under GDPR Article 17"), we can only guess.
And those best-guesses are what people complain about when they are upset that an AI model no longer does what it did last week.
There is an argument that having open access to the weights makes it easier to perform the research necessary to even be able to answer this question. That's an important counter! But it's not even remotely where you're going with your comment.
> ChatGPT has clearly caused significant negative social impact (students cheating on their essays, SEO spam, etc.) and they didn't give a shit.
None of those are significant social impact. Negative, sure, but the "significant" risk isn't SEO spam, it's giving everyone their own personal Goebbels; and it's not "students cheating" because, to the extent the AI is capable of that, the tests they can cheat on with it represent something that is now fully automated to the point that it shouldn't be tested — I mean, I grew up with the maths teachers saying "you won't have a calculator on you all the time when you're an adult", but by the time I got to my exams some of the tests required a calculator, and now I regularly do ⌘-space and it does the unit conversion as well as the maths, and even for the symbolic calculus I can usually just ask Wolfram Alpha… which is on my phone, which I have on me all the time.
And noticing your lack of argument in support of a company you confuse with a sport club is not an insult, by the way.
Altman and the safety narrative is hypocrite BS, and your clumsy words aren't convincing anyone but you that it is not.
I have no idea why you think I think OpenAI is a sports club. That's so weird I have to assume it's an auto-corrupt and not even what you intended to write.
I have no idea why you think a bunch of counter-examples isn't an argument, or why they are shallow — they demonstrate that you are just plain wrong. The only other people I've met who think that counter-examples weren't arguments, were… as I recall, two biblical literalist-fundamentalists and one politician. Oh, and someone who refused to accept that encryption was a good idea and government backdoors a bad idea, but otherwise I can't categorise them because they were a random Twitter account and it was a decade ago. And, now I think about it, someone who got themselves banned from HN for repeatedly insulting anyone who preferred electric/PV over hydrogen/nuclear.
And given that it is your previous comment which is in the grey as I write this, I think you need to look in a mirror before calling my words "clumsy".
Not understanding something is OK, taking pride of it like that is quite something though.
> I have no idea why you think a bunch of counter-examples isn't an argument
Using counter-arguments is fine, as long as they are on point, which isn't the case here. No aircraft maker, for instance, ever claimed they were designing in aircraft to protect people against the dangers aircraft cause to mankind.
And that's why no example taken from the existing world can help defend OpenAI' hypocrisy, because no other company in the world started as a non-profit with grandiose claims like they did.
> or why they are shallow
Your writing is shallow because it goes in every direction without logical structure or exhibition of coherent thinking. You're jumping from one idea to another without articulating them.
And this latest comment of yours is also a good illustration of this, with the majority of your comment being rambling about random people you've met in real life or over the internet. Don't you realize it doesn't bring anything to the point you're trying to make and does a great disservice to your argumentation?
> And given that it is your previous comment which is in the grey as I write this, I think you need to look in a mirror before calling my words "clumsy".
Oh no, one individual OpenAI fan[1] downvoted my comment, my day is ruined.
[1] an here you see the explanation for the “sport club” point I made earlier, for some reason a bunch of people seem to believe that it's sensible to be fans of particular company and defend them in all situations over the internet. Apple fandom is the canonical example of that kind of behavior, but nowadays Tesla and OpenAI now have a fanbase with similar zealotry.
The Meta offerings should be under the umbrella name "FreeAI" analogous to Free Software (e.g. FSF) vs the more commercial leaning Open-Source. Heck everyone should contribute to it and make it a movement so "Free" is both an adjective and a verb.
I don't think so. It's a company, not a person. Meta doesn't have intentions, just incentives. And if the incentives of a company are aligned with publishing open science and open software that's as good as it gets.
I don't require for profit businesses to do good things because they love world peace or are altruistic. Meta making its money from its consumer facing products that nobody is forced to use and having the models out in the open is exactly how it should be.
> As we get closer to building AI, it will make sense to start being less open. The Open in openAI means that everyone should benefit from the fruits of AI after its built, but it's totally OK to not share the science (even though sharing everything is definitely the right strategy in the short and possibly medium term for recruitment purposes).
Words like this suggest a failure of imagination. Given it's their own mouth and they're writing in their own defence, what might be a more generous interpretation?
"The article is concerned with a hard takeoff scenario: if a hard takeoff occurs, and a safe AI is harder to build than an unsafe one, then by opensorucing everything, we make it easy for someone unscrupulous with access to overwhelming amount of hardware to build an unsafe AI, which will experience a hard takeoff."
The article in question appears to be: http://slatestarcodex.com/2015/12/17/should-ai-be-open/
Which opens with:
"""H.G. Wells’ 1914 sci-fi book The World Set Free did a pretty good job predicting nuclear weapons:
They did not see it until the atomic bombs burst in their fumbling hands…before the last war began it was a matter of common knowledge that a man could carry about in a handbag an amount of latent energy sufficient to wreck half a city"""
"""Once again: The decision to make AI findings open source is a tradeoff between risks and benefits. The risk is that in a world with hard takeoffs and difficult control problems, you get superhuman AIs that hurl everybody off cliffs. The benefit is that in a world with slow takeoffs and no control problems, nobody will be able to use their sole possession of the only existing AI to garner too much power.
But the benefits just aren’t clear enough to justify that level of risk. I’m still not even sure exactly how the OpenAI founders visualize the future they’re trying to prevent. Are AIs fast and dangerous? Are they slow and easily-controlled? Does just one company have them? Several companies? All rich people? Are they a moderate advantage? A huge advantage? None of those possibilities seem dire enough to justify OpenAI’s tradeoff against safety."""
and
"""Elon Musk famously said that AIs are “potentially more dangerous than nukes”. He’s right – so AI probably shouldn’t be open source any more than nukes should.""
This is what OpenAI and Musk were discussing in the context of responding to "I've seen you […] doing a lot of interviews recently extolling the virtues of open sourcing AI, but I presume you realise that this is not some sort of panacea that will somehow magically solve the safety problem?"