The ethics problem is an artifact of googles model of trying to keep their AI under lock and key and carefully controlled and opaque to outsiders in how the sausage gets made and what it’s made out of. Ultimately I think many of these products will fail because there is a misalignment between what Google thinks you should be able to do with their AI and what people want to do with AI.
Whenever I see an AI ethicists speak I can’t help but think of priests attempting to control the printing press to prevent the spread of dangerous ideas completely sure of their own morality. History will remember them as villains.
Good researchers won't work somewhere that doesn't allow the publishing of papers. And without good researchers, you won't be on the forefront of tech. Thats why nearly all tech companies publish.
Interesting analogy. Google, like the priests, is acting out of mix of good intentions (protecting the public from perceived dangers) and self-interest (maintaining secular power, vs. a competitive advantage in the AI space). In the case of the priests, time has shown that their good intentions were misguided. I have a pretty hard time believing that history will be as unkind towards those who tried to protect minorities from biased tech, though of course that's impossible to judge in the moment.
Most of the ethicists I see actually doing gatekeeping from direct use of models--as opposed to "merely" attempting model bias corrections or trying to convince people to avoid its overuse (which isn't at all the same)--are not trying to deal with the "AI copies our human biases" problem but are trying to prevent people from either building a paperclip optimizer that ends the world or (and this is the issue with all of these image models) making "bad content" like fake photographs of real people in compromising or unlikely scenarios that turn into "fake news" or are used for harassment.
(I do NOT agree with the latter people, to be clear: I believe the world will be MUCH BETTER OFF if such "bad" image generation were fully commoditized and people stopped trying to centrally police information in general, as I maintain they are CAUSING the ACTUAL problem of misinformation feeling more rare or difficult to generate than it actually already is, which results in people trusting random people because "clearly some gatekeeper would have filtered this if it weren't true". But this just isn't the same thing as the people who I-think-rightfully point out "you should avoid outsourcing something to an AI if you care about it being biased".)
In the name of protecting [minorities, child, women, lgbt, etc] many harms will be done.
I don't think "don't let the plebes have the models" is a good stance. But neither is pretending that the ethics and bias issues aren't here.
Most people represented in photos are younger. Same story.
The problematic issue is the media has morphed reality with unreal images of people/families that don't match society so unreal expectations make people think that having white people generated from a white dataset is problematic.
There are lots of other ways you could get training data, but they might not be so cheap. You could have humans give English descriptions to images from other language contexts. I'm guessing there's interesting things to do with translation. But all the weird stuff about bodies, physical objects intersecting etc ... maybe it should also be rendering training images from parametric 3d models? Maybe they should be commissioning new images with phrases that are likely to the language model but unlikely to the image model. Maybe they should build classifiers on images for race/gender/age and do stratified sampling to match some population statistics (yes I'm aware this has its own issues). There are lots of potential technical tools one could try to improve the situation.
Implying that the whole world must change before one project becomes less biased is just asking for more biased tech in the world
It's 2022 and we can be more thoughtful. Yes there are tradeoffs between unleashing new capabilities quickly vs being thoughtful and potentially conservative in what is made publicly available. I don't think it's bad that Google makes those tradeoffs.
FWIW Google open sources _tons_ of models that aren't LLMs / diffusion models. It's just that LLMs & powerful generative models have particular ethical considerations that are worth thinking about (hopefully something was learned from the whole Timnit thing).
As for learning from the timnit thing I'm pretty sure the only thing people outside Google learned from that is that Google ai "ethicists" all seem to be crazy. Certainly that's the clear vibe on this thread.
You can let your kid use Google to look up math lectures without fearing that they would see something slightly traumatizing though, right? That wasn't the case in 1996! The point is that products have varying levels of readiness, and it's totally fair to say "the thing isn't ready, it has too many sharp edges." Especially when the thing could be used at scale.
> As for learning from the timnit thing I'm pretty sure the only thing people outside Google learned from that is that Google ai "ethicists" all seem to be crazy. Certainly that's the clear vibe on this thread.
That's a sad take, but who knows if it's true. HN commenters aren't exactly a representative sample.
Imagen and Imagen Video is not released to the public at all. You might be confusing it with OpenAI's models.
"We have taken multiple steps to minimize these concerns, for example in internal trials, we apply input text prompt filtering, and output video content filtering. However, there are several important safety and ethical challenges remaining. Imagen Video and its frozen T5-XXL text encoder were trained on problematic data. While our internal testing suggest much of explicit and violent content can be filtered out, there still exists social biases and stereotypes which are challenging to detect and filter."
I think it's gotten a ton better vs 10 years ago, and is getting better still.
More on topic -- when folks here complain that Google can't release these models, it's not like they're just sitting there using that as an excuse -- Google has entire teams dedicated to ML safety trying to figure out how to filter out bad stuff, make models fairer, and avoid situations like M$FT's "Tay" (or worse).
With Stable Diffusion I think they just didn't expect someone to produce a truly open version. There are plenty of AI models that Google have made where they've maintained a competitive advantage for many years by not releasing the code/models, e.g. speech recognition.
They won't be using the models they train to commit crimes, for example. Someone who gets access to their best models may very well do that. It'd be really funny (lol, no) if Google's abuse team started facing issues because people are making more robust fake user accounts...by using google provided models.
I'm sorry to be sarcastic. I generally try not to be, but I just can't fathom the level of naivete required to think that mega-corps act out of their moral responsibility rather than their profit-interest.
But even with that all being true, real people (typically some thoughtful researchers) build these models. And my point is: _there really are ethical reasons to keep large generative models trained on flawed data away from the general public until better safeguards are in place._ You can verify this for yourself by reading about ML bias and safety. Don't let cynicism keep you from internalizing that fact. OpenAI didn't make GPT-3 widely available for the same reason.
At the end of the day, Google doesn't need an excuse like "we have ethical qualms" to not release the models. Stuff that is really secret sauce you won't hear about until many years later when it's not a competitive advantage anymore. Google _does_ need to cover its ass and not deal with its employees yelling that it helped perpetuate algorithmic racism, or surveillance state, or increased levels of inauthenticity on the internet.
When I said "Google has a responsibility" -- I don't mean that the faceless entity feels responsibility, I mean the people who work on the specific things have a responsibility and they do feel & act on that. If you work on lifesaving drugs that could also be dangerous / addictive, it's kind of on you to be thoughtful about how to make them generally available, no?
What reasons necessitate keeping image or video generation models private that wouldn't also argue for keeping animation software or picture editing tools private? Should we somehow prevent such tools from getting better or easier or stop people from educating others on how to use them?
No, that's crazy. If the tools are so dangerous we can't trust the public to have them then they are way too dangerous to trust Google with them. If it were actually true that Google was developing AI too dangerous for the public, then we should storm the Google headquarters, kill their engineers, and burn their data centers.
Of course it's not true. Google is developing image and video generation models and equivalent versions will be open source by the year's end I expect. These models aren't especially dangerous. Yes, people will use them to be racist or mean, same as they use their phones or computers or books or whatever to be those things.
As a final note, it's obviously not true that GPT-3 was kept private for the "ethics" reason. I can buy GPT-3 generations now for 2 cents per 1k tokens generated. There is no real oversight into how these generations are used and you could absolutely use them to power social media bots or whatever you are concerned with. The reason they keep GPT-3 private but sell access to it is not because they want to be ethical, but because they want to sell access to it.
This is not a fair characterization of what's going on here. Google spent a ton of money on researchers & training infra (it's wildly expensive even just hardware-wise) to train these models. It's not different from other proprietary technologies -- they don't owe the public anything here. Providing the research findings + methodology in a paper without the implementation & data is a _tradeoff_ as a participant in the field. If someone else implements the model with their money and uses it for nefarious purposes, that's more acceptable than if they directly use Google's _already known to be flawed_ models.
> I'm curious what ethical reasons you think require that new technology only be used in secret and without oversight by trillion dollar companies. This is supposed to be AI safety?
If I make a chair and I know it's not always safe to sit on, maybe I should not sell that chair. We can talk about this proof-of-concept chair as a research subject, but if you go to build one and use it to prank someone, that's on you.
That's all that's going on here. If the model could be used to generate CSAI, maybe Google doesn't want to be part of that.
> Google is developing image and video generation models and equivalent versions will be open source by the year's end I expect. These models aren't especially dangerous.
Maybe that's the disconnect -- you don't think generative models are dangerous, but they can be, and Google would know because they have entire teams dedicated to AI fairness & safety researching this topic.
It's also not trivial to reproduce these models. Given the cost to simply train even if you had the source data, any organization releasing these models has to have a bit of money and skill. The onus will always be on the team building these models to think about what their ethics are and how they want to proceed knowing there may be negative externalities.
> Yes, people will use them to be racist or mean, same as they use their phones or computers or books or whatever to be those things.
Tools empowering large-scale inauthenticity & disinformation are not comparable to individuals making comments.
You say that Google doesn't "owe the public anything" and that may, or may not, be true from a legal standpoint, but obviously, from a norms, ethical, and moral standpoint Google does have a massive obligation to the public that they are breeching. Google uses the public's data to train, public research, and publicly shared models to iterate on. Then, after building on the shoulders of giants, Google refuses to share what they have built in contravention of the norms that they benefit from.
Regarding your chair metaphor - the "danger" of these models, if there is such, is not that they would hurt the user, like a faulty chair, but that they could be used to hurt others - e.g. a bot army to manipulate public opinion or create fake news. Google isn't building a chair that might break and hurt the user then, but a gun that might hurt others. It's true that guns shouldn't be widely available - not even a die hard libertarian would want a child to have access to a gun, but the entity that sets rules regarding availability is a representative government for the people for whom those rules are being set - not a private company. In other words, if these tools can cause harm they should be regulated by the government, not Google. If the tools are dangerous, that is not an argument that Google should keep them secret.
Google's not doing this (LLM, generative image model) research on academic datasets freely shared with them. They're doing this research on data they gathered at their expense. This is not a violation of academic norms. Again, Google shares a lot of datasets and models, just not LLMs and generative sets trained on problematic source datasets.
> As I characterized previously Google is able to do whatever they want, conceal their results, and impede progress and understanding because they aren't sharing their results. You say this isn't a "fair characterization" but it is exactly what is happening - which part is wrong?
Anyone can do research and not share back to the community. Google _does_ share back to the community in the form of papers (and again, very frequently with models and datasets). If you have the money and expertise to implement the papers, more power to you. Every technology company has some secret sauces they don't share with everyone. That Google may have some of those is not a moral failing.
> from a norms, ethical, and moral standpoint Google does have a massive obligation to the public that they are breeching. Google uses the public's data to train, public research, and publicly shared models to iterate on
From the other end: Google gets user data and has a responsibility to not proliferate that data, no? I wouldn't want them to share a dataset that has my personal data, even if anonymized because there are ways to deanonymize. There are levels to everything, and choosing "I'll release the paper but not the model + data" for some potentially sensitive models seems sane.
> Then, after building on the shoulders of giants, Google refuses to share what they have built in contravention of the norms that they benefit from.
People are building on the shoulders of Google's research all the time, and plenty of companies are doing similar things to Google and being way less open about their work. I mean, every company that trains a big model on data collected from the public -- are they all required to share their models with everyone? Is Cruise sharing their pedestrian detection model? I don't think what you're suggesting could possibly be the standard.
> Regarding your chair metaphor - the "danger" of these models, if there is such, is not that they would hurt the user, like a faulty chair, but that they could be used to hurt others - e.g. a bot army to manipulate public opinion or create fake news.
Sure, I was trying not to be hyperbolic and compare LLMs to guns since they have plenty of awesome use cases (whereas guns really don't). A faulty chair that you set out for anyone to use can hurt people other than the chair's creator / people who are aware of the specific risks. But yeah, seems like you now agree these models have the potential to cause great harm.
> In other words, if these tools can cause harm they should be regulated by the government, not Google
I agree that gov't regulation can be helpful for setting a minimum standard. But I strongly disagree that lack of laws means we should abdicate our own moral responsibilities. If I sell / provide something, I need to be able to sleep at night knowing I didn't make the world worse. Googlers typically try to do this.
Well, they do have the "special claim" of inventing the model and not owing its release to anyone.
Said "public research" didn't come with a requirement to release anything you build on top of it. This would pretty much be the research equivalent of compelled speech. Luckily, not happening.
But now with these models they have such a ridiculously heavy handed approach to the ethics and morals. You can't type any prompt that's "unsafe", you can't generate images of people, there are so many stupid limitations that the product is practically useless other than niche scenarios, because Google thinks it knows better than you and needs to control what you are allowed to use the tech for.
Meanwhile other open source models like Stable Diffusion have no such restrictions and are already publicly available. I'd expect this pattern to continue under Google's current ideological leadership - Google comes up with innovative revolutionary model, nobody gets to use it because "safety", and then some scrappy startup comes along, copies the tech, and eats Google's lunch.
Google: stop being such a scared, risk averse company. Release the model to the public, and change the world once more. You're never going to revolutionize anything if you continue to cower behind "safety" and your heavy handed moralizing.