The head of AI at Nvidia, who is very aggressive in her AI Ethics activism, recently put a screenshot of a list of names, which included mine, on a public tweet. In the tweet, she described the list as containing "fan boys" of an academic critical of "wokeness" as he calls it. I didn't know anything about the academic. I was added to the list because I commented on her tweet attacking him and asked her if she had engaged in any verbal discussions with him. That's it. I asked a question, and was put on a list proclaiming everyone in it as alt-right bigots. I found out about this because a former colleague called me and told me I was on the list. I didn't follow the guy she was attacking or anything like that. Not that it would have been justified if I had.
It should be noted that I was a leading advocate at a large enterprise software company for getting CUDA integration and helping NVIDIA sell more hardware in the enterprise. NVIDIA didn't fire her, and they should have.
I also would also everyone look at Gebru's work [1] - certainly seems like they're an established research who comes from a top university, has published in top conferences, and has a good understanding of the technical parts of the field
one's past experiences with something they are criticizing or defending is probably a good prior to have, but also shouldn't be the sole criterion. "leading" scientists are more often than not wrong on plenty of things
Should have paused the project and reached out to the sales team responsible for your company's corporate account. Would have been interesting to see who pulls more weight (sales and paying customers or ethics "researchers").
> I asked a question, and was put on a list proclaiming everyone in it as alt-right bigots.
Were you, or are you editorializing her actions? Because what you actually described is that she put you on a list
> of "fan boys" of an academic critical of "wokeness" as he calls it
That's not calling you alt-right. I believe the stated purpose of the list was to have a set of people that allies could engage with constructively if they wanted to. In other words, it was a list of people whom might be good to engage in verbal discussions with. Because of that, you're calling for her to be fired. Why?
She specifically said the things I stated, in a now deleted tweet. I know the cognitive dissonance is causing you to question my account, but that's what she did.
Look up articles where screenshots are posted.
I wasn't claiming Gebru has no accomplishments either. Many of her fellow ethicists fall into that category. She's not an ethicist, just won't shut up about it, along with race stuff, and has nasty, unprofessional habit of attacking people publicly. She's a bully, and I'm sick of people like you defending her because you've bought into her idiotic religion.
I agree that she said the specific things you quoted. I'm not debating what she actually said. However when you say she characterized you as a member of the alt-right, that is your editorializing, and not something she said (nor did you claim it was). I'm simply clarifying that she never accused you of being a member of the alt-right, and that you're projecting that accusation.
> Edit: I was mistaken in my response, since the commenter named Gebru. The head of AI at Nvidia is not Gebru, it's Anima Anandkumar. Gebru had nothing to do with this.
Right, but Gebru had a highly discussed disagreement with Yann, so when you said
> Watching highly entitled, and often narcissistic people with no tangible accomplishments attack Yan LeCun is pretty insane.
You were saying "Gebru is a narcissistic person with no tangible accomplishments". Either that, or you're making up generic events and hypotheticals, but that would be a really strangely specific hypothetical. Anandkumar also has very significant achievements, so there's no one you could be directing that comment at that wouldn't be belittling.
And let's be clear:
Gebru's accomplishments aren't shit compared to LeCun's, and everyone knows that in the AI field. Doesn't mean she's not accomplished, because she's a hell of a lot more accomplished than I am.
Which ones? I followed things at the time, most of the other people who disagreed with Yann were...also PhDs, many of whom have significant accomplishments in their own right. I mean there are certainly random twitter people saying things, but that wasn't unique to the sides of the ethicists. Random twitter people also attacked and insulted Gebru and co too. So saying "Watching highly entitled, and often narcissistic people with no tangible accomplishments attack Timnit Gebru is pretty insane." is equally correct, but you didn't say that. That's suspicious and perhaps revealing of your biases on the subject.
> Gebru's accomplishments aren't shit compared to LeCun's
Of course. She's also had her PhD for 3 years, as opposed to 30. Its no surprise LeCun has accomplished more in a career 10x as long.
He went through hell. He was rejected from conferences despite having algorithms that exceeded all others, just because neural nets had fallen out of fashion. Meanwhile, the folks who attacked him have cushy jobs earning massive salaries, working on technology that he popularized.
Is this some poor attempt at a joke? That's exactly what Twitter mobs do, right? Especially when the instigator calls the targets "fanboys" and "fanatics", and explicitly calls for them to be "cancelled"?
That is absurdly, unjustifiably optimistic. The worst activists have larger negative contributions than the best contributors have positive contributions. And as you yourself have observed, even typical activists often have worse-than-zero contributions.
In fact, it's more politically consistent to have equal representation. We already know what AI experts believe - make money in the end.
Talk about cancel culture.
I think JPKab is saying that an NVIDIA employee mistreated someone who is effectively a prospective customer for NVIDIA (or who is working to create prospective customers), and that this by itself should lead NVIDIA management to fire that employee even in the absence of public attention to the event.
When you declare your employer in your bio and talk about your field of work, it becomes relevant to your continued employment.
The way it works is you basically shame people and tech. You don't need to make a strong or valid case against a behavior or a tech, like a law would, because it's not legally binding. It's not real engineering regulations like a building or fire code. At best it's a mob made of folks who convinced other ethics PhDs to grant them a PhD that will rally against you and shame you [0].
Great move from Google to stop feeding the ethics racket.
[0] https://syncedreview.com/2020/06/30/yann-lecun-quits-twitter...
This is simply untrue. The goal of AI ethics research is (also) to build algorithms where one of the _inputs_ is a set of ethics. It doesn't matter what those ethics are. It just so happens that the "set of ethics" currently fed in tends to have a particular flavor ("woke liberal" ethics), which you seem to disagree with. Finding the algorithms matters, and it's likely that the standard ethics put in are going to come from some dominant ideology, but we still need the algorithms if we are to understand how to make AI that aligns with humanity's interests (however they are defined).
I'd call that computer science.
Yes it's Computer Science but it's implementing normative-ethics.
Why not provide the AI descriptive-ethics instead and let it decide for itself?
But we understand the principle of law-making, what laws mean and how they're applied. This is fairly uniform. What really changes from place to place is the content of those laws.
AI ethics and AI safety are attempting to give us a set of "law-making" rules but for AI. We get to decide (democratically ideally), in countries, states, cities, what "ethics" (what laws) we want, but AI Ethics as a field gives us tools to achieve that regardless of what the ethics/laws are.
For example, how do we encode the trolley problem in self-driving cars? We could decide democratically that cars should act and kill 1 instead of 5. Or the opposite. But then how do we translate that into ifs and else? No one really knows how to do that.
A thorny ethics example from my own role was a decade+ ago I was hired to do a privacy impact assessment (PIA) on an ML solution for detecting benefits fraud. It had all the ingredients for a complete debacle: "artificial intelligence"(?!), a vulnerable population, credit reporting companies, arms length government agencies, highly paid consulting firms, etc. What we delivered was modifications to the system to solve their core fraud risk problem without making peoples lives (more of) a dystopian hell.
Privacy is the currently the technology ethics department in large institutions, and I really see the AI ethics people as fancy privacy people with more math. Privacy has moved on from being a technical discipline that old hacker dudes like me did to being one largely done by women with law degrees. The tension between the security architect competency and the privacy policy analyst is still there in the field, but now we all see how they add a lot of value without needing the bottom-up view of a technologist.
The flavour of ethics in the privacy field has evolved, and I was absolutely suspicious of the change, as if you weren't a cypherpunk in privacy in 2005, you were a 5th column infiltrator looking to co-opt it for corporate ends - but of course, that's not true. Anyway a lot of text to say don't throw out babies with bathwater on AI ethics, but I'm saying they're not as new as anyone thinks, and if you think privacy is important, then you will probably recognize ethics in AI is important as well.
Read the article again - it's Google being uninvited as a sponsor. The title is poorly translated from German and should probably be changed to something like "Google no longer wanted as a sponsor of AI ethics conference".
AI or Data Science ethicists tend to be qualified data scientists who have a better understanding of concepts like differential privacy than the vast majority of software developers. Don't dismiss them all because you disagree with some of them.
Th OP refered to them as a back scratching "mob", but when I say don't dismiss them all you accuse me of being "rather thoughtless" and using a "playground dismissal" when I gave specific reasons why they are qualified for the job.
a lot of the big famous names in AI ethics (e.g. Moritz Hardt, Cynthia Dwork) also have very strong contributions outside this field.
Maybe it's trendy to say and think that? but that doesn't sound anything like ethics. Ethics is a part of philosophy, not a part of science or maths. That it is what it is, is not a "problem with ethics". (Sorry if I misunderstand you. Maybe you were just talking about "AI ethics". ...In which case, what do people who call AI ethics "ethics" call ethics, a.k.a. moral philosophy?)
Introducing AI as automation for human jobs has consequences beyond one company's profits. So there needs to be non-profits or committees that take a stance and represent themselves. There doesn't need to be some scientific papr on "AI hut my job" because all the science is on the company's part, making the AI in the first place.
One of these can be objectively evaluated for skill. There may be unwanted side effects. But the skill is there or it isn’t. That does not apply to the other, which makes their hiring and firing fundamentally different.
https://facctconference.org/2021/acceptedpapers.html
it's a mix of "soft" humanities, applied ML, methodology, algorithms/theory, and HCI.
it's not any harder to evaluate an AI ethics researcher than it is any other kind of researcher.
With AI ethicists... if you could look 100 years in the future.. there's not an objective measure of ethical outcomes you could use to evaluate the performance of your AI Ethics team. I suppose a cynic would look at the total cost spent defending and paying settlements / judgements in ethics-related lawsuits. As far as a non-cynical objective metric for AI Ethics performance, I can't think of any.
You could also look at compliance costs?
> dollars lost due to losing trust.
Trust, ethics, and public moral outrage are hopefully related, but not the same thing. Also, measuring loss of trust and allocating changes in revenue to that change is very difficult.
> compliance costs?
I think many people would agree that there's often a big difference between legal requirements and ethics.
If I understand what you saying correctly, you're saying AI ethics researchers have no skill?
It is instead related much more closely to a person's sense of right and wrong.
Of course AI ethics researchers have skills. They have skill in researching, they have knowledge of AI, and presumably education in formal logic and ethics. They have skill in writing and communicating the results of their findings.
However the product they produce is not writing and it's not research. It's the commentary on the research, the recommendations based on the results they have found. It's very subjective.
It's very different than a skilled hacker applying their skill and knowledge to uncover a zero-day exploit in your software that needs to be fixed. That's a much more objective process.
It is far easier to evaluate a security researcher's output than an ethics researcher.
I think many of us are familiar with this concept, as many of us likely preferred math over english in school as we preferred being deterministically evaluated.
However they're empowered to fix them, essentially no questions asked. So you don't often get p0 posts about unfixed bugs in Google products.
I don't think OP is claiming that there ethicists have no skill, but rather that it's fundamentally a lot more difficult to evaluate.
Then there’s the ease of changing things. In security, I’d imagine fixing a vulnerability might require lots of work, but rarely novel research. In ML that’s not the case. and after all that research, the finding might be a way to improve the worst cases by worsening the average cases, lowering revenue. And then there’s the clarity of what even is an issue. If someone shouldn’t be able to access a file, but they can, it’s clear. The goal then is convincing stakeholders that the problem should be prioritized. In AI ethics, you have to convince them that the problem even is a problem, convince them to allocate research, then convince them that the technique is worth lowering revenue for, and only then get to the question of allocating resources to implement the fix. Convincing stakeholders to prioritize fixes is a wholly different game.
I like your analogy, and there are surely lessons to be learned (e.g. maybe the ethics researchers’ primary job shouldn’t be publishing papers), but the differences are strong enough that the lessons might not extend too far.
I think the analogy breaks down a little since hackers were generally outsiders to the companies (even Michael Lynn), and hiring them and starting programs around them was the thing Microsoft did to fix their relationship to the community, whereas the folks here were already insiders at Google, but maybe that was the problem, that they were explicitly hired as academics with the expectation that they would keep doing the same work, rather than internally focussed researchers working on improving the products and systems being created.
Hackers offer solution to your loophole, AI ethnics researchers only call you out and hijacking the narrative for their own benefits.
When a professional opinionator is brought in with a bias that this data / company is racist/ transphobic / not inclusive then drafts a report proving it.
It is much harder to validate it.
It cannot even be evaluated as a legal argument with references (or missing reference) to laws.
Ethics is inherently philosophical field. From it stems understanding and ideas that may be codified in an interpretation of them into laws. Then those laws are tested and case law and valid interpretations are formulated.
Whereas "ethics" is a subjective field, so you can invent arbitrary "ethics violations", and you would - precisely because your paycheck depends on it. This is not new, see also bioethics - at best it's completely useless, at worst it's basically self-sabotage and your competitors would be happy to see you engage in it.
I have no idea how to make that work.
Maybe something like legal codification, so that publicly traded companies and the various parties are granted some temporary liability armor when vulnerabilities (or equiv) are identified.
PS- Just had a notion. Probably dumb: Maybe the liability (re)insurer gets a seat on the board of directors.
In the broadest sense, your suggestion is The Correct Answer™.
Every org should have some kind of funnel for feedback. Discoverable, transparent, accountable. Probably both public (external) and private (internal) funnels. CRMs by another name. Your notion of a GitHub repo-based issue system would leverage existing infra. Both familiar and cheap. No invention necessary.
In the cases with potential liability, it seems to me privileged access to confidential information will still be necessary. Even if that limited access is only time boxed (eg quiet periods before earnings reports).
Except that afaik game studio actually hire people that worked on cheats - e.g Riot Games.
In order to be defender you also need to know how to attack.
Hacker skills are a really long tailed distribution and the best are brilliant, the worst are just riding coattails. The ability to form an opinion about the implementation scope of a given variable for one of the many complicated functions required in ML puts you in some pretty rare air.
I work in security, where I need the breadth to argue the merits of the entropy of a diversification component or the choice of an elliptic curve in a protocol design and balance it with state level attackers - and regulatory policy and legislation, with the tedious compliance aspects as a theme.
The shape of the skills is weird. Where you need breadth first in a lot of areas, and depth first in ones that are not valuable outside very limited contexts. I can't vouch for where on the long tail the people involved in this controversy might be, but I don't like the odds on being dismissive of them.
Nuh huh. She was hired at her megacorp precisely because she was a researcher with an illustrious trajectory who had been involved in these circles for years, herself frustrated by the low technical level of the discussion.
When these people bit Google on the hand, Google got rid of them. Simple story.
Google is a monopoly and needs to be broken up under anti trust rules.
I've worked for companies who hired and then had to fire hackers for publishing vulnerabilities in their own and in customer products. Security is fundamentally a technology governance function, and AI ethics piece is also a technology governance function, where both hackers and AI ethics researchers are in-effect activists to create awareness for change, but aren't typically who you would put in the actual governance role.
An AI ethics conference has analogies to Defcon or Blackhat, in that shunning vendors because the attendees perceived that some prominent hackers were treated unfairly may have some precedents. Microsoft's relationship with Defcon in the 90's vs. post 2000 is an example of how this both happened and changed.
Extending that analogy, the disconnect in the relationship between researchers and vendors was a symptom of what a disadvantage the vendors were at to the asymmetric risk that activists/hackers posed because the media story of the conflict between giant corporations being vulnerable to scrappy hackers writes itself. In terms of how to handle it, google can probably use precedents from MSFT vs. defcon, and AI ethics researchers can look at how hackers both succeeded and failed to change security and tech governance.