Ok, so what is the solution then? Does she have a concrete set of steps or goals to address the problems she sees? Is there a list somewhere of things that would appease her, and in her mind make ML fair? Honestly interested.
When it comes to problems, particularly in complex subjects which aren’t yet well understood, and where there aren’t yet an overabundance of high caliber researchers, this comes across as dismissive of the problem.
This can be even more concerning if we know there are investors lined up who will happily sell something to the world and who will intentionally hide or minimize known ethical concerns. And then play dumb and shocked later when the very same problems manifest.
I see this dismissal of justifiable and real concerns an awful lot in conversations of all kinds lately.
And from what I’ve seen, no one here is anti-ML, no one on either side here is a luddite. But like so many conversations online, we should quit talking past each other and likely need to quit trying to paint people as if their concerns don’t have very real ethical implications which will, if left unaddressed, manifest in all kinds of negative ways throughout society.
Again, a lack of a neat and tidy solution doesn’t mean the problem doesn’t exist.
As for Timnit and LeCun: a user above notes that the epistemological frameworks they're using to analyze this problem are not aligned. I found that comment pretty eye-opening, honestly.
Shouldn't "I care about ethics" just be assumed? How many people do you know who would say "I don't care about ethics"?
Computer science programs around the country promise their young, bright-eyed undergrads that they'll change the world. Very few of them teach them the ethics they'll need to do that in a thoughtful way.
The assumption that science and technology is inherently ethical has unfortunately led to dangerous ideas over the past 150 years. Some, like geographic determinism and eugenics, have directly led to the suffering of millions of people. I hope we tread carefully and take action when we see harmful models. [0]
[0] https://twitter.com/SpringerNature/status/127547736519656652...
But I didn't say that science and technology are inherently ethical. I said that most people care about ethics. They might have different priorities or philosophy than you, but almost nobody commenting on a social issue is doing so because they want the immoral thing to happen. Right? So asking everyone to say "of course I care, of course" before everything they say is laborious.
I think that the underlying assumption is that science and technology are inherently neutral (which isn't true) and that neutrality would be inherently ethical (which also isn't true).
Because it identifies which team you are on. People only recognize the existence of a nuanced point if it's made from someone on their own side.
This is purely a power game in which some individuals have managed to blackmail everyone else into recognising their role or being cancelled. Now and then someone prominent needs to be attacked to remind the others what they risk if they don't fall in step.
Specific recommendations start at 17 min.
This is an extremely unpleasant position to take, if your point of view is empowered within the status quo. It is much extra work for no discernible benefit to the researcher.
If your point of view is subject to disproportionate suffering under the status quo, then reinforcing current practices by implicitly enshrining them in input datasets will make improving your situation even harder.
As an example, consider the case of public school funding. In a hypothetical system where school resources are provided proportionally based on student success, good schools will thrive and bad schools will get worse. If someone points out this isn't fixing the problem, you can reverse the proportions -- this will cause good schools to suffer while bad schools will get additional funding (disincentivizing student success). In cases like this, it's not enough to just have a purely abstract set of metrics on which to base resource allocation: it will always require actual investigation of why good schools produce good results and why students do poorly in specific schools.
This is sort of obvious, of course, but it isn't being translated into terms that some researchers can or will grasp. It's never enough to just tell someone to 'debias the dataset,' as determining that bias is a monumentally difficult challenge that people have failed to achieve for many generations. A key factor in fact is the propensity for this kind of research to get deployed, today, by people who are not experts in a given domain of investigation, with possibly disastrous results in policymaking. These tools are not abstractions that require a team of experts to translate from research paper to the real world; ML researchers put out results that you can shove into your nearest computer and run.
What Timnit and others are getting at is that it requires thoughtful and careful assessment to get real value out of this sort of research. Ideally, in Timnit's assessment, the researchers themselves would put effort into identifying possible calamities and put as much effort into mitigating them as they do into publicizing the work itself.
Yann LeCun and other researchers simply do not believe this is their responsibility; all they want to focus on is the mathematics themselves. I'm sympathetic to this position but I also very much do understand the opposition. One of my favorite movies from childhood, "Real Genius," deals with this sort of issue as the main plot line.
The problem here is that many stakeholders really really* want to reduce the required intervention to a mechanistic increase/decrease of some sort, never mind that different stakeholders want opposite interventions.
No matter how "intelligently" you go about it (whether the intelligence is natural or artificial), this desire to simply boil down policy making to reallocating resources or incentives/disincentives is fundamentally lazy. It reminds me of the mania for diversified conglomerates and corporate management of the firm-as-portfolio (ie. reducing management's function to determining financial allocation between corporate units and deemphasizing the need for operational knowledge) during the 1970s.