Are you at all close to this space? It sounds you may be underestimating corporate politics and the lack of rigour and ethical thought with which these systems are applied. The example Cory puts on policing -- and the many other examples you can find in Evgeny Morozov's book or "The End of Trust" -- are solid proof of this.
The deepmind team etc type of group who actually know what they’re doing and the boundaries of what they are working with
the “AI-washing” startups, corporate groups who know they are faking it and that what they’re doing is extremely limited
the corporate project team types who are just doing random tool play and honestly don’t understand what they are doing or that they are absolutely clueless with no self-awareness at all
I’ve worked with all three and they really are just totally different things that are all being lumped together. They also are listed in terms of increasing proportion. For every self-aware AI-washer team I’ve seen 50 “we are doing AI” Corp team types spinning out one trivial demo after another to execs who know zero.
You’re observing that they aren’t doing a perfect job, which is true, but my grouping isn’t related to perfection of results.
You claim that they "know what they are doing and the boundaries of what they are working with" -- and yet they recklessly make public a racist vision product?
It is easy to have a theory of what is going on, to model the processes of how things are playing out inside the system, to make external predictions of the system, and to be utterly wrong.
Not because your model is wrong, but because either the boundary conditions were unexpected, or there was an anti-pattern in the data, or because the underlying assumptions of the model were violated by the data (in my case, this happened once when all the data was taken in the Southern Hemisphere...)
In all these cases, you can know what you're doing, you can know the boundaries of what what you're working with, and you can get results that surprise you. It's called "research" for a reason.
The model can also be ridiculously complex. Some of the equations I was dealing with took several lines to write down, and then only because I was substituting in other, complicated expressions to reduce the apparent complexity. It's easy to make mistakes - and so you can know what you're doing, and the boundaries that you're working with, and still have a mistake in the model that leads to a mistake in the data ... garbage in, garbage out.
In short, this shit is hard, yo!
If instead it was a fuckup, well that seems adequately covered by “this shit is hard”.
If you are instead complaining about a lack of oversight, I don’t have a horse in that race. Ask someone else, I don’t care about the politics, I’m here for the technology.
And why would they have assumed in the first place that the model _would_ generalize across human races, or any other factor for that matter?
When this technology gets into their hands with a dev leash it will be recklessly implemented and people will die.
I am.
> The example Cory puts on policing
My most upvoted comment on this website was discussing this exact scenario. https://news.ycombinator.com/item?id=23655487
Could you perhaps clarify the generalization you're making about me and people like me so I can understand it?
Also, the lack of thought and accountability that I mention above I think is fairly general from my experience, even outside of policing. That is why I don't generally agree with the lunch statement. Guys are having a hell of a party as far as I can tell -- at the expense of horror stories suffered by the victims of these systems.
That is all part of engineering to me, so by definition, I think many in the field are in fact, out to lunch.
"Don't say that he's hypocritical
Say rather that he's apolitical
'Once the rockets are up, who cares where they come down?
That's not my department!' says Wernher von Braun
Some have harsh words for this man of renown
But some think our attitude
Should be one of gratitude
Like the widows and cripples in old London town
Who owe their large pensions to Wernher von Braun"
Many people are very resistant to the idea that their particular work can have a negative impact or that they should take responsibility for that. See Yan Lecun quitting Twitter (https://syncedreview.com/2020/06/30/yann-lecun-quits-twitter...)
Other people are very aware of the dangers of their work. But, when the money gets big enough, they take their concerns to the bank and their therapist. See Sam Altman's concerns about the dangers of machine intelligence before he invested in OpenAI (https://blog.samaltman.com/machine-intelligence-part-1) Contrast that with his decision to become the CEO, take the company private and license GPT-3 exclusively to Microsoft. (https://www.technologyreview.com/2020/02/17/844721/ai-openai...) He had reasons. He posts here. He might defend himself. But to me it seems like the kind of moral drift I've seen happen when people in silicon valley have to make hard choices about money and power.
There are also applications of ML that are generally safe and can be of benefit to society. See the many medical uses including cancer detection.(https://www.nature.com/articles/d41586-020-00847-2) Most of the work being done to expose the risks and biases of ML is being done by researchers who are at least somewhat within the field. In my math and computer science program, two and a half of the 25 students are doing their thesis in safe ML. (I'm giving myself a half because I'm working on logic based ML.) I don't think it's fair to believe that every person working in ML is participating in something negative for society.
Ultimately, I think we need some reasonable regulation and a lot more funding for research into safe ML. Corporations and governments want ML for purposes that can be unethical. Unfortunately they also control a lot of the research grants. So they have a disincentive to fund AI ethics or safe ML over pushing the boundaries of what ML can accomplish.
Finally, I think many engineers would like their work to be positive for society. Unfortunately, with what we know now, a lot of the edge cases we run into are unfixable. When Google Photos started classifying black people as gorillas, Google just removed primates from the search terms. Years later, they hadn't fixed it. (https://www.wired.com/story/when-it-comes-to-gorillas-google...) I'm sure most engineers on the project knew that was a hack. When faced with an unfixable issue like that, the engineer either tries to get the company to stop using ML for that problem, compartmentalizes and ignores the issue, or they quit. Where do you draw the ethical line? It's good to hold people accountable but it's unrealistic to expect that to solve the problem.
Your third and fourth points I think are linked. I am not exactly sure where or how you would draw the line, but I kind of think of these ML/AI applications as something that could be export-controlled or be regulated along those lines, just like certain pieces of hardware are export-controlled on the grounds that they could be used for harm, and weapons, of course (and I mean, add some salt here because governments will cause the harm regardless, but hopefully the point comes across.) Once the regulations are in place, and corporations take _substantial_ economical hits for their errors (unlike, say, GDPR violations, which Google just factors into their OPEX), those corporations will rapidly start effecting real change. Corporations understand the language of (economic) violence suprisingly well, it's an effective tool for change. But like you said, it is precisely the same governments and corporations driving the research and exercising economic and political power, so I am not entirely sure how that would start shaping into place. Like almost everything else in life, the first step will probably be to keep raising social awareness; change will emanate from us at the bottom -- if we can direct our anger correctly and if the climate catastrophe that is upon us does not wipe us all first.
I don't think this is an example of a straw man, given that his audience is readers of Locus, a science fiction magazine. While researchers and practitioners in ML understandably hold a more nuanced, informed view, the position he's arguing against is pretty common among the general public, and certainly common in science fiction.
What deep learning seems to step into more and more is time-based statistical inference.
AGI is not:
seeing that a girl has a frown on their face.
seeing that a girl has a frown, because someone said "you look fat"
seeing that a girl has a frown because her boyfriend said you look fat
seeing that Maya has generally been upset with her boyfriend who also most recently told her she is fat.
But keep going and going and going and we might get somewhere. Do we have the computer power to keep going? I don't know.
That executive and arranging function is the unknown. From whence cometh that characteristic of Dasein? That preponderance of concern with the act of being as Being?
It's a tough nut to crack, even in philosophical circles. To think that we're going to articially create it by any means other than accident or luck is hubris of the highest order.
This meta-comment of restatement in various contexts, with various amounts of story-telling and technical detail, brings up the educational burdens of communication -- to be effective you have to reach a reader where there are today .. in terms of assumptions, technical learning, and focus of topic.. since this is such a fast-moving and wide subject area, its super easy to miss the distinction between "low value, high volume audio clips recognition" and "life and death medical diagnosis for less than 100 patients". hint - that matters a lot in the tech chain AND the legal structure, and therefore combined, the "do-ability"
A ML model and most of what we currently call AI, is a statistical model. It predicts things, and in the process of "training" this model we can also learn about the world it interacts with.
Anyone who has ever spend any amount of time on the issue knows that the idea of "generalizing" or "adapting"in ML corresponds directly to causal inference. This is not some secret to be uncovered by every other pundit, it's a direct consequence of ML models being.. statistical models. If our AI can understand when statistical relationships hold and when they do not, then it is inferring causality from data. Currently, we are dealing with ML algorithms adapting or generalizing: learning which regularities to "trust" and which to "discard" when the "DGP" changes. This sounds nebulous and difficult, but that's only the case because the models are complex.
Nevertheless, the underlying statistical problem is old. Ancient. Talked to death in every scientific field that tries to do an experiment, to infer some causal parameter from observational data, or construct counterfactuals to guide policy. Similarly, the divide between quantitative and qualitative approaches, are decades (if not centuries) old.
These problems are so well understood that we can state very precisely what needs to happen to make our inference causal. The catch is, that first it depends on the DGP and our model thereof, and further that none of these things can be proven to be true in the same framework. Whether the DGP is as we need it to be, or whether the model identifies what we want, is something we can be confident about given a set of assumptions, but it is not something we can know to be true. And guess what, if the model is complex, then so is thinking about its inferential capabilities.
The discussion seems tired, because it is. It's not a deep philosophical issue. It's a practical one. That doesn't imply there is a good solution to it either, but the basic issue hasn't changed for a long time. Maybe we are not trying to predict a treatment's effect on some randomly selected population, maybe we are instead trying to land a plane under conditions we can barely foresee. But what needs to happen for these two predictions to be unbiased, consistent, low variance, whatever... has not changed.
Researchers are frustrated with these article, because pundit after pundit claims to uncover some general problem with AI or ML in face of reality. But it is not about AI. It's about statistics, and we all want to say: "Yes, we know. Now what?"
Clearly, no biologist claims that humans evolved from modern primates, just like no modern AI researcher seriously thinks that current machine learning methods will lead to "True AI".
This article is the opposite. He's treating ML as basically a simple supervised architecture that doesn't allow any domain knowledge to be incorporated and simply dead-reckons, making unchecked inferences from what it learned in training. Under these constraints, everything he says is correct. But there is no reason ML has to be used this way, in fact it is extremely irresponsible to do so in many cases. ML as part of a system (whether directly part of the model architecture and learned or imposed by domain knowledge) is possible, and is generally the right way to build an "AI" system.
I think ML has its limitations and will be surprised to see current neural networks evolve into AGI. But I also don't think the engineers working in this space are as out to lunch as the author seems to imply, and would not write off the possibilities of what contemporary ML systems can accomplish based on the flaws pointed out in relation to a very narrow view of what ML is.