Researchers propose a new interdisciplinary subfield called “Machine Behaviour”
nature.com
nature.com
What has AI policy produced so far, really? Algorithm analysis and safety research is done on the algorithmic side, not on the policy side. Every time I have read an AI policy paper, it has been listings of desirable properties with absolutely no input on how to make any real steps towards them. In the meantime, realities of algorithm driven mass surveillance and censorship is moving ahead at a rapid pace.
Possible conclusion 1: We just need way more AI policy research.
Possible conclusion 2: AI policy is done by the wrong people and deep algorithmic understanding is a pre-requisite.
Possible conclusion 3: Current AI policy is just used as a fig-leaf by tech companies who hire a few policy essay-writers without substance. The lack of progress in that area is a feature, not bug.
Not mutually exclusive, not exhaustive. Feel free to point to valuable policy research.
"The correct way of ensuring algorithmic fairness is X" where X is something they already happened to work on.
In various papers X could be robust optimization, differential privacy, causal inference, adversarial training, etc. The papers are mostly good, actually, but it's sort of putting the cart before the horse.
I think there might be legitimate value in having a bunch of lawyers, sociologists, and philosophers set a target more-or-less in ignorance, and let the people on the technical side try to hit it.
Of course even better would be interdisciplinary collaboration but that's hard, there aren't many incentives in its favor (where do you publish? any given venue is worthless to half of the authors). and it requires humility on the part of everyone involved.
Not surprising. Computer scientists have a bit of experience with computation, after all. It would actually be more concerning if a bunch of new stuff was being invented whole cloth.
> but it's sort of putting the cart before the horse.
Not at all. ML algorithms are just fucking algorithms and computer scientists have been thinking about what it means for an algorithm to be correct since... Turing.
And have been proving various theorems about ML algorithms in particular since at least the 60s.
Complaints that "AI safety looks a lot like previous CS research" are basically equivalent to observations that "neural nets have been around for a lot longer than alexnet".
> I think there might be legitimate value in having a bunch of lawyers, sociologists, and philosophers set a target more-or-less in ignorance, and let the people on the technical side try to hit it.
I disagree. This is how you end up with endless navel gazing about trolley problems while actual vehicles kill people by accelerating without control because redundant parts are too expensive and engineers don't have enough voice. Philosophers are rarely interested in honest-to-god engineering ethics, which almost always boils down to "pay well enough to hire good people, and then listen to the good people you're paying good money to have around".
Would you mind expanding on this? I would expect representatives of a field of knowledge that has an area called 'ethics' to be more concerned about ethics than your run of the mill engineer.
But the intention doesn't really matter.
What matters is the utility of the output!
In fact, somewhat ironically, I think a lot of the good work on ethics for AI is coming out of engineering, business, statistics, and economics departments. And those academic departments to do be a bit more "money grubbing" relative to philosophy :-)
I have an experience to share. I was there at the start of the web in the sense that I was building early web servers. I contributed nothing, but I was very excited and interested - as were tens of thousands of Computer Scientists.
I did not imagine any of the negative consequences of the web that have emerged. Very few people did. The community was blindsided, and I think that we have a reasonable excuse, just as a person t-boned on a junction has a reasonable excuse. We didn't see it coming, and it had never happened to us before.
With AI the fears and concerns raised in the media are mostly stupid, but there are genuine potential societal harms in terms of loss of freedom, dignity and opportunity that could easily arise if we don't set the system to favour human values over commercial and state ones.
We will not have the excuse of ignorance and surprise this time, and so I salute informed, reasonable and open people who are making the effort to engage with these issues.
The point of my post was that I did not foresee, nor did almost anyone, but we were idealistic and as a community we had not had the pervasive impact that Computer Science has now had on people's lives. We now have a far greater resource, experience and responsibility to innovate responsibly and with care. Delinquent and careless research that creates harm is unacceptable.
1. Someone with extensive knowledge of machine learning, and a much smaller (but existent) interest in the societal importance of interpretability
Or
2. A policy guy who knows tons about why these things should be interpretable, but has a very limited knowledge of machine learning?
Now, let’s say the position is policy relating to machine learning. Do you hire
1. The guy with extensive knowledge of policy who knows something of machine learning
Or
2. The guy with extensive knowledge of ML who knows a bit of policy?
It just seems like common sense, and it’s hard not to read a lot of comments as high ego engineers arguing that their tribe needs more sinecures and special treatment
I would hire both. If you want to do anything serious in the space, you need the resources to hire a well-rounded team. Maybe in five years this won't be true, but at the moment this is a space where technological capabilities are driving policy making decisions.
I had to choose one, the answer depends on if this is a soldier or a general.
If it's a soldier ($ or $$ salary), I would choose a CS PhD who has demonstrated an interest and aptitude for learning about policy. There are a lot of opportunities for that person to learn about policy including part-time fellowships with think tanks and federal agencies, sometimes even embedded in a lawmaker's staff. Conversely, taking a policy person and getting them to the point where they can adequately process the fire hose of AI fairness/safety/explainability research is going to require a lot more effort.
If it's a general ($$$ or $$$$ salary), I would choose whoever I could manage to hire that has the most influence in whatever agency or legislature is most relevant to the policies I want to push.
But again, especially for soldiers, it's a false choice, and if you find yourself in a situation where you have to make this choice then you need to focus on fundraising instead of your first hire...
Sure, a few modern fields are affected more than others, but improving voice recognition or visual object identification component by 20% doesn't fundamentally change the system using such code to a degree that it warrants a new academic discipline of study.
A new legal field of study on automation makes more sense to me. Algorithmic bias, product or service liability, baseline accountability, the standards of due diligence and safety -- these are rising concerns exacerbated by the recent increase in automation, AI-based or not. But these problems are hardly unique to AI any more than the misdirection of elections was unique to manipulation of digital social media.
I'm also doubtful that meaningful solutions to byproducts of automation by corporations and nations can be meaningfully addressed by a bunch of academic computer scientists.
Except AI is very much changing how things are done as well as what computers do. Image recognition, natural language processing, self piloted vehicles, these are all novel applications for machines which carry significant real world risks to life and property.
Moreover, AI in it's current state is a complex black box with unpredictable outputs for given inputs, which may be chaotic - see, for example, adversarial attacks. If we want a better handle on, say, regularizing outputs in a predictable way, or at least a clearer window into the occluded complexity of neural network decision making, a whole new field may very well be warranted. There's a reason we are calling these program outputs behaviors now; it's a new type of loosely -deterministic computing until we develop a deeper understanding which will not be trivial.
Obviously none of these are AI and I don't see how you're helping anything by insinuating they are.
That implies any meaningful definition to begin with. That is entirely an illusion and people are using this to mislead investors (not that I'll ever shed a tear for them).
Arguing about definitions is easily one of the least relevant parts of any discussion so I'll leave it here. I just wish AI topics didn't always have someone say it wasn't "real AI".
Real-world usage would imply people use it in a meaningful way. I have yet to see any evidence of this.
Also, feel free to look up some of the authors on this paper. They don't suck.
I'm inclined to believe a substantial number of researchers are currently being deliberately fuzzy about what "AI" can and cannot do. Why not call it algorithms and statistics? I think lay-people have a very skewed understanding of what has already been achieved through AI. They may also not understand the word algorithms but at least it doesn't make them think of Skynet. For example, if you asked a person on the street whether there exists an "artificially intelligent supercomputer" somewhere that could help you plan all aspects of a small but entertaining dinner party, they would probably say yes. They imagine that you could just ask IBM Watson to help out, and he'd tell you what to do. This is completely false. "AI" systems are very fragile, and yes, we could build something that plans dinner parties, but we'd have to start over if we needed to plan a kid's birthday party instead. It's very far from strong AI. Ten years ago when it was mostly IBM telling lies, the end result was a couple of billion dollars wasted by hapless healthcare conglomerates. That's already bad. But now we have people from MIT, Stanford, Harvard, Yale, and more embracing the term AI and relying on unfounded hype to push for funding. It would be much less sexy if we called it "Facebook/Google/etc enable unfair/discriminatory advertising by combining intensive data collection with logistic regression, sometimes in multiple layers, and with some graph algorithms thrown in". But it would be a much better starting point for a well-informed debate. I'm not trying to minimize the importance of algorithms in our world, but a healthy discussion should be based on a sound understanding of the facts on the ground, and AI hype is not helping with that. I strongly prefer the less hyperbolic terminology adopted by someone like Aaron Roth at UPenn, e.g. see the blurb for "The Ethical Algorithm: The Science of Socially Aware Algorithm Design".
Bonus AI rant: For the celebration of the new MIT Schwarzman College of Computing, which is a huge expansion of the arguably most important computer science department in the world, there was a discussion panel on AI consisting of MIT President Reif, Henry Kissinger (War criminal?, Theranos board member, wannabe AI expert), Tom Friedman (columnist of limited substance), and Stephen Schwarzman (business man, coined the "increased taxes on carried interest are like Hitler's invasion of Poland" analogy, brought the dough). How the heck is that the inaugural panel!?!
There is no "AI" and anyone who uses that term is mentally ill or selling snake oil.
I can see how a good-enough marriage of ML/NN to classic symbolic AI could yield a system capable of higher-order, intention-based “behavior” in complex circumstances. But we’re still a long way from achieving that, as far as I know.
What about non-AI powered machines that mediate our social, cultural, economic and political interactions? For example, science journal paywalls.