Yoshua Bengio: ‘The dangers of abuse are real’
nature.com
nature.com
And good luck holding the military back.
One way or another humanity is on its way to the inevitable, and it's our evolutionary destiny because we are basically hard-wired to pursue the curiosity that is AI.
I can't help but think the same is true of AI or any technology that could be used for good or bad.
We spend a lot of time trying to control the tools, but less time trying to improve culture and ethics of the operator.
How do you make the majority of the society believe in the same social goals without resorting to tales a-la Christianity during the fall of Roman Empire?
You'll either need to make the whole of the humanity live in the sorts of Star Trek universe, where everyone is equally educated and is equally capable to critical thinking.
Or you end up with the same system where the majority of people are obliged to give their absolute trust to the minority.
In my opinion this is the key question.: How do we get everybody on board with e.g. UN Sustainable Development Goals[1] or something like them?
[1] https://en.wikipedia.org/wiki/Sustainable_Development_Goals
Probably because politicians are more aware of the realities of _current_ human nature rather than some future hypothetical that may never happen at all.
Also there's an important point to be made that the powers that actually owned the majority of nuclear arsenal on the planet had enough missiles to destroy most of the human civilization.
We got treaties on nuclear and chemical weapons. It could happen for autonomous weapons and weaponized intelligent agents also. But it might require a serious case of abuse before we get there, as it did for chemical and nuclear...
We would need to consider, however, access to finished products (like trained models). In some cases in the future, having access to the appropriate model might be as dangerous as handing someone a nuke.
How hard would it be for someone to assemble a indiscriminate "killer drone," given the appropriate image recognition and flight control models?
My background is an engineering role in the military and my current path is CS. I don't think it's unfeasible that an individual could build a crude drone that would kill a single person. Building an entire swarm of death robots is a different story.
It is for now. We will get to a point where a negative utilitarian with accesses to genome synthesis technology will be able to kill tens of millions. And this seems like it could become pretty feasible, even from a basement.
The danger of AI abuse is that existing large, powerful entities will put too much faith in statistical models without understanding their limitations, and will make bad decisions that harm people. Fortunately we have a filtering function -- entities that use these techniques with more nuance will slowly replace the failing ones because they are making better decisions. There will be a long transition as governments desperately try to save the dinosaurs from the consequences of their poor decision making, but if they can't make the transition, they're going to cease being relevant.
AI abuse is just statistical abuse. If a bank has a model based on ten years of real estate data that says, "hey, real estate increases in value, we should leverage into that", then come 2008 they'll find that things that their neat statistical model stated don't work any more. That's all the "dangers of abuse" are.
Maybe you missed the section where he said "Killer drones are a big concern." and "A lot of what is most concerning is not happening in broad daylight. It’s happening in military labs, in security organizations, in private companies providing services to governments or the police."
> it's just a statistical modeling technique
...and gunpowder is just a chemical.
These days there are a lot of high school kids messing around with machine learning. They might not have the theoretical understanding, but I don't think it will require a graduate degree to get there.
I'm not sure why there aren't more hacker tools that apply machine learning? Maybe we're not quite there yet.
We can control businesses. And as for someone in their basement, it isn’t clear at all that one could develop AGI in their basement. AGIs discovery will probably require massive, massive amounts of compute. More than could fit in a basement or be powered by domestic grid connections without drawing attention.
Good luck holding the military back? You are talking about the US right? The military and every other government entity answers to the people either directly or indirectly. If enough ordinary people wanted it, the constitution of the United States itself could be amended into the ITunes terms of service. And since AGI is to the detriment of literally every human being on the planet, I think some basic legislation might be able to be passed.
Don’t succumb to these notions of hopelessness. They are wrong. And you even spread them around which just makes solving the problem even harder.
Nobody is stealing anything. What they don’t realize is that the free market will incentivize AI researchers to go to the private sector. It’s just simple economics.
For instance, the public US university I attended for grad school has an endowment of about $1B, and spent ~$150M on athletics in the past year. A post doc salary there is about $40k a year.
Why would anyone take this over $200k a year (with lots of sweet perks) at eg Google? Because they are passionate about academic research (and potentially teaching), something that the university is very aware of and has no problem using to their advantage.
Talking about level playing fields here seems disingenuous.
Athletics programs are a huge driver of both alumni donations and student enrollment. Unfortunately. Not to mention licensing revenue from apparel and the like.
> Because they are passionate about academic research (and potentially teaching), something that the university is very aware of and has no problem using to their advantage.
This was more effective when the pay disparity was smaller.
Really?
I wish we would give people a chance before we dismiss people as dumb/.. .
In a link from this quote: "You have expressed concern that corporations have ‘stolen’ talent from academia. Is this still an issue?" https://www.nature.com/news/ai-talent-grab-sparks-excitement... I can't find a word "stolen" or anything similar being used at all.
Seems like it was introduced by a journalist who wrote final article. And considering that, it's not hard to believe that this journalist may have some opinions on economics that are far from reality, that (probably unconsciously) triggered his word-twisting reflex.
> "It’s continuing. ..."
So he's not objecting to the phrasing, eh?
The linked article has this:
> The loss of expertise in academia concerns Yoshua Bengio, a computer scientist at the University of Montreal in Canada, which has also seen a surge in graduate-student applications. If industry-hired faculty members do retain university roles — as Hinton has at the University of Toronto and Ng has at Stanford University in California — they are usually only minor, says Bengio.
So it's obvious he's not concerned about corporations literally "stealing" students and faculty somehow, like unmarked vans and burly men in black turtlenecks, eh?
> And considering that, it's not hard to believe that this journalist may have some opinions on economics that are far from reality, that (probably unconsciously) triggered his word-twisting reflex.
The word "stealing" is in quotes in TFA.
Obviously facetious, but "stealing" is not some axiomatic boolean that never changes. It's judged differently by different societies depending on time, circumstances, and ideas.
"Raiding" employees was common place. Now there are non-raid agreements. If non-raid agreements became very ubiquitous it would be considered stealing.
Working employees 100 hour weeks with no overtime was commonplace. Now there are labor laws. It would be considered wage theft and exactly "stealing" if you did it today.
"Simple economics" gives you no guidelines to judge any of these things "stealing" or "not stealing".
Making a job offer to an academic that they accept isn’t a statutory offense punishable by law, and therefore is not stealing, no matter how much you play with semantics.
Don’t worry about the phenomenon itself, which is 2 entities acting legally and of their own volition. Worry about the systems at play that enable the phenomenon. Universities are clearly not compensating (whether in terms of raw dollars or benefits) AI researchers sufficiently. That’s what one should be concerned about, not “theft” of AI researchers.
Isn't misuse possible in all (or, at least, almost all) technology?
Drugs can change humor, physical condition, hormones, and I think can be miused even as a long-term weapon over the population of some nation.
Nuclear. Explosives. Social medias. TV programs. Data processing with punch cards. Drones. Knifes. Herbicides. Small plastic rejects.
I understand the reasons Bengio tries to make, but it seems that it is the same problem with all technological stuff: someone, somewhere, will find a way to use it against others, and will cause minor or giant consequences.
> "Deep learning, as it is now, has made huge progress in perception, but it hasn’t delivered yet on systems that can discover high-level representations — the kind of concepts we use in language. Humans are able to use those high-level concepts to generalize in powerful ways. That’s something that even babies can do, but machine learning is very bad at."
I read this as: "We have super-advanced skip-logic software that can produce specific results when provided a large enough data set, but "intelligence" as it is defined, does not exist."
AI is really just sophisticated software algorithms.
In my opinion, there is no true artificial intelligence, and it will be unlikely that we will ever create such a thing for quite some time, if at all. AI is being used as a buzzword to garner attention.
It seems much more likely that we will build a brain-computer interface before true AI, and it will prove more efficient than what we have today, which is effectively many computers churning through a super-long list of "if-then" statements.
"Killer drones are a big concern....dangers of abuse, especially by authoritarian governments...AI can amplify discrimination and biases".
Bengio is specifically talking about malicious actors using AI. So saying "just unplug it" isn't really applicable, especially when it comes to governments.
https://www.forbes.com/sites/jilliandonfro/2019/04/04/google...
I for one and fking sick and tired of my existence being subject to debate due in part to these groups.
When AI starts being used t decide who to hire or if someone should be offered a rental then what happens when it is biased against LGBT folk?
https://www.heritage.org/civil-society/commentary/dont-typec...
Heritage actively campaigns in the US and supports people in the U.K. who work against trans rights. Yes they do oppose my existence they just wrap it up in statements about “freedom of debate” or “women’s safety” even when those aren’t actual issues.