- 2 days ago
Category
🗞
NewsTranscript
00:00When you chair the International AI Safety Report, we'll just start there. Is AI safe?
00:05No, that's the problem. We're building systems that we don't know how to control.
00:11They will behave sometimes against our instructions, and this lack of reliability
00:19is both a scientific challenge and a problem for the deployment of AI.
00:25So how do we, okay, if we use social media as the example, we failed that experiment,
00:32I think many would say, that we did not keep ahead in terms of oversight, regulations.
00:37We could go into all the things that have gone wrong and how it's impacted society.
00:41What faith do you have that we get this right?
00:43And it can't just be the U.S. deciding or Europe deciding.
00:47It's got to be a global work or progress on this.
00:50Absolutely. Safety is a global question because a dangerous model,
00:54I mean, one that can be misused in dangerous ways, like mythos, for example,
00:59could be designed in one country, and then somebody in a second country uses it to attack a third country.
01:05And, you know, so the Americans need to make sure the Chinese models are not used against the U.S.
01:10and vice versa. It is an international question.
01:12Can I ask you, is it like missiles and nuclear weapons and that there will be a mutual deterrence
01:20because everyone will be proficient on some level when it comes to AI?
01:25There are analogies. I think one of the big differences is that there isn't as much money to be made
01:31with nuclear weapons
01:31as there is with AI, which means the incentive structure is now pushing companies towards going faster and faster
01:38and not paying enough attention to the safety issues and, you know, the public protection.
01:43But I think one thing that I think concerns a lot of people is that, okay, you could have,
01:50let's just use Anthropic as an example. You have Anthropic and what they're doing with mythos,
01:55but just because Anthropic has that right now doesn't mean DeepSea couldn't have that six or eight or 12 months
02:01from now
02:01or even something more powerful.
02:03That's right.
02:03So it doesn't matter if, like, one country is doing things a certain way or one company is doing a
02:08certain way
02:08if the entire world is not actually adhering to a certain set of framework.
02:13Exactly. That is why we need international coordination.
02:16Initially, it's probably going to be just the U.S. and China.
02:19But at some point, as other countries are also, you know, beefing up their capabilities, it has to be international.
02:24What keeps you up at night? Like, lay out the doomsday scenario for us.
02:27Well, there are, like, many catastrophic risks because intelligence gives power.
02:32And that power could be in the wrong hands.
02:33It could be used by terrorists or countries which want to destabilize us.
02:38It could be used even by the people in power to create, you know, a dictatorship.
02:44And we don't know how to make sure AIs themselves will not use that power against us.
02:48But how? Like, harnessing social media, logging in, like, taking infrastructure offline.
02:54Like, give us concrete examples of what these entities could do.
02:57So, what the companies are worried about right now are cyber attacks because these AIs really have a lot of
03:04knowledge of computing and vulnerabilities.
03:07And on their radar and the evaluations they're making is also the use of AI to build dangerous viruses.
03:13So, you could think of people using AI to create new pandemics.
03:18They're also worried about what happens when AI itself is used to create the next version of AI.
03:25So, AI doing AI research.
03:26And if we don't control these systems, they could put back doors to create AIs that are not, like, friendly
03:32to us.
03:33Okay, now I am worried.
03:35So, in other words, you're talking about AI that can think on its own and work on its own.
03:38Well, it's already the case.
03:39The whole point of agents is that they're autonomous, they have goals, and they work towards those goals.
03:45The real question is, who controls those goals?
03:48We would like to be the ones in charge.
03:50But right now, we're not, we don't have the technical answers to make sure that's the case.
03:56Yashwa, do you regret, then, the work that you did years ago?
03:59I think I should have been seeing ahead, you know, earlier, that we were building something that could become extremely
04:08powerful and that we don't know how to control.
04:10And it would impact societies in which that currently, you know, on the current trajectory, could be destructive.
04:19So, okay, but here we are, right?
04:21Yes.
04:22And it sounds like it does feel very doomsday and scary, but I do...
04:27And it also feels like it's too late.
04:29No, it's not.
04:30But, and I want to talk about that, but I also feel like there's a lot of potential to do
04:34some great things as well.
04:36Yeah.
04:37Talk to us about that side.
04:38Absolutely.
04:39And, frankly, I've been working for many years on how we can use AI to design better drugs in medicine
04:44more broadly, how we can use AI to design, like, better materials for batteries, for energy storage.
04:51And I know a lot of people working, for example, in how AI can help agriculture, can even help democratic
04:58debate tools.
05:00The question is, the same power that enables these things also enables, for example, in the case of biology, dangerous
05:07uses that could really harm a lot of people.
05:11Do the potential positives outweigh the potential negatives?
05:17Well, there's so much we don't know here.
05:20You have to weigh in the uncertainty.
05:23Because it sounds great to have better medicine.
05:25It sounds great to have more efficient batteries.
05:28It sounds great to have an ecosystem where people are able to learn more and do quicker and be more
05:33efficient.
05:33But none of that matters if the AI agents are going to kill us.
05:37Yes.
05:37Or if we lose our democratic institutions and we end up in a dictatorship.
05:42So these are...
05:44But we're already having, like, if I think about democracy, let's go there.
05:48Yes.
05:48Right? We're already having these kinds of problems.
05:50We do.
05:50And this is a pre-relate date.
05:51We do.
05:52But it could get worse.
05:53There are studies showing that AI is now getting better than people at persuasion.
05:58In other words, making people change their mind.
06:00Imagine this applied to change political opinions, to change public opinion.
06:05So this is, like, shaking the foundations of what democracy rests on.
06:11You know, I don't know.
06:12You know, people talk about the new industrial revolution.
06:16They make comparisons.
06:17Do we have any comparison in terms of the impact AI agents?
06:24It's something else that's changed our world dramatically.
06:27The light bulb.
06:27As usual, I think, you know, the past can tell us things.
06:31But you have to realize we've never been in a situation where we build technology machines that would be smarter
06:37than us in many ways.
06:39And it's already the case in, you know, various domains.
06:42It's just growing.
06:43And there's nothing in history.
06:44Like, we build machines that help us physically.
06:47They have much more powerful, like, muscle-wise.
06:50But brain-wise, this is completely new.
06:53We're speaking with Yoshua Bengio, founder and scientific advisor at Milo Quebec AI Institute, co-president and scientific director at
07:01Law Zero.
07:02It's focused on building safer AI agents.
07:05The productivity question is one that we want to get to.
07:07The investment that's been made from a CapEx perspective has just been so massive.
07:12That the promises are that this will actually make us so much more productive and will lead to higher returns
07:17on investment.
07:19Is that true?
07:21I don't have a crystal ball.
07:23But let's say there are different scenarios about how AI intelligence capabilities will grow in the future.
07:30I think the default is going to continue on the same trend.
07:33But it could also saturate.
07:34And then maybe some of these promises will not be realized.
07:38The other issue is, even if AI is more capable, it doesn't mean that it is behaving well.
07:45In fact, there is a serious scientific hypothesis that intelligence and the goals to which that intelligence is put to
07:54use are like two very different things.
07:57All right, so we talk about races in the world, and this is certainly a tech race, an AI race,
08:04right?
08:04So it sounds like for things to go better, safer, you've got to have global collaborations.
08:11So what is happening on that regard, or what needs to happen in your view?
08:16Based on your knowledge, you understand how this works.
08:18So what needs to happen to make sure that we don't have AI taking over?
08:23I would say the most important thing that needs to happen is better understanding and awareness of what we're talking
08:30about by the governments around the world.
08:33And of course, China and the U.S. in the first place.
08:36But eventually, you know, many governments who have researchers and companies building AI as well.
08:43Because once we understand the magnitude of the risks that we're taking, we'll sit down.
08:48We'll sit down at a table and negotiate.
08:51One of the things I want to go back to what you said about productivity, what about the labor market?
08:55Like, is it going to create, it's already tough for some of those starting jobs, I think, of kids coming
09:00out of college.
09:01Like, what does it mean for the global economy?
09:04What does it mean for the labor force?
09:05Does it create jobs?
09:07I mean, building data centers, yes, but how do you see it?
09:11I mean, the number of people needed for data centers is peanuts compared to the effect that AI could have
09:16on the job market.
09:18But the economists don't agree.
09:20It really depends on the curve of advances in AI in the future.
09:24If it continues at a current rate, then yes, there will be massive automation coming too quickly for society to
09:30adapt.
09:31Yashua, given the work that you've done in this, understanding the potential pitfalls of the technology, name names for us.
09:40Who's doing this well?
09:41Who understands the potential repercussions here?
09:44I don't name names.
09:46I think there are a lot of people who have good intentions.
09:50The issue that makes the current dynamics dangerous in some ways is the incentive structure.
09:59It's the competition between companies, the competition between councils.
10:02So maybe I'll ask this question a different way.
10:05What if you could have Dario Amade in a room with Sam Altman, along with Jensen Wang, along with the
10:12leaders of all the hyperscalers, and then, of course, the leaders of the Chinese companies as well?
10:17What would you tell them?
10:17What do they need to know?
10:18So they know already, right?
10:22They know what I'm talking about.
10:23Tell them.
10:24What they need to do is to agree with each other on a set of principles about safety and, like,
10:33careful deployment of AI.
10:35And they're already talking about it.
10:37What they are asking, actually, is governments to agree with each other to make sure there's a level playing field
10:44for all the companies so that every company, before they build and deploy a very powerful AI, needs to show
10:51to the public, to the governments, that their system isn't going to be, like, dangerous in the wrong hands and
10:57all these issues we're talking about.
10:58But are the financial incentives getting in the way?
11:01The race is on to spend and spend.
11:02Yeah, it seems like a prisoner's dilemma.
11:03Yes.
11:04And it's who can go fastest and get it all done.
11:08So does that prevent the right thing from happening?
11:11You've just got about 30 seconds left.
11:12Well, no, we've dealt with situations like this before.
11:15I mean, governments have regulated things like drugs, trains, meat.
11:22Because if it was just competition, we would end up with, you know, dangerous food, dangerous planes, dangerous drugs.
11:31But now we enjoy all these things that are safe and useful because government's having to feed.
11:37So in a word, yes or no, you're more hopeful than worried?
11:42Yes, hopeful or no?
11:43I'm, I'm, I don't know.
11:47I mean, I feel like it's not my vision that matters.
11:53It's the actions that we take collectively.
Comments