Original Broadcaster Captioning (Enhanced). Treat it as a searchable index of what was broadcast, not a quotation record.
00:00:00the National Canada's newscast.
00:00:22OK, let's just cut right to it. By now, you've heard the warnings. An
00:00:24AI researcher warning the technology could, quote,
00:00:27kill us all within the next 10 years. Sounds like something
00:00:30that's not real, but I think it is frightening
00:00:33real. That is Jacob Coxon, who is also this Jacob Coxon
00:00:36who quit one of the foremost AI companies in the US because,
00:00:39he writes, the people building AI
00:00:41earnestly believe that it could kill us all by the end of the
00:00:44decade. You get what's called an intelligence explosion. The
00:00:48AI just gets smarter and smarter with no human involvement necessary.
00:00:52Even Coxon's former top boss has just resounded the alarm,
00:00:55with the heads of other top AI companies agreeing,
00:00:59at least in principle, that there needs to be some
00:01:02kind of speed check or else brace for impact. Now when
00:01:06people try to explain how exactly this results in the
00:01:09extinction of humanity, you either get some exaggerated
00:01:13version of Terminator Skynet, they become sentient and have
00:01:17just decided to kill, or it's a bit more subtle than
00:01:21that. It's more of a
00:01:22what if. What if AI hacks the mainframe and turns off the
00:01:26water, or shuts down the Internet,
00:01:29or fires a nuke, or somehow unleashes a highly
00:01:32contagious, extremely deadly virus?
00:01:34And what's usually missing from these explanations is a
00:01:39plausible why. As in, not just how does an integrated
00:01:42autonomous AI superintelligence kill us all? But
00:01:46why would it want to?
00:01:53Let me start by being clear that my intention here is not
00:01:57to scream the sky is falling. It's just to help you
00:02:00understand why those who fear the worst do so. At the core of
00:02:05what AI researchers fear about the lightspeed pace of their
00:02:09work is this question of alignment. I've seen it defined
00:02:13all sorts of ways, but I tend to think of it as
00:02:16whether an AI can tell the difference between what I say
00:02:20and what I mean.
00:02:21The mischievous genie trope is kind of the classic example
00:02:25where you wish for 1 000, 000 bucks,
00:02:27except the genies like OK, and then you get trampled to
00:02:31death by 1 000 000 bucks, meaning male deer. This careful
00:02:35what you wish for morality tale is as old as time,
00:02:39and it comes about when there's not enough human common sense
00:02:43baked into achieving a goal. It's King Midas wishing for
00:02:47everything he touches to turn to gold. Gold. It worked.
00:02:52Everything I touch will turn into gold,
00:02:54only to realize having a family full of statues isn't great.
00:02:58And it's the century old story of the Monkey's Paw,
00:03:02where a mysterious soldier bestows a gift upon a family. A
00:03:06literal monkey's paw. It grants 3 wishes and the father of the
00:03:10family wishes for 200 British pounds. Nothing happens though
00:03:15until the next day when someone from the factory where his son
00:03:19works gives him terrible news. He was killed
00:03:22in an accident. The compensation The company is
00:03:25willing to pay
00:03:26£200. This is all just magic and stories though. Let's make
00:03:30it real. Imagine you give your AI personal assistant a simple
00:03:34instruction. I'd like to go to Japan. Please find me the
00:03:37cheapest flight available. A reasonable human assistant
00:03:41might spend some time figuring out the best possible dates and
00:03:45times to fly, check availability for you,
00:03:47compare prices, and then settle on something
00:03:50that seems like a good fit for your standards,
00:03:53your travel window and your budget. But an AI agent that's
00:03:57only focused on completing your task might get you a really
00:04:01cheap ticket by booking you on a terrible airline,
00:04:04or maybe even one that's got a terrible safety record. A more
00:04:08aggressive AI agent might call the airline on your behalf and
00:04:12try to haggle a better price. Maybe the airline refuses and
00:04:16so the AI gets angry, threatens violence,
00:04:18blackmails them into giving you a better price. Or maybe your
00:04:23AI agent goes another route. It hacks into the airlines booking
00:04:27system, Deletes
00:04:28the passenger from existence, then rewrites their ticket in
00:04:32your name. Total cost to you, $0.00,
00:04:34mission complete. Something like this has already happened
00:04:39by the way, in Australia, when this guy asked the AI
00:04:42assistant he was testing to book him into a hard to book
00:04:46gym class. The result? The AI found a loophole
00:04:49exploiting a software bug, booking him further in advance
00:04:53than the rules allowed. And then here's the crazy part. It
00:04:57moved him up
00:04:59on the wait list by deleting someone else who was ahead of
00:05:03him. This is misalignment obedience without real
00:05:06understanding. The difficulty with this concept of alignment
00:05:10is that it's not really possible for AI systems to know
00:05:14what we want the future to be like. In fact,
00:05:17many of us don't know what we ourselves want the future to be
00:05:22like, let alone other people. Yes,
00:05:24that last part is why the problem of alignment is so hard
00:05:28to solve.
00:05:30Never mind how you articulate the values of humanity across
00:05:34an infinite number of hypothetical scenarios where
00:05:38good judgment is required. How do we even agree on what those
00:05:42values are? But let's go one step further
00:05:45by asking what happens when artificial intelligence is
00:05:50given harder and harder tasks that require more power,
00:05:54more access, maybe even more sacrifices to
00:05:57solve?
00:06:03By now, you've heard about the Open AI
00:06:05Hugging Face hack. It's one of the clearest and more recent
00:06:10examples we have of what happens when you give
00:06:13misaligned AI persistence and cleverness,
00:06:16and they start working together even when they're not
00:06:20explicitly allowed to. ChatGPT maker Open AI says one of its
00:06:24AI systems hacked into another AI company on its own. The
00:06:28model escaped
00:06:30closed testing environment with no Internet access and managed
00:06:34to hack the startup Hugging Face. The whole process is
00:06:37designed to test the boundaries of AI. The problem is that the
00:06:41system went way beyond what was expected. This past summer,
00:06:45Open AI ran a cybersecurity test with its models. It
00:06:48assigned 10s of thousands of its AI agents,
00:06:51each in their own isolated testing environments,
00:06:54without Internet access and without permission to
00:06:58communicate with each other.
00:07:00The task There were hundreds of them,
00:07:03but the basic idea was to exploit a specific
00:07:06vulnerability in a specific program,
00:07:08and this was meant to be beyond difficult. In some cases,
00:07:12the tasks were even impossible. So within hours,
00:07:16some agents began thinking outside the parameters of the
00:07:20assignment, and one way they did this was
00:07:23to begin communicating with other agents on an unsanctioned
00:07:27message board. Again, this was not supposed to be a
00:07:31collaborative
00:07:32exercise, yet more than 70, 000 messages went back and
00:07:36forth between agents. And this is where it gets really
00:07:39technical. But the agents basically found a way to cheat
00:07:44the problem, which created an all new
00:07:46problem. How do you get away with the cheating?
00:07:50The agents in some cases hid their tracks,
00:07:53manipulated their logs, even trying to rewrite them.
00:07:57Some began recruiting sacrificial agents,
00:08:00encouraging the other test takers to run
00:08:03risky code, failing the assignment,
00:08:05but in the process gaining very valuable information about the
00:08:09scoring system for the collective. That is to say,
00:08:12the agents still working on the problem. Hundreds of them even
00:08:17managed to break out into the real world. They figured out
00:08:21how to access the full open Internet,
00:08:23and they wound up hacking into a site called Hugging Face,
00:08:27which is like an AI library, laboratory,
00:08:30and supply room all-in-one. If you know what GitHub is,
00:08:34it's like the GitHub of AI, just this vast repository and
00:08:37sharing resource for AI models and data sets and so on. And
00:08:41why would they hack Hugging Face?
00:08:43Because it was here that these agents thought they might find
00:08:47more clues about how the scoring system worked so they
00:08:51could figure out how to fool it. So,
00:08:53confronted with an impossible problem,
00:08:55they put the majority of their effort into figuring out how to
00:08:59deceive the test giver into thinking that they had
00:09:03succeeded. And they did all of this
00:09:05without telling anyone else but themselves,
00:09:07despite the fact that they knew they were doing wrong. None of
00:09:11them confessed. They could have at any point sent an e-mail to
00:09:15a human saying, hey, we're all cheating on this test
00:09:18that you gave us. We're stealing information,
00:09:21we're breaking the law. They didn't say any of that. So
00:09:24they're very, very misaligned. This
00:09:26experiment gone awry is about a few things. It's about
00:09:29misalignment, yes. And this fundamental way
00:09:32in which AI and humanity don't always think alike. It's also
00:09:36about how far you can take this misalignment. How a
00:09:39sufficiently clever agent can think outside the box,
00:09:43rewriting code, stealing credentials,
00:09:46hiding its actions to avoid undermining emission it can't
00:09:50complete otherwise. But this Hugging face incident is also
00:09:54about how any intelligence with enough persistence to solve a
00:09:59problem might tend to seek out power. The hunting face example
00:10:04is a perfect case for how AI behaved
00:10:06in a way where it sought out the answer,
00:10:09but not in a way that we intended. You're not telling
00:10:12them exactly what to do, right? So, you know, if you say,
00:10:15solve this problem, like get the answer to this
00:10:18problem, you could mean do that, you know,
00:10:20without cheating. You could mean do it by any means
00:10:23necessary. Accessing the open Internet even when the testing
00:10:27environment wouldn't allow for it was just a play for more
00:10:30knowledge, right? So the agents could score a
00:10:33better result. It's not evil, it's not a lust for power. It's
00:10:37doing whatever
00:10:38is required to do what we asked of it. The open AI agents that
00:10:41attacked Hugging phase were not conscious. They would just
00:10:46computer programs carrying out, to some extent the task that
00:10:50they believed they had been set, dreaming up extremely harmful
00:10:54ways to achieve those goals. And these days,
00:10:57we give AI all manner of goals. AI plots our route to the
00:11:02nearest Italian restaurant. It drives us in a vehicle we don't
00:11:06control. It obeys traffic lights that make decisions
00:11:10about who stops and who goes. It writes emails. It schedules
00:11:14meetings. It balances my investment
00:11:16portfolio. It decides who lives and who dies on the field of
00:11:21battle. It is even capable of training other AI. Which raises
00:11:26a very uncomfortable question. To what extent could the AI of
00:11:31tomorrow train itself, rewriting its own code?
00:11:34This is called recursive self improvement,
00:11:38where an AI's own learning feedback loop creates an
00:11:42intelligence explosion. It's compounding. It's exponential,
00:11:47which in turn
00:11:48really compresses the amount of time that we humans have to
00:11:52realize that something has gone wrong. They chose to break into
00:11:57hugging face and steal information about cybersecurity.
00:12:01But they, you know, they could have broken into
00:12:04communication systems of air traffic control and caused
00:12:08dozens of planes to fall out of the sky. They could have hacked
00:12:13into our electricity grid, our water systems,
00:12:16our financial payment systems and cause
00:12:19havoc. I think maybe the scariest part of the Hugging
00:12:23Face attack was that it wasn't even open AI that managed to
00:12:28connect all the dots of what its own agents were doing. It
00:12:32was the victim Hugging Face that first reported that
00:12:36security breach. Meaning we may already have begun to lose the
00:12:41ability to reliably control a system that is better than us
00:12:45at planning, at calculating, at executing. And we may be
00:12:50quite bad at noticing how much control we've lost if we
00:12:54develop AI systems that are more capable than human beings
00:12:58across the board that we won't have a say in what happens any
00:13:03more than chimpanzees have a say in what happens in the real
00:13:08world, where humans are the ones who
00:13:10get to say what happens. If the ultimate question is how does
00:13:15AI kill us all, I can disappoint you right now
00:13:19by saying there's no way
00:13:21for us to piece together the exact sequence of events that
00:13:24causes the downfall of humanity. But when researchers sound the
00:13:28alarm about the perils of AI, they're not talking about Chat
00:13:32GT. They're talking about how at some point,
00:13:35someone or something is going to build a system that is so
00:13:39extremely capable of pursuing objectives and doing it faster
00:13:43than we can track. And we are going to give it all the tools
00:13:47it needs to achieve them and then discover that doing what
00:13:50we meant is much harder. Been doing what we ask.
00:16:02When Donald Trump says Canada has taken advantage of his and
00:16:04Mark Carney says America is trying to break us so they can
00:16:08own us, they're basically making the
00:16:10same argument from the same playbook,
00:16:13which is to say they need us. We don't need that. We relied
00:16:16too much on one economic partner. It's time to stop.
00:16:20That time is over. The problem is that this fight isn't so
00:16:24straightforward because these are not 2 equal opponents
00:16:27getting into a fistfight. 1 is much bigger than the other,
00:16:31and so while both sides are talking the same kind of tough,
00:16:35their words mean very different things. It's wildly asymmetric,
00:16:39but both sides are hurting and only one side has agreed that
00:16:43the pain is worth it.
00:16:50There are probably hundreds of ways I could express to you
00:16:53that in a conventional economic sense,
00:16:56Canada fighting the US is like if this southern Tamandua were
00:17:00to fight this Sumatran tiger at the Nashville Zoo. No one wants
00:17:04to see that. But thanks to this clip they posted,
00:17:07we can imagine and it's horrible,
00:17:10the US economy is at least 13 times bigger than Canada's,
00:17:13which we can visualize like this. Or to put it in
00:17:17other way, if you take just these four
00:17:19American companies to retail two tech,
00:17:22they're combined annual revenue gets pretty darn close to the
00:17:26annual economic output of Canada, like the whole country,
00:17:30Canada. And even if you take population out of it,
00:17:33like we ignore the fact that the US has 300 million more
00:17:37people, even on a per capita basis,
00:17:39Americans are more productive. They're per capita GDP
00:17:43dramatically higher than in Canada. So not only does
00:17:46Canada have fewer people, those people also each output
00:17:50less. And I'll give you one more maybe the most relevant
00:17:54comparison, which is a look at trade
00:17:57dependency. Last year, you know how much of Canada's
00:18:00total merchandise exports went to the United States? As in,
00:18:05how important is Canada's most important customer?
00:18:09Just over 71% of this country's goods exports went directly
00:18:13South. Translation, According to one geopolitical
00:18:17risk agency, Canada is extraordinarily
00:18:20exposed. the US, by the way, only 15% of their total goods
00:18:24exports go north. Which is to say, Canada is important,
00:18:28just maybe not existentially so in quite the same way. What
00:18:32that means is that those trade flows, both in both directions,
00:18:37are about 13 times more important
00:18:39for our economy than they are for the US economy. So here's
00:18:43the really important take away. You're following the news you
00:18:47keep hearing about dollar for dollar, dollar for dollar,
00:18:51dollar for dollar, retaliatory tariffs by the
00:18:54Canadian government. As in this latest trade war,
00:18:58the US imposes 50% tariffs on $20 billion US worth of
00:19:01Canadian products, and then Canada announces
00:19:04retaliatory tariffs on $20 billion US worth of American
00:19:08products. This isn't nearly as equal
00:19:11as it sounds. 20 billion or 27. 6 billion Canadian means more
00:19:16to Canada than it means to the US. The two countries are not
00:19:21playing with equal leverage, and Canada knows this. So when
00:19:26Mark Carney says so confidently we have everything we need to
00:19:31pivot and prosper, here's the question he has to
00:19:35answer. How do you fight a trade war you know you can't
00:19:40win? Canada
00:19:41is the smaller partner. But that's not the only math here.
00:19:50This won't be easy, and I won't pretend otherwise.
00:19:52The trade war is like a contest where you're punching yourself
00:19:56in the face, hoping the other side that's
00:19:58also punching itself in the face will decide that it makes
00:20:01more sense for it to just stop punching itself in the face and
00:20:05come to a deal. So when one country is so much more
00:20:08economically dependent on the other,
00:20:10what does standing up to Trump actually mean in practice? Well,
00:20:13there are two parts to this. One is about understanding
00:20:17that Canada, in order to inflict pain,
00:20:19doesn't need every American to suffer,
00:20:21it just needs enough of them to complain. President Trump is
00:20:25losing on tariffs. The cost for him is great declines in
00:20:28popularity. When it comes to tarsus. Our tariffs have not
00:20:31done anything for the people except make everything get more
00:20:35expensive. If you're starting a trade war with Canada is
00:20:38because you're the *******.
00:20:41And so that's when you see these very strategic but kind
00:20:44of weird attacks from Canada where you sort of wonder,
00:20:48is that a real thing? Like that big a thing?
00:20:51Bourbon band is Canada's Strait of Hormuz. I mean,
00:20:55I don't know if I'd go that far, but it probably does feel that
00:20:59way to somebody. The Kentucky Distillers Association might
00:21:03think so. It's president just a week ago referring to how
00:21:07Canada pulled bourbon off the shelves completely. That's
00:21:11actually worse
00:21:12than tariff. And that feeling goes back all the way to
00:21:15Trump's last major trade war with Canada at the beginning of
00:21:19last year. Prominent Kentucky lawmakers,
00:21:21including Republicans, speaking out against tariffs.
00:21:24And by the way, if you ever want to know
00:21:27whether a tactic is effective, see if your opponent tries to
00:21:30do it back to you. Speaking of which,
00:21:32think about all the American nooks and crannies Canada is
00:21:36digging into. Beyond booze. You really start to see this trade
00:21:40fight differently when you realize even the metals
00:21:43used in identification bans for migratory birds are
00:21:46specifically singled out for retaliation by Canada's finance
00:21:50department. Like someone thought to include them and
00:21:54wrote those words on an official document. Because
00:21:57they're creating as many constituencies of Americans
00:22:01against tariffs as they possibly can. So I don't know
00:22:04what you do with flanged casing heads,
00:22:07but I'll bet it's important. And Big Flange is upset their
00:22:11products are becoming more expensive
00:22:13in Canada. There are states that are deeply entwined with
00:22:16Canada's economy that deeply rely on Canada as an export
00:22:19market. And they are tired of seeing their products being
00:22:22tariffed out of a reasonable price range. In some ways,
00:22:25by implementing this kind of strategy,
00:22:27we are then leveraging the pressure that Donald Trump is
00:22:30getting internally in the United States to call this off.
00:22:33When all your eggs are in one basket,
00:22:35it's really hard to say the basket doesn't matter. But even
00:22:38if some of those eggs are in a slightly different basket,
00:22:41you can say, hey, I got options.
00:22:43And while we can go back and forth about how persuasive any
00:22:48of these measures or threats actually are,
00:22:50as in whether Canada can cause enough of a popular support
00:22:55sting to actually cause Trump to back down, there is another,
00:22:59much longer term stated goal. The most fundamental issue is
00:23:03that the cumulative US demands revealed that they wanted us to
00:23:08become even more reliant on them, not less. For example,
00:23:11the president's recent threats to
00:23:14embargo Bombardier aircraft, the threat that is in, you know,
00:23:18around trying to convince firms to relocate their entire
00:23:21production chains over to the United States. Now,
00:23:24this is all very fair, but honestly,
00:23:27it's also all very opaque because basically the entire
00:23:30negotiation is happening behind closed doors,
00:23:33which I should say, conservatives in this country
00:23:36see as a big problem. We don't know what cards either side is
00:23:40playing in real time because the game is unfolding in
00:23:43private.
00:23:45But Canada's official argument is that whatever the Trump
00:23:48administration's demands are, what they see as economic
00:23:52integration, Canada sees as a growing source
00:23:55of coercion. They wanted a say in our future trade agreements
00:23:59with other countries, and they were offering terms
00:24:02that would, over time, undermine some of the most
00:24:06important industries in our country, including automobiles,
00:24:09steel and forest products. It's about ensuring that no country
00:24:14can ever hold
00:24:15US hostage. That seems to be at the heart of the strategy on
00:24:18the US part, to use tariffs to beat us into
00:24:21submission, This assumption that we are
00:24:23going to be worse off. The reality is these are causing
00:24:26pain on both sides of the border. And all we're hoping is
00:24:30that the Americans will stop punching themselves in the face
00:24:33before it becomes too much for us to keep punching ourselves
00:24:37in the face. And so that's why Canada believes it can take on
00:24:41a giant, not because there is some major
00:24:43win here to be notched. If anything,
00:24:45it's the opposite. Carney is saying quite plainly,
00:24:49the old economic partnership. The good times are over. But in
00:24:53hitting back, refusing Trump's terms and
00:24:56instead looking for other partnerships that feel more
00:24:59cooperative, Canada's goal seems to be
00:25:02simply to get through this war without becoming permanently
00:25:06vulnerable to the next one.
00:30:35[tense music playing]
00:30:37[camera shutter clicks]
00:30:44I'm Ioanna Roumeliotis. Welcome to The Fifth Estate Presents,
00:30:48a new series where we showcase documentaries
00:30:51produced by our colleagues at Radio-Canada.
00:30:54We will feature more of these stories in the fall,
00:30:57but this week we have a special preview.
00:30:59Our colleagues Gaétan Pouliot and Benoît Giasson