AI Regulation, Explained

AI just went nuclear. It's gone from a niche subject to basically all anyone can talk about. Artificial. Artificial Artificial intelligence is making rapid strides. There's talk of a new evolution that could fundamentally change life on our planet. Today, people are reacting to AI the same way they did to the development of the atomic bomb 75 years ago. And that reaction is putting the future of AI in jeopardy because there's a wave of regulation coming that could prevent AI from reaching its incredible full potential. You have a nuclear regulatory Commission that governs how you build a plant and is licensed. Do you agree, Mr. Altman, that these tools you're creating should be licensed? These two technologies, nuclear and AI, have more in common than you think. AI has been making tons of headlines, but they've mostly been about Microsoft and Google fighting it out to integrate AI into their search engines. Now, don't get me wrong, the battle between big tech is important, but the bigger story is what's happening on a global scale. Governments are fighting for control over AI systems, and politicians are mapping out plans to create new regulations. We got it wrong with nuclear, but we can still get it right with AI. And by the end of this video, you'll know exactly how. The technologies that change the world create massive questions. Should they be used at all? If so, how? And who gets to access them? This was true for nuclear innovations decades ago, and it's true today for artificial intelligence. Both of these technologies actually have two entirely separate use cases, government and civilian. You can visualize them both on an X and Y graph. So let's do it. First, let's look at nuclear. Prior to the creation of the atomic bomb, no one had access to nuclear technology. But after World War II, it became clear that nuclear had applications beyond just weapons. That clarity created a common vision of a nuclear powered future with abundant energy too cheap to even measure. In that moment of hopefulness, the world shifted to the top right quadrant of the graph. Nuclear weapons were proliferating, but so was the promise of nuclear power. Because of this early progress, people started to dream even bigger. Perhaps instead of a world with nuclear power and nuclear weapons, we could just have the former. What if all those talented nuclear scientists left the defense sector and focused on improving civilian nuclear technology? That brings us to the bottom right quadrant. A world with lots of nuclear power but no nuclear weapons. Sounds like an incredible outcome. But of course, that's not how it panned out. The Cold War led to an aggressive arms race between the United States and the Soviet Union. And the world entered a multi decade era defined by Fear of nuclear annihilation. And even though we managed to avoid complete annihilation, the number of nuclear weapons skyrocketed. While nuclear power stagnated slowly, as militaries continued to develop nuclear weapons and civilian developments slowed to a crawl, we shifted from the top right quadrant to the top left. This is the worst of all possible outcomes. Now, maybe endless nuclear power in a world free of nuclear weapons was always a pipe dream. But we did have a real shot at using the window of time where the world's most powerful nations were stockpiling nuclear weapons to also invest in nuclear power. But we didn't. And very soon, I'll explain why that's relevant for AI. But first, let's quickly recap how this happened. How did humanity discover what's basically an infinite energy glitch? And then, instead of scaling it up to provide endless power, we locked it away inside nuclear submarines and aircraft carriers. Today, nuclear power generates about 10% of global electricity. It should be so much more. But nuclear power keeps running into hurdles. The biggest one is that nuclear power plant construction costs have dramatically increased over the last few decades. In the 1980s, the state of Washington had to cancel a project to build several nuclear power plants after costs skyrocketed from $4 billion to over $24 billion in just a few years. And more recently, two reactors in South Carolina were scrapped after costs more than doubled from $10 billion to $25 billion. You'd think the team in South Carolina would have learned from the disaster in Washington decades earlier, but they didn't. So how do the people behind these projects mess up so badly? And why does it keep happening? Proven technologies normally get cheaper over time. We've seen this in semiconductor manufacturing, where the learning curve has allowed prices to plummet and cheap electronics to make their way into every household. Cars, flat screen TVs, air travel. They've all become much more affordable over time. Not nuclear power plants, though. In the late 1960s, nuclear power plants cost around $1,000 per kilowatt of capacity. Just 10 years later, that cost nearly 10x to $9,000. The reason? Regulation. Starting in the 1960s and 70s, nuclear power plant operators have been hit with a huge number of new regulations. And these regulations aren't static. They often change, even during the construction process. Imagine building a house out of wood, and then halfway through, being told that building with wood was suddenly against the law. That's what happened to companies building nuclear power plants. Just look at this chart of changes to nuclear regulations over time in every year except for 1991. Something changed. And there were dozens of changes every year during the 1970s. That's why nuclear power isn't more economically competitive. But where did these regulations come from? What happened? At 4am on March 28, 1979, the Unit 2 reactor of the Three Mile island nuclear Generating Station in Pennsylv partially melted down. The accident had a huge impact on global nuclear safety regulations and directly led to changes in how plants were built and operators were trained. Cleaning and decommissioning the site cost about a billion dollars. Sounds pretty serious. So how many people died at Three Mile Island? Well, zero. Okay, so what about injuries? Also zero. Several studies have tried to determine if there was a statistically significant increase in cancer rates in the part of Pennsylvania near the reactor, and they didn't really find anything definitive. Although a small amount of radioactive gases leaked into the surrounding environment, the reactor's containment structure successfully prevented the release of more. But here's the interesting thing. Remember that chart showing tons of new regulations on nuclear power plants? That all started right at the beginning of the 1970s, nearly ten years before three Mile Island. The accident raised public concerns about the safety of nuclear power. But people were already worried about the potential risks. And. And that fear fueled an industry of consultants and businesses that sold safety concerns to governments and safety equipment to nuclear power plants. Every time new regulations rolled out, every power plant would have to comply. And that meant millions of dollars spent on new safety equipment and oversight procedures. Don't get me wrong. These regulations weren't all bad. Nuclear accidents do happen, and they should be prevented. Places like Fukushima and Chernobyl are bywords for disaster. Disaster. But it's important to keep things in perspective. Nuclear energy is much safer than almost every other kind of energy. Nuclear energy is so safe that it doesn't even register on this graph. It's even safer than wind power. Maybe that's because of decades worth of regulations, but there's a point at which regulation becomes over regulation. If you never leave your house, you'll be really safe. But you'll have regrets. That's arguably the situation we're in with nuclear power. And as I'll explain soon, we're arguably in the same situation today with AI but humans struggle to stay rational when thinking about life and death situations. People are more afraid of flying than driving on the highway, even though planes are way safer than cars. These days, the thought of dying in a plane crash feels completely out of our control. But everyone thinks they can avoid a car crash if they're a good driver. The Same goes for the relative risks of energy production. Wind power is really only dangerous to the workers who have to climb to the tops of those massive windmills. That's dangerous work, but it feels within the control of a single person. Nuclear radiation is invisible, and that's scary, but it really shouldn't be. Because although it's serious business, nuclear radiation just isn't actually that dangerous. Here's an interesting how many nuclear bombs do you think have gone off in the past hundred years? Well, most people know about the two used during the Second World War. Clever people might think of a few test bombs that were dropped as well. So what, maybe like 10? But the answer isn't 10, it's over 2,500 nuclear bombs have been detonated all over the world. Just take a look at this map. And yet we've learned to harness this technology and avoid doom. And this relates perfectly to the current moment in artificial intelligence development. Just like nuclear, we have an incredibly powerful new technology on our hands. AI could change the economy and the global balance of power at the same time. It's hard to clearly define its impact, and most people don't understand the technical side of AI. This uncertainty causes fear, and as we know, fear leads to regulations. We've seen it in the last few weeks. At the recent G7 summit in Japan, world leaders talked a lot about generative AI. The outcome of those talks was the Hiroshima process. As part of this process, the governments were of the G7 countries. That's the US, UK, Germany, France, Canada, Italy and Japanagreed that they would each hold cabinet level discussions and present results back to the G7 by the end of 2023. The leaders cited a concern that AI could lead to a new wave of incredibly convincing deepfakes and disinformation that could put further stress on democratic systems. The European Union is already taking steps towards regulating AI, such as requiring companies to notify users they're interacting with AI and banning the real time identification of people in public. The Hiroshima process sounds reasonable. Governments should be looking into AI, but the risk is that regulations are driven by fear and lack of understanding. That would bring us back to the upper left quadrant of our matrix. A world with plentiful AI weaponry in the hands of superpowers and rogue states, while regular people don't get any of the benefits. We can avoid this outcome, but we need to understand what's actually going on. Let's start with the risks. In the recent Senate subcommittee hearing on AI regulation, the Democratic senator from Connecticut, Richard Blumenthal, opened with a Fake AI generated voice recording designed to sound like him. Too often we have seen what happens when technology outpaces regulation. The unbridled exploitation of personal data, the proliferation of disinformation, and the deepening of societal inequalities. That voice was not mine, the words were not mine, and the audio was an AI voice. Cloning software deepfakes are getting pretty good. The amount of AI generated content is already exploding, and it's not going to slow down anytime soon. Open source models like stable Diffusion make it simple for anyone with a graphics card to produce remarkable images. But the cat's out of the bag at this point. The only way to actually stop the creation of fake images would involve seriously infringing on people's civil liberties. There's a weird misconception, probably from the 2016 election, that political misinformation is new. In fact, it's literally thousands of years old. In the 13th century BC, the Egyptian Pharaoh Ramses the Great was fighting a battle against the Hittites. Neither side could gain any ground, and the fight ended in a stalemate. The Egyptians in his Hittites signed a treaty ending the battle. But Ramesses didn't feel like taking the L, so he launched a propaganda campaign to make it look like the battle was a stunning victory for the Egyptians. He had murals painted on the walls of every temple, showing himself smiting foes to make it look like a decisive victory. Or consider this photo of Stalin strolling along a canal in Moscow. Here's the original. Soviet censors doctored the photo after Stalin had the man on the right to Nikolay Yazov executed. Today, we're still dealing with political misinformation, and the only solution remains developing a strong understanding of what can be fabricated. Sam Altman, the founder of OpenAI, the company that built ChatGPT, outlined exactly how people will adjust to these new creative tools during the Senate hearing. When Photoshop came onto the scene a long time ago, you know, for a while people were really quite fooled by Photoshopped images and then pretty quickly developed an understanding that images might be Photoshopped. Fortunately, this is already happening. Most young people are extremely skeptical about anything they see or read online. Instead of needing to shut down AI content generation, we can instead develop relationships with trusted resources that we know are authentic. Disinformation isn't a big enough problem for us to justify stopping further development into AI. But what about AI putting people out of their jobs? Even if you're not worried about getting tricked by a political misinformation campaign, the idea of AI taking everyone's job is a scary one, but I don't think it should be. Fear about technology replacing human labor is nothing new. For hundreds of years people have argued that industrial progress will cause broad unemployment. You've probably heard the term Luddite to describe someone who's anti progress, but have you ever thought about where the word comes from? From during the Industrial Revolution, the Luddites were English textile workers who opposed the use of cost saving machinery, including new types of looms. They would often launch clandestine raids to destroy these machines. But the thing is, they were wrong. There has been a nearly continual rise in both jobs and wages in capitalist countries since the Industrial Revolution. But despite decades of steady job growth, there have still been been two major anti technology moral panics. During the 2000s, everyone was worried that the Internet would cause significant job losses due to outsourcing. And then in the 2010s that fear shifted to robots stealing physical jobs. So far, both of those fears have basically been unfounded. And now we're hearing the same rumblings when it comes to AI. So many people are saying this time is different. Whenever you hear that phrase, you should always be a little scared. There are a bunch of reasons why AI won't cause mass unemployment. Let me explain. First off, despite all the crazy headlines, AI isn't actually developing all that quickly. Sure, it feels like ChatGPT came out of nowhere and changed everything overnight. But it didn't. OpenAI's GPT model was first released five years ago, and the company is nearly 10 years old at this point. And the history of AI goes back decades, with a long series of incredible breakthroughs followed by long winters where progress stagnated. Now, that's not to say that we're not at an inflection point. We absolutely are. Just like the iPhone revolutionized the mobile phone industry and made smartphones a thing, AI is at a point where it will switch from niche uses to being broadly available everywhere you look. But that doesn't necessarily mean job losses. Aside from ChatGPT, the most successful AI product to date is arguably GitHub's Copilot, which uses the same GPT model to provide code suggestions. While programming Copilot has been out for two years now and is basically universally loved by software developers. It doesn't matter if you're brand new to programming or one of the best programmers in the world. Copilot improves productivity hands down. But no one is replacing programmers with Copilot. It just doesn't work like that. The clue is in the name AI models like Copilot currently work best when they're used only alongside a human pilot. The AI is just the co pilot, helping to do repetitive tasks and avoid obvious errors. Every commercial passenger plane has at least two pilots, and they're both essential. But what about less complex tasks? There are lots of people who do basic data entry for a living or simple translation, and AI models are essentially superhuman at these tasks. Now, it's true that some of these jobs will be replaced, but the thing about capitalism is that it's the best mechanism ever invented for creating new types of jobs. Already, lots of people who did manual data entry are helping train AI models to do new tasks. They've shifted over, and that trend will continue for quite some time. And there are other areas of the economy that feel ripe for disruption, but might be more resistant than you think. Just look at this chart that shows the various prices of goods and services over time. What you'll notice is that healthcare and education have gotten dramatically more expensive, While technology like TVs and software have dropped in price significantly. Technology has never been evenly applied to our economy. Certain sectors benefit significantly from advances in technology, while in others, we're basically still doing things the same way as we did them decades ago. Sometimes industries lag behind because we just haven't found a way to use technology effectively there. No one thinks ChatGPT is going to put construction workers out of a job, for instance. But sometimes there are other structural factors at play. Take lawyers, for instance. As far back as 1973, the American Bar association was worried about the effect that automation would have on the legal profession. They established a special committee and went to work studying how lawyers would use computers in the future. Their conclusion was dire. They predicted that automation would dramatically reduce the number of lawyers, as well as how much money lawyers got paid. But over the following 50 years, the exact opposite happened. Today there are more lawyers than ever, and they're making more money than ever before. I think a few different things might happen to the legal profession over the next few years. First, they might just ban the use of AI outright. The Bar association already determines who can legally practice law, so they have the authority to do that. But they could also learn to work with AI effectively and add value in other ways, just as they did during the first computing revolution. Even if AI does wind up having a significant impact on the legal profession, my bet is on it happening very slowly. After all, courtroom sketch artists are still working today, even more than 100 years after the invention of the camera. I'd be very surprised if midjourney puts them out of business. We've talked about disinformation and unemployment, but what about the ultimate bad outcome? What if AI becomes super intelligent and decides to eliminate all humanity, humans, Terminator style? Even though it sounds like pure science fiction, it's not completely unimaginable. AI safety research has been a serious field of study for more than 10 years at this point. If artificial intelligence becomes dramatically smarter than every human being, how would humanity keep it in check? Well, fortunately, even the most hardcore AI pessimists don't think this is happening anytime soon. The stat you'll hear quoted most often is that 50% of AI researchers think that AI could produce an extremely bad outcome for humanity. That sounds scary. A 50% chance of an extremely bad outcome. But when you dig into the data, the numbers tell a different story. See, that stat includes anyone who assigned more than a 10% chance of a bad outcome. So even for the researchers who do believe AI possesses a real existential threat, they generally think that there is a very low probability that that a bad outcome will happen. The important thing to remember is that no one wants an AI takeover. Every AI research organization takes this stuff extremely seriously. That's why Sam Altman is testifying in front of Congress about it. Sam has clearly thought deeply about the impact of artificial intelligence, and that's why he's asking for regulation much earlier than most other tech founders would. But I think that's a case for optimism. Remember the scenario we really want to avoid. Void is the top left quadrant of our matrix. That's the place where civilians and companies don't have access to AI but governments can do almost anything they want. Nuclear weapons have dramatically shifted the global balance of power. Many world leaders want AI weapons to do exactly the same thing in their favor. China already uses AI to conduct surveillance on their population. New developments will only make authoritarian control easier to implement. That's the concern with something like the Hiroshima process that the leaders of the G7 countries want to implement. Carefully weighing risks and benefits is good. But waiting for too long when countries like China, Russia and Iran are already exploring the potential military uses of AI, that's a real risk. There is a potential for a good outcome here, though. One where no one is fooled by deepfakes, where no one loses their job, where everyone gets an AI copilot to and make them more effective and creative. That's the world we want to fight for. So getting regulation right here will be critical. If the government goes too far, it will restrict innovation. And that's why Sam isn't advocating for totally locking down AI development. AI is only going to become a bigger topic of conversation over the next few years. We have an opportunity to avoid making the same mistakes we did with nuclear power. It won't be easy, but if we succeed, we'll be safer, richer and happier as a result. But if we fail, we'll get stuck in the top left part of that matrix. Now, in order to understand the full history of AI, you have to watch this video next. Thanks a lot.