Transcription
Time to talk about AI. Right now, we're in this weird moment where lots of smart people agree that we're on the cusp of this truly world-changing technology, but some of them seem to be saying it's going to kill us all, while others are saying it's more profound than fire. "You know, I've always thought of AI as the most profound technology, more profound than fire or electricity..."
It's clear at this point that something big is happening. But my problem is, it's all just so vague. I want to know: How specifically would AI kill me? Or how would it dramatically transform my life for the better? In this video, that's what I'm going to try to figure out, what the most extreme bad and good possible features with AI actually look like, so that you and I can get ready. And more importantly, so that we can be a part of making sure that our real future goes right. "Artificial intelligence -" "artificial intelligence -" "artificial intelligence" "the benefits vastly outweigh the risks" "eventually they will completely out-think their makers -" "AI to begin to kill humans -" "AI has the potential to change society" "and a lot of people can be replaced by this technology" "Is this depressing? I don't see why it should be..." "This will be the greatest technology humanity has yet developed."
To understand why you're seeing so many mind-blowing AI tools all of a sudden, you need to understand how they actually work. And to do that we need to play some chess. This isn't one of those "oh my god, AI beats a person" kind of games. In this game, neither of the players are human. One is a famous chess engine, a system programmed by humans with insanely complex rules for how to play the game. The other is using a very different strategy. And that second player absolutely crushed the first... "It had learned the game without any of those rules, it just watched enough games to see what winning looked like."
That is Eric Schmidt, former CEO of Google and chairman of its parent company, Alphabet. Yeah. He was chairman of the company when they created that second player, AlphaZero. "Before that moment all of the game playing was done algorithmically, move here, evaluate this, do the math that..." But that's not how AlphaZero worked... "It didn't understand the principles of what a rook and a pawn and so forth and so on, it just knew how to play because it had observed enough games and it learned how to win."
In other words, our best systems had gone from using human-given rules to win, to using observation to win. "So you can think of that as moving from algorithms to learning. That to me was a major major deal." That ability to learn changed everything. It's what makes incredible tools like ChatGPT possible today. You now know this technique as "machine learning" "Machine Learning!" "machine learning..."
The reason that it suddenly feels like "AI" is everywhere is because of the incredible success of machine learning specifically. At a basic level, the idea is that instead of giving a computer a rigid set of rules that says "if this happens, then these are the possible outcomes," instead you give a computer a set of inputs and outputs and allow it to create the rules that turn one into the other. Meaning that it might come up with rules that we didn't think of or maybe don't even understand... but making the AI models that can do all the incredible things that you see now just recently became possible. And it's because the computers training them have gotten way more powerful. Look at this graph: So you see it going up and then around 2009 the computing power behind AI models just begins to explode.
That change is largely thanks to a switch in the physical technology used to do that training, going from CPUs to GPUs. My favorite way to show the difference between CPUs and GPUs is this Mythbusters demo back in 2009. That robot right there represents a CPU and it shoots paint in these little sequential bursts. It can get the job done but it's slow. And this robot represents a GPU, so instead of shooting paint one little bit at a time it can shoot in parallel. Basically, the physical tools behind AI are extremely powerful now and they're getting even more powerful, fast. According to OpenAI, the amount of computing power used in the largest AI models has been doubling every three months. This is why you're seeing now AIs able to pass the bar exam, make more realistic images, answer more complex questions. It's why this particular type of AI technology is "the risk that could lead to the extinction of humans" "AI is a fundamental existential risk for human civilization." "How do we know we can keep control?"
So we have this technology that can learn. And it's learning fast. And so of course, in large part thanks to Hollywood, we imagine that it'll learn to kill us. "My CPU is a neural net processor, a learning computer," but as much as these systems appear to be human, they're not. Why would they want to kill us? They don't want anything.
And yet, Bill Gates, Sam Altman, and hundreds of other tech leaders recently signed a 22-word statement that shocked me. I'll just read it to you: "Mitigating the risk of Extinction from AI should be a global priority alongside other societal scale risks such as pandemics and nuclear war." That is an incredible statement, that the development of AI is in the same realm of risk and importance as destruction by nuclear war.
To better understand why they feel this way, I turn to this survey. This is the same one that's been widely reported as "half of AI researchers give AI a 10% chance of causing human extinction." The specific question that they were asked is, "what probability do you put on human inability to control future advanced AI systems causing human extinction..."
So what's going on here? Well, the surveyors summarized an argument for why AI might be so dangerous by saying "it's essentially the old story of the genie in the lamp, or the sorcerer's apprentice, or King Midas: You get exactly what you ask for, not what you want."
Imagine this: In the future, someone creates a powerful machine learning system and gives it the desired output of a very accurate climate prediction. Then the AI, using its self-created rules, figures out that the more computing hardware it can use the more accurate its prediction will be. Then it figures out that by releasing a biological weapon there would be fewer humans taking up the valuable computing hardware that it needs. So that's what it does and then it gives its climate prediction to no one left. This is the category of thing that the researchers mean when they say "a system optimizing a function of n variables will often set the remaining unconstrained variables to extreme values." In other words, it might optimize for what we tell it to do at the expense of other things that we care about. "You get exactly what you ask for, not what you want." The term that researchers use for this is "specification gaming," and 82% of the researchers surveyed agreed that it was an important or the most important problem in AI today.
Specification gaming leading to disaster becomes less likely if we work to contain AI systems and we don't let them get connected to tools that might physically harm humans. Like don't give them the nuclear codes, but how likely is anything like this to actually happen? I honestly don't know, and I think neither does anyone, which is a big reason why all of those tech CEOs signed that letter and why you might have heard people advocating for a pause on AI development. However, there are real risks to not moving forward too.
There's a fairly large and impressive group of people now advocating for a pause on AI development. What do you think about that? I think it's a terrible idea, and the reason for that is that a pause would give time for our competitors, which starts with China, to catch up. At the moment, the US is in a very strong position. We have all of the top models, we have the majority of the researchers, we have the majority of the hardware, we have the majority of the data that's being used. That's not going to be true forever, but this is a critical time for us to build this technology in American values, liberal values, not authoritarian values.
So we've created these tools that have started to become so powerful that we're concerned about how well they might do what we ask, and at the same time, every country, every company is incentivized to build them first with their own interests in mind. But why should we want AI in the first place? Like what's the goal here??
In my view, the most positive extreme case for AI that I've heard isn't how much better or faster it can do the mundane things that we already do; it's how it could leapfrog us to do things that we can't. You might be wondering, how? Because of how incredibly good machine learning systems are at pattern matching, they can sometimes give us results that we can verify are correct but we don't totally understand how it got there. It's funny, it's the same skill that scares us is the one that gives this tool such incredible potential. And if you're feeling a little bit skeptical here, that's totally fine and understandable; I was too, until I heard this example: In 2021, researchers used machine learning on a problem that had, up until very recently, been called "one of the most important yet unresolved issues of modern science." It figured out the structure of a protein from just amino acid building blocks. For decades, our best effort to do this has been to spend hundreds of thousands of dollars per protein to shoot X-rays at them all in the hopes of learning just a little bit more about our own bodies and make better medicines. This is how we got new treatments for diabetes and sickle cell disease, breast cancer, and the flu, but then researchers fed pairs of sequences and 3D structures that we already knew into a machine learning system and allowed it to learn the patterns between them. And the result was just incredible. We now have predicted 3D structures for nearly all proteins known to science, more than 200 million of them. "Deepmind's AlphaFold" "AlphaFold" "AlphFold was able to do in a matter of days what might take years!" "solving an impossible problem in biology..."
I get a little emotional just thinking about this, about how many people's lives might actually get better because of this knowledge explosion. And this is just one example of what we've already been able to do. As machine learning systems get better and better, people have extremely high hopes about what we might be able to use them for... "We have lots of problems in the world. Think about climate change, for example. Climate change will be solved to the degree it's solved by using techniques that are very complicated and very powerful that will have as their basis generative AI. And I think that we want that future."
After learning more about AI and this moment that we're in, I think I've figured out why it feels so confusing and so hard: We're living inside a trolley problem. Down one path is the status quo, life without AI. But with this incredible new tool, we can pull ourselves onto another path, one that could fundamentally change society. But we just don't know, at what cost? Will AI give us what we ask for or what we actually want?
In this video, we've only talked about the most extreme futures with AI. In other episodes, we're going to go deep into specific applications. We'll go full on Huge If True into AI in music and news and robotics and climate and food and sports and more to explore how these tools might transform our world. It's easy to dismiss it as crazy when you hear someone say that AI might be "more profound than fire or electricity," and while the cynical side of my brain wants to say that it's probably true that most of the most ambitious AI efforts will likely fail, the more optimistic Huge If True side of my brain just keeps wondering: What if they actually work?