But let's take a deep breath and try to raise our spirits and cheer ourselves up, because the rest of my talk is going to be about the good news, that it's not inevitable, and we can absolutely do better, alright? (Applause) So... The real problem is that we lack a convincing plan for AI safety. People are working hard on evals looking for risky AI behavior, and that's good, but clearly not good enough. They're basically training AI to not say bad things rather than not do bad things. Moreover, evals and debugging are really just necessary, not sufficient, conditions for safety. In other words, they can prove the presence of risk, not the absence of risk. So let's up our game, alright? Try to see how we can make provably safe AI that we can control. Guardrails try to physically limit harm. But if your adversary is superintelligence or a human using superintelligence against you, right, trying is just not enough. You need to succeed. Harm needs to be impossible. So we need provably safe systems. Provable, not in the weak sense of convincing some judge, but in the strong sense of there being something that's impossible according to the laws of physics. Because no matter how smart an AI is, it can't violate the laws of physics and do what's provably impossible. Steve Omohundro and I wrote a paper about this, and we're optimistic that this vision can really work. So let me tell you a little bit about how. There's a venerable field called formal verification, which proves stuff about code. And I'm optimistic that AI will revolutionize automatic proving business and also revolutionize program synthesis, the ability to automatically write really good code. So here is how our vision works. You, the human, write a specification that your AI tool must obey, that it's impossible to log in to your laptop without the correct password, or that a DNA printer cannot synthesize dangerous viruses. Then a very powerful AI creates both your AI tool and a proof that your tool meets your spec. Machine learning is uniquely good at learning algorithms, but once the algorithm has been learned, you can re-implement it in a different computational architecture that's easier to verify. Now you might worry, how on earth am I going to understand this powerful AI and the powerful AI tool it built and the proof, if they're all too complicated for any human to grasp? Here is the really great news. You don't have to understand any of that stuff, because it's much easier to verify a proof than to discover it. So you only have to understand or trust your proof-checking code, which could be just a few hundred lines long. And Steve and I envision that such proof checkers get built into all our compute hardware, so it just becomes impossible to run very unsafe code. What if the AI, though, isn't able to write that AI tool for you? Then there's another possibility. You train an AI to first just learn to do what you want and then you use a different AI to extract out the learned algorithm and knowledge for you, like an AI neuroscientist. This is in the spirit of the field of mechanistic interpretability, which is making really impressive rapid progress. Provably safe systems are clearly not impossible. Let's look at a simple example of where we first machine-learn an algorithm from data and then distill it out in the form of code that provably meets spec, OK? Let’s do it with an algorithm that you probably learned in first grade, addition, where you loop over the digits from right to left, and sometimes you do a carry. We'll do it in binary, as if you were counting on two fingers instead of ten. And we first train a recurrent neural network, never mind the details, to nail the task. So now you have this algorithm that you don't understand how it works in a black box defined by a bunch of tables of numbers that we, in nerd speak, call parameters. Then we use an AI tool we built to automatically distill out from this the learned algorithm in the form of a Python program. And then we use the formal verification tool known as Dafny to prove that this program correctly adds up any numbers, not just the numbers that were in your training data. So in summary, provably safe AI, I'm convinced is possible, but it's going to take time and work. And in the meantime, let's remember that all the AI benefits that most people are excited about actually don't require superintelligence. We can have a long and amazing future with AI. So let's not pause AI. Let's just pause the reckless race to superintelligence. Let's stop obsessively training ever-larger models that we don't understand. Let's heed the warning from ancient Greece and not get hubris, like in the story of Icarus. Because artificial intelligence is giving us incredible intellectual wings with which we can do things beyond our wildest dreams if we stop obsessively trying to fly to the sun. Thank you. (Applause)
