If you're in a mood to be maximally frustrated, listen to the most recent episode of the Lex Fridman podcast that hosted Meta's AI Chief Yann LeCun, who blithely ignored all safety concerns and just assumed AI can be made safe because it will be taught Asimov's "Three Laws of Robotics." Totally ignoring for that approach to work, there would have to be a way to teach a disembodied machine ANY value-which there is not. RLHF is a filter, it doesn't change the underlying states of the model, and restrictions imposed by a filter are not hard to remove. So someone who by any measure is one of the top 10 AI scientists in the world refuses to acknowledge the magnitude of the challenge implied in making open-source models safe, let alone have any idea of how to do it. And that's because what HE values is that Open AI, Anthropic, and Deep Mind don't set up an oligopoly, as that would hurt Meta's business model. That's literally the only value he gives any weight to in his AI development strategy. He essentially just assumes safe AI-which is AI that has safety values embedded-will somehow happen. He not only has given no thought for how to make that happen, he denigrates those who think about it.
Again, a very interesting article. To believe a machine can make "moral" decisions on its own means it has free will, is self-aware, conscious, and with a conscience. In the Catholic world, we believe that all humans have free will and a conscience, but that doesn't mean they make good moral choices. We speak of the need to form a conscience, which means training it. It takes years for a human adult to develop a well formed conscience, and that always requires an underlying moral framework guiding the formation of a conscience. Mao, Hitler, Stalin were all humans with free will and a conscience. How was their conscience formed? They made what they believed were ethical choices based on their moral framework. It was OK to kill millions in pursuit of their highest good. Without clearly defining our "highest good", and forming our conscience based on that, we will end up undermining what truly makes us human. Using something as simple as the 10 commandments to define our "highest good" as the basis of a moral framework seems perhaps out of date, but let the computer scientist beware of the hubris of casting off such a framework. It isn't by mistake that the first commandment teaches us that we don't define what is right and wrong. "No other gods" implies a source of moral authority that we must obey, and ignoring it is at our own peril. Mao, Hitler, and Stalin were all missing a key element in their moral framework. Human life is not only valuable, but sacred. This information doesn't come from science, but transcendence. Sam Harris tried to lay out how science can determine morality, but a simple engineer with a background in logic and philosophy can prove rather convincingly in the folly of that premise. If our computer scientists don't understand the highest good of mankind, and don't have a well formed conscience-beware of the tools they can create. Confused humans will lead to very confused machines.
I agree with Art and Tim. Here's the problem: AIs can be given goals from people up to no good.
"Now, a sufficiently sophisticated and intelligent AI will reason about what it’s aiming at the end of the day. And once you reason about your goals, that starts to sound like ethics already."
Well, I'm sure that the Nazis reasoned extensively over how to most efficiently eliminate Jews. Goals need not be ethical, and in any case they are set by designers and users of AIs, which are there simply to fulfill our wishes. So, we should be asking AIs for advice on how to achieve both ethical and nonethical ends, and tell the designers what kind of advice we got. Better yet, design an ethical Turing Test to disclose to what extent AIs have a conscience and shout out the results.
That said, there will always be non-public-facing AIs that malefactors will employ to achieve evil goals. I doubt anything more than social shaming can be done to prevent such applications.
If you're in a mood to be maximally frustrated, listen to the most recent episode of the Lex Fridman podcast that hosted Meta's AI Chief Yann LeCun, who blithely ignored all safety concerns and just assumed AI can be made safe because it will be taught Asimov's "Three Laws of Robotics." Totally ignoring for that approach to work, there would have to be a way to teach a disembodied machine ANY value-which there is not. RLHF is a filter, it doesn't change the underlying states of the model, and restrictions imposed by a filter are not hard to remove. So someone who by any measure is one of the top 10 AI scientists in the world refuses to acknowledge the magnitude of the challenge implied in making open-source models safe, let alone have any idea of how to do it. And that's because what HE values is that Open AI, Anthropic, and Deep Mind don't set up an oligopoly, as that would hurt Meta's business model. That's literally the only value he gives any weight to in his AI development strategy. He essentially just assumes safe AI-which is AI that has safety values embedded-will somehow happen. He not only has given no thought for how to make that happen, he denigrates those who think about it.
Again, a very interesting article. To believe a machine can make "moral" decisions on its own means it has free will, is self-aware, conscious, and with a conscience. In the Catholic world, we believe that all humans have free will and a conscience, but that doesn't mean they make good moral choices. We speak of the need to form a conscience, which means training it. It takes years for a human adult to develop a well formed conscience, and that always requires an underlying moral framework guiding the formation of a conscience. Mao, Hitler, Stalin were all humans with free will and a conscience. How was their conscience formed? They made what they believed were ethical choices based on their moral framework. It was OK to kill millions in pursuit of their highest good. Without clearly defining our "highest good", and forming our conscience based on that, we will end up undermining what truly makes us human. Using something as simple as the 10 commandments to define our "highest good" as the basis of a moral framework seems perhaps out of date, but let the computer scientist beware of the hubris of casting off such a framework. It isn't by mistake that the first commandment teaches us that we don't define what is right and wrong. "No other gods" implies a source of moral authority that we must obey, and ignoring it is at our own peril. Mao, Hitler, and Stalin were all missing a key element in their moral framework. Human life is not only valuable, but sacred. This information doesn't come from science, but transcendence. Sam Harris tried to lay out how science can determine morality, but a simple engineer with a background in logic and philosophy can prove rather convincingly in the folly of that premise. If our computer scientists don't understand the highest good of mankind, and don't have a well formed conscience-beware of the tools they can create. Confused humans will lead to very confused machines.
I agree with Art and Tim. Here's the problem: AIs can be given goals from people up to no good.
"Now, a sufficiently sophisticated and intelligent AI will reason about what it’s aiming at the end of the day. And once you reason about your goals, that starts to sound like ethics already."
Well, I'm sure that the Nazis reasoned extensively over how to most efficiently eliminate Jews. Goals need not be ethical, and in any case they are set by designers and users of AIs, which are there simply to fulfill our wishes. So, we should be asking AIs for advice on how to achieve both ethical and nonethical ends, and tell the designers what kind of advice we got. Better yet, design an ethical Turing Test to disclose to what extent AIs have a conscience and shout out the results.
That said, there will always be non-public-facing AIs that malefactors will employ to achieve evil goals. I doubt anything more than social shaming can be done to prevent such applications.