Transcription
What are your spiritual views? >> Um, uh, Jewish. I And I would say I have like a fairly traditional view of the world that way.
>> So, you're religious. You believe in God?
>> Uh, I don't I don't I'm not like a literal I don't believe the I'm not like a literalist on the Bible, but I I'm not someone who says like I'm culturally Jewish. Like, if you ask me, I would just say I'm Jewish.
>> But do you believe in God? Like do you believe that there is a force larger than people that created people created the earth set down a specific order for living that there's an absolute morality attached that comes from that god?
>> Um I think probably like most other people I'm somewhat confused on this but I believe there is something bigger going on than you know can be explained by physics. Yes.
>> So you think the earth and the people were created by something. It wasn't just like a spontaneous accident.
>> Um, do I would I say that it does not feel like a spontaneous accident? Yeah. I don't I don't think I have the answer. I don't think I know like exactly what happened, but I think there is a mystery beyond my comprehension here going on.
>> Have you ever felt communication from that force or from any force beyond people, >> beyond the material?
>> Not not no not really. I ask because it seems like the technology that you're creating or shephering into existence will have more power than people on this current trajectory. I mean that will happen. Who knows what will actually happen? But like the graph suggests it and so that would give you, you know, more power than any living person. So I'm just wondering how you see that.
>> Um I used to worry about something like that much more. I think what will happen I used to worry a lot about the concentration of power in one or handful of people or companies because of AI.
>> Yeah.
>> Um what it looks like to me now and again this may evolve again over time is that it'll be a huge upleveling of people uh where everybody will be a lot more powerful or that embraces the technology but a lot more powerful. But that's actually okay. that scares me much less than a small number of people getting a ton more power. If if the kind of like ability of each of us just goes up a lot because we're using this technology and we're able to be more productive and more creative or discover new science and it's a pretty broadly distributed thing like billions of people are using it. Um that I can wrap my head around that feels okay.
>> So you don't think this will result in a radical concentration in power?
>> It looks like not but again the trajectory could shift again and we'd have to adapt. I I used to be very worried about that and I think the the kind of conception a lot of us in the field had about how this might go could have led to a world like that. But what's happening now is tons of people use chatbt and other chatbots and they're all more capable. They're all kind of doing more. They're all, you know, able to achieve more, start new businesses, come up with new knowledge and that feels pretty good.
So if it's nothing more than a machine and just the product of its inputs then the two obvious questions like what are the inputs like what's the moral framework that's been put into the technology like what is right or wrong com according to jet >> GPT you mean answer that one first question um so on that one I >> someone said something early on in charge uh when that really has stuck with me, which is one person at a lunch table said something like, you know, we're trying to train this to be like a human, like we're trying to learn like a human does and read these books and whatever. And then another person said, no, we're really like training this to be like the collective of all of humanity. We're reading everything, you know, we're trying to learn everything. We're trying to see all these perspectives and and if we do our job right, all of humanity, good, bad, the, you know, very diverse set of perspectives, some things that we'll feel really good about, some things that we'll feel bad about, that's all in there. Like this is learning the kind of collective, experience, knowledge, learnings of humanity. Now, the base model gets trained that way, but then we do have to align it to behave one way or another and say, you know, I will answer this question, I won't answer this question. Um, and we have this thing called the model spec where we try to say, you know, here's here are the rules we'd like the model to follow. It may screw up, but you know, you you can at least tell if if it's doing something you don't like, is that a bug or is that intended? And we have a debate process with the world to get input on that spec. Um, we give people a lot of freedom and customization within that. There are, you know, absolute bounds that we draw, but then there's a default of if you don't say anything, how should the model behave? What should it do? What are What are how should it answer moral questions? How should it refuse to do something? What should it do? And this is a really hard problem. Um, you know that we have a lot of users now and they come from very different life perspectives and what they want. Um, but on the whole uh I have been pleasantly surprised with the model's ability to learn and apply a moral framework.
But what moral framework? I mean the sum total of like world literature or philosophy is at war with itself like the marquee is you know like nothing in common with the gospel of John. So like how do you decide which is superior? That's why we wrote this like model spec of here's how we're going to handle these cases, >> right?
But what criteria did you use to decide what the model is? >> Oh, um like who decided that? Who did you consult? Like what's you know, why is the gospel of John better than the marquee dad?
>> Uh we consulted like hundreds of moral philosophers uh people who thought about like ethics of technology and systems and at the end we had to like make some decisions. The reason we try to write these down is because a we won't get everything right. Uh b we need the input of the world. And we have found a lot of cases where there was an example of something that seems that seemed to us like you know a fairly clear decision of what to allow or not to allow where users convinced us like hey by blocking this thing that you think is an easy decision to make. um you are not allowing this other thing which is important and there's like a difficult trade-off there in in general the attention that so a principle that I normally like is to treat our adult users like adults very strong guarantees on privacy very strong guarantees on individual user freedom and this is a tool we are building you get to use it within very broad framework on the other within a very very broad framework on the other hand as this technology becomes more and more powerful. Um there are clear examples of where society has an interest that is in significant tension with user freedom and we could start with an obvious one like should chatbt teach you how to make a boweapon. Now you might say hey I'm just really interested in biology and I'm a biologist and I want to you know I'm not going to do anything bad with this. I just want to learn and I could go read a bunch of books but cha can teach me faster and I want to learn how to you know I want to learn about like novel virus synthesis or whatever and maybe you do maybe you really don't want to like cause any harm but I don't think it's in society's interest for chacht to help people build boweapons and so that's a case >> sure that's an easy one though there are a lot of tougher ones um >> I did say start with an easy one >> well every decision is ultimately a moral decision and and we make them without even recognizing them as such and this technology will be in effect making them for us >> and so >> well I don't agree with it'll be making them for us but it will have >> they'll be influencing the decisions for sure >> and um because it'll be embedded in daily life >> and so who made these decisions like who spec who who are the people who decided that one thing is better than another >> um you mean like >> what are their names >> which kind of decision I >> the the basic the specs that you oh uh that you alluded to that create the framework that that does attach a moral weight to worldviews and decisions like you know liberal democracy is better than Nazism or whatever they seem obvious and in my view are obvious but are still moral decisions. So who who made those calls?
>> Um as a matter of principle I don't like dox our team but we have a model behavior team and the people who want to >> well it just it affects the world. What I was going to say is the person I think you should hold accountable for those calls is me. Like I'm a public face eventually. Like I'm the one that can overrule one of those decisions or our board.
>> Um and turn 40 this spring.
>> I will make it's pretty heavy. I mean do you think as and it's not an attack but it's I wonder if you recognize sort of the the importance.
>> How do you think we're doing on it? I'm not sure, but I think I think these decisions will have, you know, global consequences that we may not recognize at first. And so I just wonder >> there's a lot you get into bed at night and think like the future of the world hangs on my judgment.
>> Look, I don't sleep that well at night. Um, there's a lot of stuff that I feel a lot of weight on, but probably nothing more than the fact that every day hundreds of millions of people talk to our model. And I don't actually worry about us getting the big moral decisions wrong. Maybe we will get those wrong, too. But what I worry, what I lose most sleep over is the very small decisions we make about a way a model may behave slightly differently. But it's talking to hundreds of millions of people. So the net impact is big.
So but I mean all through history like recorded history up until like 1945, people always deferred to what they conceived of as a higher power in order. Hmurabi did this. Every every moral code is written with reference to a higher power. There's never been anybody who's like, "Well, that kind of seems better than that." Everybody appeals to a higher power and you said that you don't really believe that there's a higher power communicating with you. So I'm wondering like where did you get your moral framework?
>> Um I mean like everybody else I think the environment I was brought up in probably is the biggest thing like my family, my community, my school, my religion probably that
>> do you ever think which is I mean I think that's a very American answer like everyone kind of feels that way but in your specific case since you said these decisions rest with you that means that the millia in which you grew up and the assumptions that you embibed over years are going to be transmitted to the globe to billions of people. That's like a thing.
>> I want I want to be clear. I view myself more as like a I think our the world like our user base is going to approach the collective world as a whole.
>> And I think what we should do is try to reflect the moral I don't want to say average but the like collective moral view of that user base. I don't there's plenty of things that ChachiBT allows that I personally would disagree with. Um the but I I don't like obviously I don't wake up and say I'm going to like impute my exact moral view and decide that like this is okay and that is not okay and this is a better view than this one. What I think ChachiBT should do is reflect that like weighted average or whatever of humanity's moral view which will evolve over time and we are here to like serve our users. We're here to serve people. this is like you know this is a techn technological tool for people and I don't mean that it's like my role to make the moral decisions but I think it is my my my role to make sure that we are accurately reflecting the preferences of of humanity or for now of our user base and eventually of humanity.
>> Well, I mean humanity's preferences are so different from the average middle American preference. So, would you be comfortable with an AI that was like as against gay marriage as most Africans are?
>> Um, there's a version of that like I think individual users should be allowed to have a problem with gay people. And if that's their considered belief, uh, I don't think the AI should tell them that they're wrong or immoral or dumb. I mean, you can, you know, sort of say, hey, you want to think about it this other way, but like you I probably have like a bunch of moral views that the average African would find really problematic as well, and I think I should still get to have them, >> right? I think I probably have more comfort than you with like allowing a sort of space for people to have pretty different moral views or at least I think in my role as like running Chad GPT, I have to do that.
>> So, what will be the downside? I mean, I can see the upsides for sure. efficiency. Medical diagnosis seems like it's going to be much more accurate. Fewer lawyers. Thank you very much for that. But what will what are the downsides that you worry about?
>> I I think this is just like kind of how I'm wired. I always worry the most about the unknown unknowns. If if it's a if it's a downside that we can really like be confident about and think about. Um, you know, we talked about one earlier which is these models are getting very good at bio and they could help us design biological weapons. uh you know, engineer like another co style pandemic I worry about that but because we worry about it I think we and many other people in the industry are thinking hard about how to mitigate that the the unknown unknowns where okay there's like a there's a societal scale effect from a lot of people talking to the same model at the same time this is like a silly example but it's one that struck me recently um LLMs like ours and our language model and others have a kind of certain style to them you know they talk in a certain rhythm and they have a little bit unusual addiction and maybe they overuse m dashes and whatever. And I noticed recently that real people have like picked that up and it was an example for me of like man you have enough people talking to the same language model and it actually does cause a change in societal scale behavior.
>> Yes.
>> And you know did I think that chat GPT was going to make people use way more m dashes in real life? Certainly not. It's not a big deal, but it's an example of where there can be these unknown unknowns of this is just like this is a brave new world.
So, you're saying, I think correctly and succinctly, that technology changes human behavior, of course, and changes our assumptions about the world and each other and all that. And a lot of this you can't predict. Considering that we know that, why shouldn't the internal moral framework of the technology be totally transparent? We prefer this to that. I mean, this is obviously a religion. I don't think you'll agree to call it that. It's very clearly a religion to me. That's not an attack.
>> I >> actually would love I don't take that as an attack, but I would love to hear what you mean by that.
>> Well, it's it's something that we assume is more powerful than people. and to which we look for guidance. I mean, you're already seeing that on display. What's the right decision? I ask that question of whom? My closest friends, my wife and God. And this is a technology that provides a more certain answer than any person can provide. So, it's a it's a religion. And the beauty of religions is they have a catechism that is transparent. I know what the religion stands for. Here's what it's for. Here's what it's against. But in this case, I press and I wasn't attacking you sincerely. I was not attacking you, but I'm just trying to get to the heart of it. The beauty of religion is it admits it's a religion and it tells you what it stands for. The unsettling part of this technology and not just your company, but others is that I don't know what it stands for, but it does stand for something. And unless it admits that and tells us what it stands for, then it guides us in a kind of stealthy way toward a conclusion we might not even know we're reaching. Do you see what I'm So like why not just throw it open and say chp is for this we're you know we're for suicide for the terminal ill but not for kids or whatever like why not just tell us
>> I mean the reason we write this long model spec and the reason we keep expanding over time is so that you can see here is how the here's how we intend for the model to behave um what used to happen before we had this is people would fairly say I don't know what the model's even trying to do and I don't know if this is a bug or the intended behavior tell me what this long long document of you know tell me how you're like when you're going to say do this and when you're going to show me this and when you're going to say you won't do that. The reason we try to write this all out is I think people do need to know.
>> And so is there a place you can go to find out a hard answer to what your preferences as a company are preferences that are being transmitted in a not entirely straightforward way to the globe. Like where can you find out what the company stands for, what it prefers?
>> I mean our model spec is the like answer to that. Now I think we will have to make it increasingly more detailed over time as people use this in different countries. there's different laws whatever else like it will not be a it will not work the same way for every user everywhere but and that doc so that I I expect that document to get very long and very complicated but that's why we have it
thanks for watching our YouTube channel we hope you'll subscribe to it and by the way you can hit the little bell on there and get notifications every time we produce a video we hope you'll do that also