WEBVTT - Inside OpenAI, the Architect of ChatGPT 0:00:01.240 --> 0:00:04.040 How does all this talk about, you know, relationships and AI, Like, 0:00:04.240 --> 0:00:07.040 could you see yourself developing a relationship with an AI? 0:00:07.600 --> 0:00:11.760 I'd say yes, as a reliable tool that enhances my life, 0:00:11.880 --> 0:00:12.959 makes my life better. 0:00:14.360 --> 0:00:18.040 I'm Emily Chang, and this is the circuit. We're inside 0:00:18.079 --> 0:00:20.640 a nondescript building in the heart of San Francisco where 0:00:20.640 --> 0:00:23.640 one of the world's buzziest startups is making our AI 0:00:23.760 --> 0:00:27.560 powered future feel more real than ever before. It's giving 0:00:27.600 --> 0:00:30.880 me very West World Spa vibes. It's almost like suspended 0:00:30.960 --> 0:00:32.959 in space and time a little bit. They're behind two 0:00:33.040 --> 0:00:37.640 monster hits, chat Gypt and Dollie, and somehow the biggest 0:00:37.640 --> 0:00:40.800 tech giants to market, kicking off this competitive race that's 0:00:40.840 --> 0:00:43.720 forced them all to show us what they've got. Is 0:00:43.760 --> 0:00:46.800 it magic? Is it just algorithms? Is it gonna save 0:00:46.880 --> 0:00:50.120 us or destroy us? To help us separate AI height 0:00:50.360 --> 0:00:53.400 from reality, I sat down with LinkedIn co founder and 0:00:53.440 --> 0:00:56.760 Facebook investor Reid Hoffman, who was an early backer and 0:00:56.840 --> 0:01:00.560 board member of Open Ai. He also used Chatchi to 0:01:00.600 --> 0:01:04.400 write a novel. But first, here's my conversation with Miramrati, 0:01:04.560 --> 0:01:07.920 chief technology officer from inside open Ai. 0:01:09.200 --> 0:01:10.720 Well, thank you so much. For doing us. It's really 0:01:10.720 --> 0:01:12.600 great to have you, and you've been very busy. 0:01:12.800 --> 0:01:14.600 I want you to take us back a little bit 0:01:14.920 --> 0:01:18.039 when you were making the decision about releasing chat Gpt 0:01:18.160 --> 0:01:20.040 into the wild. I'm sure there was like a go 0:01:20.240 --> 0:01:22.920 or no go moment. Take me back to that day. 0:01:23.400 --> 0:01:26.920 You know, we had chat gipt for a while, and 0:01:26.959 --> 0:01:30.640 we had been exploring it internally and with a few 0:01:30.720 --> 0:01:35.880 trusted users, and we realized that we sort of hit 0:01:35.920 --> 0:01:39.560 a point where we could really benefit from having more 0:01:39.680 --> 0:01:42.880 feedback and having more people try to break it and 0:01:43.160 --> 0:01:46.400 try to figure out how best to use it. Let's 0:01:46.400 --> 0:01:49.640 make sure that we've got some guardrails in place and 0:01:50.040 --> 0:01:53.600 start rolling it out incrementally so we can get feedback 0:01:53.600 --> 0:01:56.120 from how people are using it, what are the risks, 0:01:56.160 --> 0:02:00.600 what are the limitations, and learn more about this technology 0:02:00.640 --> 0:02:03.320 that we have created and started bringing it in the 0:02:03.440 --> 0:02:04.680 public consciousness. 0:02:04.720 --> 0:02:07.040 So you wanted people to break it or try. 0:02:06.920 --> 0:02:10.200 Yes, we definitely wanted people to try to break it 0:02:10.280 --> 0:02:15.000 and find the fragilities in the system. We had reached 0:02:15.000 --> 0:02:17.040 the point where we had done a lot of that 0:02:17.240 --> 0:02:18.280 internally and. 0:02:18.280 --> 0:02:20.320 With a small group of people. 0:02:20.880 --> 0:02:25.239 External experts as well, and we wanted more external researchers 0:02:25.280 --> 0:02:26.639 to play with it. 0:02:26.639 --> 0:02:30.160 It became the fastest growing tech product in history. Did 0:02:30.200 --> 0:02:32.400 that surprise you, I mean, what was your reaction to 0:02:32.440 --> 0:02:33.240 the world's reaction. 0:02:33.760 --> 0:02:39.000 Yeah, it was a huge surprise for us. We were surprised. 0:02:38.440 --> 0:02:42.080 By how much it captured the imaginations of the general 0:02:42.120 --> 0:02:46.760 public and how much people just loved spending time talking 0:02:46.880 --> 0:02:49.640 to this AI system and interacting with it. 0:02:49.880 --> 0:02:51.240 I want to take a step back a little bit, 0:02:51.320 --> 0:02:52.799 you know, because a lot of people still don't really 0:02:52.880 --> 0:02:56.520 understand how it works. Chatgbts trained on you know, tons 0:02:56.520 --> 0:02:57.799 and tons of data and text. 0:02:58.080 --> 0:03:00.160 It can now mimic a human. It can write, It 0:03:00.240 --> 0:03:02.360 can code at the most. 0:03:02.120 --> 0:03:05.600 Basic level and in the most succinct way that you 0:03:05.680 --> 0:03:08.120 can How does it work? How does this all happen? 0:03:08.480 --> 0:03:08.679 So? 0:03:08.840 --> 0:03:13.200 TUGBD is a neural network that has been trained on 0:03:13.760 --> 0:03:18.400 a huge amount of data on a massive supercomputer, and 0:03:18.639 --> 0:03:22.200 the goal during this training process was to predict the 0:03:22.240 --> 0:03:25.880 next word in a sentence, and we found out that 0:03:25.960 --> 0:03:30.200 by doing this we also got the ability to understand 0:03:30.320 --> 0:03:34.520 the world in text more like humans do. The goal 0:03:34.560 --> 0:03:38.640 here is to have these systems have more robust concepts 0:03:39.040 --> 0:03:43.480 of reality, similar to how we think of the world. 0:03:43.480 --> 0:03:46.880 We don't just think and reason in text. We also 0:03:47.520 --> 0:03:51.720 obviously have the world in images, visual world around us. 0:03:51.920 --> 0:03:54.320 That's been the goal over time, which is why we've 0:03:54.320 --> 0:03:58.800 been adding more and more modalities. And it turns out 0:03:58.840 --> 0:04:02.520 that as you trained larger and larger models and more 0:04:02.560 --> 0:04:08.080 and more data, the capabilities of these models also increase. 0:04:08.320 --> 0:04:13.600 They become more powerful, more helpful, and as you invest 0:04:13.680 --> 0:04:16.919 more on alignment and safety, they become more reliable and 0:04:16.960 --> 0:04:17.719 safe over time. 0:04:17.880 --> 0:04:19.800 So I'd love to hear a little bit more about 0:04:19.839 --> 0:04:22.159 your personal story. I know you grew up in Albania. 0:04:22.320 --> 0:04:25.120 What was the road like from Albania to Silicon Valley. 0:04:25.360 --> 0:04:29.440 I grew up in Albania, and when I was growing 0:04:29.520 --> 0:04:33.960 up in Albania, it was a pretty tumultuous place politically 0:04:34.000 --> 0:04:37.880 and economically, so I always knew that I wanted to 0:04:38.920 --> 0:04:43.240 study abroad. I always loved learning, and in this pursuit 0:04:43.279 --> 0:04:47.120 of knowledge, took me to Canada on a scholarship, and 0:04:47.360 --> 0:04:51.200 from there I came to the US and I've stayed 0:04:51.279 --> 0:04:52.320 in the US ever since. 0:04:52.880 --> 0:04:55.680 You've worked on aerospace, you worked at Tesla, you worked 0:04:55.680 --> 0:04:59.000 on virtual reality. How did you become CTO of Open Ai. 0:04:59.320 --> 0:05:03.760 It's been My training has been in mechanical engineering. I've 0:05:03.800 --> 0:05:08.919 always loved maths and physics and these were my favorite 0:05:08.960 --> 0:05:13.320 subjects as a kid. So my training took me from 0:05:13.360 --> 0:05:18.800 aerospace engineering to automotive engineering and then in applications in 0:05:18.880 --> 0:05:21.320 virtual reality and augmented reality. 0:05:21.960 --> 0:05:23.120 But there was always this. 0:05:23.400 --> 0:05:30.840 You know, deep technological advancement in pursuit of some problem 0:05:31.160 --> 0:05:35.920 that makes our lives a little better. And five years 0:05:35.920 --> 0:05:39.719 ago that brought me to Opening Eye because I thought 0:05:39.800 --> 0:05:42.800 that there is no other more important problem that I 0:05:42.800 --> 0:05:48.440 could be working on than artificial general intelligence. And I 0:05:48.560 --> 0:05:53.800 joined Opening Eye to help with leading research teams, and 0:05:54.040 --> 0:05:58.640 from there I went on to build a product team. 0:05:59.040 --> 0:06:03.000 And you know, after having done a few of the 0:06:03.080 --> 0:06:06.599 roles in the company and having built a lot of 0:06:06.640 --> 0:06:10.520 the technical teams, I'm now leading all of the technical 0:06:10.560 --> 0:06:11.600 teams done. 0:06:11.760 --> 0:06:15.080 So as CTO, how do you set the pace of 0:06:15.200 --> 0:06:20.279 open AIS technology development? How do you balance speed versus 0:06:20.839 --> 0:06:23.039 responsibility versus safety? 0:06:23.160 --> 0:06:26.160 Like, how do you where are your priorities? 0:06:26.240 --> 0:06:31.520 Essentially, so, I think today we are dealing with unprecedented 0:06:31.640 --> 0:06:36.680 advancement in technology, and I think the most important thing 0:06:36.720 --> 0:06:42.839 we can do is to manage its advancement and do 0:06:42.960 --> 0:06:48.080 some in a way that's going to benefit people, maximize 0:06:48.120 --> 0:06:52.600 the number of amazing applications that AI can bring, and 0:06:52.760 --> 0:06:56.720 really fuel this energy that people have about interacting with 0:06:56.800 --> 0:07:01.520 AI and making great use of AI, also giving people 0:07:01.720 --> 0:07:05.919 the tools to do so in a reliable and safe way. 0:07:06.320 --> 0:07:13.480 So Atpenyi, our safety teams and research teams collaborate very closely, 0:07:13.720 --> 0:07:18.200 and safety teams are integrated in many of our research domains. 0:07:18.720 --> 0:07:19.800 But we also. 0:07:19.640 --> 0:07:24.760 Provide more room for long term research for safety and 0:07:24.880 --> 0:07:28.000 policy research as well. It's important to work both on 0:07:28.520 --> 0:07:32.440 kind of the near term present issues that we see clearly, 0:07:32.880 --> 0:07:37.040 but also have make a lot of room for exploratory 0:07:37.120 --> 0:07:41.120 and frontiers research when it comes to safety and policy. 0:07:41.400 --> 0:07:45.600 So CHATGBT could revolutionize so many things, and obviously AI 0:07:45.640 --> 0:07:48.120 more broadly, What are the things you're most excited about? 0:07:48.120 --> 0:07:50.040 Like, what's the amazing What. 0:07:49.920 --> 0:07:54.040 I'm most excited about is how it will transform education 0:07:54.200 --> 0:07:58.360 and our ability to learn because you can really see 0:07:58.400 --> 0:08:01.800 that advancing society in a way. You know, there are 0:08:01.840 --> 0:08:04.320 many even the most advanced societies that they are quite 0:08:04.600 --> 0:08:09.360 limited when it comes to education. There is this formula 0:08:09.680 --> 0:08:14.600 on how people are supposed to learn and we all 0:08:14.720 --> 0:08:18.240 learn very differently. We have different interests, and so I 0:08:18.280 --> 0:08:22.520 think by using technologies like tragibiity the underlying models, we 0:08:22.640 --> 0:08:29.840 can really build custom virtual tutors or virtual teachers that 0:08:30.800 --> 0:08:33.040 can help us learn about the things that we are 0:08:33.080 --> 0:08:37.280 really interested in, can really push our creativity, and by 0:08:37.280 --> 0:08:42.960 pushing human knowledge and human creativity, I think we can 0:08:43.200 --> 0:08:47.480 really transform the fabric of society. 0:08:47.720 --> 0:08:48.280 What about the. 0:08:48.200 --> 0:08:50.640 Scary stuff, like what are you most concerned about? 0:08:51.040 --> 0:08:53.920 You know, whenever you have a technology that is so 0:08:54.120 --> 0:08:57.680 powerful and so general, there's always the other side of it, 0:08:57.800 --> 0:09:02.120 and there are always that we have to worry about. 0:09:02.880 --> 0:09:07.079 And we've been very vocal about this since the beginning 0:09:07.160 --> 0:09:11.400 of open AYE and very active in studying the limitations 0:09:12.320 --> 0:09:15.560 that come with the technology. Right now, one of the 0:09:15.559 --> 0:09:19.560 things that I'm most worried about is the ability of 0:09:20.080 --> 0:09:24.760 models like jubid for to make up things. We refer 0:09:24.880 --> 0:09:29.920 to this as hallucinations, so they will convincingly make up things, 0:09:30.200 --> 0:09:34.479 and it requires you know, being aware and being. 0:09:35.960 --> 0:09:37.520 And just really knowing. 0:09:37.400 --> 0:09:41.640 That you cannot fully blindly rely on what the technology 0:09:41.679 --> 0:09:42.960 is providing as an output. 0:09:43.400 --> 0:09:44.760 But on the other hand. 0:09:44.840 --> 0:09:48.840 It also makes it glaringly obvious that this is a 0:09:48.920 --> 0:09:52.760 tool with which you're collaborating. People can misuse it in 0:09:52.840 --> 0:09:57.920 various ways. They can spread misinformation, it can be misused 0:09:58.080 --> 0:10:02.800 in high stakes scenario. So from GBD three point five 0:10:02.880 --> 0:10:07.520 to GPT four we work very hard to reduce hallucinations 0:10:07.720 --> 0:10:08.880 or increase the. 0:10:08.800 --> 0:10:11.120 Factual output of the models. 0:10:12.200 --> 0:10:16.559 And we worked on GBT four for over six months 0:10:16.679 --> 0:10:21.800 just to make it more aligned, safer, more helpful, more accurate, 0:10:21.920 --> 0:10:26.360 more reliable, and held back the release of the model 0:10:26.520 --> 0:10:28.679 so that we could focus on these aspects of it. 0:10:28.720 --> 0:10:31.720 But it's far from perfect, and we're continuing to work 0:10:31.720 --> 0:10:35.680 on it and get the feedback from the daily use 0:10:35.960 --> 0:10:38.880 and make the model better and more reliable. 0:10:39.120 --> 0:10:41.200 I want to talk about this term hallucination because it's 0:10:41.200 --> 0:10:44.240 a very human term. Why use such a human term 0:10:44.440 --> 0:10:47.520 for basically an AI that's just making mistakes. 0:10:47.760 --> 0:10:52.200 A lot of these general capabilities are actually quite human like. 0:10:52.880 --> 0:10:55.640 Sometimes and we don't know the answer to something, we 0:10:55.679 --> 0:10:58.840 will just make up an answer. We will rarely say 0:10:58.880 --> 0:11:03.160 I don't know a lot of human hallucination in a conversation, 0:11:03.880 --> 0:11:06.800 and sometimes we don't do it on purpose. So we're 0:11:06.840 --> 0:11:09.960 constantly borrowing from the way that we learn the way 0:11:10.000 --> 0:11:10.640 we see the. 0:11:10.559 --> 0:11:11.600 World to. 0:11:13.280 --> 0:11:16.000 Have a more intuitive understanding of the systems. 0:11:16.160 --> 0:11:19.000 Should we be worried about AI though that feels more 0:11:19.040 --> 0:11:22.840 and more human like? Should AI have to identify itself 0:11:23.120 --> 0:11:26.160 as artificial when it's interacting with us? 0:11:26.360 --> 0:11:29.440 I think it's a different kind of intelligence. It is 0:11:29.640 --> 0:11:34.160 important to distinguish output that's been provided by a machine 0:11:34.240 --> 0:11:37.680 versus another humans. So you have that understanding, but we 0:11:37.720 --> 0:11:40.720 are moving towards a world we are collaborating with these 0:11:40.800 --> 0:11:43.440 machines more and more, and so output will be hybrid 0:11:43.880 --> 0:11:48.040 from a machine and a human and so they're almost like, 0:11:48.640 --> 0:11:54.240 you know, amplifying tools that are pushing the ability that 0:11:54.280 --> 0:11:58.800 we already have, whether that's reasoning or creativity, and these 0:11:59.000 --> 0:12:03.600 machines are helping us push the bounds of that even further. 0:12:03.720 --> 0:12:09.280 So it's going to be difficult to you know, distinguish 0:12:09.440 --> 0:12:13.400 the output once you have this collaborative engagement between the 0:12:13.520 --> 0:12:15.240 human and the missione the. 0:12:15.240 --> 0:12:16.160 Air of confidence. 0:12:16.200 --> 0:12:20.120 Obviously that CHATGPT sometimes delivers an answer with is it 0:12:20.200 --> 0:12:22.040 can take you off your toes a little bit, right, 0:12:22.160 --> 0:12:25.079 why not just sometimes say I don't know, or program 0:12:25.120 --> 0:12:26.040 that into che GIPT. 0:12:26.440 --> 0:12:30.440 So it turns out that when you're building such a 0:12:30.720 --> 0:12:34.920 general technology. Like with large language models, the goal is 0:12:34.960 --> 0:12:38.400 to predict the next word in a sentence. The goal 0:12:38.440 --> 0:12:42.360 is not to predict the next word reliably or safely. 0:12:43.240 --> 0:12:46.440 Just from this simple goal, we got the ability to 0:12:46.800 --> 0:12:51.480 understand language quite well. We got a lot of creativity, 0:12:51.960 --> 0:12:55.720 ability to even code. And it turns out when you 0:12:55.760 --> 0:13:02.240 have such general capabilities, it's very difficult to handle some 0:13:02.320 --> 0:13:07.640 of the limitations such as what is correct. Also, the 0:13:07.679 --> 0:13:11.280 model doesn't really know much about the user in terms 0:13:11.360 --> 0:13:15.480 of their context and their preferences. But it's still the 0:13:15.559 --> 0:13:19.960 early days, and that's why we're pushing out these systems slowly, 0:13:20.280 --> 0:13:24.160 in a controlled way, but in a way that allows 0:13:24.240 --> 0:13:26.679 us to get some feedback from how people are using them, 0:13:26.720 --> 0:13:30.079 so we can use that, implement it and make them 0:13:30.120 --> 0:13:33.240 better and more reliable. One thing that we did recently 0:13:33.640 --> 0:13:38.920 with child jipt is we rolled out this ability to 0:13:39.040 --> 0:13:42.520 browse the Internet so that it can become a bit 0:13:42.559 --> 0:13:48.280 more reliable on questions that have factual nature. And this 0:13:48.480 --> 0:13:51.680 is now offered as a plugin on child jipd plus service. 0:13:52.120 --> 0:13:54.640 But it's still the early days and this feature is 0:13:54.720 --> 0:13:55.640 only in alpha. 0:13:55.720 --> 0:13:58.160 Some of these texts and some of the data is 0:13:58.520 --> 0:14:01.760 highsed some of it may be correct. Isn't this going 0:14:01.760 --> 0:14:04.600 to accelerate the misinformation problem? I mean, we haven't been 0:14:04.640 --> 0:14:06.360 able to crack it on social media for like a 0:14:06.440 --> 0:14:07.280 couple of decades. 0:14:07.520 --> 0:14:11.000 Misinformation is a really complex heart problem. But you know, 0:14:11.679 --> 0:14:14.120 these systems become smarter. 0:14:14.240 --> 0:14:18.959 It's actually also easier to guide them because you can 0:14:19.000 --> 0:14:24.080 give direction in just natural language and say I don't 0:14:24.120 --> 0:14:29.400 want you to do xting. Then the system, by being 0:14:29.400 --> 0:14:33.400 more intelligent, more capable, has the ability to actually. 0:14:33.080 --> 0:14:34.760 Follow that particular instruction. 0:14:36.200 --> 0:14:40.960 Obviously, with more powerful models, you're also expanding the profile 0:14:41.120 --> 0:14:44.800 of risks, and so you have more risks that you 0:14:44.840 --> 0:14:46.320 need to understand. 0:14:45.840 --> 0:14:46.520 And deal with. 0:14:47.080 --> 0:14:51.800 There are several things that we are exploring. For example, 0:14:51.840 --> 0:14:55.680 one of the things that we've been researching and exploring 0:14:55.800 --> 0:14:59.880 is water marketing. The output where you are able to 0:15:00.040 --> 0:15:05.440 you distinguish what is AI generated output versus human generated output. 0:15:05.640 --> 0:15:06.920 There are ways to deal with it. 0:15:07.040 --> 0:15:10.720 Also from a policy standpoint, I think it's a complex issue. 0:15:10.760 --> 0:15:16.120 There needs to be addressed from research policy perspective. But 0:15:16.280 --> 0:15:19.320 on the other hand, also you know, society needs to 0:15:19.360 --> 0:15:26.120 adapt to these challenges and the capabilities that these models 0:15:26.200 --> 0:15:30.160 are bringing just like we adapted, you know, to using 0:15:30.400 --> 0:15:33.440 calculators and other technologies. 0:15:33.680 --> 0:15:36.280 There's sort of like underlying anxiety. I feel like when 0:15:36.280 --> 0:15:36.760 you talk to. 0:15:36.720 --> 0:15:39.680 Most people about AI, you know that's cool, but it's 0:15:39.720 --> 0:15:40.560 also scary. 0:15:41.120 --> 0:15:42.040 And I've heard. 0:15:42.040 --> 0:15:44.680 AI experts talk about the potential for the good future 0:15:44.800 --> 0:15:45.280 versus the. 0:15:45.200 --> 0:15:48.240 Bad future, and the bad future gets kind of scary. 0:15:48.320 --> 0:15:48.520 You know. 0:15:48.560 --> 0:15:52.120 There's talk about this lead leading to human extinction. Are 0:15:52.120 --> 0:15:53.000 those people wrong? 0:15:53.320 --> 0:15:57.240 You know, there's certainly a risk that when we have 0:15:57.360 --> 0:16:00.720 these AI systems that are able to set to then goals, 0:16:00.760 --> 0:16:04.680 they decide that their goals are not aligned with ours 0:16:05.080 --> 0:16:09.160 and they do not benefit from having us around and 0:16:09.520 --> 0:16:11.080 could lead to human extinction. 0:16:12.080 --> 0:16:13.840 I don't think this risk. 0:16:13.840 --> 0:16:17.120 Has gone up or down from the things that have 0:16:17.200 --> 0:16:19.680 been happening in the past few months. I think it's 0:16:19.680 --> 0:16:24.560 certainly been quite hyped and there is a lot of 0:16:24.600 --> 0:16:28.320 anxiety around it. Well, this risk is important, and we 0:16:28.400 --> 0:16:32.120 need to work on frontier research to figure out how 0:16:32.160 --> 0:16:36.280 to deal with super intelligent AI alignments. We are dealing 0:16:36.360 --> 0:16:39.800 with a lot of risks today that are very real, 0:16:40.280 --> 0:16:47.280 very present, very high probability that they impact us, and 0:16:47.480 --> 0:16:50.160 I think if we cannot figure out how to handle 0:16:50.480 --> 0:16:53.240 and deal with these risks while the stakes are low. 0:16:53.920 --> 0:16:57.240 Then you know, we wouldn't have much hope to deal 0:16:57.280 --> 0:17:01.200 with it when things are more complex. So my view 0:17:01.280 --> 0:17:04.320 is a bit more pragmatic that one, where you know, 0:17:04.400 --> 0:17:06.520 we really need to figure out how to deal with 0:17:06.600 --> 0:17:11.399 the present risks that the systems pose, and coordinate among 0:17:11.680 --> 0:17:18.119 developers and work with regulators, legislators and governments various countries 0:17:18.600 --> 0:17:24.160 to come up with reasonable policies and regulation. 0:17:23.800 --> 0:17:24.919 Around the are elon. 0:17:25.000 --> 0:17:28.439 Steve Wosniak, a bunch of other you know experts have 0:17:28.600 --> 0:17:31.760 called for six months pause on AI development. Do you 0:17:31.800 --> 0:17:34.640 have any intention of slowing down or what's your response 0:17:34.680 --> 0:17:35.480 to that letter? 0:17:35.920 --> 0:17:40.440 So the letter from FLI makes a lot of good 0:17:40.480 --> 0:17:45.360 points about the risks that the technology poses, and we've 0:17:45.359 --> 0:17:46.720 been talking about some of them. 0:17:47.119 --> 0:17:48.360 Opening Eye is being. 0:17:48.240 --> 0:17:52.280 Very vocal about these risks for many, many years, and 0:17:52.560 --> 0:17:55.720 we've been doing active research on them. One of them 0:17:55.880 --> 0:18:00.320 is acceleration. I think that's a significant risk that we 0:18:00.640 --> 0:18:04.200 as a society need to grapple with. The Private companies 0:18:04.280 --> 0:18:10.520 and governments need to work together to figure out the 0:18:10.640 --> 0:18:17.880 risks that acceleration brings building safe AI systems that are 0:18:17.880 --> 0:18:23.159 in general is very complex. It's incredibly hard, and I 0:18:23.200 --> 0:18:28.200 don't think that it can be reduced to a parameter 0:18:28.520 --> 0:18:33.080 set by a letter. The question then becomes, you know 0:18:33.240 --> 0:18:39.720 who is abiding to this letter? Is it all different 0:18:39.840 --> 0:18:43.120 countries in the world? How is that happening? I think 0:18:43.119 --> 0:18:46.680 the issuing reality is far more complex, and it requires 0:18:47.240 --> 0:18:54.280 coordination from private companies, from governments, and figuring out how 0:18:54.320 --> 0:19:00.560 do you deal with these advancements in technology versus blocking advancement. 0:19:01.000 --> 0:19:03.679 There have been parallels drawn to the Manhattan Project, which 0:19:03.880 --> 0:19:05.040 you know, they gathered the. 0:19:04.960 --> 0:19:09.960 Best scientific minds to develop nuclear weapons, and Robert Oppenheimer, 0:19:09.960 --> 0:19:13.680 who led that project, said when he saw the first detonate, 0:19:13.920 --> 0:19:16.520 a line from Hindu scripture ran through his head. Now 0:19:16.560 --> 0:19:20.119 I am become death, the destroyer of worlds. I realize 0:19:20.119 --> 0:19:22.520 the sound dramatic, but if we're talking about the risk 0:19:22.600 --> 0:19:25.639 for human extinction, you know, not being totally out of 0:19:25.640 --> 0:19:28.879 the question. Like in your development of AI, have you 0:19:28.880 --> 0:19:32.800 had a moment like that where you're just like, Wow, this. 0:19:31.440 --> 0:19:32.600 Is this is big. 0:19:32.840 --> 0:19:35.440 I think a lot of us at Open AI joined 0:19:35.560 --> 0:19:38.119 because we thought that this would be the most important 0:19:38.520 --> 0:19:41.240 technology that humanity would ever creates. 0:19:42.240 --> 0:19:43.280 I suddenly think. 0:19:43.119 --> 0:19:48.720 That now with that comes a lot of responsibility. Of course, 0:19:48.920 --> 0:19:51.400 I think AI is going to be amazing. 0:19:51.520 --> 0:19:52.280 It already is. 0:19:52.520 --> 0:19:59.760 It has this incredible potential to extend our creativity and 0:20:00.080 --> 0:20:04.400 human knowledge and make our lives better in so many vectors. 0:20:05.640 --> 0:20:09.000 But of course the risks, on the other hand, are 0:20:09.040 --> 0:20:13.560 also pretty significant, and this is why we're here. 0:20:13.800 --> 0:20:16.600 I just rewatched the movie her which has this very 0:20:16.680 --> 0:20:19.639 vivid depiction of life with AI in ten years. 0:20:19.800 --> 0:20:21.960 How will our lives be different? How will daily life 0:20:22.119 --> 0:20:22.600 be different? 0:20:22.800 --> 0:20:29.080 I haven't watched a movie herd a lot aboutime. I 0:20:29.119 --> 0:20:31.520 hope ten years is a long time, but I hope 0:20:31.560 --> 0:20:34.679 that in the next few years we will have a 0:20:35.040 --> 0:20:41.400 future where we use AI systems as tools to amplify 0:20:41.960 --> 0:20:46.840 a lot of our own abilities. And I hope that 0:20:47.119 --> 0:20:52.840 we have systems that help bring customized education to as 0:20:52.880 --> 0:20:56.040 many people out there as possible, and I hope that 0:20:56.400 --> 0:21:01.280 they You know, we can build tools, diagnostics, tools or 0:21:01.440 --> 0:21:07.240 ways to understand diseases and the problems in healthcare much 0:21:07.320 --> 0:21:09.360 much earlier and figure out how to deal with them 0:21:09.400 --> 0:21:13.760 at scale. And you know, we're dealing with massive problems 0:21:13.800 --> 0:21:18.480 in climate change, figuring out new solutions figuring out ways 0:21:18.520 --> 0:21:22.639 in which we can help reduce the risks that climate 0:21:22.680 --> 0:21:23.360 change processes. 0:21:23.440 --> 0:21:27.439 Could you put what you're developing here inside robots and 0:21:27.520 --> 0:21:29.440 could they combat loneliness? 0:21:29.520 --> 0:21:33.159 I think bringing these systems into the physical world is 0:21:33.760 --> 0:21:35.600 a pretty significant step. 0:21:36.280 --> 0:21:38.840 Feels like we're a bit far from that. 0:21:39.720 --> 0:21:43.159 But also, you know, just having a chatbot that you 0:21:43.280 --> 0:21:48.280 can ask for advice suddenly, not in high stake scenarios 0:21:48.400 --> 0:21:51.639 right now, seems like that would be helpful for a 0:21:51.680 --> 0:21:52.320 lot of people. 0:21:52.640 --> 0:21:57.960 That's quite profound that we could someday have relationships with computers. 0:21:57.520 --> 0:21:59.560 In a way we already do. 0:21:59.720 --> 0:22:03.120 Right We're spending so much time on our computers, We're 0:22:03.119 --> 0:22:06.800 always on our phones. We're almost like enslaved to this 0:22:06.920 --> 0:22:10.159 interaction that we have with the keyboards and with the 0:22:10.240 --> 0:22:10.919 touch screen. 0:22:11.320 --> 0:22:14.080 I think a lot about my kids and them having 0:22:14.080 --> 0:22:17.480 relationships with AI someday and this thing that has much 0:22:17.520 --> 0:22:19.679 more time to spend with them than I do. How 0:22:19.720 --> 0:22:22.040 do you think about what the limits should be and 0:22:22.080 --> 0:22:24.919 what the possibility should be when you're thinking about a child. 0:22:25.320 --> 0:22:28.720 I think we should be very careful in general with 0:22:29.400 --> 0:22:35.840 putting very powerful systems in front of more vulnerable populations, 0:22:36.000 --> 0:22:39.840 people under thirteen cannot access it, and even under eighteen 0:22:40.080 --> 0:22:44.760 requires parents of supervision. So there are certainly checks and 0:22:44.800 --> 0:22:49.159 balances in place because it's still early and we still 0:22:49.200 --> 0:22:52.600 don't understand all the ways in which this could affect people. 0:22:52.920 --> 0:22:56.480 There's also some business interests here, and by releasing chatgbt 0:22:56.960 --> 0:23:00.199 open ays kind of turbo charge this competitive frenzy. Do 0:23:00.240 --> 0:23:01.920 you think you can bet Google at its own game? 0:23:02.119 --> 0:23:04.800 Do you think you can take significant market share and search? 0:23:05.200 --> 0:23:10.000 You know, we didn't set out to dominate search when 0:23:10.359 --> 0:23:14.280 we build child jepet. In fact, it actually started as 0:23:15.160 --> 0:23:19.760 a project around understanding and dealing with truthfulness of large 0:23:19.840 --> 0:23:23.440 language models, and then it evolved. But I think what 0:23:23.600 --> 0:23:29.200 child gipt offers is a different way to understand information 0:23:29.960 --> 0:23:34.160 and a different way to interact with the same tool. 0:23:34.520 --> 0:23:37.399 And you could be, you know, searching, but you're searching 0:23:37.560 --> 0:23:41.720 in a much more intuitive way versus keyword based. That 0:23:41.880 --> 0:23:46.159 is definitely an outcome that we saw afterwards, and we 0:23:47.040 --> 0:23:51.200 built an interface that would allow people to interact with 0:23:51.240 --> 0:23:54.520 it much more smoothly, and as we can see, it 0:23:54.600 --> 0:23:59.400 is pushing other people to build more assistant like products 0:24:00.160 --> 0:24:03.679 companies and small companies. I think the whole world is 0:24:03.720 --> 0:24:05.360 sort of now moving in this direction. 0:24:05.960 --> 0:24:07.320 I think our focus will. 0:24:07.119 --> 0:24:13.000 Remain on building these general technologies and figuring out how 0:24:13.240 --> 0:24:15.439 we can bring them to the public in ways that 0:24:15.520 --> 0:24:16.080 are useful. 0:24:16.560 --> 0:24:19.159