{"url": "https://www.youtube.com/watch?v=-p-ad4AOsFw", "page_url": "https://www.youtube.com/watch?v=-p-ad4AOsFw", "title": "Interview of Sam Altman and Daniela Rus by Teddy Lee for the 2023 MIT AI Conference", "show_or_channel": "Teddy Lee", "date": null, "duration_s": 1118, "type": "interview", "speakers": ["Sam Altman", "Daniela Rus", "Teddy Lee"], "confidence": "medium", "notes": "2023 MIT AI conference interview.", "people": ["sam-altman"], "id": "0000_teddy-lee-interview-of-sam-altman-and-daniela-rus", "audio_duration_s": 1117.6345, "worker": "jc-w15", "transcribe_s": 15.0, "model": "large-v3", "segments": [{"start": 0.0, "end": 30.53, "text": "Okay, good afternoon, Danielle and Sam. Thank you for being here today. And hello, attendees of the 2023 MIT AI Conference. My name is Teddy Lee, and I'm the president of the MIT Club of Northern California and a product manager at OpenAI. I'm joined today by Danielle Roos, the director of MIT's Computer Science and Artificial Intelligence Laboratory, and Sam Altman, the CEO of OpenAI. Danielle and Sam, tuning in for this conversation, are members of the global MIT community, including..."}, {"start": 30.53, "end": 58.69, "text": "Some of the over 14,000 members of the MIT Club of Northern California, which is about 10% of the global MIT alumni population. They're very excited to hear from you today. Now for our audience, a few more words about Danielle and Sam before we dive into the conversation with them. Danielle Roos is the director of MIT's Computer Science and Artificial Intelligence Laboratory and the Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science. Her pioneering research focuses on advancing robotics, AI,"}, {"start": 58.69, "end": 87.94, "text": "And machine learning to enable greater autonomy and collaboration between humans and machines. Daniela has led key projects on perception, learning, and planning for robotics. She serves as director of the MIT Toyota Joint Research Center on AI for Intelligent Vehicles and is a member of the Global Partnership on AI. Among her honors, Daniela is a member of the National Academy of Engineering, a MacArthur Fellow, and a recipient of the Engelberger Robotics Award. She earned her PhD in computer science from Cornell University. Sam Altman."}, {"start": 87.94, "end": 117.18, "text": "is the co-founder and CEO of OpenAI, the AI research and deployment company behind ChatGPT and DALI. Sam was president of Y Combinator from 2014 to 2019. In 2015, Sam co-founded OpenAI as a nonprofit research lab with a mission to build general purpose artificial intelligence that benefits all of humanity. OpenAI remains governed by the nonprofit and its original charter today. With that, let's jump into questions. We'll alternate between Danielle and Sam. Our first question."}, {"start": 117.18, "end": 137.66, "text": "Both of you are advocates for the transformative potential of advanced AI systems. Could you elaborate on what some of the most exciting societal impacts might be? Danielle, we'll start with you. So I am so excited about this question, and I think we can spend the rest of the time talking about it. At the high level, AI can do so much for humanity."}, {"start": 137.66, "end": 161.26, "text": "With AI we can better diagnose, monitor and treat disease. We can synthesize medicine that is individualized to people. We can use AI to ensure that there are no road fatalities and that you will get to your destination much faster in a more optimal way. We can create"}, {"start": 161.26, "end": 188.19, "text": "Educational programs that are personalized to the learner. We can ensure that we keep your information private and safe. We can transport people and goods much more effectively. Basically, we can yield routine tasks to machines so that people can focus on the higher level, more rewarding tasks. And if I can dive into..."}, {"start": 188.19, "end": 217.94, "text": "One particular area that I'm especially excited about that is not so much talked about, I would like to say that with AI, we can really enhance accessibility. We can empower people who have disabilities to use AI-driven tools for so many different capabilities that they do not have at the moment. And in fact, I'm talking about an actual project that we have done together with Andrea Bocelli."}, {"start": 217.94, "end": 234.34, "text": "Where we have used AI and we have used essentially a wearable version of a self-driving vehicle to empower Andrea, who is blind, to see and experience the world in ways that he simply cannot do with a white walking stick."}, {"start": 234.34, "end": 264.98, "text": "And the last thing I want to say is climate. We can use AI to have a better understanding of our planet and to think ahead, to design the kind of mechanisms and solutions that will help us reverse global warming. How about you, Sam? Obviously, there's so much to say, but I'll focus on just two things. One, the amplification of individual human will and desire and ability, I think is going to be tremendous."}, {"start": 264.98, "end": 293.49, "text": "I'm a huge believer that if you give people great tools, they'll do things that astonish you. And with the tools of AI, even with the limited capabilities of today, people can be multiple times more productive and in new ways they just couldn't before. And so what humans will be able to create using these tools I think is going to be quite remarkable. And then the second area is the advancement of science. So AI that can help us discover new knowledge I think will be a great thing. While we're discussing the benefits, it's crucial to address the other side of the coin."}, {"start": 293.49, "end": 320.5, "text": "What do you consider to be the most concerning risks associated with the development of highly intelligent automated machines? Sam, we'll start with you this time. I think there'll be a long, you know, there'll be a spectrum over time. But humans using these systems to do bad things, designed bioweapons, hacking the computer systems, these are all quite scary. And then eventually, as the systems get really powerful, some of the more sci-fi risks come into play. Daniela?"}, {"start": 321.71, "end": 340.22, "text": "So I like to think about Sam's answer to the first question, all the empowerment that comes from AI where we can get productivity, we can get knowledge, we can get insight, we can get creativity, we can get foresight, mastery, empathy, we can get all these things with AI."}, {"start": 340.22, "end": 365.7, "text": "All these great things that we can put forward for the greater good, they can also empower supervillains. And so worrying about how to ensure that the tools we develop are used for good is important. Putting guardrails, putting barriers in front of the supervillains is really important."}, {"start": 365.7, "end": 394.18, "text": "Bad things that supervillains can do. I would like to focus on misinformation and disinformation. This is a very important topic because the more powerful AI and generative AI gets, the more accurate the deepfakes and the other types of misinformation and disinformation we spread is. And so we really need to think about methods that allow us to use the tools for good while stopping"}, {"start": 394.18, "end": 405.17, "text": "Things like misinformation and disinformation. And here we have technical solutions. We can take all these problems and we can approach them in a way that..."}, {"start": 405.17, "end": 434.3, "text": "That a proud MIT engineer is taught to do, is taught to solve problems. So for instance, for misinformation and disinformation, we can begin to develop watermarking technologies. We can begin to develop technologies that ensure the trust of the information sources. Thank you. Open source and closed source has been a point of considerable debate in the evolving AI landscape."}, {"start": 434.3, "end": 459.79, "text": "At OpenAI, we've made efforts to contribute to the open source community, but we also keep some of our technologies proprietary. And Danielle, I know MIT CSAIL has been a proponent of open source research and development. How do you both reconcile the trade-offs between open and closed source? Danielle, we'll start with you. So I'm a big proponent of open source as a way of really understanding technology that is not yet understood."}, {"start": 459.79, "end": 475.1, "text": "And so, for example, we are deploying very large language models and they are magical in their operations. They also hallucinate and make mistakes and we don't really understand the mechanisms that get us there."}, {"start": 475.1, "end": 502.51, "text": "If we were to open source the models, for instance, like in the case of LAMA, then the scientific community would have access and the scientific community could make inroads, could make progress towards better understanding what is happening inside the black box. And so I think that for where we are with our understanding of AI, especially of the huge models,"}, {"start": 502.51, "end": 526.61, "text": "Today, the community collaboration towards transparency is super important. Sam? Yeah, I think open source is very important. There's obviously a balance. Not everything should be open source and everything should be closed source. But the role of open source in scientific progress and understanding these systems, I think, is very important. And we expect to continue to do more there."}, {"start": 527.12, "end": 534.99, "text": "Can I say something more about open source? Because I will say that while with open source,"}, {"start": 534.99, "end": 563.01, "text": "How do you both envision the evolution of autonomous agents?"}, {"start": 563.01, "end": 591.6, "text": "And what are the key challenges and opportunities you see in deploying agents ethically and effectively? Sam? Look, I think this is one of the things that people most want from AI. How to do it, how to make it safe, that remains to be seen. But this idea that you have an autonomous agent, that companies have autonomous agents, that my agent can talk to this company's agent, or you can talk to my agent, I can ask my agent to go do things, that is another step."}, {"start": 591.98, "end": 618.77, "text": "Towards what people really want out of computers. I think we made a significant leap forward in terms of computer interfaces with the idea of language interfaces. But language interfaces that can control agents is certainly another logical step that people really want. The answer will be like start slow and set the training wheels very low. And then as we understand the risk surface and where it's okay and where it's not, we can make..."}, {"start": 618.77, "end": 642.26, "text": "Thoughtful, risk-adjusted decisions about how to move those training wheels up. But the upside here is going to be tremendous for people. If you think about having an agent that you can give a task to, to help you with whatever you're trying to do and be confident it'll go do it, that's awesome. Daniela? I really like the idea of agents, whether embodied or not."}, {"start": 642.26, "end": 671.79, "text": "That can help us with cognitive and physical work. The embodied agents would help us with cognitive work, sorry, with physical work and the non embodied agents would be like our super assistants. And so the question is, how do we get to the point where these agents can actually deliver on the tasks that we have to do? Many of these tasks happen in the physical world. And part of the"}, {"start": 671.79, "end": 697.44, "text": "Capability that we don't quite have yet has to do with interaction, has to do with interaction with people and with the physical world. And so in the future, I really imagine progress in connecting the tools that we have right now with more extensive capabilities that will give the agents"}, {"start": 697.44, "end": 728.05, "text": "Shifting gears, what implications do you see for the job market, both positive and negative? Daniela?"}, {"start": 729.1, "end": 750.03, "text": "Well, I'm really excited about how AI can help people and companies. And when I think about what AI can do in a job setting, I think of the agents really as helping as assistants or as teammates."}, {"start": 750.03, "end": 778.4, "text": "Or as autonomous agents. So what do I mean by this? An assistant is kind of like an AI agent that can go off in the digital world and explore, find patterns, bring you data that is processed and synthesized in ways that are too complicated and too expensive for people to do. And so with this data, the person is empowered."}, {"start": 778.4, "end": 808.29, "text": "to make better decisions. So I like to think about an AI agent as an assistant running around doing errands for me and bringing me data to act on. That's assist. Augment means the AI system could complement me. And I will give you an example. There is a study at Harvard where doctors and machine learning were given scans of lymph node cells and they were asked to"}, {"start": 808.29, "end": 836.78, "text": "Label them as cancer or not cancer. The doctors made 3.5% error as compared to the AI system where the error was 7.5%. But when they worked together, the doctors and the people achieved 80% improvement and made only 0.5% error. And that's because humans see the world in some ways, the AI systems see the world in different ways when humans and"}, {"start": 836.78, "end": 854.61, "text": "Machine learning comes together. They take the best of both worlds. And then there is the autonomous mode, where the idea is that people could yield some low-level tasks to agents. So that's on the positive side. On the negative side,"}, {"start": 854.61, "end": 879.34, "text": "It is true that some of the jobs that are currently done by people will be done faster and more efficiently by machine learning and by AI agents. And so we really need to prepare for this kind of future by empowering people, by teaching people the new tools and by elevating the work of people."}, {"start": 879.86, "end": 908.59, "text": "The fact is that in every job we do different types of tasks and not all tasks are automatable. It's really the low level routine tasks that are currently automatable. And so with AI tools, I think we can really empower people to focus on the more interesting aspects of their jobs that cannot be done by machines yet. Sam? I think it's going to be a lot of great things and I'll talk about those before I do."}, {"start": 909.04, "end": 937.38, "text": "I think it is very important to acknowledge that some jobs are just going to go away and other jobs are going to be worse than they are today. And although that is the way of technology and although it's a net good for society, it doesn't mean it's not an incredibly painful thing for the people impacted by that. Obviously it is. And so even when we talk about all of the wonderful new jobs and prosperity that are going to be created and the jobs of today that will get much better and higher impact, there is going to be real pain here."}, {"start": 937.38, "end": 968.13, "text": "Like there is with every technological revolution. And it is easy to be too dismissive of that. We'll do all sorts of things, you know, we'll have, who knows what, maybe we have things like UBI and retraining programs and all these wonderful things, but it's going to be a disruption to people's lives. And I think that deserves tremendous empathy that our industry has not done a great job of always holding. On the positive side, like with any other technological revolution, maybe more than any before,"}, {"start": 968.13, "end": 997.26, "text": "The amount of work that one person will be able to do, the amount of leverage on ideas and creative spirit that will come from this, the ability to go create amazing new things, to do existing jobs way better, and to think of entire new categories of jobs that we can scarcely imagine today. Would've been hard to think of an AI programmer as a job before the computer revolution. That's gonna be amazing. We've got a ton of work to do to make sure that"}, {"start": 997.26, "end": 1023.76, "text": "These tools are widely accessible to people and that we come up with like a new socioeconomic contract that still makes sense for the AI age. But the upside here is tremendous. I think to this because I really agree with Sam that it's important to acknowledge the displacement and the fact that some people will indeed lose their jobs. But I do want to highlight a recent study from MIT."}, {"start": 1023.76, "end": 1051.74, "text": "Where one of our economists has looked at the impact of automation and has asked the question, does it make economic sense? And what he has found is that for some automation jobs, like for instance, for visual inspection, which you could now offload to machines, the cost of bringing in the compute, the expertise, the software, the data."}, {"start": 1051.74, "end": 1079.92, "text": "This cost is actually quite high and right now it does not make economic sense for many of these types of inspection jobs. Perhaps in the future it will, but my point is that we have some time to adjust to the inequality that the AI solutions are bringing into the workplace and we could take on"}, {"start": 1079.92, "end": 1109.49, "text": "Active programs to make sure that the people whose jobs are in danger can get equipped with new tools so that they can move on and focus on other types of roles. Thanks for sharing that, Daniela. So that brings us to the end of our time. Daniela, Sam, thank you so much for your time and for sharing your insights. And attendees of the MIT AI conference, I hope you enjoy the rest of your experience. Thank you. Thanks for having us."}, {"start": 1109.49, "end": 1112.24, "text": "Thank you. Bye bye."}]}