[0:00:05] Hello, hello. All right. Okay, this is better. Cool, cool, cool. All right, so, yeah. Okay, so, I mean, it's incredible, right? The excitement that we can see here. I mean, not only in this hall, but also the overflow halls are full.
[0:00:57] I mean, as we know that there are so many people across India sort of virtually joining us on Twitter and Digital India platform. Thanks to them as well. So the man of the moment is here, Sam Altman, the chat GPT guy. Let's welcome him with a round of applause.
[0:01:35] I mean, we are already running behind time, so let me begin with requesting the chairman, Board of Governors, IIIT Delhi, Mr. Kiran Karnik, to say a couple of words to welcome Sam. Good afternoon, and it's a pleasure and privilege to welcome Mr. Harkman and his team from Hawaii to India, and particularly to the Indraprastha Institute of Information Technology, known, Sam, as IIIT Delhi.
[0:02:18] A little different from our cousin or uncle down the road who has only two eyes. We pride ourselves, like you, with being more agile, more innovative, quicker. And that's the speciality of IIITD. I won't waste time trying to introduce Sam to you. All of you know him. All of you know him well enough. But let me take a moment, Sam, if I may, to introduce this audience to you.
[0:02:50] This is mostly young people from the IIIPT. Students from here who will be future pathbreakers and pathmakers and for whom we hope you will be an inspiration and they will take inspiration from you in taking further the work that you have done over these years. Also interspersed among them are slightly older people, the faculty who we are very proud of. They are a very distinguished faculty doing great work.
[0:03:19] Some of them are at ages not too distinguishable from students, but some of them are a little older. And we are very, very happy to have them with us today. We are looking forward to a great interaction and discussion. I won't take more time. I know all of us want to hear Sam rather than hear me. But let me take a moment more to request our director, Professor Ranjan Bose, to say a few words of welcome. Over to you, Ranjan. The excitement is really great in this room. We were so excited to have...
[0:03:56] Sam with us. Thank you, Sam, for joining. And we are really privileged and excited to have you here. We'll not take more time. Sam is in the business of inventing the future. AI is going to affect the way we do lead our lives the next five years, ten years. So let's get on with the event. And we are looking forward to a very exciting fireside with Sam. Over to you, Sam. All right. So Sandini and...
[0:04:29] Aati will be accompanying Sam here this afternoon. Let's welcome them as well. All right, so now we are good to begin with our really tough Q&A session with Sam. And the format is going to be that you are not supposed to be using JAT GPT to ask what to ask Sam, right? Perfect. It's okay. So with that...
[0:05:07] We are going to take questions from audience here, as well as questions from audience who are joining us virtually, either from the overflow rooms or from somewhere else watching live stream. So if you have to ask any question to Sam, right, so please use this Twitter hashtag, OpenAI at IIIT Delhi, right? So this is the Twitter hashtag for this.
[0:05:35] Afternoon. OpenAI at IIIT Delhi. Hashtag OpenAI at IIIT Delhi. Okay? So we are going to read out questions from Twitter, as well as we'll take questions from the audience in this hall. Okay. So I ask Pravesh, my colleague, to begin with the first question. Sure. Hi, Sam. So you met Prime Minister Modi this morning. How was your meeting? It was great. It was really fun. He was so enthusiastic, really thoughtful about AI.
[0:06:04] The benefits of it. We asked why India has embraced Chachi PT so much and so early. It's really been fun for us to watch. He had great answers about that. We talked about the opportunities in front for the country, what the country should do. Also the need to think about global regulation to make sure we prevent some of the downsides from happening, but it was a great hour. Awesome, awesome. So let me start with the first, as you said, the first difficult question. It's not that difficult a question actually.
[0:06:33] We have talked about a lot of misuse of AI, and I don't think we have to talk about that here because there are more opportunities than misuse probably. But if you are in India, and the first question that when people use ChatGPT, the first question they ask, anybody, whether it's a call center person or anybody, would say that, hey, you know what? This is going to take my job away. So while that may be true or not true, that's a different thing, but how do you see it playing in a country like India, which is so, so much into...
[0:07:03] Services sector. So I think- So I think with this, what we're seeing with things like ChartGPT in our models, jobs will change. How people do their work day-to-day might change and it will evolve. Something we've seen happen is people might go from like doers of small tasks to actual reviewers, where perhaps the smallest task of something people might use ChartGPT for it. So that's coming. Things will change. I think the way to approach it is probably to take care of it to kind of
[0:07:35] Keep up with most of the tools, see what's happening and learn to adapt with it. But it's definitely coming. So, it's something to keep on the horizon and in the mind. That's right. So, I think if you can talk about the positives in the sense, how would the country which is more or less 30 and under 30, and a lot of huge developer ecosystem here. I mean, most of the folks here are developers in fact.
[0:08:05] Talk about their opportunities from the AI point of view, not just OpenAI, but in general AI point of view. Yeah, I think this is going to be the most transformative technology humanity has yet invented. And the countries that embrace it, the people that embrace it early, are going to be particularly rewarded. It really is amazing to watch what's happening in India with the embrace of AI, not just OpenAI, other technologies too. And I think we're going to look back 30 years, 50 years, whatever.
[0:08:34] And say, wow, we live like barbarians. You know, life is so much better. When intelligence becomes close to free, and not only that, but when it becomes qualitatively different at what it's capable of, you know, we view our role as we build the tools, we give to all of you, you all amaze us with your creativity and what you can create for the world. So I think that's really going to happen. I think India is already, like, one of the centers of this, and I'm sure that will continue.
[0:09:08] Okay, so I think I'm audible. All right, so my first question or a follow-up question to what Pravesh asked is on the optimistic side, right? So Pravesh talked about job losses. I'm going to ask when are we going to see OpenAI India? Like an office in India? In an office as well as maybe a research center or something like that? I think the first thing we'll do is fund startups in India.
[0:09:39] Before this job, I used to work at Y Combinator. We were always amazed and quite grateful to the quality of Indian startups. And I think we are now heading towards what will be the greatest startup boom, at least since the internet. It's going to be bigger than mobile. And the energy we've seen from startups here, we're very excited to figure out how to start funding. So we'd be ready to do that now.
[0:10:03] OK, perfect. So being a professor, I always have a concern about placement of my students, right? And when it's placement season here, believe me, they prepare throughout the night for coding interviews, right? So when should the students of IIIT Delhi be expecting to prepare for OpenAI interview? That would be a corollary.
[0:10:40] I think if you guys amaze us with using the API, build some amazing products, contribute to open source repos, not just OpenAI, all of the AI ecosystem, and then just email sam at openai.com, you'll get a job. All right. Cool, guys. So please note it. All right. So before we move to maybe Twitter, we can take a couple of questions from audience now. So who first? Okay. Hello. Hello.
[0:11:13] Hi, I'm Janvi, and my question to Sam is, so do you feel that the impact of GPT-4 is being limited by access to hardware? So for example, there are very less homes in India which use robot cleaners, etc. So in that sense, AI does not clean my home. So what are OpenAI's efforts on making hardware that GPT runs on versatile and accessible? You mean hardware like robots? Yeah, robots are hardware on which GPT can run on.
[0:11:44] Like computer chips or robots? Yeah, both. So like what are OpenAI's current innovations which are in this direction? We started with robots. We'd like to get back to it someday. Robots were hard but not hard because they were helping us figure out how to do good ML research. They were hard because they break and the simulators are imprecise and all of that. So that will come back. We've been watching what's developing with robots with great interest. I don't know exactly when.
[0:12:12] In terms of hardware, yeah, we definitely, in terms of computing hardware, like chips that can run these systems, we are super limited, like the whole field is, and we want to figure out a way to do more of that. Hi, Sam. My question is, like, GPT-3 is 176 billion parameters, and GPT-4, we are lessening one trillion or so, or you yesterday talked about GPT-10. I don't know how many trillions. So it's, don't you think it's stressing the hardware capacity?
[0:12:51] Computing capacity, so how far will it go? Or we should really or OpenAI should really look into the optimizations of this so that parameter remains doesn't go so high and computing of the hardware is limited and you can solve more problems. Well, I think we should do all of those things. I think we should push the hardware as much as we can. I think we should make the most efficient algorithms we can. I think we should develop new techniques to go beyond what we're currently doing. What people care about is not the number of parameters.
[0:13:20] For the number of exa-flop months that go into a model, what they care about is what it can do that's useful to them. And that's the metric that matters. So we will push on that, and we will optimize that as much as we can. And that'll probably involve all of those elements all coming together to stay at the forefront. Thank you. I have a question. Okay, so let me put a corollary to that question. So well, I mean, the context is that I am a professor of distributed machine learning here, right?
[0:13:50] And when I teach, I also have to sort of refer to energy efficiency, right? So when you are saying that we are going to push the limit of size of the model as well as the hardware, I mean, how do you see the trade-off between energy efficiency and the model size? We use vision. Right, so. I have a question here, hi. Hi, I'm Akanksha here, hi. I have a question from the policy side of it. You've been doing great work.
[0:14:20] Everyone at OpenAI has been doing great work. Thank you so much. We've all used the models. But I have a question. When it comes to open source AI projects, how do you think, I mean, I do know that you've yourself said multiple times that the door has to be let open for small projects so that they can grow up. But there has to be regulation in some sense because the growth has been exponential as we're seeing. Do you see how can policymakers and governments
[0:14:48] How do you sort of drain in regulation at least when it comes to open source projects? So I'd say open source is a very new and evolving field for sure. It's one of the unknowns in some ways. OpenAI strategy here, what we really try to do by being in the lead is norm set. So every time we release a big model, we focus a lot around how are we releasing it, how are we educating the world about it, what policies are we releasing it with.
[0:15:15] So that's been one piece of our strategy. The second piece is exactly what Sam has been doing right now, going on tour, talking to a lot of regulators. So hoping that this piece of education will also inform how regulators think about regulating models. All of that being said, open source makes it much harder to regulate in some ways. So one of the open challenges of the field, and we'll see how it evolves. Sam, what's your take?
[0:15:43] I agree with all that. I think open source is important. As it gets more powerful, we will need some guardrails on it. But I worry about over-regulating and stopping what is an incredible, incredible Cambrian explosion of talent and creativity. So there are many questions on Twitter now on the same question as the lady started on regulations part. So one of the questions that is asked just now by multiple people is that regulations don't catch up with disruptors soon enough.
[0:16:11] And we have talked about, you guys have actually been very vocal about the misuse part. Also, you know, in many ways, elections being, you know, I saw the Warsaw thing. So would you be willing to pause the development for regulations to catch up? Or what is the right way? I don't think a pause, a pause sounds good, but I don't think it quite works in practice. Because, you know, what do you do at the end? And what about people who don't pause? And also, a lot of our insights about how to make these systems safer.
[0:16:41] It comes from working with the latest models. It doesn't work to sort of sit in the ivory tower and think as well as it seems like. What I think instead of a pause makes sense are norms like external audits and red teaming and tests, dangerous capability evals that models have to pass. So I think that'll work better than a calendar pause. It'll also help us move towards a regulatory framework that works. Anything to add? And a follow-up question.
[0:17:09] Not on regulations but on somebody wrote about danger to GPT. So, they're asking this question that there's been a lot of research and innovation around open source models and which are intended to be fine-tuned with custom data. Do you think that the custom data fine-tuning would make things better than GPT itself? Is it a threat? I think the AI model capabilities come in two places. One is the pre-training phase.
[0:17:42] Which is where you collect large amounts of data and that's where the core knowledge of the model comes from. And then fine-tuning comes in the post-training phase where we sort of align the model, make sure it follows the instructions that you give it, or is particularly good at a specific capability like creative writing or coding, etc. So, I think the two actually go hand in hand. It's very important to have large foundation models with large base training and number of parameters.
[0:18:08] As well as fine-tuning to actually make the model useful for specific applications. So, I think the opportunity actually for the open source ecosystem and for the world at large and developers is to build on the foundation models but fine-tune them for specific use cases so that the benefits of AI are spread to as many applications and industries as possible. So, now we can go one more round. Hi, Sam. I'm Bhakesh. I'm a fourth-year undergrad at IIIT Delhi.
[0:18:37] And I had a question from perspective of a student. Many students like me who are studying at IIIT Delhi, participating in deep learning and computer vision research. We often face a very big challenge when we go and apply to positions that we really want to fill in the industry, especially in the research domain. Many big companies like Microsoft, Google, are not welcoming to undergrad students. Do you think that this is something which can be tackled by companies like yours, by OpenAI and other companies?
[0:19:06] Initiative should be taken in this direction to promote people who are from undergrad background or like yourself who dropped out of college and want to pursue what they are really passionate about. Two things. One, we definitely hire undergrads. In fact, many of our best researchers came out of undergrad or dropped out. And I think so many of us share this belief that you certainly don't need a PhD to do great work. But OpenAI is like an unusually welcoming place. So I would say we're like a good fit. The other is like,
[0:19:36] You can always just start your own thing. I did drop out, and I wouldn't have got a job, but I was able to start something. So I think until the world catches up with the fact that you don't need an advanced degree, that's always an option. But in the meantime, yeah, give us a call. So there's more question from Twitter. Again, so sorry, guys. What I'm trying to do is I'm trying to cluster the questions here and ask questions from as many people as possible. But we can go back there.
[0:20:02] There's a very nice question from Ishan Goswami. He says that he saw an interview where you said that while building OpenAI, and this is a startup advice he's asking for, you went against a lot of startup advice you yourself give. That's what he claims you said. That's for sure. So he wants to know your framework to which advice to take and which not to. Honestly, I feel so bad about the advice I gave while I was running YC. I've been thinking about deleting my entire blog. I might do it.
[0:20:32] But there were a lot of things that we really held dear as YC advice. You've got to launch right away. You've got to launch a first version you're embarrassed about. You raise very little capital up front. You don't take big R&D risk. You've got to immediately find product market fit with something. OpenAI raised like a billion dollars of capital before we had any product at all. It took us four and a half years after we started to release something.
[0:21:00] When we released it, we really didn't talk to users at all for a while, and it took us a while to figure that out. We just didn't do it the same way, and it still worked. And I think it would be bad to make a habit of that, but there's existential proof that at least my advice to not do that is not always accurate. And I think the only certain thing I can say about startups, which is like the thing that we put on t-shirts, YCT shirts, is make something people want.
[0:21:30] We took an unusual path to get there. We eventually did, but it wasn't how we normally advise other companies to do it. And I think that's kind of like the one rule that really matters. And as long as you figure out a way to do that, even if you take an unconventional path, it's okay. But yeah, maybe I will delete the blog, because now I feel bad. No, no, because I'm here and I have this opportunity, I'll tell Sam that.
[0:21:59] Make something people want is what we have been telling everybody here. So that's a great idea. Thanks so much. So I am Abhinav, and I represent the development sector. And one of the things that I see when I look at OpenAI is the tremendous potential that is there to impact things like education, healthcare, and so on and so forth. But towards the bottom of the pyramid. So we in India, for example, around 300 million children which are there in approximately a million and a half schools.
[0:22:30] Education is one of the areas that we are most excited about impacting. I'll let Sandhya talk about it because she's looked into it.
[0:23:08] I think more than I have. We had a good chat about it with the Prime Minister earlier today. We'd love to work with India on this. I think the one thing that I'd add there is people here who are working with students who are seeing what the need is are closer to the problem than we are at the end of the day. So this is an area where actually we have to follow your lead. So we have the API precisely for this reason to enable innovation in spaces like this and let others innovate and build these products.
[0:23:34] That's kind of what we're excited about, seeing innovation happen in these challenging spaces through our API and doing what we can to empower that. Yeah, I think the resource question is an interesting and important one. I think that's a challenge, but it's like, how do we get the right resources? At OpenAI, I think we do think a bit about these problems as well, but we're not as close to it as it needs to be, and we do try and remain focused as a company.
[0:24:23] But it's again, it's a potential, it's an area for possible growth. Okay. So now again, a representative question from a cluster of questions on Twitter. So the question is that out of all these technologies out there, language, hardware, so forth, how did you exactly decide to pick up RLHF or generative models?
[0:24:54] We have an approach which is whatever it takes to accomplish the mission. So our mission is to build and figure out how to deploy AGI that is safe and has maximum benefit. Within that, we just do whatever works. Sometimes we do things that don't work for long periods of time, but we try to stop that. And so what we found, we didn't know this was going to be our path. In fact, per the earlier question, we started with robotics and then systems that play video games. And what we found is that...
[0:25:24] Language models were doing something quite powerful. At the time, they were totally neglected, even when we put out GPT-1 and really even GPT-2 also. No one really cared, but we saw in there a path to AGI. After we developed that, we found a bunch of other problems, and so we invented RLHF. We also came up with scaling laws, and I think those were two massive contributions. The fact that scaling laws work, the fact that not only we have an algorithm that learns, but that it gets predictably better with scale.
[0:25:53] And we can forecast ahead of time before we spend the hundreds of millions of dollars on a training run how it's going to perform. I think that is one of the most amazing facts discovered in the last decade. RLHF, the fact that with such little human data, we can get the models to perform as well as they do, also quite amazing. We probably need to make one to five more discoveries of that magnitude. Maybe bigger, like one more really big discovery, let's say, 20 little ones.
[0:26:22] And I think we'll have something quite extraordinary. But we just try to do whatever is in service of the mission. Right, so because you mentioned AGI, let me ask a question on that. So, well, you talk about AGI, and then there are people like Ben Gosdell who talk about AGI as a collaborative technology, right? So, for example, there are such, you know, sort of ideas that let's collaborate all together and let's build a really, really huge AGI. So do you think that idea of picking up over, let's say...
[0:26:51] The efforts inside, or maybe efforts by consortium, or even by organizations like OpenAI. I think, actually Sam, I think you're best positioned to answer this. I sort of think what we think of as super intelligence, the framework I used to have is it would be this like one super brain in the sky, and we would like hope that it did the thing we wanted and answered our question. And that was scary for a bunch of reasons, difficult to imagine.
[0:27:25] What I now think is that there will be many AI systems in the world that are part of society in the same way that we are and that contribute knowledge and ideas and the technology tree and the infrastructure. And what makes us different than our great, great, great grandparents is not the biological improvement. Our neural networks didn't get that much bigger. They didn't get that much more powerful. Sure, we eat a little better. We take better care of ourselves. We have modern medicine. But that doesn't explain the difference.
[0:27:54] What explains the difference is the infrastructure, the sort of societal scaffolding that is built up between all of us. You know, all of the people that came before that worked super hard to contribute, all of the pieces of technology that it took to make this microphone or that computer that we all now get to use to build even more impressive tools on top of. And in that sense, the super intelligence is not any neural network, but it is the infrastructure that exists between them.
[0:28:21] And I think that's how things are going to go. I think that's much better. It's a safer world. You know, humans get, I think, a more fun role for sure and one that's more in control. But that's sort of what I think happening. There will be many efforts. People will contribute many AI systems into the world. And the super intelligence will be the product of all of it outside of any neural network. Hi, Sam. Let's take a couple of questions. Hello. Hi. Yes. One by one. Hello. Sam.
[0:28:51] As you say that AI is future, and yes, we definitely agree and we believe it, but there are concern rising that there are climatic degradation due to it. One Microsoft data says that there were 0.2 million gallons of water were used to train GPT-3. So going forward, how do you think to balance climatic action and also to develop AI? I'm not sure if that number is right. It might be. That's more than I would have guessed for the cooling.
[0:29:22] We want to be net zero immediately. And I think we are on our training runs. And we want to be all on renewable soon. Water for cooling, we're looking at a lot of cooling ideas that don't evaporate fresh water. We're still doing active research there. I think at the current scale of systems, they do not meaningfully contribute to global energy usage. But it's a fair question, because eventually they will. And we're designing systems now that have quite significant energy usage.
[0:29:50] I believe we're going to have fusion working as a planet in the next few years. If not, I think we can, although it'll be more expensive, still get to where we need to go with solar and storage systems. It's very important to us to transition to clean energy and also to think about the other impacts like water usage. But I think it does get overstated at the current system scale by quite a lot. Hi, Sam. So quick question. So you just mentioned that we want to build technology.
[0:30:20] I think one thing that's different about this field than any other field of technology I've seen is people working on it.
[0:30:53] At least at the leading edge, understand what the stakes are. No one wants to destroy the world. No one wants to cause grievous harm. And it is unusual that in this field, it is the AI developers that are calling the most for regulation of superintelligent systems and are doing the most to build out proactive testing. So I hope it keeps working like this. I think the VW example was definitely a shock to many of us. But I see something very different happening this year. I don't know if either of you have anything to add.
[0:31:25] Hey, Sam, this is... Okay, if I may interrupt. Okay, look, so please don't sort of abduct the mic, right? Please keep on passing it. All right, yeah, please go ahead and then finish. Hi, Sam. By the way, huge fan. I'm co-founder of Preclapse. We're building our own products. And just wanted to say that ChatGPT is something that has made the software development process itself cheaper.
[0:31:55] For sure. And, you know, we are building something sort of hardware, some kind of embeddings. And would you say that, you know, to kind of find my differentiability in a more longer term, I need to think about products in a much more different way and my own skill sets in a much more different way than people used to think, like, for the last 20 years in itself? As a programmer myself, the first time I used GPT-4, I was definitely a little bit scared for my job.
[0:32:29] But I think the beauty of these systems is that they actually have the ability to 3x, 10x, maybe eventually 100x your productivity. And I think human creativity is still there. We haven't lost that yet to the AI. And I think we'll be very excited. I think we're seeing using these systems to achieve a lot more means you build better products faster, get to market quicker. And I think that's the way things are going to go for a while at least. Hey, Sam. This is Dave from 42 Media.
[0:33:01] I just had a quick question about the startup OpenAI fund. So I just wanted to understand in coming future or already have you had any discussion with Indian startups or maybe possible in the future coming months that you might be investing in Indian startups? We had some conversations yesterday. Be excited to do more. We would love to invest in Indian startups.
[0:33:23] Right, so especially a request to friends from media. I mean, you perhaps would have already watched, right, Sam's interview last evening, so please don't repeat the questions here. Thanks. Okay, so question from you. Yeah. Hi, Sam. At Triple IT Delhi, we have been building a new area called computational gastronomy, where we are trying to use data science and AI for capturing culinary creativity. And we have also used GPT-2 as a pre-version. One of the...
[0:33:53] Question that has been with us for a while is hallucinations. I guess hallucinations in chat GPT are still acceptable, but when a recipe comes out, hallucinating it becomes a serious problem. A quick thought on that in terms of where major problems are being solved and generative approaches are used and hallucination-like problems arise, that is going to be a major problem as far as we think. So, what's your take on it?
[0:34:19] I think we will get the hallucination problem to a much, much better place. It will take us, my colleagues weigh in, I think it'll take us a year and a half, two years, something like that. But at that point, we won't still talk about these. There's a balance between creativity and perfect accuracy, and the model will need to learn when you want one or the other. But it's clearly, in addition to speed and cost, one of the biggest complaints that people have, and we're trying to improve it. Either of you have a different opinion?
[0:34:51] I think the only thing I'd add, that use case is really interesting because it needs to be both creative and at the same time not hallucinate. So it's a very interesting use case. And I'd say one challenging aspect there is if you're looking for something novel and if that's something you can't verify because there's no ground truth, there might just be some need for experimentation there as well because the models will be getting what is truth from something that's ground truth at the end of the day. Right. Thank you.
[0:35:19] Quick request, please do tweet out questions. There's a hashtag up there, both for people in this room, other rooms on campus, as well as online, joining us virtually. Thank you. Go next. Yeah. Okay. Hi, Sam. In your very recent talk at Tel Aviv University, you said that sometimes for searching for certain results, rather than turning to Wikipedia, you actually turn to ChatGPT. What do you think are the implications of this kind of a statement, particularly when ChatGPT is not linked to any kind of...
[0:35:50] You know, search venues, any kind of links which Wikipedia actually rests on. And what implications this has for next generation of researchers who very, very clearly follow you a lot. Can you reflect on that? Thank you. Oh, the version that I always use it with web browsing. And so it does have links that come out. Because I probably trust the answers that come out of ChatGPT the least of anybody on Earth. So I use it in web browse. OK, thank you. OK, so a fine question. Let me take that from Twitter.
[0:36:22] So I remember, Sam, in one of your tweets, perhaps last week or sometime, you mentioned that you watched Ex Machina recently. I did. Okay, cool. So a question on Twitter is that, does Chad GPT have a kill switch? Because people have seen, right, Ex Machina, what she did. I mean, yeah, we can power down the data centers, but I think it's dangerous to take too much inspiration from sci-fi. You know?
[0:36:52] I don't think it's quite... I think things are going to go wrong, but they're not going to go wrong in quite the same way. And the physical kill switch idea, I think it's the wrong thing to depend on for safety. The problems that we face are going to be much more subtle and less obvious than that. Okay, so again, sort of building on that, do you think that someone can actually download a chat GPT on a robot and then it becomes XMachine? No. I think that...
[0:37:23] It's very easy to mix up, whether this is a tool or a creature. And the more you put language, to some extent, but certainly into a robotic body, it feels more like that. But no, I don't think if you put GPT-4 into a robot, it would become that thing. Hi, Sam from Triple IT Delhi. So human societies clearly are much more than just conversations. They're also about ethics and creativity.
[0:37:48] So I'm sure you've all already thought about these dimensions which are beyond conversations. And what is OpenAI targeting next to make it safer? To make it safe? Yes. We don't believe there's any one single answer to safety. We have an approach that has many layers. So we train the models to refuse certain behavior. We have monitoring on top of it. We have our post-training steps. And we use internal, external red teams. We have these audits. We have these tests that we run.
[0:38:18] At the current knowledge that we have of how these systems work, defense in depth and I think good safety systems engineering is the only practice that will work. Thank you. Hey Sam, I have one question. It is more related to your other startup, WorldCoin. So it completely makes sense that you're trying to solve the problem of identity and defects using cryptocurrency and blockchain. But why do you think iris scan is the solution? Don't you think in the era of AI, even iris scans can be faked? Eventually.
[0:38:48] The company has put a lot in place to make that harder for now. I think there are other solutions that'll work, and I think it's good as a world for us to try many approaches to this problem. So that's one idea. If there's other people doing others, I think we will somehow need to be able to verify unique humans in a privacy-preserving way, and also have an ability with that to equitably share the benefits, the access to these systems, and eventually the governance too. But if there's a, you know...
[0:39:16] People will try many ideas here. Hi. So Sam, you talked about not getting too much inspired from the sci-fi and everything. But I've read articles wherein you yourself have mentioned you were ready for the doomsday with guns and the gas masks and the gold. So what's about that? I mean, I was like a Boy Scout as a kid. You know, be prepared is our motto. It was very fun to do survival drills. Let me tell you, if AI goes wrong, no gas mask on Earth is going to help you.
[0:39:43] It's a little bit silly. It was like a fun hobby that I stupidly mentioned to a reporter 10 years ago or something like that. It's not something I actively think about or like spend time practicing. Hi, Sam. I'd like to ask a question that a major problem that ChatGPT has is context. So at this moment, you have a maximum window of 32K tokens, right? And I guess you are working on increasing that. I'd like to ask...
[0:40:12] One of the reasons why it might be the case because of the one square requirement of the transformer model. Have you guys looked into techniques such as quantization or low rank optimization or even approximation of transformer models like performer or reformer, etc. All right. I think we will find something other than n squared attention. I don't know how long it'll take us. I think there's a lot of approximations that could work.
[0:40:43] I don't know the latest on all of our research there, but we're actively pushing on it. And we want to get to a world where effectively, like in spirit, you can have as long of a context as you want. It may not literally happen by you passing 100 million tokens to a context, but something that has that effect. And we're hard at work on that.
[0:41:07] First of all, a big fan of yours, sir. Basically, my question is related to a lecture of Yanli Kahn. Previously in a lecture, he perfectly pointed out that these LLMs, not just chat GPT, are prone to hallucinations or to the reasoning task. And he said that it's due to the problem of the autoregressive nature of transformers. So first question is that, what's the limitation behind making these transformer models non-autoregressive?
[0:41:37] Second is that if we are able to make these transformer models non-autoregressive, so using the hidden layers of transformer networks can be made a certain context on which these LLMs can condition on it and have some logic behind whether they are getting a correct answer or not. Autoregressive models have certainly gone much further than Yann LeCun thought they were going to.
[0:42:06] Also much further than I thought they were going to. I still do share his opinion that something feels inelegant if this is the approach all the way. And we're looking at new architectures all of the time. And I hope we find one. Or I hope we find, if not a new architecture, just an alternative to this autoregressive way. Obviously the whole world has taken note of diffusion models and what those can do, but it hasn't yet worked quite the same way for text.
[0:42:34] Maybe it does someday, maybe it doesn't. I think what I would say is all of these are details, and it's so easy to lose the forest for the trees here. What matters is we have an algorithm that learns, and it gets predictably better with scale. And if we can make it a little bit more efficient, great, but I think everything is capable in this paradigm. Okay, guys, before the next question.
[0:43:03] Just a remark. I mean, let's not go too technical, you know, into the technology in this chat. Okay. Furthermore, so... May I? Just a sec. So we will have a hard stop at 3 p.m. We are anyway running sort of 10 minutes late. I'm sorry, 4 p.m. I'm sorry. Okay, so 4 p.m. And we will now take...
[0:43:31] Perhaps only a few questions here. So if you could club your questions somehow. So one question from each columns. I have a very generic question doubling down on what you exactly said, Professor, if I may ask.
[0:43:50] So Sam, for quick context, I'm Aditi. I've been working with early-stage founders and also have briefly spent time in the venture capital to understand how technologies scale at a vast level. And I think while we have been hearing your thoughts upon OpenAI, GPT a lot in your interviews, your articles, even in the academic institution today, I would really like to briefly touch upon your incredible trajectory of having been a founder, an investor, supporting early-stage companies, being on the front line with Y Combinator.
[0:44:19] Then again, going back to building from scale and building a company for the future. So be it with AI or crypto or blockchain technology as a whole, I think these are technologies which are really going to...
[0:44:33] Hi, could you please start your question? Sure, yes. I would just like to ask what are the specific...
[0:45:01] Challenges you identify in early stage building journey of these founders and also of investors with AI specifically into the future. What do you think could be some of the challenges? Let's take those many questions answered. Okay, yeah, thank you. I think my model is just work on the problems that are important. It doesn't matter as much how you work on it. So you can be a founder, you can be an investor. What matters is like working on something that you're passionate about that you think will be important with people you want to work with.
[0:45:31] I happen to like operations more than I like being an investor, but you can do it either way. I think it should be all about the what. So guys, here's a solution. Just, you know, we want as many questions as possible. So we just ask the question. Maybe we can hear five, six questions at a time, and then you can pick some. Okay, we'll try. Yeah, so please be very, very concise. We don't have much time. Thank you. So as students, we should double down on things that AI would not be able to do.
[0:45:59] Are there some AI-proof skills or ways of thinking that we should now work on? Okay, thanks. Yeah. Hi. Hello. Hi, Sam. So I'm Ramya. I'm one of the founders here at IIIT. Actually, in 2020, you had launched with your brothers Apollo Projects, and we were one of the 19 startups that were shortlisted for them, and we actually pitched to you before. So we are building neural interfaces for different brain health challenges.
[0:46:30] Being in this area of brain health and eventually mental health, ChatGPT poses a lot of potential issues when people are sort of asking ChatGPT questions. So what are the future plans in place that take care of the potential medical risks associated with questions that are being asked to the interface, especially when it comes to mental health?
[0:46:57] So long context, right, for that. Sam, so question. You have met so many head of states on your tour. Have you suggested them personally about the IAEA model that you suggested for this AI? And second, AI does not have a physical footprint, you know, when you, you know, if a country buys more compute or something like that, right? So how do you track that? Do you think that bad actors can be prevented?
[0:47:25] Okay, let's do AI proof skills first. Adi, you want to take that one? Okay, if I may interrupt. Okay, volunteers, please don't let anyone leave the room before the event gets over, right? So there will be a soft close at the doors. Please don't leave the room. Okay, please continue. So about skills you should learn as AI becomes better and better, how do you stay relevant? Sam actually had a great answer at the developer roundtable earlier, which is the skill you need to know is...
[0:47:53] Just how to keep up with change and how to keep learning. And as long as you're learning at a faster rate than the models, I think you're good. So just, I think the meta skills actually are more useful in the long run. I'll take the mental health one. I'd say on mental health, it's actually, it has two parts to it. One, actually verifying, does the technology actually work for this use case? And then two, it's this fundamental question of should language models, should AI...
[0:48:22] Hi.
[0:48:52] And then on the IAEA question, we've been really pleasantly surprised by how much enthusiasm there's been from really all of the world leaders we've talked to to thinking about an idea like this. I'm optimistic we can get something done. And I do think given the scale of computing hardware and energy it takes, it will be sort of trackable. Hi. I have a question about geopolitical weaponization of technology. It was seen recently during the Ukraine-Russia war.
[0:49:20] How safe is OpenAI from the influence of superpowers and their attempts to... Okay, all right. All right, so let's grab a question from... Yeah, I have a question. So, you know, like when we say that GPT models are like, you know, creative, but creativity comes with responsibility. So isn't it contradictory that you're asking for regulation from governments and not self-regulating?
[0:49:48] Because that would be the ideal form of creativity is to take responsibility for what we create. So we do self-regulate. I mean, we spent almost eight months on GPT-4 making sure that it was safe enough to release. We built new technology. We worked with external organizations to figure out what the limits should be and to test that the model met them. We think that's great. We won't be the only players here. We do think that coordination is important. So self-regulation.
[0:50:19] It is important. It's something that we want to offer. But I don't think the world should be left entirely in the hands of the companies either, given what we think of as the power of this technology. Sam, hi. So just one interruption. So I mean, like Claude, one question at this point. I think after this one, we won't be able to take questions online. So there is a from ..
[0:50:46] So what is Sam's view on hiring cheap labor for data correction? Cheap labor for what? Data correction. Data labeling, like? Yeah. I mean, I think we pay more than any other company in the world by a lot. Like, we really care about high-quality data labeling and treating people well. Okay. Right. So Sam, right here.
[0:51:06] I was asking about data sets. And India has been priding itself about its diverse set of data sets for the past few years now. So thinking about data sets, how is OpenAI going to interact with the kind of data sets that the government data sets, also India stacks, that we have available in India? And how is that going to help with OpenAI? And I think you haven't answered the geopolitical question.
[0:51:34] I'd say for us, more data is usually just always good. So all the data sets that we get, we do validate. And if it's good, we usually use them to train models. Like more data is always good. I'd say one additional thing here, though, is like different languages and how we sort of balance that. I do think there's a lot of potential here that you all could help contribute as well in terms of creating better evaluations in different diverse languages.
[0:51:59] Also just getting higher quality data. Even right now, English just has the biggest volume of data. So it's still a lot of work needed to catch up to it. So there's definitely a lot of room for innovation here. On the superpower question from earlier, we believe we can withstand a lot of attacks. But if we continue to succeed to the degree we think, we don't really know what the geopolitical attacks, the superpower attacks look like. Nor does anyone else. We'll be in uncharted territory at some point.
[0:52:28] So we're trying to think of what you do for a world like that and how you defend our IP against that. Quite tricky. Hey, Sam. Royal here. I'm founder of a SaaS called Spur, also a graduate from IIIT Delhi. I'm particularly intrigued by the concept of chat GPT plugins. Could you shed some light on when do you plan to make the access of creation of chat GPT plugins public? And what kind of impact can this create when you provide this access to a wider audience that can
[0:52:59] Contribute to the shape of ChatGPT through plugins. Yeah. We have the expert here. Oh, awesome. Plugins is one of the most exciting projects I've worked on at OpenAI. I think connecting ChatGPT to the external world, to APIs, et cetera, unleashes a bunch of new use cases. That said, I think we're still trying to fine-tune the model to be really great at calling plugins. I think it does a good job, but not an amazing job. So it's still, in a way, pre-product market fit. And so we're working towards it.
[0:53:29] I would love to get your email. Send me an email and we'll get you developer access immediately. Okay, one very large question. So I'd say the first part of that question, so that's something that we call jailbreaks, which is where people find creative ways to break the model. Actively working on it, like robustness, is just a very challenging problem. I think the second part of that around sort of emotions and like chat GPT possibly demonstrating this.
[0:54:51] It's really tricky to say what's happening in the model and whether it even has understanding, like far less emotions. Depending on which researcher you talk to, there's like a wide breadth of opinions there right now. I'd say at the very least, like right now with something like GPT-4, it probably doesn't have this nuanced understanding. And what we're seeing is just mimicking behavior, like it's mimicking what it's seen in the training data.
[0:55:15] Which is things people have put on the internet. So it's probably seeing that people respond this way to threats. People respond this way when you speak to them this way. So I think that's what's happening there, where it's mimicking that behavior. I think this question the world is still trying to answer. And it's going to need to be, like with many other industries, dependent on the facts of a particular case. The main thing I would say is we should all work together.
[0:56:14] The fundamental model developer, the foundation model developers, the people building applications on top of that, the way we educate our users to try to make that not happen. But things will go wrong, and there will need to be a framework for that. And it'll be some combination of, you know, has the foundation model developer appropriately disclosed the limits and addressed what they know about? Did the application developer kind of try to get around those? Was the user using it in a way that it was not intended? Like, you know, you still can't drive a car drunk 100 miles an hour, even.
[0:56:44] That's just not allowed. That's on you. So, there's going to have to be a complicated thing there, and I don't think any of us know the answer yet. Thanks a lot, guys, for all your questions. This is the last one. I'm here. So, this is the last question, and we won't be able to take more questions now. Hi, Sam. So, basically... Okay. So, guys, it was really thrilling, believe me. I mean, the last one hour was... Let's get his question, and then we'll see. Yeah, yeah. Okay.
[0:57:17] Sorry, they got behind you. Please pass the mic if someone is close. Hi, Sam. So basically, I wanted to ask, since a lot of AI, basically, now new problems are facing in the area of bias and things like that, so what plans does OpenAI have for tackling that in a diverse country like ours and like many others?
[0:57:42] So I'd say probably one of our main responses right now is what we're calling democratic inputs and a call for public input into how the behaviors of these models should be shaped. We recently put out a blog post around a few weeks ago. If you're interested in the problem, you can actually work on it right now because we've put out this public call for it. And the goal is to really learn from people, see what they want, and then encode that into the models. Okay, thank you all very much.
[0:58:11] I wish we had more time, but this was a lot of fun. These were great questions. OK. Thanks. Thanks a lot for joining us. And thank you very much to all others who joined us even online. And let me request on behalf of all of you, if Sam and team could
[0:58:34] Sort of have some kind of answering to the questions which were posed online in the last one hour. So with that request, please don't leave the hall again. And the request is also to the volunteers. Please manage that, including those who are in the overflow rooms. And once the volunteers announce that you can leave the room, you can then start leaving the room. Thank you very much again. So yes. All right. Thanks.