OpenAI CTO Mira Murati on Chatbots and Artificial General Intelligence | The Atlantic Festival 2023

for a conversation about Ai and the future of humanity please welcome mea Mora the chief technology officer for open AI here to lead the discussion is the Atlantic staff writer Ross Anderson Mera welcome thank you thanks for having me here um you know I've gotten to know you a little bit and I want our audience to get to know you a bit um and so I thought that we'd take you uh 5 years into the past um it's 2018 uh great year pre pandemic um and You by that time had worked at Leap Motion amazing augmented reality company um you'd worked at Tesla for three years and at the time open AI is like this little dinky kind of think tanky company in Silicon Valley nobody's heard of it except like nerdy types like me and you decide to sort of stake your F future on it um tell us about that choice yeah so maybe on a quick biographical note I've always been interested in math and science in general and so you know in a way I I was always interested in um thinking and how how thinking develops but computers didn't really come into my life until later um I I grew up in Albania and uh so you know I I was really interested interested in actually solving problems and figuring out how to build Technologies um that would help us shape our lives in some positive way in some form and I started by working in Aerospace um after studying engineering and then went on to Tesla uh where I worked on Model X um and you know after getting a taste for how to apply AI to self-driving cars uh I started thinking about other applications of AI and um particularly in you know how we interact with technology in general and started thinking about spirital Computing virtual reality augmented reality um and how the the great potential that that has in you know expanding our thinking and helping us think better about very abstract Concepts like you know Chaos Theory or um the shape of molecules and you know imagine if you could just intuitively grasp these Concepts the way that we do you know a football by just interacting with with these Concepts uh in in a natural way just like we interact with our physical space and so so that's what took me into virtual reality and augmented reality and I think that was a bit early um in terms of the Readiness of technology to do something real with it uh and and so but that kind of got me thinking more about um General applications of AI and how how far you can push this this narrow applications um and so started getting really interested in uh thinking about uh General AI what we now call AGI um and at the time Deep Mind and open AI were working on AGI and to me it was very obvious that uh open AI was going to be the place where I want it to be because of it Mission which was to and still is to develop General artificial intelligence in a way that it benefits Humanity um so I joined when it was a nonprofit and since then we've sort of changed the structure a bit um so that we can raise enough funding to get the compute that we need to actually develop these Technologies but we're still a capped profit so that kind of limits um the The Profit the company makes um and then everything above a certain amount goes still to the nonprofit and you know at the time in 2018 openi was very much a research lab there are a lot of exciting projects going on um but it was still you know there wasn't thinking going on about building developing a product or deploying these Technologies out there it was very much a research lab and just before that in 2017 um deep Minds had developed um uh Alpha go and uh so you know we saw these systems that were really great at strategy games this AI systems that were created at really very hard games for the first time but AI was very much in the game Space we were still kind of operating in uh from Atari games into to strategy games and that was really the space we had not gone too much into language understanding that was yet to come in 2019 yeah and uh you've gone quite a bit ways of ways into language understanding and I know one of the things that falls under your personal fum at openai is chat GPT and this week you all announced that uh very soon we'll all be able to talk to chat gbt and have it talk back to us and uh maybe this is just me but I sort of hate talking to Siri and Alexa um and so I'm wondering why I will like talking to chat gbt so I wonder why you hate talking to those other systems it's probably because you know the interaction is not very natural they don't really talk back seamlessly or the understanding it's difficult so it kind of breaks your experience and the the immersion that you have with the technology and maybe even the trust that you have that you know it will respond back it will actually understand you it will be this very natural high bandwidth interaction um you know very similar to how you talk to a colleague or how we're talking right now and so that is really the goal you know to interact with these AI systems in a way that's actually natural that in in a way that You' collaborate with someone and it's high bandwidth you know you could talk talk in uh text just exchange messages or we could talk like this or I could show an image and say hey look uh I want to I got all these business cards when I was in the hill in these meetings can you just put them in you know my contacts list something like this yes and um you know you have this very rich and natural way of interacting with the technology so that is really the goal we want to kind of move further away from our current interaction which is very you know we sort of slave to the keyboard and this the the touch mechanism of the phone and if you really think about it that hasn't really been revolutionized in decades and we have these technologies that are capable of so much more and it seems like the first thing to unlock this capabil ities is how we interact with them and so this week we we introduced a few more uh Technologies and really a different way of interacting with AI systems through audio a much more improved Del system and bringing those in ch GPT and CH GPT provides this interface that's very um easy to interact with you can kind of go back and forth with the system and you can say hey you got got this wrong can you try again and you can kind of get a little bit of insight into what's going on even when the system gets the answer wrong so it's very powerful sort of this like teacher student method a little bit of you know Socratic type of method can you tell us about like a fun conversation that you've had with chat gbt so presumably you've been using the voice version yes I I usually use the voice version when I'm you know walking around and I'm not typing or uh when I'm in the car um so yeah fun fun conversations that uh let me see what I mean usually I use it for work uh and so it's you know interesting and fascinating and uh I I'll usually ask it to maybe explain something uh or help me understand what's going on with something um kind of go back and forth yesterday I was doing a demo and I asked it to make an illustration of um kids story that had uh you know dinosaurs and unicorns that it was in space and he kind of made this really crazy stuff and then it was like well can you edit the dinosaurs and make it more uh interactive and it was making this like really fun and crazy illustrations really creative really cool uh it was sort of entertaining yeah is there uh if it's the case um as you say that this uh this experience of talking to AIS is about to get a lot more compelling is there some risk that companies um not open AI of course but uh other companies might um sort of tune these up and optimize them for engagement the way that say Instagram and Tik Tok are optimized for engagement and that they might actually become addictive definitely yes even more addictive than the systems that we have today because you know we're introducing things like longer memory or personal personalization and so these systems will have a much better grasp at um you know the things that you want the things that are relevant to you and the things that will ultimately actually help you in day-to-day life um I I would love a system like that but obviously with the capability and this enhanced um capabilities comes also the other side the possibility that we design them in the wrong way and they become extremely addictive and we sort of become enslaved to them so we have to be extremely thoughtful in how we develop them really study how people are using them how how we're engaging what what it does to our daily lives to society and respond to that so I think we have to be very careful with how we deploy these systems um and really learn from this iterative engagement um that we have with them because we really don't know out of the box we we have to discover and we have to learn and we have to explore but there is certainly a pretty significant risk in making developing them wrong in a way that really you know uh doesn't enhance our lives and in fact it it introduces more risk you uh so far we've been interacting with chat GPT with our phones and on our computers um earlier this week it was reported that uh you all are talking to Johnny I um the Apple designer about the possibility of a device um what should we be imagining is that like a speaker is that an earbud um okay so I won't really speculate on the hardware uh stuff or on the news this week but what I can give you a sense of is sort of what the future might look like and what what we should expect yeah uh so you know we in in 2019 we had gbd3 and there was the first time that we had AI systems that kind of showed some sense of language understanding before that we didn't think it was really possible that AI systems would get this language understanding um in fact we were really skeptical that was the case and we were kind of exploring only the the games World um and now it's 2023 it's not you know uh it's been a really short time so you can kind of see the the rate of technological progress has been incredibly steep um and in fact when we had gbd3 in in uh 2019 we we're sort of still questioning whether you know there and even today we're still questioning whether these systems are really reasoning or there is real creativity going on um but we can all see that they're actually useful in our daily lives and they're being introduced in to our business world they're creating value in many domains they're entertaining they're fun to interact with um and I think this became kind of more obvious to the public Beyond just the tech world when we introduced chat GPT which is an interface that's super easy to interact with and at that moment in my mind you know if if if you sort of travel in the future and look back at mil stones for AI there was there was a really uh significant inflection point because it it made the capabilities of these AI systems more accessible to everyone out there and uh not only the capabilities but also you know the risks and the limitations now a lot of people understand what a hallucination of an AI system means for example um and so that is very significant because as we introduce this systems into our daily lives into our infrastructure we all need to be very thoughtful about how that actually evolves and um not just the technical infrastructure but the social infrastructure to make that go well um and so looking ahead um I I I think there's going to be these systems will become increasingly more capable we're looking at bringing all the modalities of text and code and video and images you know everything into one system and the idea behind that is that we want um higher we want these systems to to have a higher understanding uh ultimately that's where that's where where we're going we're trying to build this generally intelligent system and what's missing right now is new ideas A system that can generate uh new idea that means that you know to generate a completely new idea um uh like theory of general relativity you need to have um the capability of abstract thinking and so that's that's really where we're going towards this systems that will eventually be able to help us with extremely hard problems not just collaborate alongside us but do things that you know today we're not able to do at all um I didn't hear about a device there um uh maybe you could like blink once for a speaker um uh look since we've hauled you out here to our imperial capital um I do want to ask you a question about politics um you know uh you may have noticed that the last few Federal elections we've had here have been quite contested um the results and you all um along with other companies have put out software that can generate genuinely photorealistic images about just about anything um and that could be a rival candidate um you know committing some nefarious act or someone throwing away ballots um at a polling station and I just wonder at at this time when democracy you know when people's attitudes towards democracy is that it is so fragile uh what are you all doing to make sure that uh your technology does not undermine people's uh feelings about uh elections yeah so I think even if we were not in this context it would still be incredibly important to roll out the systems um sort of you know slowly and understand what the risks are what the limitations are try to uh build in more robustness into the mitigations and then expand from there I I don't think it's possible to aim for zero risks uh I don't think that's realistic but what we want to do is really minimize the level of risk that this technology poses or these Technologies pose um and you know the way that I think about it is usually in terms of tradeoffs like how much value is this techn techology providing in the real world and how much can we really mitigate the risks so if we think about um say gbt 4 in ch gbt or in some other deployment uh yes certainly the biggest one of the biggest risks is hallucination so the ability of the model to really make up stuff and it can do so in in quite a convincing way quite persuasive um and so we really want to minimize that and there is a ton of focus at open AI but also more broadly in the industry to do just that how do you increase the reliability and the accuracy of these models and uh you know when we launch gbd4 we took about six months to just focus on safety which means that's kind of a broad uh term to to describe a lot of things in AI U and we need to build more accurate uh taxonomy around actually what safety means but you know the in this in this case what it means is really focusing on the accuracy and the reliability of the model and one way to do that actually there are a few ways but one way is through um reinforcement learning with human feedback so we bring our feedback to this AI systems and uh you know you can ask a question and you can kind of get go back and forth forth and you can correct the system when it's uh getting it wrong and you can sort of guide it towards the right answer similar to you know really teacher student interaction that kind of dynamic and these systems respond really well to that and so in fact this is actually why we launched CH GPT initially we launched it because we wanted this feedback from researchers to help us make gbd4 more accurate more reliable people um less prone to hallucinations and uh you know obviously then afterwards we discovered that people actually really enjoy interacting with the technology and viral um which also led us into other directions and in fact the whole industry but reinforcement learning with human feedback is very powerful in um making these systems more accurate more reliable less prone to hallucinations and uh you know another another lever is to actually introduce fresh information so often you know these models are done training that is all data so it doesn't have information from today and so if you ask a question about what's going on um and you know like summarize what's happening in the world today that would be pretty difficult and in that case maybe the model should say I don't know instead of making up stuff um but you know we hallucinate all the time too we make up stuff uh and and so sometimes people also have very unrealistic expectations of what AI should do that we don't have for ourselves um but can I ask you like a like specific question about that which is uh with some combination of browsing which is what I I assume you're getting at uh with uh the models being able to go on the internet and find new information and training them on massive ly new large data sets is it possible that say 2 3 years from now we could have um not perfect accuracy as you say but like Wikipedia level accuracy um I mean it should do better than that right uh it should do better than that it should be more scientific level accuracy um where you can really Trace back the pieces of information ideally or at least understand why through reasoning sort of like Chain of Thought understand why the system got to that kind of answer um so I would hope that we're going for for something much better than that but today we do have the ability to kind of browse what's going on uh and and bring this freshness of information which obviously increases accuracy there is there is a lot more happening but um I think reinforcement learning with human feedback can get us really far along with addressing ination for tax and then you know going back to your original question around elections and misinformation um we're also working on on basically addressing um provence of information so understanding whether a piece of information is AI generated or not and you can do that with text you can do that with pixels it doesn't really matter what the modality is um so we're we're doing a lot of research both open AI but also I know other companies are working on this as well it's a really important problem uh that of actually understanding the provence of information and then how you roll that out and the policy around that that's a bit different because it affects both users and non-users and so we're going to have to make some hard decisions for how we actually bring this to the public I know I want to ask you uh question about education I know it's a passion of yours um uh this is one of the areas where we've seen like an actually really immediate impact from AI uh from chat GPT in particular uh so much so that I gather that chat GPT usage actually dips in the summer when school is out um and a lot of those students that are using it are using it the way that you and I use it which is sort of as a like kind of um uh an expert in every topic right that we can talk to at any time um and can explain things to us but some of them are also using it on their homework right like uh just to um manufacturer assignments and I just wonder when you look at the future um how do you think that Nets out like uh what's what does that trade-off look like a few years from now so I think in many ways we'll probably have to kind of change how we teach uh but the upside is that it really changes how we learn and this probably the most important thing it's actually very difficult to learn how to learn it takes us so many years and some people figure it out after decades that's that's the most important thing and depending or of where you study where you go to school that's you know often not the main focus and it's very difficult thing it requires a lot of um personalized and customiz type of teaching and right now you've got a teacher in a classroom of 30 students it's impossible to customize the learning the information to how they best learn and this is what AI can offer it can offer this you know personalized tutor um that customizes learnings and teachings to you to how you best perceive and understand the world so that's incredibly powerful because it can really increase our you know just collective intelligence um but of course there is this downside that you know if you bring these Technologies to the way that we currently learn and study then yeah it does create these issues of people using it to you know copy homework and things like that so I I do think we need to figure out how how we do this you know how how do we change the education system the way that we teach to kind of minimize the the downsides and encourage people to actually use it in a way that um is beneficial and it's great and it's enhancing um I've already heard from um uh University professors that they're thinking of removing uh concept of grades from some of their classes because then this would allow students to focus more on actually learning and using and and I heard from them that they would actually encourage use of chbt so um it's you know students would actually use it to learn and to do a much better job and would not focus so much on actually getting the the grades a higher grade so I mean I I'm not a uh I'm not an expert in education but there I I think we have to be creative and we will have to adapt Society will have to adapt to this new technologies um it occurs to me that uh an AI that can write convincing College papers um might be quite competitive at a lot of jobs um maybe even mine um and uh you know when I talked to Sam your CEO uh earlier this year he was quite Frank about this right he says you know like jobs are going to go away um and I wonder I know that you all spent a lot of time thinking about the implications of this technology and so internally just like what does that look like is that like oh don't worry worry people in DC will sort of um develop new social programs that'll blunt that cuz if so we should talk about DC um or is there some are you imagining some other way out of this problem um so I think this is actually the first point I would make is that this is why it's so important to bring these Technologies into the collective Consciousness so people understand the impact not from papers not from articles but from leaving them and really seeing how it affects your job how it affects your life and you know adapting to this change because it's incredibly rapid it's it's completely unprecedented um and it affects everything because intelligence is such a you know Universal unit it it will really change everything and so we need everyone to understand what we're what we're doing and what we're building and sort of ort of make this available to people uh and make it accessible sort of slowly and do so safely and thoughtfully but it's important to actually bring it out there um and yes it will you know jobs will be lost some jobs will be created but just like every major Revolution I think a lot of jobs will be lost there will be maybe probably a bigger impact than in on jobs and in any other Revolution and we have to prepare for this new way of life like maybe we work much less uh maybe the work week changes entirely uh okay so then you know how do people leave their ideas being explored around Ubi we probably have to be creative and think about other social programs as well um I know there is a lot of engagement right now with DC on these topics and understanding the impact on Workforce and such but um we don't have the answers we're going to have to figure them out along the way um and I think it's it's going to require a lot of work and thoughtfulness Mera thank you for being with us today that's great thank you so much

2023-10-06 21:19

Other news