Sam Altman Reveals EVEN MORE About GPT-5! (Sam Altmans New Interview)

in which he talks about project Stargate I mean I think other people are talking about that but we're probably like looking at it through a lens that no one else is quite imagining yet um I mean we're definitely wrestling with how we when we make not just like grade school or middle schooler level intelligence but like PhD level intelligence and Beyond the best way to put that into a product have a positive impact with that on society and people's lives we don't know the answer to that yet so I think that's like a pretty important thing to figure out if you believe which we increasingly do at this point that AI infrastructure is going to be one of the most important inputs to the Future this commodity that everybody's going to want and that is energy data centers chips chip design new kinds of networks it's how we look at that entire ecosystem um and how we make a lot more of that and I don't think it'll work to just look at one piece or another but we got to do the whole thing so I know that this first clip was actually quite vague samman actually does do that sometimes because the details of what he is essentially doing is quite under locking key so to speak and that just essentially means that whilst Sam mman is talking in this interview he doesn't want to give away too many key details but what we do know he's most likely referring to here is of course the infamous project Stargate for those of you not familiar with project Stargate it's A1 billion data center that would likely be built over several years for the purpose of building artificial general intelligence and beefing up Microsoft and open ai's joint collaboration effort to basically become the AI Powerhouse that they could potentially be in the future once AGI does arrive now this is what samman is talking about here and he basically you know later on the interview although I don't inter include the clip in the video but he actually does talk about how in the future this is pretty much going to be one of the most valuable resources on the planet and I think that is something that is quite incredible to State because increasingly in the future as the technology gets better and as the intelligence is more increasingly reliable and more increasingly intelligent the capabilities are too going to increase which means that access to this kind of technology is going to become something I guess you could say quite like the internet you know how right now we use the internet for a lot of our daily needs and if you don't have the Internet it's pretty hard to function I think in the future AGI is definitely going to be a resource like that which is why he's talking about right now building really big computers is of course a problem and it's something that they're working on and does it gross so you know

in terms of just like the compute cost uh correct me if I'm wrong but chat gbt 3 was I've heard it was $100 million to do the model um and it was 100 175 billion parameters gbt 4 was cost $400 million with 10x the parameters it was almost 4X the cost but 10x the parameters correct me adjust me you know it I do know it but I won't oh you can you're invited to this is Stanford Sam okay um uh but the even if you don't want to correct the actual numbers if that's directionally correct um does the cost do you think keep growing with each subsequent yes and does it keep growing multiplicatively uh probably I mean now the reason I've actually included this hip clip here is because this actually touches on a fundamental aspect of AGI and AI level systems in the future that we are yet to build the kind of infrastructure that is probably going to power the economy and the Western hemisphere is something that we truly haven't seen yet and I think this interview gives us the first Glimpse on how much is actually going to be going into this very interesting future now the reason I say that is because he talks about you know increasingly more investment in Ai and of course as you're going to see later on in the video he does talk about certain figures which I'm yet to disclose now this is something that is pretty incredible because you know GPT 4 was allegedly $400 million to train and apparently the future models the future Frontier models are going to be even more expensive than that and there have been some industry people stating that models are going to cost a billion or even 10 billion to train in the future and some people and the reason this is why I brought this up some people are stating that you know models being more expensive in the future could mean that we're using the wrong architecture to train these models but I would kind of disagree because we've definitely achieved a lot by scaling up our Technologies as many of you may know about the better lesson which actually touches upon this subject but I think it's so interesting to note that 400 million for an AI system and we've seen that Google spent like literally 194 million I do believe on Google Gemini and it seems like companies are going to have to even spend more in order to deploy and train future models which is definitely something that is pretty incredible because it means that potentially we're about to enter a new age of I openi is phenomenal chat gbt is

phenomenal um everything else all the other models are phenomenal it burned you've earned $520 million of cash last year that doesn't concern you in terms of thinking about the economic model of how do you actually where's going to be the monetization source well first of all that's nice of you to say but Chachi PT is not phenomenal like Chachi PT is mildly embarrassing at best um gp4 is the dumbest model any of you will ever have to use again by a lot um but you know it's like important to ship early and office so I mean it's clear why I included that clip he literally stated that GPT 4 is the dumbest model that we will ever have to interact with again now the reason I included this one and I'm pretty sure maybe this is even included in the title somewhere is I don't think people fundamentally understand what this means and I think that's because people are forgetting how much of a shock it was when GPT 4 was released if you were around at the time of the GPT 4 demo that thing literally broke the internet in terms of how people were so surprised on what it could do now remember Sam here is clearly stating that GPT 4 is dumb this is not the only time he said this if you've watched leex Friedman podcast he frequently talks about you know well he only mention it once but he did state that you know gbt 4 is quite dumb he looks up to the ceiling and he's like ah yeah gbt 4 is is quite dumb you know it's unreliable it's pretty bad and he's talking about it you know a product that many people including myself use on a day today basis for a variety of I guess you could say intellectually demanding work in some very light regards of course not many intensive regards but in some you know light regards such as analysis and stuff like that but I think what it shows us is that the next Frontier of model is going to be completely different from GPT 4 because for samman to say not that things are going to be a slight Improvement but to call GPT 4 actually Dum shows us that whatever they are currently training whatever new system they currently using whatever Frameworks they've implemented in their AI system however they've trained it however they find you the model it's clear that they've made some kind of internal leap because stating that a model is dumb means that potentially whatever they were doing beforehand they've realized the clear mistakes and they've realized that wow what we were doing was actually dumb and what we did is bad so that means that whatever they've got now is clearly much far superior and the reason that I think this is so surprising as well is because if we take a look at some of the other companies that are actively participating in the space companies like Claude companies like Google they are really trying to compete with GPT 4 and many people are even like oh wow this model is just as good as GPT 4 or even better this model I'm going to use on a day-to-day basis this is so amazing but the CEO stating that gbd4 is dumb I mean the writing is literally on the wall for him to say that you know I think it's going to be clear to us that gbt 5 although they don't want it to shock us I think it's definitely going to be a major surprise um and a major surprise to of course many different AI critics because many people are stating that llms aren't accept towards AGI and are fundamentally not even that good because they hallucinate and they do XY Z but I think whatever open AI are cooking up and whatever they are doing I think that it is truly going to surprise us and I find it truly hard to believe right now what kind of system would make GPT 4 look dumb I understand hallucinations reliability but I'm guessing that they're probably going to use something you know really Advanced that's definitely going to take the industry by storm so I think it's

important given what we believe is going to happen to express our view about um but more than that the way to do it is to put the product in people's hands um and let Society co-evolve with the technology let Society tell us what it collectively and people individually want from the technology how to productize this in a way that's going to be useful um where the model works really well where it doesn't work really well um give our leaders and institutions time to react um give people time to figure out how to integrate this into their lives to learn how to use the tool um sure some of you all like cheat on your homework with it but some of you all probably do like very amazing wonderful things with it too um and as each generation goes on uh I think that will expand and that means that we ship imperfect products um but we have a very tight feedback loop and we learn and we get better um and it does kind of suck to ship a product that you're embarrassed about but it's much better than the alternative um and in this case in particular where I think we really owe it to society to deploy iteratively um one thing we've learned is that Ai and surprise don't go well together people don't want to be surprised people want a gradual roll out and the ability to influence these systems um that's how we're going to do it and there may be there could totally be things in the future that would change where we think iterative deployment isn't such a good strategy um but it does feel like the current best approach that we have and I think we've gained a lot um from doing this and you know hopefully the larger world has gained something too so one of the key points that samman actually talks about here is the fact that there's going to be no more surprises now I don't mean this in the literal sense that the next AI systems might not surprise us I think what he's talking about as we saw with Zora because that was definitely a huge shock Ai and surprise just don't go well for the human you know psyche like it's just something that you know we can't really comprehend and it's something that our mental health doesn't really take well because it's something that is essentially a new frontier something that we haven't really seen before and it kind of warps our understanding of what technology is and how we've come to understand it which is why Ai and surprise don't really go well so essentially here this is something that we've talked about many different times before is that of course there might be like model releases and model upgrades but unlike previous releases they don't aim to do stuff that actually shock us that put us in a sense of you know wow this is crazy and evoke some kind of social response because that would essentially mean that you know a public outcry could force you know government regulations or certain policy makers to enact certain laws that I guess would be based out of fear rather than actually regulating technical technological progress in a normal manner so this is something to where we're not going to get any more shocking surpris prizes and I don't think that this is a bad thing I think that basically what he's stating here is that models over time will gradually get better incrementally and it's going to allow you to you know adjust to the model allow your workflows to adjust to it and of course allow the public to adjust to what is now the new normal and what was crazy was that an open a employee actually said that sora's release was basically getting us ready for what's to come because they wanted to show us how crazy AI capabilities are and even then Sora definitely did break the mold in terms of what we saw because it was their first video model and it truly did surpass every other one that was on the platform which pretty much goes to show just how crazy open AI is especially when they put their mind to something so I think it's important to understand here that there won't be any surprises but these incremental updates that you do see on open eyes Twitter feed where you're seeing them post about small update here and a small update there that is something that you can be expected to have but of course there will be the model updates but they just won't be surprising I think what would likely to be happening is that in the future we're likely going to see incremental updates for these models so with the GPT 5 release I suspect that we won't have one entire release like we did before we're going to have GPT 5 released and there's probably going to be checkpoints along the way where they update that model incrementally and I think what this is probably going to be as well is it's probably going to mean that they are probably on like GPT 6 probably training gpt7 right as G GPT 5 is you know in its later stages which means that they're going to be far ahead internally than they are externally if they're releasing things much more incrementally and remember open ey can do this because they are the market leaders so they're not going to have a problem at all whether we burn 500 million a year or 5 billion or 50

billion a year I don't care I genuinely don't as long as we can I think stay on a trajectory where eventually we create way more value for society than that and as long as we can figure out a way to pay the bills like we're making AGI it's going to be expensive it's totally worth it now most people actually look at this and I think they look at this the wrong way I think samman is looking at this from a winner takes all scenario and I've said this before but some people still don't get it the point here is that he's stating that look we're making AGI it's going to be expensive and it's going to be worth it the point he's trying to make here is that you shouldn't really try to like scrimp and save when it comes to building AI because the initial bet here is that if you get to AGI a true AGI system that can do autonomous research okay and let's say you can account for 5 to 7% of the economy's total GDP okay which is pretty crazy but let's just say that is something that AGI could capture and as of 2022 the global GDP amounted to over 100 trillion us that could mean okay that this is a winner takes all scenario where if they get like 5% of that that's $5 trillion of economic value that one company is capturing that is absolutely incredible so if they spend a few billion ion dollar making the AGI system they could surely after get to artificial super intelligence which is winner take s scenario meaning that you know they could pretty much take over they could do robotics if they and if they get an AGI system to efficiently scale Robotics and do robotics they can pretty much do anything okay like if a human is able to do a lot of the things that we do now if they're able to effectively efficiently scale robotics after achieving AGI and Asi then it's pretty much open season of openi on any industry that they want they could just choose the most profitable one and then go from there now I do think interestingly enough there's definitely going to be some new Industries maybe like not on the level of manufacturing Dyson spheres but I do think that with that kind of Technology the breakthroughs that occur after that would definitely have the technology prioritizing some crazy stuff and not just you know the standard basic jobs that we have now but I think that what stamman is stating here that like look he's going to spend $50 billion it doesn't matter because once they get to AGI that's it okay they've won the game it's over now and well not essentially over but those are the companies how on Earth are they going to even try and catch up when a company is advancing exponentially it's pretty hard to do that so I think it's a win take Soul scenario and I think the investors know that and I think Microsoft are definitely prepared to invest as much as they can in order to ensuring opening ey get their first and of course they reap the benefits and so do you have a I hear

you vision in 2030 of what if I say you crushed it Sam it's 2030 you crushed it what does the world look like to you um you know maybe in some very important ways not that different uh like we will be back here there will be like a new set of students we'll be talking about how startups are really important technology is really cool we'll have this new great tool in the world it'll feel it would feel amazing if we got to teleport forward six years today and have this thing that was like smarter than humans in many subjects and could do these complicated tasks for us and um you know like we could have these like complicated program written or This research done or this business started uh and yet like the Sun keeps Rising the like people keep having their human dramas life goes on so sort of like super different in some sense that we now have like abundant intelligence at our fingertips and then in some other sense like not different at all and you mentioned artificial general intellig AI I think Sam mman might be downplaying this one here or it's probably realistic because he's always often said that like when you get to AGI it's probably going to happen and then everyone's going to go about their daily lives and I think what he's stating here that like even if they do crush it by 2030 and they achieve AGI by 2030 or in 2030 I don't think Life Changes instantly and I think that is probably kind of true I know that people might think boom we're just snap your fingers and we're in some Singularity type world with flying cars you know these tall white buildings that are completely smooth over and there's like these insane structures I think even if that is our future I think that will take some time so I do think that by the time AI is achieved in 2030 which many do predict it might be 2028 or 2029 I think that it's still going to take some time to build out the infrastructure for future societies the new social contracts and how things are going to work after that so I mean I guess yes this does make sense but um it's something that's definitely very hard to predict I think we need a more precise definition of AGI for the timing question um because at this point even

with like the definition you just gave which is a reasonable one there's that's your I'm I'm paring back what you um said in an interview well that's good cuz I'm going to criticize myself um it's it's too loose of a definition there's too much room for misinterpretation in there um to I think be really useful or get at what people really want like I kind of think what people want to know when they say like what's the timeline to AGI is like when is the world going to be super different when is the rate of change going to get super high when is the way the economy Works going to be really different like when does my life change and that for a bunch of reasons may be very different than we think like can totally imagine a world where we build PhD level intelligence in any area and you know we can make researchers way more productive maybe we can even do some autonomous research and in some sense like that sounds like it should change the world a lot and I can imagine that we do that and then we can detect no change in global GDP growth for like years afterwards something like that um which is very strange to think about and it was not my original intuition of how this was all going to go so I don't know how to give a precise timeline of when we get to the Milestone people care about but when we get to systems that are way more capable than we have right now one year and every year after and that I think is the important point so I've given up on trying to give the AGI timeline but I think every year for the next many we have dramatically more capable systems every year so I think there were two things that were really important in that clip and I know that was a little bit longer than normal but I think that one was pretty crazy because he literally said that every year from this year there will be increasingly capable models and that is something that I think most people can't really comprehend including myself because humans don't have a great time at comprehending exponential Improvement because it just doesn't fit into how we naturally see things growing so it's something that's going to be pretty hard to comprehend but I'm guessing maybe as time increases we may have a better understanding of how that looks in our day-to-day lives because if you were to show GPT 4 to someone 2 to three years ago you they might like be like what on Earth is this magic and this is of course a recurring problem with AGI is that many people have always said that there's not a definitive state of where AGI is and I agree I agree with this completely because AGI is not a thing it is a scale of intelligence that is increasingly better and I think the lines are going to keep getting blurred because right now as many would argue gp4 is a lowlevel AGI system but in the future when we have Vision systems audio systems uh video model systems I mean embodied AGI as well I think we're going to increasingly get further down that scale where we get useful AGI and samman's personal definition was when a system can do autonomous AI research and make scientific breakthroughs so that's samman's definition and I think that definition is important because it gives you guys a benchmark on where he thinks that a true level has been reached fundamentally where breakthroughs are I don't know yet um I mean this sounds like a copout answer but I think the most important thing about GPT 5 or whatever call that is just that it's

going to be smarter and this sounds like a Dodge but I think that's like among the most remarkable facts in human history that we can just do something and we can say right now with a high degree of scientific certainty GPT 5 is going to be smarter than a lot smarter than GPT 4 GPT 6 is going to be a lot smarter than GPT 5 and we are not near the top of this curve and we kind of know what to do and this is not like it's going to get better in one area this is not like we're going to you know it's not that it's always going to get better at this eval or this subject or this modality it's just going to be smarter in the general sense and I think the gravity of that statement is still like underrated so the reason I've included this clip and I think if you're going to take one thing from this video take this from the video okay and because I think most people don't understand what Sam mman means and like he said most people think that this is a copout answer this is not a copout answer at all I think him stating that the model is going to be smarter shows us that potentially there might even be some new architectures involved but I think the main thing that we can take away from this is not that the model is just going to be smarter cuz that is kind of vague but that is of course a definition but when I break it down for you basically this means that this model is just going to be so much better at reasoning understanding comprehension context awareness all of these things are truly important to basically being useful like so many times we ask gp4 to do complex tasks and it just really can't and I think in the future we're going to see how smart A system can get in terms of actually being able to respond to exactly what we want and I think this is what Sam mman is talking about when he says that systems are going to be smarter because GPT 4 has very limited reasoning abilities although it does do some reasoning tasks very well I think future tasks like planning agentic Behavior you know trying to model the world theory of mind all of these things are increasingly intellectual tasks and for an AI system like gbt 4 those are areas where it is pretty rudimentary so I think those areas in the future mean that this kind of system is going to be more widely used because it's going to allow for essentially more wild skill deployment if there's less ulation a lot more planning and a lot more complex reasoning it's definitely going to be used a lot more because it means that this system is just truly going to be smarter and this isn't something where you know you're looking at the benchmarks and trying to see that but definitely we could be reaching an area where these systems really do you know max out the benchmarks in terms of what's possible and I think that Sam mman stating that this is essentially going to be where we're at in the future I think this actually gets me really excited for what this Future model is going to be because I think that open AI are on a completely different path to many of the other AI companies that we look at on a day-to-day basis because they're still playing catch up and opening I are playing let's set the bar higher so having a model that's it smarter and them guessing that you know they already know where gbt 6 is going to be is of course something that you should you know watch out for so I would say that if you're that's building a company um definitely plan with in mind that gbt 5 is actually going to be smarter which means it's going to be applicable for a lot more things and maybe there are certain intellectual tasks that might disappear with this and I definitely think they will because if this model gets even smarter than it is now it's definitely going to be doing a lot more cognitive tasks which means jobs uh cognitive intense jobs are definitely going to be on the line so that is something that you do need to prepare for because I think it's something that you know it wasn't really talked about in this interview but you know as some who's looking at things across the Spectrum it's just a heads up for me increasingly worried about how we're going to do this all responsibly I

in which he talks about project Stargate I mean I think other people are talking about that but we're probably like looking at it through a lens that no one else is quite imagining yet um I mean we're definitely wrestling with how we when we make not just like grade school or middle schooler level intelligence but like PhD level intelligence and Beyond the best way to put that into a product have a positive impact with that on society and people's lives we don't know the answer to that yet so I think that's like a pretty important thing to figure out if you believe which we increasingly do at this point that AI infrastructure is going to be one of the most important inputs to the Future this commodity that everybody's going to want and that is energy data centers chips chip design new kinds of networks it's how we look at that entire ecosystem um and how we make a lot more of that and I don't think it'll work to just look at one piece or another but we got to do the whole thing so I know that this first clip was actually quite vague samman actually does do that sometimes because the details of what he is essentially doing is quite under locking key so to speak and that just essentially means that whilst Sam mman is talking in this interview he doesn't want to give away too many key details but what we do know he's most likely referring to here is of course the infamous project Stargate for those of you not familiar with project Stargate it's A1 billion data center that would likely be built over several years for the purpose of building artificial general intelligence and beefing up Microsoft and open ai's joint collaboration effort to basically become the AI Powerhouse that they could potentially be in the future once AGI does arrive now this is what samman is talking about here and he basically you know later on the interview although I don't inter include the clip in the video but he actually does talk about how in the future this is pretty much going to be one of the most valuable resources on the planet and I think that is something that is quite incredible to State because increasingly in the future as the technology gets better and as the intelligence is more increasingly reliable and more increasingly intelligent the capabilities are too going to increase which means that access to this kind of technology is going to become something I guess you could say quite like the internet you know how right now we use the internet for a lot of our daily needs and if you don't have the Internet it's pretty hard to function I think in the future AGI is definitely going to be a resource like that which is why he's talking about right now building really big computers is of course a problem and it's something that they're working on and does it gross so you know

in terms of just like the compute cost uh correct me if I'm wrong but chat gbt 3 was I've heard it was $100 million to do the model um and it was 100 175 billion parameters gbt 4 was cost $400 million with 10x the parameters it was almost 4X the cost but 10x the parameters correct me adjust me you know it I do know it but I won't oh you can you're invited to this is Stanford Sam okay um uh but the even if you don't want to correct the actual numbers if that's directionally correct um does the cost do you think keep growing with each subsequent yes and does it keep growing multiplicatively uh probably I mean now the reason I've actually included this hip clip here is because this actually touches on a fundamental aspect of AGI and AI level systems in the future that we are yet to build the kind of infrastructure that is probably going to power the economy and the Western hemisphere is something that we truly haven't seen yet and I think this interview gives us the first Glimpse on how much is actually going to be going into this very interesting future now the reason I say that is because he talks about you know increasingly more investment in Ai and of course as you're going to see later on in the video he does talk about certain figures which I'm yet to disclose now this is something that is pretty incredible because you know GPT 4 was allegedly $400 million to train and apparently the future models the future Frontier models are going to be even more expensive than that and there have been some industry people stating that models are going to cost a billion or even 10 billion to train in the future and some people and the reason this is why I brought this up some people are stating that you know models being more expensive in the future could mean that we're using the wrong architecture to train these models but I would kind of disagree because we've definitely achieved a lot by scaling up our Technologies as many of you may know about the better lesson which actually touches upon this subject but I think it's so interesting to note that 400 million for an AI system and we've seen that Google spent like literally 194 million I do believe on Google Gemini and it seems like companies are going to have to even spend more in order to deploy and train future models which is definitely something that is pretty incredible because it means that potentially we're about to enter a new age of I openi is phenomenal chat gbt is

phenomenal um everything else all the other models are phenomenal it burned you've earned $520 million of cash last year that doesn't concern you in terms of thinking about the economic model of how do you actually where's going to be the monetization source well first of all that's nice of you to say but Chachi PT is not phenomenal like Chachi PT is mildly embarrassing at best um gp4 is the dumbest model any of you will ever have to use again by a lot um but you know it's like important to ship early and office so I mean it's clear why I included that clip he literally stated that GPT 4 is the dumbest model that we will ever have to interact with again now the reason I included this one and I'm pretty sure maybe this is even included in the title somewhere is I don't think people fundamentally understand what this means and I think that's because people are forgetting how much of a shock it was when GPT 4 was released if you were around at the time of the GPT 4 demo that thing literally broke the internet in terms of how people were so surprised on what it could do now remember Sam here is clearly stating that GPT 4 is dumb this is not the only time he said this if you've watched leex Friedman podcast he frequently talks about you know well he only mention it once but he did state that you know gbt 4 is quite dumb he looks up to the ceiling and he's like ah yeah gbt 4 is is quite dumb you know it's unreliable it's pretty bad and he's talking about it you know a product that many people including myself use on a day today basis for a variety of I guess you could say intellectually demanding work in some very light regards of course not many intensive regards but in some you know light regards such as analysis and stuff like that but I think what it shows us is that the next Frontier of model is going to be completely different from GPT 4 because for samman to say not that things are going to be a slight Improvement but to call GPT 4 actually Dum shows us that whatever they are currently training whatever new system they currently using whatever Frameworks they've implemented in their AI system however they've trained it however they find you the model it's clear that they've made some kind of internal leap because stating that a model is dumb means that potentially whatever they were doing beforehand they've realized the clear mistakes and they've realized that wow what we were doing was actually dumb and what we did is bad so that means that whatever they've got now is clearly much far superior and the reason that I think this is so surprising as well is because if we take a look at some of the other companies that are actively participating in the space companies like Claude companies like Google they are really trying to compete with GPT 4 and many people are even like oh wow this model is just as good as GPT 4 or even better this model I'm going to use on a day-to-day basis this is so amazing but the CEO stating that gbd4 is dumb I mean the writing is literally on the wall for him to say that you know I think it's going to be clear to us that gbt 5 although they don't want it to shock us I think it's definitely going to be a major surprise um and a major surprise to of course many different AI critics because many people are stating that llms aren't accept towards AGI and are fundamentally not even that good because they hallucinate and they do XY Z but I think whatever open AI are cooking up and whatever they are doing I think that it is truly going to surprise us and I find it truly hard to believe right now what kind of system would make GPT 4 look dumb I understand hallucinations reliability but I'm guessing that they're probably going to use something you know really Advanced that's definitely going to take the industry by storm so I think it's

important given what we believe is going to happen to express our view about um but more than that the way to do it is to put the product in people's hands um and let Society co-evolve with the technology let Society tell us what it collectively and people individually want from the technology how to productize this in a way that's going to be useful um where the model works really well where it doesn't work really well um give our leaders and institutions time to react um give people time to figure out how to integrate this into their lives to learn how to use the tool um sure some of you all like cheat on your homework with it but some of you all probably do like very amazing wonderful things with it too um and as each generation goes on uh I think that will expand and that means that we ship imperfect products um but we have a very tight feedback loop and we learn and we get better um and it does kind of suck to ship a product that you're embarrassed about but it's much better than the alternative um and in this case in particular where I think we really owe it to society to deploy iteratively um one thing we've learned is that Ai and surprise don't go well together people don't want to be surprised people want a gradual roll out and the ability to influence these systems um that's how we're going to do it and there may be there could totally be things in the future that would change where we think iterative deployment isn't such a good strategy um but it does feel like the current best approach that we have and I think we've gained a lot um from doing this and you know hopefully the larger world has gained something too so one of the key points that samman actually talks about here is the fact that there's going to be no more surprises now I don't mean this in the literal sense that the next AI systems might not surprise us I think what he's talking about as we saw with Zora because that was definitely a huge shock Ai and surprise just don't go well for the human you know psyche like it's just something that you know we can't really comprehend and it's something that our mental health doesn't really take well because it's something that is essentially a new frontier something that we haven't really seen before and it kind of warps our understanding of what technology is and how we've come to understand it which is why Ai and surprise don't really go well so essentially here this is something that we've talked about many different times before is that of course there might be like model releases and model upgrades but unlike previous releases they don't aim to do stuff that actually shock us that put us in a sense of you know wow this is crazy and evoke some kind of social response because that would essentially mean that you know a public outcry could force you know government regulations or certain policy makers to enact certain laws that I guess would be based out of fear rather than actually regulating technical technological progress in a normal manner so this is something to where we're not going to get any more shocking surpris prizes and I don't think that this is a bad thing I think that basically what he's stating here is that models over time will gradually get better incrementally and it's going to allow you to you know adjust to the model allow your workflows to adjust to it and of course allow the public to adjust to what is now the new normal and what was crazy was that an open a employee actually said that sora's release was basically getting us ready for what's to come because they wanted to show us how crazy AI capabilities are and even then Sora definitely did break the mold in terms of what we saw because it was their first video model and it truly did surpass every other one that was on the platform which pretty much goes to show just how crazy open AI is especially when they put their mind to something so I think it's important to understand here that there won't be any surprises but these incremental updates that you do see on open eyes Twitter feed where you're seeing them post about small update here and a small update there that is something that you can be expected to have but of course there will be the model updates but they just won't be surprising I think what would likely to be happening is that in the future we're likely going to see incremental updates for these models so with the GPT 5 release I suspect that we won't have one entire release like we did before we're going to have GPT 5 released and there's probably going to be checkpoints along the way where they update that model incrementally and I think what this is probably going to be as well is it's probably going to mean that they are probably on like GPT 6 probably training gpt7 right as G GPT 5 is you know in its later stages which means that they're going to be far ahead internally than they are externally if they're releasing things much more incrementally and remember open ey can do this because they are the market leaders so they're not going to have a problem at all whether we burn 500 million a year or 5 billion or 50

billion a year I don't care I genuinely don't as long as we can I think stay on a trajectory where eventually we create way more value for society than that and as long as we can figure out a way to pay the bills like we're making AGI it's going to be expensive it's totally worth it now most people actually look at this and I think they look at this the wrong way I think samman is looking at this from a winner takes all scenario and I've said this before but some people still don't get it the point here is that he's stating that look we're making AGI it's going to be expensive and it's going to be worth it the point he's trying to make here is that you shouldn't really try to like scrimp and save when it comes to building AI because the initial bet here is that if you get to AGI a true AGI system that can do autonomous research okay and let's say you can account for 5 to 7% of the economy's total GDP okay which is pretty crazy but let's just say that is something that AGI could capture and as of 2022 the global GDP amounted to over 100 trillion us that could mean okay that this is a winner takes all scenario where if they get like 5% of that that's $5 trillion of economic value that one company is capturing that is absolutely incredible so if they spend a few billion ion dollar making the AGI system they could surely after get to artificial super intelligence which is winner take s scenario meaning that you know they could pretty much take over they could do robotics if they and if they get an AGI system to efficiently scale Robotics and do robotics they can pretty much do anything okay like if a human is able to do a lot of the things that we do now if they're able to effectively efficiently scale robotics after achieving AGI and Asi then it's pretty much open season of openi on any industry that they want they could just choose the most profitable one and then go from there now I do think interestingly enough there's definitely going to be some new Industries maybe like not on the level of manufacturing Dyson spheres but I do think that with that kind of Technology the breakthroughs that occur after that would definitely have the technology prioritizing some crazy stuff and not just you know the standard basic jobs that we have now but I think that what stamman is stating here that like look he's going to spend $50 billion it doesn't matter because once they get to AGI that's it okay they've won the game it's over now and well not essentially over but those are the companies how on Earth are they going to even try and catch up when a company is advancing exponentially it's pretty hard to do that so I think it's a win take Soul scenario and I think the investors know that and I think Microsoft are definitely prepared to invest as much as they can in order to ensuring opening ey get their first and of course they reap the benefits and so do you have a I hear

you vision in 2030 of what if I say you crushed it Sam it's 2030 you crushed it what does the world look like to you um you know maybe in some very important ways not that different uh like we will be back here there will be like a new set of students we'll be talking about how startups are really important technology is really cool we'll have this new great tool in the world it'll feel it would feel amazing if we got to teleport forward six years today and have this thing that was like smarter than humans in many subjects and could do these complicated tasks for us and um you know like we could have these like complicated program written or This research done or this business started uh and yet like the Sun keeps Rising the like people keep having their human dramas life goes on so sort of like super different in some sense that we now have like abundant intelligence at our fingertips and then in some other sense like not different at all and you mentioned artificial general intellig AI I think Sam mman might be downplaying this one here or it's probably realistic because he's always often said that like when you get to AGI it's probably going to happen and then everyone's going to go about their daily lives and I think what he's stating here that like even if they do crush it by 2030 and they achieve AGI by 2030 or in 2030 I don't think Life Changes instantly and I think that is probably kind of true I know that people might think boom we're just snap your fingers and we're in some Singularity type world with flying cars you know these tall white buildings that are completely smooth over and there's like these insane structures I think even if that is our future I think that will take some time so I do think that by the time AI is achieved in 2030 which many do predict it might be 2028 or 2029 I think that it's still going to take some time to build out the infrastructure for future societies the new social contracts and how things are going to work after that so I mean I guess yes this does make sense but um it's something that's definitely very hard to predict I think we need a more precise definition of AGI for the timing question um because at this point even

with like the definition you just gave which is a reasonable one there's that's your I'm I'm paring back what you um said in an interview well that's good cuz I'm going to criticize myself um it's it's too loose of a definition there's too much room for misinterpretation in there um to I think be really useful or get at what people really want like I kind of think what people want to know when they say like what's the timeline to AGI is like when is the world going to be super different when is the rate of change going to get super high when is the way the economy Works going to be really different like when does my life change and that for a bunch of reasons may be very different than we think like can totally imagine a world where we build PhD level intelligence in any area and you know we can make researchers way more productive maybe we can even do some autonomous research and in some sense like that sounds like it should change the world a lot and I can imagine that we do that and then we can detect no change in global GDP growth for like years afterwards something like that um which is very strange to think about and it was not my original intuition of how this was all going to go so I don't know how to give a precise timeline of when we get to the Milestone people care about but when we get to systems that are way more capable than we have right now one year and every year after and that I think is the important point so I've given up on trying to give the AGI timeline but I think every year for the next many we have dramatically more capable systems every year so I think there were two things that were really important in that clip and I know that was a little bit longer than normal but I think that one was pretty crazy because he literally said that every year from this year there will be increasingly capable models and that is something that I think most people can't really comprehend including myself because humans don't have a great time at comprehending exponential Improvement because it just doesn't fit into how we naturally see things growing so it's something that's going to be pretty hard to comprehend but I'm guessing maybe as time increases we may have a better understanding of how that looks in our day-to-day lives because if you were to show GPT 4 to someone 2 to three years ago you they might like be like what on Earth is this magic and this is of course a recurring problem with AGI is that many people have always said that there's not a definitive state of where AGI is and I agree I agree with this completely because AGI is not a thing it is a scale of intelligence that is increasingly better and I think the lines are going to keep getting blurred because right now as many would argue gp4 is a lowlevel AGI system but in the future when we have Vision systems audio systems uh video model systems I mean embodied AGI as well I think we're going to increasingly get further down that scale where we get useful AGI and samman's personal definition was when a system can do autonomous AI research and make scientific breakthroughs so that's samman's definition and I think that definition is important because it gives you guys a benchmark on where he thinks that a true level has been reached fundamentally where breakthroughs are I don't know yet um I mean this sounds like a copout answer but I think the most important thing about GPT 5 or whatever call that is just that it's

going to be smarter and this sounds like a Dodge but I think that's like among the most remarkable facts in human history that we can just do something and we can say right now with a high degree of scientific certainty GPT 5 is going to be smarter than a lot smarter than GPT 4 GPT 6 is going to be a lot smarter than GPT 5 and we are not near the top of this curve and we kind of know what to do and this is not like it's going to get better in one area this is not like we're going to you know it's not that it's always going to get better at this eval or this subject or this modality it's just going to be smarter in the general sense and I think the gravity of that statement is still like underrated so the reason I've included this clip and I think if you're going to take one thing from this video take this from the video okay and because I think most people don't understand what Sam mman means and like he said most people think that this is a copout answer this is not a copout answer at all I think him stating that the model is going to be smarter shows us that potentially there might even be some new architectures involved but I think the main thing that we can take away from this is not that the model is just going to be smarter cuz that is kind of vague but that is of course a definition but when I break it down for you basically this means that this model is just going to be so much better at reasoning understanding comprehension context awareness all of these things are truly important to basically being useful like so many times we ask gp4 to do complex tasks and it just really can't and I think in the future we're going to see how smart A system can get in terms of actually being able to respond to exactly what we want and I think this is what Sam mman is talking about when he says that systems are going to be smarter because GPT 4 has very limited reasoning abilities although it does do some reasoning tasks very well I think future tasks like planning agentic Behavior you know trying to model the world theory of mind all of these things are increasingly intellectual tasks and for an AI system like gbt 4 those are areas where it is pretty rudimentary so I think those areas in the future mean that this kind of system is going to be more widely used because it's going to allow for essentially more wild skill deployment if there's less ulation a lot more planning and a lot more complex reasoning it's definitely going to be used a lot more because it means that this system is just truly going to be smarter and this isn't something where you know you're looking at the benchmarks and trying to see that but definitely we could be reaching an area where these systems really do you know max out the benchmarks in terms of what's possible and I think that Sam mman stating that this is essentially going to be where we're at in the future I think this actually gets me really excited for what this Future model is going to be because I think that open AI are on a completely different path to many of the other AI companies that we look at on a day-to-day basis because they're still playing catch up and opening I are playing let's set the bar higher so having a model that's it smarter and them guessing that you know they already know where gbt 6 is going to be is of course something that you should you know watch out for so I would say that if you're that's building a company um definitely plan with in mind that gbt 5 is actually going to be smarter which means it's going to be applicable for a lot more things and maybe there are certain intellectual tasks that might disappear with this and I definitely think they will because if this model gets even smarter than it is now it's definitely going to be doing a lot more cognitive tasks which means jobs uh cognitive intense jobs are definitely going to be on the line so that is something that you do need to prepare for because I think it's something that you know it wasn't really talked about in this interview but you know as some who's looking at things across the Spectrum it's just a heads up for me increasingly worried about how we're going to do this all responsibly I

think as the models get more capable we have a higher and higher bar we do a lot of things like uh red teaming and external Audits and I think those are all really good but I think as the models get more capable we'll have to deploy even more iteratively have an even tighter feedback loop on looking at how they're used and where they work and where they don't work and this world that we used to do where we can release a major model update every couple of years we probably have to find ways to like increase the granularity on that and deploy more iteratively than we have in the past and it's not super obvious to us yet how to do that but I think that'll be key to responsible deployment and also the way we kind of have all of the stakeholders negotiate what the rules of AI need to be uh that's going to get more complex over time too so here we have the last clip where s actually talks about and that's just not where I want to stop the video but he talks about you know responsible AG I think this is probably the hardest thing to do because they might actually have to have some kind of AGI system where it's mathematically impossible for it to be jailbroken or go wrong I know that sounds like sci-fi and whatever but I don't think there's any way currently you know whilst looking at the future any current way to safely uh deploy an AGI but like they said they're probably going to have to do it granularly which means narrowly like some narrow versions of AGI which are you know maybe only working for Finance restricted to finance related tasks or one that are only related to marketing related tasks one that aren't really General because I think you know as models have been jailbroken before if an AGI system is jailbroken again if it's that smart considering if we extrapolate out to like gbt 8 then it could definitely be used for some pretty heinous stuff so I think that is something that is uh it's it's a problem that I'm glad I'm not working on because it would definitely keep me up at night but I think that this interview was really insightful and I think we can all look forward to the GPT 5 release in potentially 2 months or less so let me know what you thought about this you know project Stargate what you thought about the compute poor and compute Rich GPT 4 being completely dumb which is pretty crazy uh no more surprises of course the winner takes all scenario AGI by 2030 the actual AGI timeline and gbt 5 and gbt 6 getting increasingly smarter