OpenAI Researchers Prove AGI Is Closer Than We Think

so here we have a very interesting blog post by someone who works at openai as a research engineer since 20122 this is called general intelligence by James beter and it is a fascinating read on the next few years for the timeline of AGI and artificial intelligence now I think this is really important because after you view this video and after you view some of the things that are stated in this video you're going to realize that there are two main takeaways from this number one things are probably going to change more rapidly than you think and number two that there is a certain year SL dat that consistently gets mentioned when people are talking about AGI so without further Ado let's dive into this blog post because it's truly intriguing on some of the things that it talks about so he speaks about how folks in the field like to make predictions for AGI and I have thoughts and I've always wanted to write them down so let's do that he said since this isn't something I've touched on in the past I'll start by doing my best to Define by what I mean by general intelligence a generally intelligent entity is one that achieves a special synthesis of three things so we got three things here number one being a way of interacting with and observing a complex environment typically this means embodiment the ability to perceive and interact with the natural world this is of course something that current AI systems are not really there yet with but of course basically this is what humans have a true understanding of we can perceive and interact with the natural world with all of our sensory inputs like touch and smell and sight and all of these crazy different things that we have a robust World model covering the environment this is the mechanism which allows the entity to perform a quick inference with reasonable accuracy World models in humans are generally refer to as intuition fast thinking or even system one thinking and system one thinking is basically where you just intuitively know exactly what to do for example recognizing someone's face recognizing a familiar face in a crowd without you know conscious effort you know like when you're driving on an empty road stuff like that and for example if you heard a loud bang you would immediately look towards that bang and you'd immediately think oh my God what on Earth was that you know these are just like your initi intuitive reactions and these things are you know genuinely there in humans and even some animals now he also talks about a mechanism for performing deep introspection on arbitrary topics this is the thought of in many different ways it is reasoning or slow thinking or system two thinking so system two thinking this is where you know you have a hard math problem or for example there's a deviation in your environment you have to kind of think about how you can solve that problem so for example you know let's say you are going to work and today the bridge is closed that you usually cross you have to think about you know a new way to get to work what road to take is it safe is it fast can you get there on time those are the kinds of things where you engage in your system to thinking where it's not just like an intuitive reaction you have to actually think okay what is going on here we need to think about this you know and it's kind of like a slow level of thinking because it just requires a lot more thought and then basically he says that with all of these three things so we've got key like three important things here if you have these three things you can build a generally intelligent agent now it's important to note here that what he doesn't say is that this is Agi but what he does a generally okay generally intelligent agent okay and it's important to make that distinction because and this is a little different from AGI and he actually talks about this later on in the video I'll explain but it's basically different because AGI is kind of like a scale so we're not just going to have you know Ai and then boom we get AGI it's more going to be like okay we can have something that's generally intelligent and it's an agent that can do a decent amount of things and then we get to like another iteration that's going to be you know a step up from that so it's not just going to be Leaps and Bounds every single time as much people think it's going to be like you know one giant staircase like boom and then boom it's more going to be like you know a gradual increase of the capabilities of whatever system they're building so this is where he talks about how using these three key components the you know the perceiving and interacting with the natural world um a world model covering the environment you know where you're able to truly understand what goes on immediately and of course a mechanism for performing deep thinking on different topics which is you know

referred to as system to thinking so he says if you can build these three things you can have a generally intelligent agent and here's how so first you need to you know seed your agent with one or more objectives so have the agent use system to thinking which is your deep level of thinking in conjunction with its World model to start ideating ways in order to optimize for its objectives so if your AI system or whatever you know agent that you did have its goal was to you know uh steal a car let's use that cuz it's more engaging you would have to think okay how could I start to steal this car without getting CAU do I want to do at night time make sure it's in a dark alley yada y yada and then you think about the best plan and of course it picks the best idea and builds a plan and then it uses this plan to take an action on the world and of course it might you know think okay night time is the best so I'm going to wait until night and then I'm going to take an action to scout out the location and then of course it obser observes the result of this action and Compares that result with the expectation that it had based on its World model so the world model is basically its understanding of how the world works and of course and this is like true for humans as well as you take your actions in the world you're going to update your understanding and of course your knowledge of the world so it might update its World model here with the new knowledge gained and it uses system to thinking to make alterations to the plan and then you rinse and repeat this is true for like a lot of things that people do anyways like you know even for example do doing YouTube you might think okay you you record videos in a certain fashion then you upload them and then they don't do well so you update your world model and then of course with the new knowledge gained you become better so this is what people do to uh you know using system to thinking for a lot of different things and this is like the basic definition rinse and repeat of how these you know generally intelligent agents should work provided that they have these three things right here so he says my definition for general intelligence is an agent that can coherently execute the above cycle repeatedly over long periods of time thereby being able to optimize any given objective so this is his definition of general intelligence and I think that is really important because you know we're seeing a lot of different definitions one of the key things that's going on in AI is that no one can really agree on what the definition of you know AGI is so it's important to see what a research engineer from open aai his perspective is on exactly what's going on so it's kind of interesting to see exactly how that kind of works and then um what we do have here is of course we have you know his further explanation so he says that the capacity to actually achieve arbitrary objectives is not a requirement some objectives are simply too hard and adaptability and coherence are the key that can the agent use what it knows to synthesize a plan and continuously act towards a single objective over long time periods and this is of course going to be a truly capable system at least I think it is because as much as we like to think that humans are you know truly like these crazy creatures and that we're so special and so different from robots like sometimes you can think about how people act and they truly don't like literally update their world model and they truly don't you know use system to thinking to even make changes to their plan like if you about some people who and of course there are like a bunch of different you know things that you could talk about regarding this but I'm just talking surface level like for example people who still do the wrong thing and for whatever reason they get the same result and they don't update their plan like they don't make alterations at all like if you were someone who was you know trying to I guess you could say um you know lose weight or you know learn a new skill you know if you try one way um of course if that way doesn't work you have to then you know make the alterations and then of course update the plan and then rinse and repeat and you'd be surprised at like if you just and I know this might be like a weird concept but if you like look at how AI learns and how AI is being more effective and if you actually like kind of apply these to your life like it really does help in terms of being able to achieve certain goals like if you literally just make alterations to your plan after trying it if it doesn't work you will eventually get there but of course this is something that is much easier said than done when navigating the world and an AI agent that's able to do this remarkably effectively I think it's going to be a little bit more powerful than the average kind of human working around so this is the section where we actually talk about world models so he says we're already building World models with auto regressive Transformers same architecture that we've been using recently and then of course particularly of the omn model variety how robust they are is Up For Debate of course you've got hallucinations all these kinds of different things and of course recently the GPT 40 the Omni model which is you know a new edition and he actually did

another blog post about how crazy that is so he then States there's good news though in my experience scale improves robustness okay so this is where like comput and those kind of things kind of improve the not capability but like the overall improve robustness of the model and humanity is currently pouring Capital into scaling autor regressive model so essentially what we do have is we do have a situation on our hands that was kind of you know sparked by open AI the GPT Series where now people are pouring billions and billions of dollars Microsoft of course is pouring a lot of capital into project Stargate in conjunction with open AI an100 billion supercomput Sam Alman seeking $7 trillion in capital of course that is a clickbait headline but um you know over the course of the next 10 to 20 years it might not be that crazy and this is where the large majority of investment is going because a lot of that investment is based on the current transf former autoaggressive architecture so what we have here is that as long as the scale is increasing up to a certain extent we don't know if we're just like we don't know where we are on the sigmoid curve is basically just the kind of growth area so we don't know where we are like we could be you know here and things are about to just taper off with GPT 5 coming into the mix we could be like somewhere down here and things are about to get really crazy so that's why this entire thing um of course if we've seen you know some early results from pouring more Capital into scale this is going to be something that we continue to do so of course over the next you know few years we can expect robustness to improve but remember this is literally just based on the scale okay this is and that's not just based on the efficiency of course of the many different other things that go into AI systems such as you know the algorithms software all those kind of you know ways that you know you can make those systems even better especially other things on top of llms as well so with that said I suspect that the world models that we have right now are sufficient to build a generally intelligent agent so you can see here he says that with the world models that we have right now okay right now that's kind of like in italic are sufficient to build a generally intelligent aident and he says that I suspect and this is a pretty crazy statement like I don't know if everyone would agree with this but um you know we don't really have a generally intelligent agent just yet so um but he's talking about the world models that we do have because I would argue that you know the systems that we do have a decent understanding of the world they just don't have the embodiment right now at least the effective embodiment that we really do need so he says I also suspect that robustness can be further improved via the interaction of system to thinking and observing the real world this is the Paradigm we haven't really seen in AI yet of course but still happens all the time for living things and it's a very important mechanism for improving robustness so you know the real world um you know observing the real world is something that you know we haven't really seen in AI yet of course this is going to happen in the next future it's going to be a lot harder to do than just making AI systems better because robotics is really hard like a lot harder than you know traditional just working you know on something that's you know software based because robotics is you know it's Hardware based so there are basically physical you know limitations like you have to you know look at the laws of physics and when you're testing things it's a lot harder to get you know feedback and stuff like that um this just it's just harder basically so um this is why you know we haven't really seen that yet and slowly and surely of course we are going to get there now one of the things he does talk about is of course the Skeptics like Yan Lan so he says while llm say we haven't yet achieved the intelligence of a cat this is the point that they are missing yes llm still lack some basic knowledge that every cat has but they could learn that knowledge given the ability to self-improve in this way and such self-improvement is doable with Transformers and the right ingredients okay so what he's basically saying here if you haven't you know seen the clip of yan Lun Yan Len is you know someone who's very respected in the AI Community for his contributions to the field and essentially the reason that Yan Lan gets mentioned so much and I even did a recent video where I spoke about lot of his ideas and what they mean for the space the brain of a house cat is uh is about 800 million neurons you have to multiply this by about 2,000 to get the number of uh synapses the connections between neurons which is the equivalent of number of parameters in an llm the biggest llms that we have at the moment that are practical are have a few hundred billion uh parameters the equivalent of synapsis um so we we're maybe at the size of a cat but why is it that those systems are

not nearly as smart as a cat you know a cat can do uh can remember first of all understands the physical world can plan complex actions um can do some level of reasoning actually much better than the biggest llms and uh so what that tells you is that we're missing something really conceptually something really big but to summarize that 30 minute video basically yanan is stating that you know llms are autoaggressive and that kind of architecture just doesn't work with you know humans and how you know if you're trying to get to AGI it's just not going to work basically so um he's basically saying that you know uh the current llm systems they just haven't yet achieved the intelligence of a cat but he's arguing here that you know llms you know they could learn that knowledge and given the ability to self-improve in that way um it's doable with Transformers and the right ingredients so of course this is a bold claim because I'm not going to say that this guy has some secret information open AI but I think that based on you know the current information okay it being doable I would love to see that because that would be a huge step up in terms of the capabilities so I think the future is definitely going to be interesting in this part because I think some theories are going to be disproved or they're going to be proven right which means either you know we're we're like an off-ramp to AGI and right now we're going on the offramp and this is just a huge tangent which was created by open Ai and we're going down the wrong you know I guess you a architecture we're pouring billions of dollars into the wrong thing and eventually we find the new architecture or we're barreling down the right way and we're about to get some very interesting stuff because either way I think this is going to be interesting so this is where he talks about you know the reasoning and he says there is not a well-known way to achieve system 2 thinking this is where you know systems have you know a long thought process but I'm quite confident that it is possible within the Transformer Paradigm with the technology and compute we have available to us right now basically that we can achieve system to thinking the long-term thinking that AI systems need in order to achieve goals that are quite effective in the actual world I've seen some systems do that like I've seen a few demos here and there like a few agents being able to plan you know things like Devon and such this kind of system to thinking is you know it's there but it's not there to the point where it's remarkably effective so basically you know within two to three years we're going to be able to build a mechanism which is you know sufficiently good enough for the cycle described above so this is one of the first things so you can see that this is 2 to three years away now 2 to 3 years this is kind of important because 2 to 3 years is 2026 2027 and that date also lines up with Leopold Ashen Brenner's date of AGI in 2027 and I think you guys really need to like understand that system to thinking is incredible because what we've seen like from llms is that a lot of the times when we give an llm the ability to think with whatever you know kind of um you know architecture that we kind of implement you know not architecture but whatever kind of prompting strategy just on the base level that we use whether it be you know hey think step by step or you know Monte Carlo Tre search or whatever you know Chain of Thought prompting whatever kind of way that we um think like that allows the AI to truly improve the results um and that shows us that you know if we can get a very effective system to thinking and I think a lot of people are working on this then we can get something that's good enough for the cycle above which is an intellig in generally intelligent agent and that's truly going to change everything because it's going to improve the reasoning and that's going to improve you know how effective it is even in its embodiment and just the overall you know accuracy and robustness of the model so I think it's important to know that like this system to thinking area um is pretty crazy because I can't remember which paper was I was reading but they basically said that look if someone asked you to you know a math question and you had to give it an them an answer immediately without thinking like within 0. 1 seconds your answer would probably be wrong and that's essentially what we're doing to llms when we ask them a question and they respond immediately but when we you know give them the time to think and they are able to you know debate on you know what it is that they're able to do they're able to reason over a longer period of time and this is something that Sam Alman said in an interview a long time ago not you know 8 months in the AI industri is a long time ago he said that it's something that he is working on um the responses do get a lot better so I think this is going to be exacerbated in future models like they're really going to you know hammer down on that um as well as scale and I think that's going to drive a lot more improvements than you know a lot more people think and um here's where he comes to embodiment so the embodiment is of course something we're still figuring out with AI this is of course you know stuff like figure you know the humanoid robots the Tesla Optimus um and of course he says it's

once again something I'm quite optimistic about near term in the near-term advancements there is a convergence currently happening between the field of Robotics and llms that is hard to ignore of course the recent figure demo where they combined the knowledge of GPT 4 or whatever AI system it was with the you know fluidity of the new figure robots providing us with an very impressive demo that showcased what the future is about to you know become he says robots are becoming extremely capable able to respond to very abstract commands like move forward get up kick the ball reach for object for example see what figure is up to or the recently released unitary H1 which is um I guess an AI agent Avatar which is you know going to be doing a lot of interesting stuff in the future because it's interestingly enough it kind of looks like Boston Dynamics robot but I know that Boston Dynamics have been working on that for quite some time so I don't know how unry managed to produce that robot that quickly like it's truly incredible if they did just you know look at boss and Dynamics for inspiration but it was really quickly that they managed to get that done so you can see here on the opposite end of the spectrum large omnimodels give us a way to map arbitary centor inputs into commands which can be then sent to these sophisticated robotic systems and that's of course you know these models that you know have so many different inputs and outputs that allow us to you know use them in a way that we haven't really pered he says I've been spending a lot of time lately walking around outside talking to gp40 while letting it observe the world through my smartphone camera I like asking it questions to test its knowledge of the physical world and it's far from perfect but it is surprisingly capable we're close to being able to deploy system systems which can commit coherent strings of action on the environment and observe and understand the result I suspect we're going to see some really impressive proc in the next 1 to two years here he says this is the field of AI I'm personally most excited in and I plan to spend most of my time working on this over the coming year so of course some impressive progress is going to be happening in the next 1 to two years you know coherent strings of actions and he's talking about GPT 40 of course if you don't know GPT 40 has that update where you can you know talk to it through a camera and that stuff so I'm guessing maybe he just has access to the one that open AI haven't released yet but you know he does a summary here and I think this summary is really cool he says we've basically solved building World models we have two to three years on system two thinking and I think one to two years on embodiment the latter two can be done concurrently once all of the ingredients have been built we need to integrate them together and build the cycling algorithm I described above and I'd give that another 1 to two years so my current estimate for AGI is 3 to 5 years I'm leaning to three for something that looks an awful lot like a generally intelligent embodied agent which I would personally call an AGI then a few years to get it to the point that we convince Gary Marx of the world so basically he's stating that you know a generally intelligent agent I think I was wrong at the beginning where I said that this is not AGI but I guess that this would be AGI but of course on the scale you know it might not be on the scale that you know some people um of course would accept it and that's why I've added here not I've added but he's added you know the Gary Marcus and that's why I said it might not be AGI to some people um but of course that could be AGI and then of course you know after you get the generally intelligent agent that's able to do those kinds of things um you're going to have to refine it for some other years and basically Gary Marcus I wouldn't say he's an AI skeptic but he is someone that you know criticizes um AI quite a lot so he can be seen as someone that is like you know just a skeptic and someone that just you know kind of criticizes a lot of the advancement so it's going to be interesting to see because he does make a lot of good points but like I said before it will be interesting to see if some of the points he makes about the future and I'm talking about Gary Marcus here if they are proven wrong because um we're at that point where it's either going to be exponential Improvement or maybe we were wrong about this so um yeah one to two years for robotics so to solve that because I think robotics has done really well but I do think that you know the main thing that we need to think about of course is system to thinking um and of course the how the world models are going to be you know interacting with them so he says we've basically solved that and of course two to three years on system two thinking and one to two years on embodiment so I am kind of intrigued about this you know prediction on why he thinks that you know two to three years on system two thinking will take a lot longer than embodiment when you know traditionally you know Mor morex Paradox suggests that like you know robotics is going to take a lot longer than you know software but um of course he's the research engineer and I'm the person making the video so um I'm guessing that when you actually think about it you know planning things in the real world is actually really hard and when we do look at a lot of the agent systems that we do have right now they aren't that good at system to

thinking which is the ability to plan um in a longterm Horizon in a way that's very effective I remember I was looking at another paper on GPT 4 on deception and it wasn't able to effectively plan when there were multiple steps involved when you had one layer it was fine but as soon as you got to two layers uh the accuracy just dropped to like 10 to 15% but I mean I guess we're going to have to see with GPT 5 GPT 6 if there are any like special models that open ey build just based on that and they integrate them in because we know they previously did you know this mixture of experts and that's how they got to like GPT 4 that like how they got it to you know be so good but um yeah overall you can see 3 to 5 years for AGI and I'm leaning for towards something three that looks like an awful lot like an embodied agent which I would personally call an AGI so we got three years to something that's pretty much AGI which would put us at 2027 like I said that date i' I've heard that date so many times now so I would say the 20 7 might be like the first demo for AGI and I think the thing is as well is that like open AI isn't going to be the only one leading the charge here remember like a lot of companies now like the light bulb moment has hit them so they're going to be pouring you know millions and millions of dollars in companies nations are going to be pouring billions and billions of dollars in so there's going to be a lot of money flowing into this industry because it is definitely a race because um the PO of gold at the end is so big that they are definitely willing to do that so 3 to 5 years for in 3 years we'll get the first embodied agent um and of course 2 to 3 years for system thinking and 1 to two years on embodiment because we are pretty far ahead for robotics because if you actually think about like where bosson Dynamics Atlas is like if we just took like every single company on Earth and just put them in a list if you've seen how effective that robot is at moving I mean combine that you know with a world model and system to thinking that is insane so um yeah I I think this article was really you know insightful on the future and I think it goes to show like some of the ideas that are floating around now about you know general intelligence and where we're headed I think they're kind of converging and overlapping which is a good sign because you know a lot of times what we have in AI is like a lot of different you know contrasting ideas but overall I think 20 27 to 2030 I think those three-year period that threeyear period provided there's no National tragedy I think it's going to be super interesting to be in this space and actually paying attention so with that being said if you did enjoy the video I'll leave a link to this down below let me know what you think about this and if you enjoyed the video don't forget to check out the school um and I'll see you guys in the next one