# GPT-5 Delays, Superintelligence, Humanoid Robotics and GPT-4 Is Not As Smart As You think

## Метаданные

- **Канал:** TheAIGRID
- **YouTube:** https://www.youtube.com/watch?v=9IzAp4acajo
- **Дата:** 06.06.2024
- **Длительность:** 30:28
- **Просмотры:** 25,715

## Описание

Join My Private Community - https://www.patreon.com/TheAIGRID
🐤 Follow Me on Twitter https://twitter.com/TheAiGrid
🌐 Checkout My website - https://theaigrid.com/


Links From Todays Video:

00:07 GPT-4'S Performance
04:18 OpenAI + GPT5 Delays
08:26 OpenAIs New Voice Mode
11:21 Deepfakes
14:34 Text To Video AI
16:50 Humanoid Robotics
23:06 Superintelligence
28:11 Gemini Coding


Welcome to my channel where i bring you the latest breakthroughs in AI. From deep learning to robotics, i cover it all. My videos offer valuable insights and perspectives that will expand your knowledge and understanding of this rapidly evolving field. Be sure to subscribe and stay updated on my latest videos.

Was there anything i missed?

(For Business Enquiries)  contact@theaigrid.com

#LLM #Largelanguagemodel #chatgpt
#AI
#ArtificialIntelligence
#MachineLearning
#DeepLearning
#NeuralNetworks
#Robotics
#DataScience

## Содержание

### [0:07](https://www.youtube.com/watch?v=9IzAp4acajo&t=7s) GPT-4'S Performance

lot to cover so let's not waste any time one of the things that was actually reported a lot when GPT 4 was initially released was the fact that it passed the Bor exam and not only it passed the bar exam it passed the exam in the 90th percentile so this was a report from Stanford of 2023 well they were basically stating what does this mean right now for AI tool in the legal profession because it's clear that you know it not only passed but in the 90th percentile this is pretty crazy and this is going to have a lot of impacts but why am I talking about something from 2023 well apparently this result on GPT 4's performance wasn't exactly accurate you see there was a recent paper where they decided to reevaluate the claims of gbt 4's bar exam performance so they state that perhaps the most widely touted thing of gbt fors launch at zero shot capabilities has been its reported 90th percentile performance on the uniform bar exam this paper Begins by investigating the methological challenges in documenting and verifying the 90th percentile claim presenting four sets of findings that and here we go indicates that opening eyes estimates of GPT 4S percentile are in fact apparently overinflated okay so basically they're stating that they done some research that I will dive into in a moment and they state that the results are quite overinflated so they state that although GPT 4's score uh the ubbe scores near the 90th percentile when examining approximate conversations from fedy administrations for the Illinois bar exam these are heavily skewed towards repeat test takers who failed the July Administration and score significantly lower than a general test taking population so it basically says here that second data from a recent July administration of the same exam suggests that GPT 4's overall Ube percentile was below the 69th percentile and 49th percentile on essays and I think this is rather important because whilst yes this is still very impressive if the results are actually overinflated as this paper suggests it means that maybe this can I guess you could say realign a perspective on where AI systems are probably going to be because whilst yes this is just not that bad I think it's important because this was one of the things that a lot of news articles and a lot of AI researchers were looking at when they said look this system is very smart and this is something that is incredibly gamechanging whilst yes it is gamechanging I think we have to understand that sometimes there can be a little bit of I guess you could say overin inflation or a little bit of hype that leads us to believe these systems are a bit more intelligent that they are and the reason that uh this result was really important is because they're basically stating that look if we overestimate GPT 4's capabilities this might lead to misuse in legal tasks which result in poor legal outcomes and it actually did people that weren't using the specialized version of GPT 4 actually were subject to I guess you could say hallucinations by the model when it repeated cases that just simply hadn't existed so this is uh something that you know shows us that there is an importance to you know at least verify and check whatever comes out of these Labs because whilst yes it's important to you know look at you know what opening eyes says It's always important to do your independent evaluations because yes of course these companies are always going to tout that the model is good at something so I think that this was important because not only does it you know show us the true trajectory of where we are I think it's important because we can now look at where gbt 4 is in certain areas and be like okay the capabilities aren't in the 90th percentile they are actually in the 69th percentile so I think this is something that is uh you know it's important because we can now be like okay this is how the graph is going uh and this is where things are going to head so one of

### [4:18](https://www.youtube.com/watch?v=9IzAp4acajo&t=258s) OpenAI + GPT5 Delays

the articles that was also rather interesting was an article by the information and there was something that a lot of people didn't really see but I thought it was really cool so essentially they spoke about how long opening ey's First moover Advantage can last uh and I think this is really important because they spoke about how you know a lot of uh other companies are actually starting to pay for other models and basically what they're showing us is that open aai slowly since January 2024 or since the release of Claude they've kind of been losing their lead and this is a little bit interesting because it shows that the race between these giant companies Google anthropic and eye and of course meta with llama 3 show us that you know it's pretty crazy um in terms of how much of lead the anthropic has but you can see right here that the share of cruise consultings client startups paying for llms from respective providers so of course openai is at 67% but what they're showing us is that anthropic is rapidly gaining market share under a new you know their new model which is of course Claude 3 now this is important because it shows us that other companies are also managing to take the market share from openi which means that potentially openi might want to step things up if they want to keep their market dominance now competition I believe is of course good but I think it's really important for of course the customer so that the customer can get the best product on whatever it is now another thing that they said in this article and that's the only reason I included this article because otherwise it would be a little bit boring but was that this article actually gave us some relevant information regarding the GPT 5 dat so it said and of course there's still much AED GPC 5 which we're hearing could come out in December and could blow today's models out of the water only time will tell now the reason I'm taking the information's information into consideration when discussing the future of models releases is because they've been remarkably accurate on what's going on in the AI Community behind the scenes like they reported on Sam's firing uh and a few of other things like before a lot of the uh I guess you could say big uh industry media Giants managed to get to it so essentially what we have here is we have them stating that the GPT 5 release could be apparently you know come out in December now this is a little bit surprising because the general elections are on November the 5th uh and I was thinking that they would be in November because according to a recent conference open uh basically said they didn't really say it but on a graph the date that was plotted to when the next sample the next AI model was scheduled to come out it wasn't December it was November so I'm guessing that the November December time is remarkably different from what was going on in the summertime so I'm guessing that potentially what we will have in the summer you know now that we know that this is going to be coming out in December after the general elections okay or potentially late November it means that it's looking like a gbt 40 summer so potentially the new voice model which open and I have spoke about quite a lot it seems like that's what's going to be their summer play now if you don't know why I'm basically saying that's their summer play it's basically because openi realized that in the summer months the traffic for openi dipped quite a lot because there were people that weren't using this in academics anymore so their traffic went down quite a lot but this year they wanted to launch something during the summer so that people would continue to use the software so it seems that we got an updated timeline for gbt 5 so apparently this is going to come out in December so you can update your timelines cuz there are some people saying that it's literally going to come out tomorrow which is the Thursday and that would be rather surprising so I mean depends but I would say the information's information is really solid so don't take it with a grain of salt so now in regards to GPT

### [8:26](https://www.youtube.com/watch?v=9IzAp4acajo&t=506s) OpenAIs New Voice Mode

40 they also released a new video in which they showed the dynamic range of character voices with the new AI system basically what they showcase is how an AI system can actually realistically sound very humanlike so I'm going to show you guys exactly what they're talking about hey um I'm writing a story I'm going to have you practice a couple voices with me for like you know different characters so for the first one I'm thinking we're going to have like a majestic lion he's kind of like an old King and it's done um and I want you to see something like who goes there uh be the lion I really want you to like embody what that might feel like who goes there that's pretty good um now I want to add a second character maybe a mouse that snuck into the cave uh the mouse could see something like oh it's no one how's that for a little that's not bad could you make it a little squeakier more like tiny more like Mouse like it smells how's that that's pretty good okay uh let's toss in an owl uh make the owl sound like wise a little bit stoic um and the owl could kind of be like an adviser to the lion or something who Dar enter the king's Den okay yeah that's pretty good um now let's think about what the villain might be I don't know what animal would work best but let's start with like some kind of laughter give me like a nice evil maniacal laugh not bad too squeaky make it more like so I think the reason this is so interesting like this entire demo was so fascinating to me was because like I said before the dynamic range of what this AI system can do in terms of its voice is really surprising so I think the fact that this AI system can do laughter it can do sound effects it can do like a robot voice which we saw in the previous demo these are capabilities that I've long wondered why a lot of the other you know voice to AI voice uh companies why they weren't doing them because it was something that if they could do it I just knew it would truly change the game so I was truly surprised that companies like 11lbs where you had that hyper realistic sounding voice they didn't manage to do this and another feature that most people didn't pay attention to was the fact that you can quickly and easily interrupt the AI when you're talking so um this shows us that GPT 40 I think that might tomorrow as in Thursday so that might be what people are waiting for but it will be uh interesting to see what is released tomorrow cuz there is a lot of I guess you could say uh information floating on about what is going to be released and there's a Gemini event SL update scheduled for today being the Wednesday so I'm wondering if you know they're going to try once again to outdo Google there was

### [11:21](https://www.youtube.com/watch?v=9IzAp4acajo&t=681s) Deepfakes

a new clip where Reed Hoffman was stating that deep fake videos which are indistinguishable from the real thing are only months away and this is a great statement because I'm going to show you guys something that was recently released that shows us uh just how like crazy things are about to get there is definitely technological solution that creates a kind of like what is the real Providence of like Reed is really here talking with squawkbox and there's a certificate that that asserts that goes okay we trust that certificate right I don't think that the ongoing Battle of I have a video technology that can tell you looking at if it's deep fake or not that's an uncertain ground how far are we away from making this feel really real so Becky when you sat down Becky was saying maybe that it didn't sound exactly like play at the te's sounded a little flat it didn't have the same inflections that you have I will say the first answer that I saw sounded closer to you yes um we're months away from just months yes it because it improves right so the lip sying gets a little better the gesturing ability to go sof or loud or aggressive all of that that's just improving at the pace that software is improving so yeah he's uh literally talking about GPT 40 and I think this is really important like one of the things that really prevents people from falling for AI generated scams is because when you're talking to an AI system like on the phone or something like if it's a rooc caller the problem is that the voice range is non Dynamic at all it doesn't you don't doesn't go like low like this and then it doesn't go high like this so you know that you're talking to a robot system when it literally just has a monotone tone like this and it's very easy to understand but what happens when that changes when the base Norm okay goes from A system that you know is monotone and simple the base Norm goes to a system that is constantly Dynamic constantly changing it speeds up it slows down it laughs it's able to you know converse with you it's able to understand jokes and references about your culture that only a few people would and it's able to personalize uh you know thoughts and ideology to be able to scam you a lot more effectively I think that is uh going to be a real problem for many individuals because many of us okay you know you might be young or knowledgeable in the field of technology and Ai and you might think that I have no idea how my grandma would get scammed by this Ai call but I think that uh the same thing might happen to us uh when these Technologies evolve even further and this isn't just something that is like just a piece of news this is uh like you know if you haven't seen this before this is Microsoft Versa one lifelike audio driven talking faces generated in real time um and we've seen a lot of different real time AI you know systems go crazy so this is not something that I think is uh going to be you know I guess you could say not happening at all but it's just like if even if I don't have the audio you can see that these are literally just pictures and how many pictures do you post online of yourself do you have a picture somewhere that someone could use maybe an old Facebook photo I mean the possibilities and remember guys the one thing that you have to consistently remember is that this is of course the worst it will ever

### [14:34](https://www.youtube.com/watch?v=9IzAp4acajo&t=874s) Text To Video AI

be then we have the startup Hig field introducing Nova one our creative video model this wouldn't be without the support of nebis official so they were able to train the model and this was kind of fascinating so this is Hicks field this is their Nova 1 model and I'm pretty excited about another video generation company managing to get into the space because I think companies like pabs are really good like they actually push forward and they actually publish things and they're pretty active whereas companies like Runway I don't know if they're working on anything really big at the moment they did state that they were going to be working on something that was a lot better than Sora but this Nova one model looks a little bit decent like it looks okay it doesn't look completely awful but it seems like there is a lot of movement in these uh you know different clips that we're being able to see so like I've always said competition is good and with this I do expect that there is going to be of course a lot more competition in the space when it does come to software systems like this so that's why I'm highlighting Nova one because HiFi is a smaller company but I do think that you know in the future it's going to be a lot more impactful and I don't know why P collabs didn't actually uh do a bigger announcement about this but they actually did make some announce an ments SL improvements to their video model so they added certain things where there is essentially a lot more dynamic range and a lot more movement previously what would happen is there wouldn't be a lot more movement forwards backwards or rotation just simply because it's pretty hard to do but they quietly shipped this small update uh and I think they should have been a bit louder about it because it's a lot better but I do think that considering the fact that Google also did their recent V up update where they showcased a model that was really good I made a video on it it's literally really good um I think that you know once we get a model that can match Google's vo or open I Sora I think that that's where you know wild scale implications are truly going to come now the onx backed by openai and

### [16:50](https://www.youtube.com/watch?v=9IzAp4acajo&t=1010s) Humanoid Robotics

another you know just a bunch of other companies managed to show off their final update not their final update a recent update do apologize for audio there but they managed to show off their update where essentially they showcased an autonomous robot swarm that was able to do uh a bit you know not a bit like a lot of different tasks I'm going to show you guys the video because this is rather impressive in fact it's remarkably impressive for what they've been able to do and I think this is going to be one of the companies that does lead the robot Revolution because they actually do have the brains and they've been working on this problem for quite some time so take a look at how the Rob OTS are actually you know coming into fruition in terms of being able to act together to achieve a certain goal hey could you and your friends tidy up here we have some visitors coming oh a can you please pick up the cup so I think that entire demo was a lot more impressive than people are giving it credit for because it was truly impressive to see what the future is going to hold so one of the things that I thought was really cool about this entire thing was of course the voice commands so these robots can now accept voice commands based on what someone simply says so he was like hey can we tidy up this entire area and I'm guessing that what they have is they have a system where your voice command is then interpreted into a you know policy I'm not sure if it's with an llm or whatever and then the robot you know interprets the voice command selects the right policy and then it gets to work based on what it should believe the desired outcome is and you can see that all of these systems are seemingly working together like some kind of hive mind and what they also stated was that this is where you're telling an eve to do multiple autonomous tasks back to back so you can see it's cleaning up the spilled coffee it's also pushing the chair inwards and it's also managing to do things like bring the drinks to the meeting room so it's a very impressive setup that they have here and I think it was really cool how the robot was just in there and then she was like Hey can you just pick up this mug and it literally uh picked up the mug and went back now I think this is you know impressive because it shows us what the future like this is probably I would say the first realistic glimpse ever that we've seen of robots actually walking around somewhere and well not walking you know rolling around somewhere and then actually doing tasks autonomously from a human voice command and this is all uh back to back this is not like cut scenes at all this is all one giant take which makes this even more impressive people don't know how hard that really is to accomplish and like I said before this is probably how the future of certain spaces are going to be where we've got you know five six seven different robots in one a certain area and a human giving them certain voice commands and the robot you know then interpreting that and then selecting the correct policy and executing the action based on whatever it is so I think the only thing stopping this from becoming a reality are of course the two things number one being the system the AI you know the software uh that actually makes this feasible and of course the hardware the problem with the hardware is that even if right now we had the software to do this uh the biggest problem is that robots like Eve which is onex's robot the hardware that they've built to you know move the software the only problem is that it is very expensive so these robots can I'm not exactly sure on the price but I remember somewhere looking online and the prices range from you know six figures so it's basically like the 1990s early 2000s where you know flat screen TVs and you know huge giant computers were really expensive but over time you know these things come down now you can get a flat screen TV pretty much rolling them out the door for free um on computers you know the more compute that you can get for like the more bang for buck the more you can get for your money um is uh of course increasing because the rate at which the hardware has improved uh has far outpaced uh the price of these things so I think that will happen in robotics as well we also did recently get a robot that was you know uh only $116,000 but it's just about applying this software to these so it will be interesting to see where 1X is a couple years from now because this is genuinely an impressive demo there was also a

### [23:06](https://www.youtube.com/watch?v=9IzAp4acajo&t=1386s) Superintelligence

fascinating interview with Roman yapi okay and the reason is so fascinating is because he actually believes like his percentage of Doom as in his percentage of AI you know killing us is 99. 99% so this guy is arguably stating that if we create General super intelligence we're screwed and in the video he makes a very important argument okay now the video is an interview with Alex Friedman I have to say it's just it's kind of mind-blowing at you know his his kind of uh thought process but I think it's also uh I guess it's worth a watch because the interviews over two hours long and I watched the entire thing and I want to say that uh it definitely you know brushed me up on some of the problems with AI safety so it's important you know just put it on in the background but the thing that I liked most about this interview was that he spoke about how you don't actually need to create a general super intelligence like that is something that humans uh are just inherently barreling towards that for whatever reason we don't need to actually do that so basically he was stating that you know how currently we're trying to say AGI um he's basically just stating that with the current you know systems and the current plans and software that we're developing we should just you know create super intelligences I guess you could say or you know narrow AI in certain areas like maybe maths like maybe you know for driving and just focus on those areas because if we create a general intelligence that one is going to be far worse for Humanity because the outcomes are going to be uh I guess you could say things that we literally cannot predict because with software every time we scale up these systems there are just a million different things that we literally can't predict so with this he's basically saying that you know if Humanity doesn't realize that we don't need to create a general super intelligence but if we have something thing and for example I guess he's really right because we have things like Alpha fold that can you know predict the protein folding problem and that is like a super intelligent system that focuses on just protein folding he says that's the only thing that we need to do and we shouldn't be focusing on General super intelligences uh and I'm going to show you guys that clip but there's also this clip right here cuz I think this actually solves another problem but I'm going to come back to this one right uh we can definitely keep up for a while I'm saying you cannot do it indefinitely at some point the cognitive Gap is too big the surface you have to defend is infinite but attackers only need to find one exploit so to you eventually this is we're heading off a cliff if we create General super intelligences I don't see a good outcome long term for Humanity the only way to win this game is not to play it okay well we'll we'll talk about possible solutions what not playing it means um but what are the possible timelines here to you what are we talking about we're talking about a years decades centuries what do you think I don't know for sure the prediction markets right now are saying 2026 for AGI and I thought that clip was really insightful because of course you know it's the game that we probably shouldn't play uh and I think you know and I hope in the future that you know some of these AI labs are like wait a minute you know let's just actually just focus on narrow ASI and not do that but I guess that you know you kind of need it to do a lot of different things but anyways um I guess what we have here is another clip okay and this is one I do want to include because he talks about how one of the only ways to solve the human conflict problem is to have your own virtual Universe where everyone is just happy and it's kind of an interesting concept that I didn't even really think about I knew that personalized experiences was going to be something that is uh huge in the future like it's going to be something that is just incredibly big in the future but he basically talks about how there's a lot of conflict with human values because in this Physical Realm there's a lot of things that other people want that we only have and there's a lot of conflicts because certain people don't believe our belief systems or whatever but the only way that we could potentially solve that is through creating our own virtual uh systems or environments where I guess you could say uh there there's no problems and uh there's no frustration to solve value alignment problem I'm trying to formalize it a little better usually we're talking about getting AIS to do what we want which is not well defined are we talking about creator of a system owner of that AI Humanity as a whole but we don't agree on much there is no universally accepted ethics morals across cultures religions people have individually very different preferences politically and such so even if we somehow managed all the other aspects of it programming those fzy Concepts in getting a to follow them closely we don't agree on what to program in so my solution was okay we don't have to compromise on room temperature you have your Universe I have mine whatever you want and if you like me you can invite me to visit your Universe we don't have to be independent but the point is you can be and virtual reality is getting pretty good it's going to hit a point where you can tell the difference and if you can tell if it's real or not what's the difference and one of the last things I

### [28:11](https://www.youtube.com/watch?v=9IzAp4acajo&t=1691s) Gemini Coding

want to show you guys that uh was really fascinating that you can actually use is Gemini UI to code so this is a small but powerful app that uses an agentic framework to convert an image to code and this was really cool because it uses Gemini 1. 5 Pros increased Vision capabilities and one of the crazy things was about this was the fact that um they were able to uh recreate GPT 4's UI simply from just using the image of it and it was remarkably uh you know incredible what it was able to do I'm going to show you guys that in a second but uh right here you can see this is uh what it's able to do so it's based on an image right here so it has this image and I'm guessing that was of uh this you can see this is the playground it looks pretty similar to opening eyes one I'm pretty sure it actually is but just on Mac and then we can see here that what the system is doing it's describing the UI so it's converting the vision what it's seeing into text so you can see right here it you know lines out where everything is it's got the coordinates of where everything is uh and it's stating what all the styles are and then after it has all of that data then converts all of that into code then converts that into a website you can then download the HTML and then right here you can see click to upload the image and then we get you know a working prototype for whatever we want to use it for so I feel like this is really cool because a lot of the times what tends to happen is there's this disconnect between people who design user interfaces and then people who actually code them so this could be something that Bridges the gap between those two markets cuz I know that is something that is uh you know always been some kind of issue so this is something that was uh you know of course really cool and then you can see here I've improved the framework even more he gave it GPT 40 and then this is exactly what it came up with so like I said before this is pretty wild and it gives you guys kind of like a small Insight onto how in the future quickly things are going to be built like you could just have an image or something um and it's going to be able to build it really quickly so the capabilities are increasing and let me know if there's anything I did miss that I should cover because there's still a lot to come on this channel

---
*Источник: https://ekstraktznaniy.ru/video/14265*