10 Confirmed Features Likely Coming To GPT-5

things we expect from GPT 5 one of the first things that we do expect is of course a longer context window the reason we expect is because Google's Gemini they actually did increase their context window by up to 10 million tokens we can all see right here that Google's gem 1. 5 Pro is actually up to 10 million tokens okay and you can see that it has a variety of different use cases and I don't think open I like to be at the bottom when it comes to being state of thee art system but we can all see here that we have 1 hour of video 11 hours of audio 30,000 lines of code and 700,000 words now of course with any advanced super system the compute is an issue which means that this system isn't available to the public but the point is that in the future there will be successor systems that do have an increased context window and GPT 5 is likely to be one of those now you can see the GPT 4 Turbo's context window is at 128,000 tokens and Claw 2. 1 is at 200,000 tokens now Gemini 1. 5 Pro is still the market leader in this space and of course gp4 turbo is lagging behind so it is not going to be surprising if they increase the context window of GPT 4. 5 or GPT 5 now what we can also see here is we can see the numerous applications for GPT 5S or GPT 4. 5s longer context window for the context window that long what you can do is you can analyze really long transcripts you can analyze movies you can analyze entire code bases and easily find issues and fix them and this brings an entire new domain an entire new playing field when it comes to using these AI systems for Advan now you can see right here there's about 350 pages that you can put into Claude Pro so the new version of Claude the one that was just released the context window is about 200,000 tokens and is really good and we are seeing many many different people talk about how they're being able to use Claude you can see this guy right here says I'm able to put 10,000 line code based in and have it output modifications to multiple files for a high level request I'd say it's noticeably better than gbt 4 the context alone is a selling point but the code is also high quality from what I've seen so this is something that we do know is really really impressive you can see that this larger context window is something that is going to be coming and I think it's at least going to go up to 200k that is just my estimate based on you know what is capable now with llms and based on all the surrounding research and the research papers that have been going around I wouldn't be surprised if they had some kind of version of GPT 4. 5 Elite or something like that or GPT 5 uh you know. 5 or something like that you're able to have all of this so there is new technology being developed and clearly these other air companies are using it as well now another thing here is of course Advanced reasoning capabilities

now this one right here is not speculation this one comes from Sam Alman in an interview Sam Alman actually did talk to Bill Gates and he actually spoke about these Advanced reasoning capabilities that will be coming in the successor system known as GPT 5 or GPT 4. 5 so let's take a listen to what Sam Alman actually said about this Advanced reasoning capabilities and how it's going to impact the future systems which will be gbt 4. 5 or gbt 5 we'll be able to push that much further but maybe the most important areas of progress will be around reasoning ability right now gp4 can reason and only we'll be able to push that much further but maybe the around reasoning ability right now gp4 can reason in only extremely limited ways and also reliability you know if you ask gp4 most questions 10,000 times one of those 10,000 is probably pretty good but it doesn't always know which one and you'd like to get the best response of 10,000 each time so that'll be that that increase in reliability will be important so that you can see Sam Altman is of course looking to increase the reliability this model and make sure that it has advanced reasoning capabilities all that this means is that the model is just going to be a lot smarter that just means that the applications for this are now going to be increased because if you have a model that gives you the correct answer 99% of the time not 90% of the time that means you can start to roll this out in more and more applications because some Industries have a very low margin for error and if you are in one of those Industries using GPT 4 or any kind of system like that just doesn't make sense but if we get to Advanced reasoning capability where it's able to answer your question and you don't have to put like a thousand-word prompt in for it to understand exactly what you want and it's able to understand uh you know knowledge and able to move forward in a very Advanced way this is going to be something that we are going to be seeing in these future systems now this isn't going to come as any surprise as previously we've seen you know 3 months ago Gemini Ultra surpassed the GPT 4 on its reasoning capabilities you can see on the big bench hard it surpasses it in 83. 6% 82. 4 on variable shots and you can see on the H swag it didn't actually do better than that on the reasoning was 95% but of course you can see in all of these other categories we're going to see some Advanced reasoning capabilities now in addition to that we also got to see Claude 3 once again surpass GPT 4 and you know if we're looking at all of these kind of things you know the reasoning over text this kind of score mixed evaluations all of these ones we're going to see that this is going to really really improve in terms of what we're able to see because reasoning is one of the main points of GPT 4 if you didn't realize GPT 4 was always seen as the super smart system that really understands exactly what you need to do and so far it seems that the other systems have overtaken this so Advanced reing capabilities that's what Sam mman has recently said and I would not be surprised if we do see GPT 4 surpassing all of these and I mean maybe we've seen that you know like GPT 4 whilst it was still really good like whilst GPT 4 was very good of course I do think that the future models are going to be a notch up and I think reasoning is one of the biggest things because like I said it unlocks a whole host of capabilities so like Sam Alman said Advanced reasoning capabilities are going to be there now here is something that Sam Alman did

actually talk about with Bill Gates again and this is of course increased personalization so in the next version which is gbt 4. 5 or gbt 5 I'm just going to say gbt 5 from that one I think we are going to get increased personalization but like I said this video isn't going to be speculation take a listen to what Sam Alman said in this interview where he dropped some cool gems customizability and personalization will also be very important people want a very different things out of gp4 different styles you know different sets of assumptions we'll make all that possible and then also the ability to have it use your own data so the ability to know about you your email your calendar how you like appointments booked connected to other outside data sources all of that those will be some of the most important are so you can see that increased personalization is going to be something that's coming and this is something that I think is going to be changing the game currently the other systems like clae and Google jeed aren't personalized at all these systems don't really know you they don't know your birthday name they don't know where you come from they don't know like how you like your responses there is no personalization on these models you just talk to them and they give you a response so what we need now is we need increased personalization now one kind of thing that was kind of like a cool way to like kind of look at this uh was chat with RTX so it's kind of like a cool demo app that lets you personalize a GPT large language model connected to your own content docs notes and other data leveraging R retrievable augmented generation and RT acceleration you can custom a chat mod quickly to get context relevant answers and because it runs all locally on your Windows PC you get fast and secure results so essentially we have track with RTX and that is something that you can download uh you can upload your own data and it's extremely accurate um the only downside to this is that it's running on llama 2 which isn't that good of a model like gbd4 I mean it's decent but isn't that good but it's able to understand your data pull stuff from your data using RG and it's really accurate now of course Sam talks about how everyone has different uses of GPT 4 and some people would like their you know responses personalized and not sure how it work but maybe there'll be something in the future like you know when you sign up it might be like are you male or you female uh you know do you like your responses longer or shorter are you left leaning are you right leaning what do you think about this issue that issue then it's generating a chatbot that can kind of you know use your responses what are you going to be using the chatbot for mainly and of course we've seen that with like custom gpts and stuff but it is I wouldn't say it's limited but I would say the custom gpts has some limitations because of how much data you're putting in before the prompt the AI system doesn't really work that effectively because since there's so much information before the system prompt it just doesn't always stick to your prompts effectively so personalization is going to be something that I will expect and that I do expect and I wouldn't be surprised if it comes in GPT 5 because it likely is a big feature that many of these AI systems while they're you know making them a lot smarter more efficient personalization is something to where the AI system you know remembers you it's looking at picture of you it's like oh I remember you know we talked about this before um you know it's just going to be something that's really easy to do in terms of you know being able to use the AI system much more effectively because I don't know about you guys but I always have to constantly remind the AI system you know how I like my responses uh when I start a new chat I have to say look I have to you know put a whole a thousand word prompt in just to get you know certain things the way I want them and that is something that is really frustrating so personalization with your data it's going to be possible maybe we're going to get you know some kind of uh model to be able to download I don't think open hour is going to do that you know cuz they're no longer open but I do think that in the future um we're going to get some kind of secure personalization now uh one of the things that I think is going to be really effective is the inference speed SL latency so I think that you know Sam talked about this and I can't find the clip but I know he talked about this he talked about how uh latency will be

faster when talking to the AI so essentially when you have conversation with chat gbt if you speak to the AI and you know have a conversation with it with the phone app and you are using your voice it's actually pretty slow but this is something that s said that they're going to improve the future models now I'm not sure if this is going to be done with GPT 5 I know it's going to be done in the future models it might be done with like GPT 4 because of course they're going to probably reduce the cost for that and I think it's going to be really effective because literally yesterday I was playing around with an AI demo and it was so fast and how it spoke back to me I was truly shocked because I was like w this actually kind of seems like there's another person on the line here and it doesn't seem like the AI is just sinking uh that slowly so they're going to be able to do this and another uh cool trick that they are now starting to use with these AI systems is that they're now like instead of you know having the AI just blank out they're going to make the AI system go uh like you know those thinking words that you do where you're trying to think of something uh that's essentially what these AI systems are going to be doing uh spitting that out as a voice instead of just you know being completely you know not talking so I think the latency is going to be really faster but I don't think that you know gb25 is any faster than what we've got because I think that open AI they they do have a compute struggle I'm not sure what they're building tons and tons of stuff but you know all the stuff that they are building it's very compute intensive so they do have a limited uh compute like to be able to divy it up between super intelligence between Sora between future gp5 models between visual models so I think that if we do get gb5 and it is super smart I don't think it's going to be that fast I think it's going to you know take quite a time because Sora they spoke about Sora being quite slow as you know you could go get a coffee while it's generating the video another thing for that as well with Sora is that uh we did also get Google's Gemini 1. 5 Pro and even with 1. 5 Pro when you're entering really large code bases it does some sometimes take up to like 5 to 10 minutes to get a response from the model uh that shows us that the smarter these systems are the more data that they're handling you know the more complex these systems are the longer it does take to get a response so until you know M's law kind of takes into effect and we get those gpus that are really fast we get those transition that are you know super efficient uh this thing I still think that you know inference for the smartest model won't be fastest but I think that the latency when talking to an AI system you know they're going to kind of develop that grock style latency where it's really quickly now the message cap I think this is going to go I think for gbt 4 and gbt 4. 5 this is going to go

because uh be one of the things that most people have a gripe with but um you know I think this does have a really big problem maybe it's just to get people to pay for teams now that I think about it because you know if you are stuck with 40 messages every 3 hours you can sign up for teams you get 100 messages every 3 hours which is you know 2. 5 times the amount but uh yeah this is going to be something that I think does go in the future so that would be something that I don't think hangs around forever and of course we do have this right here increased Vision capability so right now gbt 4 vision is quite limited

it's able to do quite a lot of stuff but the problem is number one it's very expensive but which means that you can't use it for certain applications because it's like you know it's just too expensive to run like some people were testing the GPT 4 Vision API and they found that you know just doing like a quick video analysis it costs like around $10 which is completely unfeasible for any real application um the point the thing that you seeing on screen right now is us see GPT 4 versus fet now if you don't know what fet is Apple's model and apple released a vision model that is much more capable than gbt 4's one so I'm guessing that in the future they're likely to have a model version of gbt 4 Vision gbt 5 Vision that's able to really understand exactly what's going on in an image and really decipher it and uh truly understand I think it's going to be cheap as well because Claude just launched a hiq which is a new model it's really cool it's really fast it's a vision model and it's much more effective than I think it's much more cost effective than gbt 4 as well so I wouldn't be surprised if they you know have a new model that is uh a lot better than what we seen here so for me this is going to be something that I'm really excited for as well because Vision really does take AI to the next level and of course we did also see that AI demoed a new vision model in their recent demo with the figure robot so I'm guessing that one is going to be rather effective too so that is going to be something that is really really gamechanging if they do manage to get this right in terms of the cost down and stuff like that so it would be interesting to see if Vision comes out soon now of course like I said now this is another thing that I did think about okay and this is of course increased memory capability so this was chat gpts

you know small kind of way to increase the personality of course remember how before I talked about personalization keep the conversation going improves over time manage what it remembers this is stuff that you know you can kind of do right now but it isn't that effective of course you can see that you know this is what chat TBT have introduced but I do think that in the future like I said along with personalization this is going to be something that is rather effective I'm not sure how it's going to work uh of course you can see that it's just going to be like a document that we can upload and then it uses probably RG to retrieve that with some kind of advanced system so it's really person personalized and then I think across the board we're then going to see other companies race to kind of get this deployed as well but I think that this is going to be something that is rather effective uh and yeah increase memory capabilities I think over the long form chat is going to be able to store what you have and I'm going to give you guys some examples of that I'm going to include a snippet from a previous video where I talks about M GPT and how that is going to work uh it's not exactly how it's going to work but it's a possibility on how it could work so I going to include a clip of that in

addition we also do have uh you know multimodality which is Dar 4 but before I talk about D 4 take a listen to what Sam Alman said you know when you look at the next two years what do you think some of the key Milestones will be multimodality will definitely be important weed which speech in speech out images eventually video clearly people really want that we launched images and audio and it had a much stronger response than we expected so you can hear from just there Sam Alman clearly stating that image generation capabilities are going to be coming now if you're wondering what is the screenshot you're looking at if you don't know Sora when it released Open the Eyes video uh model essentially what they had was they also managed to make a image generation model on top of this and it's a photo realistic image generation model that you can see right here it's able to generate photo realistic images quite like mid Journey and it's something that really effective so I wouldn't be surprised if we do get Dary four that is able to do this but with a lot of restrictions not as free as mid journey to be able to generate pictures of you know Elon Trump whoever you really want to generate but um I think it's going to be really effective because it is going to be something that samman talks about how people do want images and of course uh usually when we do have updates to the system we're likely to get these major upgrades too so I think multimodality is going to be coming of course speech in speech out I think things are going to be more natural in the terms of you know you talking to AI it's going to probably ask you questions back ask you how you're feeling how you're doing so I think that multimodality is going to really be there now one thing I actually did forget to add to this presentation when I'm talking about Advanced reasoning capabilities another thing I think that's really going to be there you know so this is going to be 10 because uh we've done points all the way up until 8 so this is going to be point no this is going to be0 n essentially what I think we're going to be able to really add here is of course Advanced coding capability so the code on the human eval you can see gbt 4 is 67% at zero shop where clae 3 is open us is 84. 9% at zero s so across the board they beat gb4 in terms of coding and I think in the future we are going to be getting a system that's you know closer to 90% because tb4 they're really going to make sure that they beat every single other AI system so I wouldn't be surprised if their reasoning takes this off uh you can see right here that you know Gemini Ultra 74. 4 and of course uh gb4 is 67% so I wouldn't be surprised if they managed to completely dominate the coding uh I wouldn't you know there was Al code 2 which is essentially 84% uh more than it's better than 84% of coders although that is really computer intensive but I do think that what we have here is Advanced coding capabilities coming in the next one so that's going to be 0. 9 Advanced coding capabilities we've already seen Devon being able to do some really cool stuff um and I wouldn't be surprised if there was a separate model that just focuses on code because we're seeing with Devon there's clearly an entire application for that so in the future if there is just you know an entire code GPT like you know open working on a completely coding model that's just able to code better than you know 90% of all coders I wouldn't be surprised and if it's able to you know go on the recent benchmarks of what we saw with Devon I wouldn't be surprised if that's there as well now I want to talk to you guys about three things that I don't think are coming and one thing that is you know a bonus Point okay so one of the things I don't think is coming in the future okay for gbt 5 okay is Advanced agentic capability so if you look at the trademarks on opening eyes uh you know if you look on the public records you can see what the future trademarks are now trademarks essentially describe what the next future models are going to have and only in GPT 6 did we see testing artificial intelligence agents and we also saw that in gpt7 so it seems as if agented capabilities are being you know delayed until GPT 6 because there's a whole host of things that they want to put in GPT 5 and then they're going to release that in GPT 6 now this could change if a new company comes out of the works and has some Advanced agentic coding stuff that we've never seen before okay maybe Claude does it maybe Google does it maybe meta does it they all are working on advanced agentic stuff but we can see that stuff is likely to be until GPT 6 now what this means is that in GPT 5 I wouldn't expect any kind of agentic capabilities of course trademarks are just trademarks but then again if you look at the Sora trademarks it does talk about video how they're going to be generating videos um and that is something that perfectly describes what Sora is so this you know a opening ey shifts Battleground to software that operates devices and automate tasks I think this is going to be coming in the future maybe next year with GPT sh and I do think that gbt 5 which is likely to be you know coming this year I do think that one gbt 5 is going to be having not advanced intelligence you know capabilities but it's going to be a really good personalized system that is just you know better than anything across the board another thing that I don't think is coming is I don't think music generation is coming when I looked at once again the trademarks we saw GPT 6 and gpt7 after GPT 5 have music generation whereas gbg 5 doesn't have any mention of Music generation in the trademark so I don't think music generation is coming in the next model it might they could always change things they can do that they're allowed to do that but I just think that is something that is not on their Focus right now because uh it's not something that people are really asking for now here's something that seemed to be pretty crazy okay they are speaking about a feature that will change everything I made a video on this okay but here's the quote from a video that I made before this is a random feature and I don't know what this is like I said it could be code GPT it could be agent anything but you can see last month Ben new house an open ey employee who has worked on computer using agents at the startup according to a person familiar with his role posted on Twitter that he was hiring for his team and building what I think could be an industry defining 021 product that leverages the latest and greatest from our upcoming models he didn't elaborate um and he described the product that will change everything so one thing that we know will change everything as of course agents but this is something that they're actively working on so I think that this random feature this random product that is going to you know throw a spanner in the works so to speak is going to be something that's a future agentic system now these are my predictions what do you think else what do you what else do you think is coming from gbt 5 we've talked about longer context windows so you can analyze code bases we've also talked about how there's going to be Advanced reasoning cap abilities how these systems going to be able to think a lot better um you know sample over you know 10,000 responses to get the best one increase personalization these systems going to remember us we've also got inference SP inference speed we're going to be able to talk to these systems like we're talking to a person the message cap might be gone depending on if they solve the compute problem increase Vision capabilities because you know Apple have already beaten them you know increase memory capabilities it's going to remember more stuff we've also got multi modality coming which is of course like they said speech and speech out and I likely upgrade to doly 3 but you know AI agents isn't coming uh likely isn't likely to come uh you know music generation and a random AI agent product is likely to be uh next year sometime next year but then again timelines are increasing things are happening we have no idea but if you did enjoy this video um I am quite ill I didn't really want to record this video but I thought an unperfect video is better than a perfect video that never got released so it's better to do this video while I'm me so you guys can get this information because gp5 uh there was some rumors that it was happening today so I thought I might as well release this to see how accurate it is when the future does come so that being said if you did enjoy the video see you on the next one