This GPT-5 NEWS Could Change EVERYTHING...

So there has been a rumor floating around about GPT5 and in this video I'll explain the rumor articulated by Alberto Romero in a long winded article that dives into the details of where exactly GPT5 is. And this isn't just a video on GPT5, it's more about where exactly the AI industry is headed as a whole. Because there could be something secret that these AI labs are doing that you ought to know. So you can see right here, the article starts that this rumor is about, you know, GPT5 and it does change everything. Now it isn't an over exaggeration, it actually does change everything when I get into the details because it means that the way that these models are going to be released is going to be completely different. So let's take a look at what this article says. So it talks about GPT5 internally and it says that what if I told you that GPT5 is real? Not just real, but already shaping the world from where you can't see it. So this is the hypothesis, okay, this is the entire hypothesis for, for this entire video and credit goes to this user. But it says that OpenAI has already built GPT5 but has kept it internally because the return on investment is far greater than if they released it to millions of ChatGPT users. And also the ROI they're getting is not money, but something else. As you see, the idea is simple enough, but the challenge is connecting the breadcrumbs that lead up to it. And this article is a deep dive into why I believe it all adds up. So overall, this article is basically stating that, look, OpenAI have created GPT5 and it is a stunning model. Like the model is truly stunning. But they've decided not to release this model because of a reason that we're going to get into, into this article. And that reason is particularly interesting. So let's go ahead and take a look at this. So one of the things they actually talk about here is they talk about the mysterious disappearance of Opus 3.5 says before going into GPT5 we have to pay a visit to its distant cousin, also missing in action, Anthropic's Claude Opus 3point. And as you know, the top three AI labs, OpenAI, Google and DeepMind and Anthropic offer a range of models designed to span the price latency versus performance spectrum. Now OpenAI provides options like GPT4O, GPT4O mini as well as O1 and O1 mini. And Google DeepMind offers Gemini Ultra Pro and Flash 1 Tropic has Claude Opus, Sonnet and Haiku and basically talking about how with all of these different models, you're trying to cater to as many customers as possible. So some prioritize top tier performance no matter the cost, while others seek affordable solutions which are, you know, good enough. Now they're talking about the mysterious disappearance of this model, the Opus 3.5 model that was meant to be coming in the series. You see right here, it says something strange happened in October 2024. Everyone was expecting Anthropic to announce Claude 3.5 Opus as a response to GPT4O. Instead, they released an updated version of Claude 3.5 Sonnet that people started to call Sonic 3.6. Now, Opus 3.5 was nowhere to be found. Remember, this was supposed to be like the GPT5 type model and it says this model was nowhere to be found, seemingly leaving Anthropic without a direct competitor to GPT4O. But when we take a look at actually what happened with Opus 3.5, it's rather fascinating. So they talk about how, you know, previously in October, he wrote in his Nuclear Weeks post that there are rumors that Sonnet 3.6, which is, you know, a really good model, is just, you know, a checkpoint of a failed training run on the much anticipated Opus 3.5. And that was because at the time, Claude Opus 3.5, according to the web pages, a lot of people did think at the time that it was scrapped because previously they were stating that, look, Claude 3.5 opus is coming, it's coming. And then they kind of just removed the fact that the model was going to be coming from a lot of different web pages. And recently there was actually an interview with the CEO of Anthropic. And this stuff about Opus 2.5 is actually all going to relate to GPT5 in a moment. But the reason we're Talking about Opus 3.5 is because that is the model that is supposed to be a GPT5 type model. So just prefacing, what we're about to go into here is that we're looking at how Opus 3.5, which is supposed to be the GPT5 type competitor, how Anthropic have done that and how revealing what Anthropic have done reveals kind of what OpenAI has done. So they talk about how Dario Amade basically says that on Opus 3.5, not giving you an exact date, but as far as we know, the plan is still to have a claw 3.5 opus, cautious, yet ambiguous, yet valid. Well, Ridiculous timeline question, when IS Cloud Opus 3.5 coming up? Not giving an exact date, but, you know, as far as we know, the plan is still to have a Claude 3.5 opus. So right here you can see that this is where we talk about the fact that, you know, in Bloomberg, there were many articles at the time that were, interestingly enough, talking about, you know, OpenAI, Google and Anthropic are struggling to build more advanced AI. And this was going from the jump from GPT4 to a GPT5 type model. And this article is really fast, fascinating because I remember the time I was reading it and it was like Gemini 2 was disappointing. They were saying the GPT5 was disappointing and the Opus 3.5, they all performed essentially better, but not better enough. So he talks about here. Bloomberg wrote an article that after training it, anthropic found that 3.5 opus, which is GPT5 competitor, performed better on evaluations than the older version, but not by as much as it should, given the size of the model and how costly it was to build and run. And it seems that Dario refrained from giving a date because although the Opus 3.5 training run hadn't failed, its results were underwhelming. And basically what they're actually stating about this was that a lot of people missed was the fact that these next iteration of models, GPT5, Gemini 2, those models were good, but the only bad thing about them was that they were just quite expensive for what they were. And trust me, you want to remember this fact because it ties into a part of why they didn't release these models. Now, here's where we get some insider information, some leaks, and it says, On December 11, semiconductor expert Dylan Patel and his semianalysis team delivered the final plot twist, presenting an explanation that weaves all the data points into a coherent story. It says anthropic finished training Claude 3.5 opus, and it did perform well with the scaling appropriately, yet for some reason Anthropic didn't release it. So Anthropic managed to get their GPT5 model performing well, doing very well, but apparently they didn't release it. And he says this is because instead of releasing it publicly. And this is where we get to the crazy theories that anthropic used Claude 3.5 opus to generate synthetic data and for award modeling to improve Claude 3.5 sonnet significantly alongside user data. Now, that's a crazy, crazy statement, but this is from a reputable source within the industry that has previously spoken about things before. They've been released and later it's come out that that information has been true. And this is basically one of the key theories driving behind why many people think that GPT5 is alive and well internally, but isn't currently yet being released to the public. And I mean, it kind of makes sense when we actually take a look at more things because we all know just how good Claude 3.6 Sonnet is, or Claude 3.5 Sonnet is whatever you want to call it. But many people can't figure out why is Claude 3.5 sonnet so good, or Claude 3.6 sonnet so good? But they haven't released Claude Opus just yet. And so you see here it says in short, anthropic did train Claude 3.5 opus. They dropped the name because it wasn't good enough. And Dario, confident a different training run could improve the results, avoided giving a date. And Bloomberg confirmed that the results were better than existing models, but not enough to justify the inference cost, which is of course, how much these models cost to run when people are using the models. And it says here that Dylan and his team uncovered the link between the mysterious Sonnet 3.6 and the missing Opus 3.5. The latter was being used internally to generate synthetic data to boost the former's performance. And so this is where we have this kind of thing right here. And trust me, it's going to link to GPT5 in a moment. But we can see here that this GPT file style model from Anthropic, one of the biggest rivals, to one of the leading labs, we can see that that model right there is going down into Sonnet 3.6 and this is basically where they're distilling the model. So essentially you could say that they internally built a really, really smart AI system, but they probably haven't released it to the public just yet. And we can see right here that publicly sonic 3.6, that's what's going on. But yeah, it's pretty crazy now it gets crazier here. So this is where they start to explain model distillation. And this is what they say, better, but also smaller and cheaper. The process of using a powerful, expensive model to generate data that enhances the performance of a slightly less capable model is known as distillation. And it's a common practice where this technique allows AI labs to use their smaller models beyond what could be achieved through additional pre training alone. So one way that a lot of these companies are managing to get these smaller models to be better and better, better is through model distillation. So you have a teacher model, then you distill that knowledge and you transfer it to a student model. And this is something that just performs a lot better than the standard pre training and that entire paradigm that we used to do. So this is something that has been going pretty well for AI companies and it's going to be something that we continue to do. Now you can see right here that it says that there are various approaches to distillation, but we're not getting into them. What you need to remember is that a strong model acting as a teacher turns student models from small, cheap and fast and weak into small, cheap, fast and powerful. So distillation turns a strong model into a gold mine. And Dylan explains why it made sense for Anthropic to do this with Opus 3.5 and Sonnet 3.6. So it talks about how the inference costs of the new Sonnet versus the old sonnet, they didn't change drastically, but a model's performance did. Why release Opus 3.5 on a cost basis when it does not make economic sense to do so relative to releasing a 3.5 sonnet with further pros training from said 3.5 opus? Basically saying that look, there's no point us going ahead and releasing Opus 3.5 when it's so expensive and costly to run. Why don't we just, you know, train, distill a lot of the capabilities from this Opus model, which is so amazing, into a 3.6 sonic model that they would love anyways. So this is super, super fascinating here. And of course this is like what I said already. Anthropic chose not to release this, you know, not because of the poor results, but because it's just more valuable internally for them to use. So they basically use AI to train the next set of AIs. And this is why this is kind of changing everything because a lot of people have speculated for quite some time that this is going to be the moment where these AI systems are going to just keep getting better and better and better because you use one to train the next version and use that version to train the next version and then so far you're just going to continue to expand in terms of the intelligence. So what's crazy about this is that he also says that Dylan guy, the one that we're referencing here, he says that that is why the open source community caught up to GPT4 so quickly. They were basically taking the gold straigh from OpenAI's mind. So people were essentially distilling GPT4's capabilities into smaller models. And that's pretty interesting because GPT 4.0 is a smaller model, it's much faster and it's much better, but it's also really, really useful. And that's essentially what these other companies are doing as well. They're using those capabilities and that's how they're getting these open source projects up to GPT4's level. So you can see here that one of the things that we all know is that Sonnet 3.6 wasn't just good, it was state of the art good and better than GPT 4.0. And of course Anthropic's mid tier model outperformed OpenAI's flagship mod, thanks to distillation from Opus 3.5 and of course, other reasons as well. So that was something that was like pretty crazy. And then this is where we talk about bigger and better is no longer the paradigm. So once these top labs are no longer talking about higher parameter counts being better, and the last time we got knowledge about the parameter counts, we actually knew that GPT 3.5 was 175 billion parameters and GPT 4, there were rumors saying that GPT 4 was 1.8 trillion parameters in a mixture of experts. But the craziest thing about this is that they actually now speculate that the Future models like GPT5 and Sonnet 3.3.6, or like whatever distilled models that we're getting, like GPT4 and Sonic 3.6, are significantly smaller than GPT4, despite them both being better than GPT4 across both benchmarks. So GPT4 is 1.8 trillion parameters. That's what I'm trying to explain to you guys. GPT4, 1.8 trillion parameters. Great model at the time of release, I think around two years from now or one year and a half ago. And at the time that model was considered crazy, steady up. But now we've got smaller models like GPT 4.0 and Sonnet 3.6 that are significantly smaller than this large model, but they are better because of distillation. And that just means that the knowledge in them is a lot more efficient and the models are a lot more smarter. So you can see right here that it says, you know, current models such as the original GPT4O are probably an order of magnitude smaller than GPT4, with 4.0 having around 200 billion and 3.5 sonnet having around 400 billion parameters. Though this estimate could be off by a factor of two, given the rough way I've arrived at it. And the point here is that, like, the thing that you want to pay attention to is that both these companies are following a similar trajectory. Their latest models are not only better, but also smaller and cheaper than the previous generation. We know how Anthropic pulled it off by distilling Opus 3.5 into Sonnet 3.6. But this is where they get into something interesting. So what did OpenAI do? Because of Anthropic, we know that they trained Opus 3.5 and put the knowledge of that model, distilled it down into Sonic 3.6. What on earth did OpenAI do? And this is where the GPT5 rumor comes into account. So we can take a look at this diagram and basically it shows that we have the teacher models and we have the distillation models. And it's basically saying that maybe OpenAI has a very secret internal model. It might be GPT5, it might be a much smarter model, considering all the things we've heard recently. But we can see here that these models are going to be models that are very smart, but they're only used for distillation to public models that are better, smaller and cheaper to actually run. And this is pretty crazy when you think about it. So internally at Anthropic, it's quite likely they have Opus 3.5, which would be incredible to use, but they managed to distill it down into Sonnet 3.6. And Sonnet 3.6, as you already know, is pretty credible. And this is where they're saying that, look, with OpenAI, it's potentially true that we have a GPT5 type model that's distilling all of this information that these models that we're getting now, these much smaller ones that we're currently using. And I do know that there was some kind of distillation going on with the strawberry model. So OpenAI are definitely familiar with that distillation process. It's quite like how we had the 01 model and we had the O1 previews models and how effective those ones were. Now, when we get onto chapter three, it says that one might assume that Anthropic's distillation approach was driven by unique circumstances, an underwhelming training run for Opus 3.5. But that is something that all of these companies have experienced. And the key thing about this is that the news kind of flipped on everyone's head because everyone took the wrong thing away from the news. They said they had subpar training ones, but subpar. It doesn't mean that it was worse. Because think about it like this. You can have something that is subpar, but it doesn't mean that it was worse than you did before. It just means that it didn't live up to your expectations. Like imagine you're expecting to win $10,000 in a competition and you only won $1,000 more than you usually win. That's probably subpar, but it doesn't mean that it wasn't an improvement. And you have to remember that with intelligence, every inch that you gain unlocks a whole new host of capabilities. So it states here that the causes for this don't matter to us. Diminishing returns for lack of data. Whatever the case is, it doesn't really matter because all of these companies are going through at the same time. And basically this is why they talk about. One of the key things for these companies that most people don't understand is that, you know, if 300 million people are using your product weekly, the operational expenditures can, you know, suddenly kill your company, which is really important. Like costs matter. Costs really do matter for these companies because there are so many people using them, there's so many servers, and you know, this is something that continually is expanding week on week. So it's pretty hard to keep up with said demand. And so this is why they talk about how distillation was really good because whatever drove anthropic to distill sonnet 3.6 to opus 3.5 is affecting OpenAI several times over. And it basically just talks about hey, how distillation works because it bridges these two universal challenges into an advantage. You solve the inference cost problem by serving people a smaller model and avoid the public backlash for underwhelming performance by not releasing a larger one. Now, of course, as most of you guys might know, one of the things that you can't do anymore is of course over training. So these AI labs have actually exhausted all the high quality data sources for pre training. And this is actually something that Elon Musk and Elias have admitted in recent weeks. He says we're back at distillation. I think that both GPT4O and Claude 3.5 Sonnet have been distilled down from larger models. So this is something that they both actually talk about as a realistic thing to the point where they're saying that the way how we now get to a next level of model is that we can't just put more and more data in. New innovations are needed and they're basically Saying that, look, distillation is probably the only way that these models are going to be getting better. So you can see here, it says every piece of the puzzle so far suggests that OpenAI is doing what anthropic did with Opus 3.5. Train the model and hide the model in the same way through distillation and for the same reasons, because of course there are poor results in terms of the cost. And that's quite the discovery because Opus 3.5 is still hidden. But we have to think about it. Where is OpenAI's and Nagala's model? Is it hiding in the company's basement? Care2 venture a name. And this is why OpenAI are currently distilling that model. Now what's really interesting as well is that they talk about how he who blazes the trail must clear the past. So he said, I started this by analyzing Anthropic's Opus 3.5 story because it's the one where we have more information. Then I traced a bridge to OpenAI with the concept of distillation and explained why the underlying forces pushing Anthropic are also pushing OpenAI. But there's a new obstacle in the theory, because OpenAI is the pioneer. They might be facing obstacles that like Anthropic simply haven't found yet. Because if you're innovating, you're going to, you know, face problems that nobody else has seen yet. So you need to clear that path. And it says here that one obstacle is the hardware requirements to train GPT5. Sonnet 3.6 is comparable to GPT 4.0, but it was released with a five month lag. We should assume GPT5 is on another level, more powerful and bigger, also more expensive, not only to inference, but also to train. We could be talking about a half billion dollar training run. Would it even be possible to do such a thing with current hardware? And yes, that would be possible, but the crazy thing about this is that you wouldn't be able to have inference of that model. So basically what they're stating here is that these companies are probably still scaling the models internally and doing insane training runs, but the only way that they could actually provide the inference to get these models out into the public is to dist those capabilities down into a smaller model. And it says in principle, our current hardware is good enough to serve models much bigger than GPT4. For example, a 50 times scaled up version of GPT4, having around 100 trillion parameters, could probably be served at $3,000 per million token. And 10 to 20 tokens per second of output speed. However, for this to be viable, those big models would have to unlock a lot of economic value for the customers using them. So of course this is the kind of reason why they don't release it. And this is super interesting because we know that these companies are always currently struggling for inference because they're trying to do a lot of research, all these kind of things. But it will be interesting to see what is actually going on behind the scenes here because if they are training up a model that is that big, then they could do a lot of things with that. But the only thing I would say that this article doesn't cover at the moment is the fact that GPT4O I remember reading a few posts about this and I do remember reading that GPT4O is a model that is an omni model that's built from the ground up to be multimodal in and multimodal out. So I'm not sure like it was just an LLM, but maybe there are more details on that. But I do remember reading about GPT4O being like this omni model, so it was trained on audioin audio out. And I do remember reading the research paper, the entire full thing where I did 3040 minute details going into it where you could actually see that the model is capable of a lot of different things. So I do think that if they do have a GPT5 type model, I don't think they've just dealt it down into GPT4O yet. I think that model is definitely going to be coming in the future and they talk about spending that kind of inference. Money is not even just a viable for Microsoft, Google or Amazon because of course they need to unlock a lot of economic value if they plan to serve this several trillion parameter model to the public. So they don't. So they train it, they realize it performs better than their current offerings, but they have to accept it as it hasn't advanced enough to justify the enormous cost of keeping it running. And that's essentially what the Wall street journal said on GPT5amonth ago and what Bloomberg said about Opus 3.5. Now another thing that they said, if OpenAI were hypothetically withholding GPT5 under the pretext that it's already, they would achieve one more thing besides cost control and preventing public backlash. So if OpenAI were hypothetically withholding GPT5 under the pretext that it's not ready, they would achieve one more thing besides the cost control and preventing the public backl that actually sidestep the need to declare whether or not it meets the threshold for being categorized as AGI. As, you know, they have a contract with Microsoft that says, you know, AGI is a system that can generate at least 100 in profits. Maybe Microsoft would say that if people are able to build wrappers out of that and that's able to get them to 100 billion, maybe they wouldn't mind triggering the AGI clause and parting ways with Microsoft. So, you know, potentially, if they were looking at 100 billion in annual recurring revenue from GPT5, maybe they wouldn't care. Now this is where the theory starts to evolve into something really crazy and it starts to finalize. So this is where the theory starts to basically say that, look, they might not need us. So even if that were true, no skeptic has stopped to think that OpenAI may have a better internal use case than whatever they'd get from it. Externally, there's vast differences between creating an excellent model and creating an excellent model that can be served cheaply to 300 million people. If you can't, you don't. But also if you don't need to, you don't. Now it says here this is crazy, okay, that they were giving us access to their best model because they needed our data, but not so much anymore. They're not chasing our money either. That's Microsoft, but not them. They want AGI and then they want ASI and they want a legacy. They're basically stating that, look, before they needed to train on user data and they needed to figure out ways to scale the model, but now that they have GPT5 and whatever internal models they're using to distill down the knowledge into simple products that we can use and get a decent amount of value from on a day to day basis, they don't really need to provide us with the cutting edge pieces of technology anymore simply because they don't need to sell serve them anymore. They just need to use them themselves in order to develop better products and better technology. And that's kind of like saying that when OpenAI develops their own internal AGI systems, then they're just going to use that themselves in order to actually make money. So this is where you can see here that this is what they're talking about. Internally their private models are Opus 3.5 and privately they have GPT5 and they distill those capabilities down into these models. I'm not sure about distilling it into GPT4. Oh, but maybe it's definitely being used to help with some of the reinforcement learning for future models and some of the synthetic data generation. And I have heard that that is the case with 0103 and you can see right here it says we're nearing the end. I believe I laid out enough arguments to make a solid case. OpenAI likely has GPT5 working internally, just as Anthropic does with Opus 3.5. But it's quite plausible that OpenAI never releases GPT5 at all. The public now measures performance again 0103, not just GPT4 or Claude Sonnet 3.5. Now with the test time, new scaling laws, the bar for GPT 5 to clear keeps rising. How could they ever release a GPT5 that truly outshines 01 and 03 and the coming O series models at the pace they're producing them? And besides that, they no longer need our money or our data anymore. Basically saying that. Look, when we take a look at how the fact that these O1 series models are just so incredible when it comes to the raw capabilities of reasoning, why on earth would they release a GPT5 level model at all when those smaller models that we're currently getting with GPT four type systems are just a lot more expensive for incremental games. So you can see right here it says training new base models like GPT5, GPT6 and beyond will always make sense for OpenAI internally, but not necessarily as products. That part might be over. The only goal that matters to them from now on is to keep generating better data for the next generation of models. From here on, the base models may operate in the background, empowering other models to achieve feats that they couldn't on their own, like an old hermit passing down wisdom from a secret mountain cave. Except that cave is a massive data center, and whether or not we meet him or not is something we're going to have to see. So it's quite likely that maybe we're going to have these internal models, GBT6, GPT7, GPT5, producing these synthetic data for these future models to be trained on. It could definitely be the case, considering the fact that it's quite likely that we might not get these models. And so what about if GPT5 suddenly gets released? They bas state that even if GPT5 is eventually released, OpenAI and Anthropic have already initiated the operation of recursive self improvement with humans in the loop. And it doesn't really matter what they give us publicly, they're going to be pulling further and Further ahead, like the universe expanding so fast that distant galaxies can no longer reach us. And that's probably how they jumped from 01 to 03 in barely three months. And that's how they're going to jump to 04 to 05. So it's probably why they've been so excited on social media, because they've implemented a new way to scale incredibly. So you can see right here, they actually talk about something that I actually spoke about quite a long time before, like ages ago when I first launched my ever AI community, I spoke about the fact that, you know, even if AGI does arrive, we probably won't get access to it because the economic value for the average person just simply doesn't make sense. And you can see right here that they state, did you really think approaching AGI would mean gaining access to increasingly powerful AI at your fingertips, that they'd release every advancement for us to use? Surely you don't believe that they meant it when they said their models would push them too far ahead for anyone else to catch up. And each new generation model is an engine of escape velocity from the stratosphere. They're already waving goodbye and they're basically saying that, look, with every time that they make a new model, it's going to become harder and harder to catch up to OpenAI because they have something else that can generate synthetic data and that can also help them further the entire cycle of increasing intelligence within those models. Now, recently there was also this very cryptic tweet that has been going viral on Twitter and it says, just got read into some info. What's happening globally, happening internally at OpenAI and Holy Mother of God, I don't even know how to express my feelings without sounding like hype, but I don't know what to say. But I will share this. The innovators are coming. The problem is we don't know how they got there. Now, I will say this isn't like a Jimmy Apples kind of post. I haven't really seen any people corroborating this statement, but not just this person has been stating this thing. I've been seeing time and time again from people at OpenAI stating crazy things about superintelligence. OpenAI on their blog have said that, you know, they're now chasing artificial superintelligence instead of AGI and they know exactly how to get to AGI. So of these statements coming around the same time at this new paradigm isn't honestly surprise. And as I was making this video, there was actually a tweet about GPT5, you can see someone said, can you comment Something about GPT5? We know you won't be able to say anything, just anything at all. And he responds saying what would you like to know? And chubby comments back saying when any time. In terms of the estimate of time arrival and of course the performance, how much better is this going to be than GPT4O? And will this GPT series merge with the O series? And he says he's still figuring out when the estimated type of arrival is and of course the performance. And in 2025 they actually talk about merging the 01 series and the GPT series. So overall, still a very vague response. Not much to decipher there, but it definitely does make sense considering the fact that they did create these models, they were going to make them anyways and of course we know that they haven't released them, so it's quite likely that they are using them internally to generate data and do many other things. Just imagine a version of Claude 3.6 that's even better than it is now. Imagine what they could be using that for. And we've seen literally how much Claude 3.6 sonnet has changed entire industries in terms of like cursor and coding and what people are able to do with that. So it's super, super interesting to see where things go from here. So, that being said, hopefully you guys enjoyed this video and I'll see you in the next one.