Transcription
So, Elon Musk just revealed the Grok 5 plan for AGI. So, let's talk about it.
So, I'm not sure if you're actually aware of this, but Elon Musk recently introduced Grok 4.3 beta. I know that might seem surprising because, you know, model releases I usually cover them, but this is one of the times where I didn't even know that this was a thing. Now, the reason being is that this is just a beta. And not only that, it seems like right now the company xAI doesn't really have much market share.
Now, essentially this tweet that you're seeing as well, I do have to be honest with you guys, this is completely AI generated. And the reason I actually did that is because xAI themselves didn't even announce this. And the reason I'm talking about Grok 4.3 beta is because you need to understand where we are to understand where we're going to go.
So, essentially what they're now doing is something completely different. Grok 4.3 is not being framed as a completely new generation. It's essentially the first step in a line of much bigger releases. This is the model that starts to transition from the current Grok generation into the larger models Elon Musk has been talking about. So, this will be rolled out very soon, but let's get onto the actual roadmap that you need to be aware of because I think xAI are going to do something crazy.
So, the Grok 4 roadmap is where we get interesting. This is where Elon Musk lays out the entire thing. He says, "Supplemental training has been added to Grok 4.3, which means the model is still being improved after the initial beta release." And then he says, "Grok 4.4 will be twice the size, 1 trillion parameters, with training data through early April. That's probably ready for release in early May, which is literally next month." And then he gives the next step, "Grok 4.5 will be 1.5 trillion parameters and hopefully out by late May."
So, the timeline here is insanely aggressive. We've got Grok 4.3 already being out right now. Grok 4.4 is expected to jump to 1 trillion parameters, and Grok 4.5 is expected again to jump to 1.5 trillion parameters. And all of that is supposed to happen in, let's say, a matter of weeks, maybe even just a month.
Now, this is important because Elon isn't just talking about a vague future model that might arrive someday. He's describing the near-term rollout, okay? And xAI appears to be moving from 4.3 to 4.4 to 4.5 extraordinarily quickly, and every step gets larger. That means that Grok 4.3 is essentially the starting point, and Grok 4.4 is the first major scale jump that we will see. Grok 4.5 will be the next one after that, and after that the roadmap points towards something much bigger, which I'll get on into a moment.
So, if you were thinking that xAI has been awfully quiet, it's because of this reason. They are locked and loaded with these models. And we all know that the scaling laws are still holding when it comes to intelligence, so it seems like they might be about to catch up.
Now, let's take a look at Grok 5, 10 trillion parameters, okay? Which is absolutely insane. So, essentially someone asked, okay, and this just gives you a context for how big the jump is. A user essentially says that Grok 4.2 is running on 500 billion parameters, and Grok 5 is expected to get a much larger 6 trillion parameter model and an even bigger 10 trillion parameter model. And then Elon replies saying, "Yeah, Grok 4.2 is just 0.5t and is missing some important training data."
And that sentence is really important because Elon Musk is essentially downplaying the current model. The current model that most people don't really use anyways. He's saying that Grok 4.2 is only 500 billion parameters and that it doesn't even have all the most important training data yet. When you think about it, that sets up the next release perfectly. If 4.2 is 0.5 trillion parameters, then 4.4 at 1 trillion parameters is twice the size. And 4.5 at 1.5 trillion parameters is three times the size. And then this is where we get into the big stuff. The models being discussed after that entire run is Grok 5, 6 trillion and 10 trillion parameter models. Those are on a completely different level. And you can see here he actually gives us the dates for these new models. He said about two to three weeks for 1t and four to five weeks for 1.5 trillion parameters. So, that's maybe around a month and a half, okay?
Now, when you think about this, that is an insane rate of speed. It seems like they're going to catch up.
Now, you might be wondering, how on earth is Elon Musk able to do this? How does he have access to this much compute? Well, if you haven't been paying attention, Elon Musk has been the king of compute when it comes to AI. We know that other companies like Anthropic and an OpenAI are having to cut programs and run out of compute and run around their hedges just to, you know, serve their users, but that is not a problem that Elon Musk faces. The thing is about Elon Musk is that he already has Tesla, which already has massive GPU clusters. X already has data and infrastructure and distribution, and SpaceX already have the engineering talent and the capital flywheel. And of course, you have to understand that they do have that thing called Colossus, which is a massive training cluster, and they built that thing insanely fast. They built that thing in months, not years, and I couldn't genuinely believe how quickly they did that.
And so, when we look at what is going on right here, SpaceX AI Colossus now has seven models in training, and I don't think you understand how insane this is. You've got Imagine V2, which is essentially the video model, which we all know that it's actually not that bad. It does pretty well in certain cases. For my specific edge cases, it doesn't perform that well, but it does do very well in terms of leaderboards and what people do like. Now, of course, you can see there's two variants of 1 trillion parameters, two variants of 1.5 trillion parameters, 6 trillion parameters, and 10 trillion parameters. So, seven models in training, which is absolutely insane. And he ends the post with saying, "Some catching up to do."
Now, this is basically xAI's scale strategy in one post. They're not training one model, they're training multiple models at the same time across different sizes and different use cases. The two variants of 1t and 1.5t are the things that we just discussed. That is probably going to be the Grok 4.5 and the Grok 4.4. And then the 6t and 10t, those are going to be connected to Grok 5. Now, I think what he's trying to [snorts] say here is that he's moving very aggressively because he's trying to catch up to OpenAI, trying to catch up to Google, trying to catch up to Anthropic, Meta, and DeepSeek. And I think they've just decided to scale aggressively.
So, when you look at this screenshot, I think it's very clear that he's not trying to compete through small model improvements. He's trying to compete through massive compute, multiple parallel training runs, and rapid scaling. So, when those new models come out, it's going to be really interesting to see what kind of checkpoints those are because those should not only be on the level of GPT-5.4 and Opus-4.6, but if he is training a 6t and 10t model, those ones are going to make AGI come into this conversation.
And so, you can see if you're wondering about the 10t model. And remember guys, 10t Grok 5, Elon Musk has said on countless occasions that this will be an AGI level model. You can see here someone said, "Just curious, how long does a training run for 10t will that take?" And the pre-training phase is going to be two months. So, this is kind of like a huge detail because this means that the 10 trillion parameter model isn't just a random future idea. Elon Musk is giving us a specific timeline. Meaning that the pre-training phase alone is around two months. And pre-training is only the first major release. After that, xAI still has to do post-training, alignment, testing, evaluations, safety work, inference optimization, product integration, and probably multiple rounds of improvement before that model is going to be ready for the likes of us.
So, I mean, when we look at this, this tells us a few things. First, the 10t model is serious enough that Musk is talking about the training timeline. Second, it's not immediately going to be released like Grok 4.4 or Grok 4.5. Those are coming much sooner. And the 10t model is more likely connected to the next big jump after the Grok 4 series, and that points directly towards Grok 5. So, when you think about it, that's why the two-month timeline matters. It puts the 10t model into a real development window, and it suggests that we're not going to get this model years from now. They're trying to push this extremely large model through the training phase quickly, which could set up a much bigger release later towards the end of the year.
Now, you can see that someone explicitly asks Elon, "Do you think we will have achieved AGI with one of these models?" And Elon replies with just two words, "Grok 5." That completely changes how we should think about the previous screenshots. The 4.3 beta, the 4.4 roadmap, and the 1t and 1.5t releases, the seven models in training, the 6t and 10t models, and the two pre-training two-month timeline, that points towards Grok 5 as the model that Elon Musk wants everyone to watch. He's not saying Grok 4.3 AGI, nor Grok 4.4 AGI, nor saying 4.5 AGI. He's pointing at towards Grok 5.
Now, it doesn't mean necessarily that Grok 5 will definitely be AGI, and it doesn't prove that they've solved artificial general intelligence. But he is associating Grok 5 with AGI in question. So, that makes Grok 5 the moment of truth. If Grok 5 launches and it is just a normal chatbot upgrade, people will say that this is all hype. They will say that xAI confused bigger models with just smaller models, and they will say that Mr. Musk over-promised again. But if Grok 5 launches and it's genuinely frontier level or meaningfully beyond current models, then the entire xAI narrative changes. Suddenly, that 10 trillion parameter training run looks less like marketing and much more like preparation. Suddenly, Colossus 2 looks like a serious advantage. Suddenly, xAI is no longer catching up. It becomes one of the companies that could define the next stage of AI.
So, the full story here is pretty simple. Grok 4.3 was released. It's in beta. Most people haven't used it because it's on the Grok heavy tier, which starts at a whopping $300 a month. Grok 4.4 and 4.5 are the fast scale-up. Colossus 2 is training the large frontier models, and the 10 trillion parameter model takes around 2 months to pre-train, and Elon Musk is saying that Grok 5 will be the AGI possible moment. So, that's going to be super interesting.
And this is not the first time he claimed this. He claimed this back in October of 2025, where he said this will be indistinguishable from AGI. Now, of course, I quickly want to add this here because I might leave a link to this video where I actually speak about Google's new AGI definition, and I think this is probably the best thing that covers it because they've been trying to make the AGI conversation less vague, and it's a paper where they titled Measuring Progress Towards AGI, and they argue that AGI should not be treated as one magical finish line where company simply declares victory. Instead, it should be measured by a broad cognitive profile, things like reasoning, memory, learning, attention, problem-solving, and cognition. The point is is that the model shouldn't just count as AGI because it's huge, or because it beat one benchmark, or because it feels like an impressive demo. It has to show broad human-comparable performance across many different cognitive abilities.
And that makes Elon Musk's claim much more interesting because under Google's definition, the question is not is Grok 5 10 trillion parameters. The question is, can it perform across the range of human cognition consistently, reliably, and at a level that actually compares to humans? Who knows? We will have to see.