"This person said, Cursor has tons and tons of user programming data, and they use it for a composer, which is actually not a real foundation model, but they could also use it to improve Grok, and that's why Elon bought Cursor. So, Cursor famously lied about having their own models, and they just used Quen's models and then fine-tuned them with their training data, but they're still, they're making it worse, they're not making it better, this data isn't valuable. And this is one of the big lessons that people who don't really understand the technology, like Elon, will never learn, is that that kind of information that user prompts is not valuable data. It's valuable data for things like, what kind of response do you want? It's not valuable data for like, become competent and capable at software engineering. And I have a really interesting story for you about what the real answer is and who's actually doing that, but let's look at the comments here, because there were a lot of people who said things like this, and I said, somebody smart could probably do that. This is sort of foreshadowing what I'm going to tell you next. And they said, generate the data from a million customers. I said, no, that kind of data is actually not very valuable. You need like a billion times more than that for it to really be valuable as training data. Like, all of the people who have ever used Cursor is a tiny, tiny drop in the ocean of training data that these models are already using. Quen's models are trained on like over 30 trillion tokens, right? So, a few million customers with however many prompts they've done is nothing. It has no impact on improving the quality. Another person said, I didn't want you to be right here, but I'm afraid that Cursor's data in Elon's Compute might be enough to become competitive. Elon's Compute is actually not connected to the power grid. His data center in Memphis is just a warehouse full of H100 cards that are not plugged into the wall in a building that is not plugged into the power grid. It's a scam to get stupid people to buy his stock from him. They said, I'm more upset that competent people at Cursor are choosing to stay and give Elon credibility, probably only until they're allowed to sell their shares. So then this person said, told you that was for, that was what it was for, and they showed this chart. So Cursor says, we partnered with SpaceX to, SpaceX AI, is that a new, is that what we're calling it now, to train Grok 4.5. This is our most powerful model yet and the first we've built for more than software engineering, for more than software engineering. First of all, all of these benchmarks are software engineering benchmarks. And if you look at the first three, they're like things you've never heard of that people are not using. And they're not even very good at any of these things, right? Like they're showing you that they're worse than everyone else on all of these benchmarks. But look at the last one here, which is the real software engineering benchmark that all the models are trying to be good at. SWE Bench Pro, which is something that can be independently verified. And they are the worst of everyone, okay, at the real benchmark. So what this tells you is that for $60 billion, Grok went from being a fifth tier model to being a fifth tier model. They're telling you that they are worse than everyone else. And these are using different compute budgets. So they're fudging the numbers here and they're still worse than anyone else. The only person who is going to succeed at this project, at the project of using AI to do software engineering, is my friend Isocant. And this is a longer story."