Transcription
I know you started off as an intern at Microsoft. You were a Windows developer, and of course, you're a big PC gamer still. You want to just talk about even your early days with Windows and the kinds of things you built?
Yeah. Well, actually, I started before Windows with DOS. I had one of the early IBM PCs with MS-DOS, and I think it had like 128K in the beginning, and then it doubled to 256K, which felt like a lot. So I programmed video games in DOS, and then later in Windows—remember Windows 3.1?
No, it's wonderful. I mean, the last time I chatted with you, you were talking all about everything, the intricacies of Active Directory, and so it's fantastic to have you at our developer conference. Obviously, the exciting thing for us is to be able to launch Grock on Azure. I know you have a deep vision for what AI needs to be, and that's what got you to get this built. It's a family of models that are both response and reasoning models, and you have a very exciting road map. You want to just tell us a little bit about sort of your vision, the capability? You're pushing on both capability and efficiency. So maybe you can just talk about a little bit of that.
Sure. So yeah, with Grock, especially with Grock 3.5 that is about to be released, it's trying to reason from first principles. So apply kind of the tools of physics to thinking. If you're trying to get to fundamental truths, you you boil things down to the axiomatic elements that are most likely to be correct, and then you reason up from there, and then you can test your conclusions against those axiomatic elements. In physics, if you violate conservation of energy or momentum, then you're either going to get a Nobel Prize or you're certainly wrong. Basically, that's really the focus of Grock 3.5: fundamentals of physics and applying physics tools across all lines of reasoning, and to aspire to truth with minimal error. There's always going to be some mistakes that are made, but we aim to get to truth with acknowledged error, but minimize that error over time. And I think that's actually extremely important for AI safety. So, I've thought a lot for a long time about AI safety, and my ultimate conclusion is the old maxim that honesty is the best policy. It really is for safety. But I do want to emphasize, you know, we have and will make mistakes, but we aspire to correct them very quickly, and we are very much looking forward to feedback from the developer community to say, like, what do you need, where are we wrong, how can we make it better, and to have something that the developer community is very excited to use and where they can feel that their feedback is being heard and is improving and serving their needs.
Yeah, I know it's in some sense, you know, cracking the physics of intelligence is perhaps the real goal for us to be able to use AI at scale. And so it's so good to take that first-principles approach that you and your team are taking. And also, you're deploying this—I mean, one of the things about what you do is you're doing unsupervised FSD on one side, you're doing robotics, of course there's Grock, so you're deploying Grock across all of your businesses, from SpaceX to Tesla, obviously at x. I would love to even—you know, one of the themes for this developer conference, Elon, is we're building pretty sophisticated AI apps—right, it's not even about any one model, it's about orchestrating multiple models, multiple agents—just anything that you are seeing in the real-world application side, even inside of your own companies when you think about even a Tesla or a SpaceX where you put Grock and these other AI models you're building. It's incredibly important for an AI model to be grounded in reality—reality. Physics is the law, and everything else is a recommendation, which is—I'm not suggesting people break the laws made by humans. We should generally obey the laws of humans, but I've seen many people break human-made laws, but I have not seen anyone break the laws of physics. For any given AI, grounding it against reality, and reality, for example, as you mentioned with with the car, it needs to drive safely and correctly; the humanoid robot Optimus needs to perform the task that it's being asked to perform. These are things that are very helpful for ensuring that the model is truthful and accurate because it has to adhere to the laws of physics. So I think that's actually somewhat overlooked or at least not talked about enough is that to really be intelligent, it's got to make predictions that are in line with reality. In other words, physics—that's a really fundamental thing—and being able to ground that with cars and robots is very important. We are seeing it be very helpful in things like customer service. The AI is infinitely patient and friendly, and you can yell at it, and it's still going to be very nice. So that's good. I think in terms of improving the quality of customer service and sort of issue resolution, AI's already—Grock is already doing quite a good job at that at SpaceX and Tesla, and we look forward to offering that to other companies.
No, that's fantastic. Really thrilled to get this journey started, getting that developer feedback and then looking forward to even how they deploy these language models. I think over time we will have this coming together of language models with vision with action, but to your point, being really grounded on a real-world model, and that I think is ultimately the goal here.
Thank you so much, Elon, for briefly joining us today, and we're really excited about working with you and getting this into the developers' hands.
Thank you very much. I can't emphasize enough that we're looking for feedback from you, the developer audience. Tell us what you want, and we'll make it.