xAI Grok 3 Launch in 9 Minutes

--- type: transcript date: 2025-02-18 youtube_id: BDseU-kmDYY --- # Transcript: xAI Grok 3 Launch in 9 Minutes the proclaimed smartest AI on Earth is now available grock 3 just tonight grock 3 was announced by the team over at xai in this video I'll do a quick overview on some of the key aspects from the announcements first up in terms of the hardware to initially train Gro 3 this was widely reported to be 100,000 gpus and they built this and wired all of this up in just 122 days now one of the announcements that hadn't been reported before today is that they actually expanded this to 200,000 gpus in just 92 additional days that 200,000 GPU cluster is the most up-to-date number that we know of that's reported in terms of how much compute capacity that they have to train these models and also presumably to host the inference as well now to dive into some of the specifics so this was more than 10 times the compute of grock 2 now in terms of the performance and some of the benchmarks the first benchmarks that they pulled up were grock 3 as well as grock 3 mini they compared these to Gemini 2 Pro as well as GPT 40 Sonic 3.5 the previous generation of models before the reasoning models that are coming to be popular such as 01 and 03 from openi or R1 from Deep seek but that's one of the key distinctions with this model is they're very clearly saying that this grock 3 Model is not independent from the reasoning model the reasoning model is essentially a layer on top of the capabilities of grock 3 itself they announced that the chocolate model that was on the chatbot Arena was actually grock 3 this was an early version of grock 3 and if you're not familiar and you didn't follow on X about two weeks ago or so there was this model that popped up on the chapot arena and if you haven't used the chapot arena it's effectively two raw llm responses streaming side by side and you decide which one that you prefer has the better response what was interesting when this chocolate model was announced is even anecdotally a lot of people thought that this was a frontier model I saw people thinking it was maybe an anthropic model or some people thinking it was a model from open AI but now we know it was in fact grock 3 within the chatbot Arena and just to put this into perspective so we can take a look at 03 mini here right at the bottom of the chart here all right next is the reasoning and I think this is arguably where a lot of people were most interested in these are the most capable models they do take a little bit longer to respond with if you've used o1 or R1 you'd know from having to wait through the thinking process before it gives you a response in terms of that reasoning and test time compute model they did demonstrate a number of examples they had an example of generate code for an animated 3D plot of a launch from Earth to landing on Mars and then back to Earth at the launch window and that was effectively a python diagram of different orbits of spinning essentially going out and then Landing from there they demonstrated what the UI look like now this is going to be the key piece that I think a lot of people are going to be interested in now in terms of the UI this is what the Standalone gro.com website is going to look like within here the notable new features are deep search think as well as big brain for really hard problems and problems that you want to unlock more compute for being able to reason for longer and use effectively more GPU power to try and solve that problem that's going to be what this big brain interfac is for now there is also deep search which I'll touch on in just a moment and then there is that thinking capability for those harder questions or if you don't need a response right away to stream back you can leverage that as well next in terms of the benchmarks for grock 3 as well as grock 3 Mini for the reasoning and test time compute across Math Science and coding this model does outperform all of its peer so 03 mini High 01 deep seek1 as well as Gemini 2 flash thinking that both the minir reasoning model as well as the full reasoning model outperform basically all of its peers now you will see this light shaded bar at the top of the graph on each of these charts similar to O3 mini where open AI allows you to control how much compute you want to allocate to the test time compute there is a similar measure here where if you're going to be using the low setting that's going to be indicative of that solid bar and for the light portion of each of these bars at the top here that's going to indicate more compute for those particular tasks and as we can see here is plotted against O3 mini 01 R1 as well as Gemini 2 flash thinking and it basically outperforms across the board now one of the interesting things here is we do see that grock 3 mini reasoning does outperform even the grock 3 reasoning beta on some Tas next they briefly touched on this idea on whether the models are overfitted to these particular benchmarks to show that this wasn't the case and that these were generalized skills from the models they performed it on this Aime exam that had just come out that wouldn't have been within the training data and you can see the respective scores across the board here with the most compute allocated both of these models will perform all of the competitors across the board next just briefly to show you what the examples look like this is what the space mission from Earth to Mars and back look like was effectively these spinning consecutive circles at different intervals to show the answer to that question and this one I found a little more interesting it was a Tetris Beed mix within the training data of models there's going to be Tetris right those will be on GitHub 10 times over but the more interesting example was one that they used to demonstrate this big brain feature and that was for this Tetris ped mix and why this was interesting that they touched on is because there's going to be examples out there within training data for Tetris there's going to be examples for Beed but being able to combine those two and have a creative solution for a new game that's where it is really interesting effectively how this worked is as different blocks were falling if three were in a row similar to Beed it would delete those blocks and they didn't demonstrate how it would work so effectively it was a combination of Tetris as well as ble so next they didn't touch on this too much but they did mention that it is coming agents are going to be the next Frontier and that is something that they're working on internally next they demonstrated the Deep search capability this is very similar to something like Gemini deep research open ai's deep research deep seek has a deep research version as well and effectively what these are is they almost create reports where it will search the internet it will reason about the question that you have asked and by the end of it it could take several minutes but it will give you a very detailed report on whatever you're asking for it will go and search for the internet and find the different information and by the end of it it will synthesize all of that for you say if you ask for tabular data it will put things within tables for you now in terms of this interface this is going to be available within gro.com and just to show you how this works you will be able to see the reasoning if you do expand the thoughts you'll see the thoughts on the right hand panel similar to chat GPT in terms of the interface itself that was one thing that I was particularly impressed with I do think that this interface is quite nice even though it is very reminiscent of something like chat GPT now in terms of being able to access it you will be able to access grock 3 on x premium if you are a paying X subscriber you'll be able to get it through that alternatively there is the grock interface so gro.com where you'll be able to access these models where you can unlock deep search as well as think and then you also get early access to new features and higher image generation limits now the other thing that they did note is they do have a voice mode that is coming this is going to be similar to what open AI released where you can have a conversation back and forth with the model it can understand intonation and emotions and Cadence of your speech and be able to talk back to you or you could whisper to it or all of neat things that open AI demonstrated with their voice capabilities now in terms of the availability within the API they did mention that this is coming within a few weeks and in terms of some of the other announcements they mentioned that grock 2 is going to be open sourced in the coming months as soon as grock 3 is fully released the general plan going forward is they're going to open source the previous series of models as soon as the latest model is effectively done is how they described it so now at time of recording I am recording this right after the announcement I do not yet have access to grock 3 within the X interface or at gro.com quite yet in terms of personal access I am an ex premium subscriber I don't quite have access I'm recording this video literally minutes after that announcement had gone live so just be mindful of that I also checked gro.com where I still do see grock 2 but I'd imagine by the time you see this video if you are an ex paying subscriber you should be able to try Oak Rock 3 but otherwise that's pretty much it for this video I just wanted to do a really quick overview going over all of the different things that they announced today if you found this video useful please like comment share and subscribe otherwise until the next one