BREAKING: Tesla Now Has the Smartest AI Ever!! Elon’s Grok 3 Demo Just STUNNED Everyone!
Sintesi
TLDRThe Grok 3 presentation by XAI discusses their mission to understand the universe through AI, specifically Grok, which has significantly improved in capabilities and reasoning. The team highlights their success in building a data center that hosts over 100,000 GPUs to train Grok 3, making it much more powerful than previous models. They showcase Grok's performance across various benchmarks and introduce a new search feature called Deep Search, designed to enhance user experience. The presentation emphasizes ongoing improvements and invites smart individuals to join their team for further advancements in AI technology.
Punti di forza
- 🤖 Grok 3 is a significant advancement in AI, more capable than Grok 2.
- 🚀 Building a massive data center was essential for training Grok 3.
- 💡 Continuous improvements are being made to the AI model daily.
- 📈 Grok has shown exceptional performance in several benchmarks.
- 🔍 Deep Search is a feature that enhances user interaction and information retrieval.
- 🎮 Grok can create games and solve complex problems like physics trajectories.
- 🗣️ A voice assistant feature is on its way to improve user experience.
- 💻 The team faced various challenges in building the infrastructure for Grok.
- 🤝 The mission is to explore fundamental questions about the universe with AI.
- 📅 Access to Grok is being rolled out to premium users, with plans for expansion.
Linea temporale
- 00:00:00 - 00:05:00
The mission of XAI and GRO is to investigate fundamental questions about the universe, such as the meaning of life and the existence of aliens. They aim for truth, even when it contradicts political correctness, and are excited to introduce GRO 3, a significant advancement over its predecessor, GRO 2, thanks to a dedicated team.
- 00:05:00 - 00:10:00
GRO is an AI tool developed by XAI. The team has worked diligently to enhance GRO's capabilities and improve user interaction. GRO is named after a term from a novel meaning to deeply understand something, and the team has charted significant progress in AI development over a short period, achieving unprecedented performance improvements in various benchmarks.
- 00:10:00 - 00:15:00
The development of GRO encountered challenges, particularly in scaling up GPU training capabilities. After initial trials with fewer GPUs, the decision was made to build a proprietary data center to accommodate the training requirements, which saw rapid progress in a short time, culminating in the largest fully connected GPU cluster, significantly enhancing AI training capacity.
- 00:15:00 - 00:20:00
The advancements in GRO 3 are marked by over tenfold increases in computational capacity compared to GRO 2. The performance improvements have been tested across multiple categories such as mathematical reasoning, general knowledge, and coding, establishing GRO 3's position as a leading model in its class, capable of surpassing competitors in intensive tasks.
- 00:20:00 - 00:25:00
The team conducted blind tests on GRO 3, demonstrating its ability to outperform other AI models across various categories with impressive ELO scores in a competitive assessment. Continuous improvements are being made to GRO 3, which shows promise in advancing reasoning capabilities and solving complex problems effectively.
- 00:25:00 - 00:30:00
GRO has been tested with advanced reasoning tasks like plotting trajectories and generating creative games, displaying its potential to conceptualize new ideas and integrate problem-solving skills. The AI's capability to think and analyze deeply about solutions is evident in its coding abilities, which aim for innovative outputs rather than merely copying existing work.
- 00:30:00 - 00:38:37
The introduction of 'Deep Search,' a next-generation AI search engine, aims to enhance user experience by providing thorough, context-aware information retrieval. This feature promises significant time savings for users, enabling them to receive accurate answers swiftly, while continuing development efforts aim to expand GRO’s capabilities further in real-world applications.
Mappa mentale
Video Domande e Risposte
What is the mission of XAI and Gro?
To understand the universe and explore fundamental questions.
How much more capable is Grok 3 compared to Grok 2?
Grok 3 is claimed to be an order of magnitude more capable.
What improvements have been made to Grok?
Significant enhancements in reasoning capabilities and training efficiency.
What is Deep Search?
A next-generation search engine that helps users find information quickly.
How can I access Grok?
Starting with premium plus subscribers on X, with a dedicated Grok app available.
Will there be a voice assistant feature?
Yes, a voice assistant feature is in development.
What types of tasks can Grok perform?
Grok can handle advanced reasoning, coding tasks, and game creation.
Is Grok open source?
The team plans to open-source Grok when the next stable version is ready.
How quickly is Grok being improved?
The model sees improvements daily, with continuous updates.
What challenges did the team face in building the AI infrastructure?
They faced issues with power, cooling, and ensuring coherent GPU communication.
Visualizza altre sintesi video
- 00:00:02all right welcome to the grock 3
- 00:00:04presentation so the mission of xai and
- 00:00:07Gro is to understand the universe we
- 00:00:10want to understand the nature of the
- 00:00:11universe so we can figure out what's
- 00:00:12going on where are the aliens what's the
- 00:00:14meaning of life how does the universe
- 00:00:15end how did it start all these
- 00:00:17fundamental questions were driven by
- 00:00:19curiosity about the nature of the
- 00:00:20universe and that's also what causes us
- 00:00:23to be a maximally truth-seeking AI even
- 00:00:26if that truth is sometimes at odds with
- 00:00:28what is politically correct
- 00:00:30in order to understand the nature of the
- 00:00:32universe you must absolutely rigorously
- 00:00:34pursue truth or you will not understand
- 00:00:36the universe you'll be suffering from
- 00:00:38some amount of delusion or error that is
- 00:00:40our goal figure out what's going on and
- 00:00:43we're very excited to present gr 3 which
- 00:00:45is we think an order of magnitude more
- 00:00:47capable than gr 2 in a very short period
- 00:00:49of time and that's thanks to the hard
- 00:00:52work of an incredible team and I'm
- 00:00:55honored to work with such a great team
- 00:00:57and of course we'd love to have some of
- 00:00:58the smartest humans out there join us
- 00:01:00team with that let's go hi everyone my
- 00:01:03name is Igor lead engineering at xci I'm
- 00:01:06Jimmy Paul leading research and Tony
- 00:01:08working on the reasoning Team all right
- 00:01:10you I don't do
- 00:01:12anything I just show up occasionally
- 00:01:15yeah like I mentioned Gro is the tool
- 00:01:17that we're working on Gro is our AI that
- 00:01:19we're building here at XI and we've been
- 00:01:20working extremely hard over the last few
- 00:01:22months to improve Gro as much as we can
- 00:01:24so we can give it to all of you so we
- 00:01:25can give all of you access to it um we
- 00:01:27think it's going to be extremely useful
- 00:01:29we think it's going going to be
- 00:01:30interesting to talk to funny really
- 00:01:31funny and we're going to explain to you
- 00:01:33how we've improved Gro over the last few
- 00:01:34months we've made quite a jump in in
- 00:01:36capabilities yeah actually we should
- 00:01:38explain maybe also what is why do we
- 00:01:39call it Gro so Gro is a word from a
- 00:01:41Highland novel Stranger in a Strange
- 00:01:43Land and it's used by a guy who's raised
- 00:01:46on Mars and the word Gro is to fully and
- 00:01:49profoundly understand something that's
- 00:01:51what the word Gro means fully and
- 00:01:52profoundly understand something and
- 00:01:54empathy is important true
- 00:01:58yeah yeah if we charted xas progress in
- 00:02:01the last few months has only been 17
- 00:02:03months since we started kicking off our
- 00:02:06very first model grock one was almost
- 00:02:09like a toy by this point only 314
- 00:02:11billion parameters and now if we PR the
- 00:02:14progress the time on x-axis the
- 00:02:17performance of favorite Benchmark
- 00:02:18numbers at mlu on the y- axis were
- 00:02:21literally progressing at unprecedent
- 00:02:23speed across the whole field and then we
- 00:02:26kick off grock 1.5 right after grock 1
- 00:02:29released after November 2023 and then gr
- 00:02:322 if you look at where the all the
- 00:02:34performance coming
- 00:02:36from when you have a very correct
- 00:02:38engineering team and all the best AI
- 00:02:40taligent the only one thing we need is a
- 00:02:44big intelligence comes from big cluster
- 00:02:47we can reconvert the entire progress of
- 00:02:49X now replacing the Benchmark and the y
- 00:02:51axis to the total amount of training
- 00:02:53flops that is how many gpus we can run
- 00:02:56at any given time to train our large
- 00:02:58language models to impress the entire
- 00:03:01internet so after all human all human
- 00:03:03knowledge really that's right yeah
- 00:03:05internet being part of it but it's
- 00:03:06really all human knowledge all
- 00:03:08everything yeah the whole internet fits
- 00:03:09into a USB stick at this point it's all
- 00:03:11the human tokens yeah that's right yeah
- 00:03:14very soon into the real world yeah so we
- 00:03:16had so much trouble actually training
- 00:03:18grock 2 back in the days we kickoff the
- 00:03:20model around February and we thought we
- 00:03:23had a large amount of chips but turned
- 00:03:25out we can barely get AK training chips
- 00:03:27running coherently at any given time
- 00:03:30and we have so many Cooling and power
- 00:03:33issues I think you were there in the
- 00:03:35data center yeah it was like really more
- 00:03:37like 8K tips on average at 80%
- 00:03:40efficiency more like like 6,500
- 00:03:42effective uh h100s training for you know
- 00:03:46several months but now now we're at 100K
- 00:03:49yeah that's right more than 100K that's
- 00:03:51right so what's the next step right
- 00:03:53after gu to so if we want to continue
- 00:03:56accelerate we have to take the matter
- 00:03:58into our own hands we have to solve all
- 00:03:59the ings all the power issues and
- 00:04:02everything yeah so in April of last year
- 00:04:04Elon decided that really the only way
- 00:04:06for XI to succeed for XI to build the
- 00:04:08best AI out there is to build our own
- 00:04:10data center we didn't have a lot of time
- 00:04:12that because we wanted to give you gr
- 00:04:13free as quickly as possible so really we
- 00:04:16realized we have to build the data
- 00:04:17center in about 4 months and turned out
- 00:04:20it took us 122 days to get the first
- 00:04:22100K gpus up and running and there was a
- 00:04:24Monumental effort to be able to do that
- 00:04:27it's we believe it's the biggest fully
- 00:04:29connected h100 cluster of its kind and
- 00:04:32we didn't just stop there we actually
- 00:04:33decided that we need to double the size
- 00:04:35of the cluster pretty much immediately
- 00:04:37if we want to build uh the kind of AI
- 00:04:39that we want to build so we then had
- 00:04:42another phase which we haven't talked
- 00:04:43about publicly yet so this is the first
- 00:04:44time that we're talking about this where
- 00:04:46we doubled the capacity of the data
- 00:04:48center yet again and that one only took
- 00:04:50us 92 days we've been able to use all of
- 00:04:53these GPS use all this compute to
- 00:04:54improve grock in the meantime and
- 00:04:56basically today we're going to present
- 00:04:58you the results of that the the fruits
- 00:05:00that came from that that's yeah all the
- 00:05:03path all the RADS leads to gr 3 10x more
- 00:05:06compute more than 10x really yeah really
- 00:05:08maybe 15x is yep compared to our
- 00:05:11previous generation model and gr
- 00:05:13finished the pre-training early January
- 00:05:16and we start you know the model still
- 00:05:18currently training actually this is a
- 00:05:19little preview of our Benchmark numbers
- 00:05:22so we evaluated gr 3 on three different
- 00:05:26categories U General mathematical
- 00:05:28reasonings on general knowledge about
- 00:05:31stem and Science and then also on
- 00:05:34computer science coding Amy uh American
- 00:05:37Invitational math examination host it
- 00:05:40once a year uh and if we evaluate the
- 00:05:43model performance we can see that the gr
- 00:05:463 across the board is in a league of its
- 00:05:48own even his little brother gr 3 mini is
- 00:05:52reaching the froner across all the other
- 00:05:55competitors you will say at this point
- 00:05:58all these benchmarks you just evaluating
- 00:06:00the memorization of the textbooks
- 00:06:02memorization of the GitHub repost how
- 00:06:04about Real Time usefulness how about we
- 00:06:06actually use those models in our product
- 00:06:08what we did instead is we actually
- 00:06:11kicked off a blind test of our gra 3
- 00:06:14Model code named Chocolate it's pretty
- 00:06:16hot yeah hot chocolate and I've been
- 00:06:18running on this platform called CH arena
- 00:06:21for two weeks I think the entire X
- 00:06:24platform at some point speculated this
- 00:06:26might be the next generation of a uh AI
- 00:06:28com me your way how this chat Arena
- 00:06:31works is that it strip away the entire
- 00:06:34product surface right it just raw
- 00:06:35comparison of the engine of those AGI
- 00:06:38the language models themselves and place
- 00:06:40interface where the user will submit one
- 00:06:42single query and you get to show two
- 00:06:44responses you don't know which model
- 00:06:46they come from and in then you make the
- 00:06:47vote so in this blind test gr 3 an early
- 00:06:51version of gr 3 already reached 1,400 no
- 00:06:55other models has reached an ELO score
- 00:06:57had to have comparison to all the other
- 00:06:59models at this score and it's not just
- 00:07:02one single category it's 1400 aggregated
- 00:07:05across all the categories in chall
- 00:07:07capabilities instruction following
- 00:07:09coding so it's number one across the
- 00:07:12board in this blind test and it's it's
- 00:07:13still climbing so we actually to keep
- 00:07:15updating it so it's it's 14,400 about
- 00:07:1814400 in climbing yeah in fact we have a
- 00:07:20version of the model that we think is
- 00:07:21already much better than the one that we
- 00:07:22tested here yeah we'll see how far it
- 00:07:24gets uh but that's the one that we're
- 00:07:26working on we talking about today yeah
- 00:07:28so actually one thing if if you're if
- 00:07:30you're using grock 3 you I think you may
- 00:07:31notice improvements almost every day um
- 00:07:33because we're we're continuously
- 00:07:35improving the model so Lally even within
- 00:07:3824 hours you'll see
- 00:07:39improvements yep but we believe here at
- 00:07:42XI getting the best pre-training model
- 00:07:45is not enough that's not enough to build
- 00:07:47the best AI and the best AI need to
- 00:07:49think like a human you to contemplate
- 00:07:51about all the possible
- 00:07:53solutions self-critique verify all the
- 00:07:56solutions backtrack and also think from
- 00:07:59the first principle that's a very
- 00:08:01important capability so we believe that
- 00:08:04as we take the best pre-train model and
- 00:08:06continue training with reinforcement
- 00:08:08learning it will elicit the additional
- 00:08:10reasoning capabilities that allows the
- 00:08:12model to become so much better and scale
- 00:08:15not just in the training time but
- 00:08:17actually in the test time as well we
- 00:08:19already found the model is extremely
- 00:08:20useful internally for our own
- 00:08:22engineering saving hours of time
- 00:08:24hundreds of hours of coding time equ
- 00:08:26you're the power user of our graic
- 00:08:28reasoning model what are some use cases
- 00:08:30yeah so like Jimmy said we've added
- 00:08:31Advanced reasoning capabilities to Grog
- 00:08:33and we've been testing them pretty
- 00:08:34heavily over the last few weeks in order
- 00:08:36to give you a little bit of a taste of
- 00:08:37what it looks like when Gro is solving
- 00:08:39heart reasoning problems so we prepared
- 00:08:41two little problems for you one comes
- 00:08:43from physics and one is actually a game
- 00:08:45that gr is going to ride for us when it
- 00:08:47comes to the physics problem what we
- 00:08:48want gr to do is to plot a viable
- 00:08:50trajectory to do a transfer from Earth
- 00:08:53to Mars and then at a later point in
- 00:08:55time a transfer back from Mars to Earth
- 00:08:57and that requires some some Physics that
- 00:08:59Gro will have to understand so we're
- 00:09:01going to challenge Gro come up with a
- 00:09:02viable trajectory calculate it and then
- 00:09:05plot it for us so we can see it and yeah
- 00:09:08this is totally unscripted by the way
- 00:09:10this is the that's the entirety of the
- 00:09:12prompt which should be clarify is that
- 00:09:13there's nothing more than that yeah
- 00:09:15exactly this is the gro interface and
- 00:09:17we've typed in this text that you can
- 00:09:19see here generate code for an animated
- 00:09:213D plot of a launch from Earth landing
- 00:09:24on Mars and then back to Earth at the
- 00:09:26next launch window and we've not kicked
- 00:09:28off or the query and you can see Gro is
- 00:09:29thinking part of grock's advanced
- 00:09:32reasoning capabilities are these
- 00:09:33thinking traces that you can see here
- 00:09:35you can even go inside and actually read
- 00:09:37what gr is thinking as it's going
- 00:09:38through the problem as it's trying to
- 00:09:39solve it yeah which we are doing some
- 00:09:42obscuration of the thinking so that our
- 00:09:44model doesn't get totally copied
- 00:09:45instantly so there's more to the
- 00:09:48thinking than is displayed in yeah and
- 00:09:52because this is totally unscripted
- 00:09:54there's actually a chance that grock
- 00:09:55might made a little coding mistake and
- 00:09:57it might not actually work just in case
- 00:09:58we're going to launch two more instances
- 00:10:00of this so if something goes wrong we
- 00:10:02were able to to switch to those and show
- 00:10:05you something that's presentable so
- 00:10:07we're kicking off the other two as well
- 00:10:09and like I said we have a second problem
- 00:10:11as well and yeah actually one of the
- 00:10:13favorite one of our favorite activities
- 00:10:15here XI is having Grog right games for
- 00:10:17us and not just any know any old game
- 00:10:21any game that you might already be
- 00:10:22familiar with but actually creating new
- 00:10:23games on the spot and being creative
- 00:10:25about it so one example that we found
- 00:10:27was really fun is create a game that's a
- 00:10:30mixture of the two games Tetris and B so
- 00:10:34this is that maybe an important thing
- 00:10:35like this obviously if you ask an AI to
- 00:10:38create a game like Tetris there's there
- 00:10:39are many examples of Tetris on the
- 00:10:40Internet or game like J whatever there
- 00:10:44it can copy it what's interesting here
- 00:10:46is it achieved a creative solution
- 00:10:49combining the two games that actually
- 00:10:51works and and is a good game yeah that's
- 00:10:54the we're seeing the beginnings of
- 00:10:57creativity yeah fingers crossed that we
- 00:11:00can recreate that hopefully it works
- 00:11:01hope so actually because this is a bit
- 00:11:03more challenging we're going to use
- 00:11:05something special here which we call Big
- 00:11:06Brain that's our mode in which we use
- 00:11:09more computation which more reasoning of
- 00:11:11our gr just to make sure that there's a
- 00:11:13good chance here that it might actually
- 00:11:14do it so we're also going to fire off
- 00:11:16three attempts here at at solving this
- 00:11:19game at creating this game that's a
- 00:11:21mixture of Tetris and B yeah let's let's
- 00:11:24see what go comes up like I've played
- 00:11:25the game it's pretty good like it's like
- 00:11:28wow okay this is something yeah um so
- 00:11:31while Gro is thinking uh in the in the
- 00:11:33background um we can now actually talk
- 00:11:34about some concrete know how how well is
- 00:11:36Gro doing across tons of different tasks
- 00:11:38that we've tested on um so we'll hand it
- 00:11:40over to Tony to talk about that yeah
- 00:11:43okay so let's see how Grog does on those
- 00:11:46interesting challenging benchmarks so
- 00:11:48yeah so reasoning again refers to those
- 00:11:50models that actually thinks quite for
- 00:11:52quite a long time before it tries to
- 00:11:54solve a problem in this case around a
- 00:11:56month ago the graph 3 pre-training
- 00:11:58finishes after that we work very hard to
- 00:12:01put the reasoning capability into the uh
- 00:12:03current graph 3 Model but again this is
- 00:12:06very early days so the model is still
- 00:12:08currently in training right now what
- 00:12:09we're going to show to people is this
- 00:12:12beta version of the gry reasoning model
- 00:12:14alongside we also are training a mini
- 00:12:16version of the reasoning model
- 00:12:18essentially on this plot you can see the
- 00:12:20gr 3 reasoning beta and then gr 3 mini
- 00:12:22reasoning the grth reason mini reasoning
- 00:12:24is actually a model that we train for
- 00:12:26much longer time and you can see that
- 00:12:28sometimes it actually perform study
- 00:12:29better compared to the gr three
- 00:12:31reasoning this also just means that
- 00:12:33there's a huge potential for the grth
- 00:12:35three reasoning because it's trained for
- 00:12:36much less time all right so let's
- 00:12:38actually look at what how it does on
- 00:12:40those three benchmarks so Jimmy also
- 00:12:42introduced already so essentially we're
- 00:12:44looking at three different areas
- 00:12:46mathematics science and coding and for
- 00:12:48math we're picking this high school
- 00:12:50competition math problem for science we
- 00:12:52actually pick those PhD level science
- 00:12:54questions and for coding it's also
- 00:12:56actually pretty challenging it's
- 00:12:57competitive coding and also some leod
- 00:13:00which is some cold inter interview
- 00:13:01problems that people usually get when
- 00:13:03they interview for companies so on those
- 00:13:05benchmarks you can see that the gr 3
- 00:13:07actually perform quite well across the
- 00:13:09board compared to other competitors um
- 00:13:12yeah so it's pretty promising these
- 00:13:14models are very smart so Tony what what
- 00:13:16what are those shaded bars yeah so okay
- 00:13:19so uh I'm glad you asked this question
- 00:13:21so for those models because it can
- 00:13:23reason it can thinks you can also ask
- 00:13:25them to even think longer uh you can
- 00:13:27spend more what we call test and compute
- 00:13:31which means you can spend more time to
- 00:13:33reason to think about a problem before
- 00:13:35you spit out the answer so in this case
- 00:13:38the Shaded bar here means that we just
- 00:13:41ask the model to spend more time you can
- 00:13:43solve the the same problem many times
- 00:13:45before it it tries to conclude what is
- 00:13:47the right solution and once you give
- 00:13:49this compute or this kind of budget to
- 00:13:51the model it turns out the model can
- 00:13:53even perform better so this is
- 00:13:55essentially the Shaded bar in in those
- 00:13:57BX so this is really exciting right
- 00:14:00because now instead of just doing one
- 00:14:01chain of thoughts with AI why not do
- 00:14:04multiple once yes so that's a very
- 00:14:06powerful technique that allows to
- 00:14:07continue scale the model capabilities
- 00:14:09after training and people often ask are
- 00:14:12we actually just over fitting to the
- 00:14:14benchmarks yes so how about your oration
- 00:14:16so yes I think yeah this is definitely a
- 00:14:18question that we are asking ourselves
- 00:14:20whether we are overfitting to those
- 00:14:22current benchmarks luckily we have a
- 00:14:24real test so about 5 days ago Amy 2025
- 00:14:28just finished this is where high school
- 00:14:30students compete in this particular
- 00:14:32Benchmark so we got this very fresh new
- 00:14:35competition and then we asked our two
- 00:14:37models to compete on the same Benchmark
- 00:14:39at the same exam and it turns out very
- 00:14:41interestingly the grth three reasoning
- 00:14:43the big one actually does better on this
- 00:14:46particular new fresh exam this also
- 00:14:48means that the generalization capability
- 00:14:50of the big model is stronger much
- 00:14:52stronger compared to smaller model if
- 00:14:54you compare to the last year's exam
- 00:14:55actually this is the opposite the
- 00:14:57smaller model kind of learned
- 00:14:59the the previous exams better yeah so
- 00:15:02this this actually shows some kind of
- 00:15:03true generalization from the model
- 00:15:05that's right so 17 months ago our gr
- 00:15:07zero and Gro one barely solves any High
- 00:15:09School problems that's right and now we
- 00:15:11have a kid that just already graduate
- 00:15:13the gro Gro is ready to go to college is
- 00:15:15that right yeah it won't be long before
- 00:15:18it's simply perfect the human exams
- 00:15:19won't be hard they be too easy yeah and
- 00:15:22internally we actually as gret continue
- 00:15:24evolves we're going to talk about what
- 00:15:26we're excited about but very soon there
- 00:15:29will be no more Benchmark left
- 00:15:31yeah yeah one thing that's quite
- 00:15:33fascinating I think is that we basically
- 00:15:35only trained Rock's reasoning abilities
- 00:15:36on math problems and comparative coding
- 00:15:39problems right so very specialized kinds
- 00:15:41of tasks but somehow it's able to work
- 00:15:44on all kinds of other different tasks so
- 00:15:46including creating games no lots lots
- 00:15:48and lots of different things and what
- 00:15:50seems to be happening is that basically
- 00:15:51Gro learns this ability to detect its
- 00:15:54own mistakes and its thinking correct
- 00:15:55them persist on a problem try lots of
- 00:15:57different variants pick pick the one
- 00:15:59that's best so there are these
- 00:16:00generalized generalizing abilities that
- 00:16:02Gro learns from mathematics and from
- 00:16:04coding which it can then use to solve
- 00:16:06all kinds of other problems that's
- 00:16:07pretty reality is the instantiation of
- 00:16:09mathematics that's right and one thing
- 00:16:12we're actually really excited about that
- 00:16:13going back to our funing mission is what
- 00:16:15if one day we have a computer just like
- 00:16:17deep thought that utilize our entire
- 00:16:20cluster just for that one very important
- 00:16:22problem in the test time all the GPU
- 00:16:24turned on right so I think back then we
- 00:16:26were building the GPU clusters together
- 00:16:28you plug
- 00:16:29cables and I remember that when we turn
- 00:16:32on the first initial test you can hear
- 00:16:34all the GPS humming in the hallway
- 00:16:37that's almost feel like spiritual yeah
- 00:16:39that's actually a pretty cool uh thing
- 00:16:40that we're able to do that we can go
- 00:16:42into the data center and Tinker with the
- 00:16:44machines there so for example we went in
- 00:16:46and we unplugged a few of the cables and
- 00:16:49just made sure that our training setup
- 00:16:50is still running stably so that's
- 00:16:52something that I think most uh AI teams
- 00:16:55out there don't usually do but it's
- 00:16:56actually totally unlocks like a new
- 00:16:58level of reliability and what you're
- 00:17:00able to do with the hardware so okay so
- 00:17:02when when are we going to solve
- 00:17:04remon the easiest solution is to
- 00:17:07numerate over all possible strains and
- 00:17:10as long you have a verifier enough
- 00:17:11compute you'll be able to do it okay my
- 00:17:14projection will be what's your guess
- 00:17:16what is your neural n calculate my my Bo
- 00:17:18prodiction so three years ago I told you
- 00:17:20this I think in now two years later two
- 00:17:23things going to happen we're going to
- 00:17:24see machines win some medals yes touring
- 00:17:28award absolutely
- 00:17:29Fields metal Nobel Prize with probably
- 00:17:32some expert in the loop right so the
- 00:17:34expert uplifting do you mean so this
- 00:17:35year or next year oh okay that's what it
- 00:17:39comes down to really yeah so it looks
- 00:17:42like Gro finished know all of its
- 00:17:43thinking on on the two problem so let's
- 00:17:45take a look at what it
- 00:17:47said all right so this was the little
- 00:17:50physics problem we had no we've
- 00:17:51collapsed the thoughts here so they're
- 00:17:53they're hidden and then we see grock's
- 00:17:55answer below that so it explains it
- 00:17:56wrote a python script here using M plot
- 00:17:58Li then gives us all of the code so
- 00:18:01let's take a quick look at the code
- 00:18:02seems like it's doing reasonable things
- 00:18:04here not totally of the mark solve
- 00:18:07Kepler says here so maybe it's solving
- 00:18:09Kepler's laws cap Kepler law numerically
- 00:18:12um yeah there's really only one way to
- 00:18:14find out if this thing is working I'd
- 00:18:16say let's give it a try let's run the
- 00:18:17code all right and we can see yeah gr is
- 00:18:20animating two different planets Earth
- 00:18:22and Mars here and then the green uh ball
- 00:18:25is the vehicle that's transiting the
- 00:18:27spacecraft that's transitioning between
- 00:18:29Earth and Mars and you could see the
- 00:18:30journey from Earth to Mars and looks
- 00:18:32like yeah indeed the astronauts return
- 00:18:35safely at the right moment in time now
- 00:18:38obviously this was just generated on the
- 00:18:39spots now we can't tell you if that was
- 00:18:41actually correct solution so we're going
- 00:18:42to take a closer look now maybe we're
- 00:18:43going to call some colleagues from space
- 00:18:45X ask them if if this is legit um it's
- 00:18:48pretty close it's it's uh I mean there's
- 00:18:51a lot of complexities in the actual
- 00:18:53orbits that have to be taken into
- 00:18:54account but this is pretty close to to
- 00:18:55what it what it looks like awes in fact
- 00:18:57I have that on my pend here got the
- 00:19:00Earth Mars home and transfer on
- 00:19:03it when are we going to install groc on
- 00:19:06a
- 00:19:07rock I suppose in two years two years
- 00:19:12everything is two years away Earth and
- 00:19:14Mars Transit can occurs every 26 months
- 00:19:17the next we're currently in a Transit
- 00:19:18window approximately the next one would
- 00:19:20be November of next year roughly end of
- 00:19:24next year and if all goes well SpaceX
- 00:19:27will send a Starship Rockets to Mars and
- 00:19:30with Optimus robots and and Gro
- 00:19:34mhm yeah I'm curious about this
- 00:19:36combination of Tetris and B looks like
- 00:19:39the tetris as we've named it internally
- 00:19:43okay we also have an output from go here
- 00:19:45it says Ro python script explains that
- 00:19:47it's what it's been doing if you look at
- 00:19:49the code there are some constants that
- 00:19:51are being defined here some colors then
- 00:19:53the trinos the pieces of Tetris are
- 00:19:56there obviously very hard to see and at
- 00:19:59one glance if this is good so we got to
- 00:20:00run this to figure out if it's working
- 00:20:02let's give it a
- 00:20:03try fingers crossed all right right so
- 00:20:06this kind of looks like Tetris uh but
- 00:20:08the the colors are a little bit off
- 00:20:10right the colors are different here and
- 00:20:12if you think about what's going what's
- 00:20:14going on here the J has this mechanic
- 00:20:17where if you get three jws in a row you
- 00:20:19know then they they disappear and also
- 00:20:22gravity activates right so uh what
- 00:20:24happens if you get three of the colors
- 00:20:26together okay so something happens so so
- 00:20:28I think what SC did in this version is
- 00:20:31that once you connect three at least
- 00:20:33three blocks of the same color in a row
- 00:20:35then gravity activates and they
- 00:20:38disappear and then gravity activates and
- 00:20:40all the other blocks fall down curious
- 00:20:42if there's still a Tetris mechanic here
- 00:20:44where if the line is full does it
- 00:20:46actually clear it or what happens then
- 00:20:49it's up to interpretation who knows yeah
- 00:20:51I mean it'll do different variants when
- 00:20:53you ask it it doesn't do the same thing
- 00:20:54every time exactly we've seen a few
- 00:20:56other the tetris that work very
- 00:20:58differently but this one seems cool yeah
- 00:21:01are we ready for game Studio at x. a yes
- 00:21:04so we're launching uh an AI gaming
- 00:21:06studio at xci if you're interested in
- 00:21:08joining us and building AI games please
- 00:21:10join XI we're launching an AI gaming
- 00:21:12studio we're announcing it tonight let's
- 00:21:15go epic games but right that's an actual
- 00:21:19games yeah yeah all right so I think one
- 00:21:24thing is super exciting for us is that
- 00:21:26once you have the best pre train model
- 00:21:29you have the best reason model right we
- 00:21:31already see that we actually give the
- 00:21:33capability for those model to think
- 00:21:34harder think longer think more broad the
- 00:21:38performance continue improves and we're
- 00:21:40really excited about the next front here
- 00:21:42that what happen if we're not only allow
- 00:21:44the model to think harder but also
- 00:21:45provide more tools this I call real
- 00:21:47humans to solve those problems for real
- 00:21:50humans we don't ask them to solve reman
- 00:21:52a hypothesis just with a piece of pen
- 00:21:54and paper no internet with all the basic
- 00:21:57web browsing search engine and code
- 00:22:00interpreters that builds the foundations
- 00:22:03and the best reasoning model builds the
- 00:22:05foundations for the gr agent to come
- 00:22:08today we're actually introducing a new
- 00:22:11product called Deep search that is the
- 00:22:13first generation of our gr agents that
- 00:22:16not just helping the engineers and
- 00:22:17research and scientists to do coding but
- 00:22:19actually help everyone to answer
- 00:22:21questions that you have day today it's
- 00:22:23like a Next Generation search engine
- 00:22:25that really help you to understand the
- 00:22:26universe you can start asking question
- 00:22:29like for example hey when is the next
- 00:22:32Starship launch day for example let's
- 00:22:34try that get the answer on the left hand
- 00:22:37side we see a high level progress bar
- 00:22:39essentially the model now is going to do
- 00:22:41one single search like the current rack
- 00:22:43system but actually thought very deeply
- 00:22:45about hey what's the user intent here
- 00:22:47and what are the facts I should consider
- 00:22:49at the same time and how many different
- 00:22:51website I should actually go and read
- 00:22:52their content right so this can really
- 00:22:55save hundreds hours of everyone's Google
- 00:22:58time if you want to really look into
- 00:22:59certain topics and then on the right
- 00:23:02hand side you can see the bullet
- 00:23:04summaries of how the current model is
- 00:23:06doing what websites browsing what
- 00:23:08sources is verifying and often time
- 00:23:10actually cross validate different
- 00:23:11sources out there to make sure the
- 00:23:13answer is actually correct before it's
- 00:23:14output final answer and we can at the
- 00:23:16same time fire up a few more queries um
- 00:23:19how about you know you're a gamer right
- 00:23:21uh sure yeah so how about what are some
- 00:23:23of the best builds and most popular
- 00:23:25builds in path Excel hardcore right
- 00:23:27hardcore League you can technically just
- 00:23:30look at the hardcore ladder might be a
- 00:23:33fast way to figure it out yeah we'll see
- 00:23:34what model
- 00:23:36does um and then we can also do
- 00:23:39something more fun for example how about
- 00:23:41make a prediction about the March
- 00:23:42Madness out there yeah so this is go fun
- 00:23:44one where Warren Buffett has a billion
- 00:23:47dollar vet if you can exactly match the
- 00:23:50I think the the sort of the entire
- 00:23:52winning tree of marsh Madness you can
- 00:23:54win a billion dollars from Warren
- 00:23:55Buffett it would be pretty cool if AI
- 00:23:57could help you win a billion dollars
- 00:23:59from
- 00:24:00Buffett that seems like a pretty good
- 00:24:02investment let's go yeah all right so
- 00:24:05now let's fire up the query and see what
- 00:24:07model does so we can actually go back to
- 00:24:09our very first one how about the buffet
- 00:24:11wasn't counting on this it's sry done
- 00:24:14that's right okay so we got the result
- 00:24:16of the first one the model thought
- 00:24:17around one minute uh so okay so the key
- 00:24:19Insight here the next Starship is going
- 00:24:21to be on 24th or later so no earlier
- 00:24:24than February
- 00:24:2524th it might be sooner
- 00:24:29yeah so I think we can go down scroll
- 00:24:31down what what the model does so it does
- 00:24:32a little research on the fight 7 what
- 00:24:34happened got grounded and actually it
- 00:24:36look into the FCC filing from this data
- 00:24:39Collections and then actually make the
- 00:24:42new conclusion that yeah if we continue
- 00:24:43scroll down let's see yeah so it makes
- 00:24:46the little table I think inside xai we
- 00:24:49often joked about the time to the first
- 00:24:51table is the only latency that matters
- 00:24:54yeah so that's how to model make
- 00:24:56inference and look up all the sources
- 00:24:58and then we can look into the gaming one
- 00:25:00so how about
- 00:25:04the for this particular one we look at
- 00:25:07hey the the build is
- 00:25:10light with the The Infernal is but if we
- 00:25:13go down the surprising fact of all the
- 00:25:15other builds look into the 12 classes
- 00:25:18yeah we'll see that the Min build was
- 00:25:20pretty popular whenever the game first
- 00:25:21came out and now the invokers of the
- 00:25:23world took over invoker monk invoker for
- 00:25:26sure yeah that's right yeah by the stone
- 00:25:28wavers and that's really good at mapping
- 00:25:30yeah and then we can see the the match
- 00:25:33manness about that one one interesting
- 00:25:35thing about the Deep search is that if
- 00:25:36you actually go into the panel where it
- 00:25:39shows what are the subtasks you can
- 00:25:41actually click the bottom left and then
- 00:25:44in this case you can actually scroll
- 00:25:45through actually reading through the
- 00:25:47mind of grock what informations does the
- 00:25:49model actually think about are
- 00:25:51trustworthy what are not how does they
- 00:25:52actually cross validate different
- 00:25:53information sources so that makes the
- 00:25:56entire search experience and information
- 00:25:57retrieval process a lot more transparent
- 00:25:59to our
- 00:26:01users and this is much more powerful
- 00:26:03than any search engine out there you can
- 00:26:06literally just tell it only use sources
- 00:26:08from X will try to respect that yeah and
- 00:26:10so it's much more steerable much more
- 00:26:12intelligent than it really should save
- 00:26:14you a lot of time so something that
- 00:26:15might take you half an hour or an hour
- 00:26:17of researching on the web or searching
- 00:26:19social media you can just ask it to go
- 00:26:21do that and and come back in 10 minutes
- 00:26:23later it's done an hour's worth of work
- 00:26:25for you that's really what it comes down
- 00:26:26to exactly and maybe better than you
- 00:26:28could have done it yourself yeah think
- 00:26:30about you have INF of interns working
- 00:26:32for you now you can just fire up all the
- 00:26:34tasks and come back a minute later so
- 00:26:36this is going to be interesting one so
- 00:26:37March M had not happened yet so I guess
- 00:26:40we have to follow up with a next live
- 00:26:42stream yeah it seems like pretty good
- 00:26:45the $40 might get you a billion dollars
- 00:26:47$40 subscription that's right my work
- 00:26:51yeah so when are the users going to have
- 00:26:53their hands on gr to yeah so the the
- 00:26:55good news is we've been working
- 00:26:56tirelessly to actually release all of
- 00:26:59these features that we've shown you the
- 00:27:00Grog free base model with amazing chat
- 00:27:02capabilities that's really useful that's
- 00:27:03really interesting to talk to the Deep
- 00:27:05search the advanced reasoning mode all
- 00:27:07of these things we want to roll them out
- 00:27:09to you today starting with the premium
- 00:27:12plus subscribers on X so it's the first
- 00:27:14group that will initially get access
- 00:27:16make sure to update your X app if you
- 00:27:18want to see all of the advanced
- 00:27:19capabilities because we just released
- 00:27:21the update now as we're talking here and
- 00:27:23yeah if you're interested in getting
- 00:27:24early access to go then sign up for
- 00:27:26premium plus and also we're announcing
- 00:27:28that we're starting a separate
- 00:27:30subscription for grock that we call
- 00:27:31Super Gro for those who those real grock
- 00:27:34fans that want the most advanced
- 00:27:35capabilities and earliest access to new
- 00:27:38features so feel free to check that out
- 00:27:40as well this this is for the dedicated
- 00:27:42grock app and for the website ex website
- 00:27:44so our our new website is called gro.com
- 00:27:46yeah and you'll also find you never
- 00:27:47guess yeah you never guess and you can
- 00:27:50also find our grock app in the IOS app
- 00:27:52store and that gives you like a more Pol
- 00:27:55even even more polished uh experience
- 00:27:56that's totally grock focused if you're
- 00:27:58if you want to have Gro easily available
- 00:28:00one Tap Away yeah the version on gro.com
- 00:28:03on on a web browser is going to be the
- 00:28:04most the latest and most advanced
- 00:28:06version because obviously takes us a
- 00:28:07while to get thing get something into an
- 00:28:10app and then get it approved by the app
- 00:28:11store and then it's if something's on a
- 00:28:13phone format there limitations what you
- 00:28:15can do so the most powerful version of
- 00:28:16grock and the latest version will be the
- 00:28:18web version at gro.com yeah so watch out
- 00:28:20for the name grock free in the app did
- 00:28:22giveaway yeah exactly that that's that's
- 00:28:24the giveaway that you have groe and if
- 00:28:26it says gr through then GR hasn't quite
- 00:28:28arrived for yet but we're working hard
- 00:28:30to roll this out today and then to even
- 00:28:32more people over the the coming days
- 00:28:34yeah make sure you update your phone app
- 00:28:36too where you're actually going to get
- 00:28:37all the tools we're showcase today with
- 00:28:39the thinking mode with the Deep search
- 00:28:42so yeah really looking forward to all
- 00:28:43the feedbacks you have yeah and I think
- 00:28:45we we should uh emphasize that this is a
- 00:28:48beta meaning that it's you should expect
- 00:28:50some imperfections at first but we will
- 00:28:52improve it rapidly almost every day in
- 00:28:54fact every day I think it'll get better
- 00:28:56if you want a more polished version I'd
- 00:28:57like maybe wait a week but expect
- 00:28:59improvements literally every day and
- 00:29:01then we're also going to be providing a
- 00:29:03voice interaction so you can have
- 00:29:05conversational in fact I was trying it
- 00:29:06earlier today it's working pretty well
- 00:29:08but not we need these a bit more polish
- 00:29:10the sort of way we can just literally
- 00:29:11talk to it like you're talking to a
- 00:29:12person it's that's awesome it's actually
- 00:29:15I think one of the best experiences of
- 00:29:16gr but that's probably about a week
- 00:29:19away yeah with that said well I think we
- 00:29:23might have some audience questions sure
- 00:29:25yeah all right let's take a look yeah
- 00:29:28let's take a look the the audience from
- 00:29:30the as platform yeah so the first
- 00:29:33question here is when grock voice
- 00:29:35assistant when is it coming out yeah as
- 00:29:37as as soon as possible just like Elon
- 00:29:39said just a little bit of polishing away
- 00:29:41from being reled to everybody obviously
- 00:29:44it's going to be released in an early
- 00:29:45form and we're going to rapidly iterate
- 00:29:47on it Y and the next question is like
- 00:29:49when will gr 3 be in the API so this is
- 00:29:52coming in uh the gr 3 API with both the
- 00:29:56reasoning models and deep is coming your
- 00:29:58way in the coming weeks we're actually
- 00:30:00very excited about the Enterprise use
- 00:30:01cases of all these additional tools that
- 00:30:03now gr has access to and how the test
- 00:30:05time compute and to use car to really
- 00:30:07accelerate all the business use cases
- 00:30:09another one is Will voice mode be native
- 00:30:12or text to speech so I think that means
- 00:30:13is it going to be one one model that is
- 00:30:16understanding what you say and then
- 00:30:18talking back to you or is it going to be
- 00:30:19some system that has text to speech
- 00:30:21inside of it and the good news is it's
- 00:30:22going to be one model like a variant of
- 00:30:24gr free that we're going to release
- 00:30:26which basically understands what you're
- 00:30:28say what you're saying and then uh
- 00:30:30generates the audio directly from that
- 00:30:32so very much like Grog free generates
- 00:30:34text that model generates audio and that
- 00:30:36has a bunch of advantages I was talking
- 00:30:38to it earlier today and it said hi igore
- 00:30:40reading my my name from probably from
- 00:30:42some text that it had um and I said no
- 00:30:44no my name is Igor and it remember that
- 00:30:47you know so it could continue to say
- 00:30:48Igor just like a human word and you
- 00:30:51can't achieve that with with Tex of
- 00:30:52speech yeah oh here's a question for you
- 00:30:54pretty spicy um you know is Gro a boy or
- 00:30:58girl and are they sing C is whatever you
- 00:31:00want it to
- 00:31:02be yeah yeah are they
- 00:31:05single
- 00:31:07yes all right the shop is open um so
- 00:31:11honestly people are going to fall in
- 00:31:12love with crocet since it's 1,000%
- 00:31:15probable yeah MH uh the next question
- 00:31:18will Gro be able to transcribe audio
- 00:31:20into text yes so we'll have this
- 00:31:22capability both the app and also the API
- 00:31:25we found that gr should just be your
- 00:31:26personal assistant looking over your
- 00:31:27shoulder
- 00:31:28right and follow you along the way learn
- 00:31:30everything you have learned and really
- 00:31:31help you to understand the world better
- 00:31:33become smarter every day yeah the voice
- 00:31:36metag doesn't isn't simply it's not just
- 00:31:38voice text it understands tone
- 00:31:40inflection pacing everything it's wild
- 00:31:42it's like talking to a
- 00:31:44person okay yeah so any plans for
- 00:31:47conversation memory yeah absolutely
- 00:31:50we're working on it right now not really
- 00:31:54forg that's right um let's see what are
- 00:31:58the other
- 00:31:59ones so what about the the DM features
- 00:32:04right so if you have personalizations
- 00:32:06and if if you have remembers your
- 00:32:08previous interactions yes should it be
- 00:32:11one Gro or multiple different grocs
- 00:32:13except to you you can have one Gro or
- 00:32:15many GRS I suspect people will probably
- 00:32:17have more than one yeah I want to have a
- 00:32:20do Gro yeah the gro
- 00:32:23dog that's right all right cool so in
- 00:32:27the past open source grock one right so
- 00:32:30somebody's asking is are we going to do
- 00:32:31that again with gr 2 yeah I think one
- 00:32:34once gr our general approach is that we
- 00:32:36will open source the last version when
- 00:32:38the next version is fully out like when
- 00:32:41gr 3 is mature and stable which is
- 00:32:43probably within a few months then we'll
- 00:32:46open source gr too okay so we probably
- 00:32:48have time for one last question what was
- 00:32:50the most difficult part about working on
- 00:32:52this project I assume gr 3 and what I
- 00:32:55most excited about I think me looking
- 00:32:57looking back getting the whole model
- 00:32:59training on the 100K h100 coherently
- 00:33:03that's almost like battling against the
- 00:33:05final boss of the universe the entropy
- 00:33:07because any given time you can have a
- 00:33:09cosmic rate that beaming down and flip a
- 00:33:11bit in your transistor and now the
- 00:33:13entire gring update if it's fit Mana bit
- 00:33:16the entire grading update is out of
- 00:33:18whack and now you have 100,000 of those
- 00:33:20and you have to orchestrate them every
- 00:33:22time any at any given time any of gpus
- 00:33:24can go down yeah it's worth breaking
- 00:33:27down like how were we able to get the
- 00:33:29world's most powerful training cluster
- 00:33:31operational within 122 days because when
- 00:33:34we started off we actually weren't
- 00:33:35intending to do a data center ourselves
- 00:33:37we were going to just we went to the
- 00:33:39data center providers and said how long
- 00:33:40would it take to have 100,000 gpus
- 00:33:43operating coherently in a single
- 00:33:45location and we got time frames from 18
- 00:33:47to 24 months so like 18 to 24 months
- 00:33:50that means losing as a certainty so the
- 00:33:52only option was to do it ourselves so
- 00:33:55then if you break down the problem I
- 00:33:56guess I'm doing like reasoning here like
- 00:33:59makes you think um one single chain
- 00:34:01though exactly we needed a building we
- 00:34:03can't build a building so we must use an
- 00:34:04existing building so we looked for for
- 00:34:07basically for factories that had been
- 00:34:09were that had been abandoned but the
- 00:34:11factory was in good shape like a company
- 00:34:13had gone bankrupt to something so we
- 00:34:14found an electrox Factory in memph in
- 00:34:16Memphis that's why it's in Memphis home
- 00:34:18of Alvis and also one of the oldest I
- 00:34:20think it was the capital of ancient
- 00:34:21Egypt and it was actually very nice
- 00:34:24Factory that I know for whatever reason
- 00:34:26that electrox had left and uh that that
- 00:34:29gave us shelter for the computers uh
- 00:34:32then we needed power the we needed um at
- 00:34:35least 120 megawatt at first but the
- 00:34:37building only had 15 megawatts and
- 00:34:39ultimately for 200,000 me 200,000 gpus
- 00:34:41we needed a qu gwatt so we um initially
- 00:34:45uh leased uh a whole bunch of um
- 00:34:47generators so we have generators on one
- 00:34:49side of the building just one trailer
- 00:34:51after trail trailer of generators until
- 00:34:53we can get the utility power to to come
- 00:34:55in um and then but then we also need
- 00:34:57Cooling so on the other side of the
- 00:34:58building it was just trailer after
- 00:34:59trailer of of cooling so we leased about
- 00:35:01a quarter of the mobile cooling capacity
- 00:35:03of the United States uh on the one other
- 00:35:05side of the building um then we needed
- 00:35:07to get the gpus all installed and
- 00:35:09they're all liquid cooled so in order to
- 00:35:11achieve the density necessary this is a
- 00:35:13liquid cooled system so we had to get
- 00:35:14all the plumbing for the liquid cooling
- 00:35:16nobody had ever done a liquid cooling uh
- 00:35:18data center at scale so this was a
- 00:35:22incredibly dedicated effort by a very
- 00:35:23talented team to achieve that outcome
- 00:35:25now may think now now it's going to work
- 00:35:27nope the the issue is that the power
- 00:35:29fluctuations for GPU cluster are
- 00:35:32dramatic so it's it's like a this giant
- 00:35:34Symphony that is taking place imagine
- 00:35:36having a symphony with 100,000 or
- 00:35:40200,000 participants in the in the
- 00:35:42symphony and the whole Orchestra will go
- 00:35:44quiet and loud you know 100 milliseconds
- 00:35:47and so this caused massive power
- 00:35:48fluctuations so then uh which then
- 00:35:51caused the generators to lose their
- 00:35:52minds and they they weren't expecting
- 00:35:54this to buffer the power we then used
- 00:35:56Tesla Mega packs to smooth out the power
- 00:36:00so the mega packs had to be reprogrammed
- 00:36:03so with xai working with Tesla we
- 00:36:05reprogrammed the MEAP packs to be able
- 00:36:07to deal with these dramatic power fluctu
- 00:36:10fluctuations to smooth out the power so
- 00:36:12the computers could actually run
- 00:36:13properly and that that worked was quite
- 00:36:16tricky and and then but even at that
- 00:36:19point you still have to make the
- 00:36:20computers all communicate effectively so
- 00:36:22all the networking had to be solved and
- 00:36:24debugging a zillian network cables a
- 00:36:27bugging nickel at 4: in the morning we
- 00:36:30sold it like roughly 4:20 a.m. yes than
- 00:36:34was figured out like there's some there
- 00:36:36were a whole bunch of issues one there
- 00:36:37was like a bios mismatch bios was not
- 00:36:40set up correctly yeah we had d r LS PCI
- 00:36:45outputs between two different machines
- 00:36:47one that was working yeah one that was
- 00:36:49not working yeah many other things yeah
- 00:36:51exactly this would go on for a long time
- 00:36:52if we actually listened to all the
- 00:36:53things but know it's like like it's not
- 00:36:54oh we just magically made it happen you
- 00:36:56had to break down the problem just like
- 00:36:57gr does for reasoning uh into the
- 00:36:59constituent elements and then solve each
- 00:37:00of the constituent elements in order to
- 00:37:03achieve uh a a coherent training cluster
- 00:37:06in a period of time that is a small
- 00:37:08fraction of what anyone else was could
- 00:37:09do it
- 00:37:10in and then once the training cluster
- 00:37:12was up and running and we could use it
- 00:37:14now we had to make sure that it actually
- 00:37:15stays healthy throughout which is its
- 00:37:16own giant Challenge and then we had to
- 00:37:19get every single detail of the training
- 00:37:20right in order to get a gr Free level
- 00:37:23model which is actually really hard we
- 00:37:25don't know if there are any other models
- 00:37:26out there that have gr's capabilities
- 00:37:28but whoever trains a model better than
- 00:37:30gr has to be extremely good at the the
- 00:37:32science of deep learning at every aspect
- 00:37:33of the engineering so it's not so easy
- 00:37:36to pull this off and this is now going
- 00:37:37to be the last cluster we build and last
- 00:37:39Model we train oh yeah we've already
- 00:37:41started work on the next
- 00:37:43cluster which will
- 00:37:45be yeah about five times the power so
- 00:37:47instead of a quarter gaw roughly 1.2
- 00:37:51gaw what's the Back to the Future
- 00:37:54War what's the power you does like the
- 00:37:57Back to the Future car yeah don't anyway
- 00:38:00the Back to the Future power car it's
- 00:38:02it's like roughly in that order I think
- 00:38:03and these will be the sort of the gv200
- 00:38:06SL300 cluster it it once again it will
- 00:38:08be the most powerful train cluster in
- 00:38:10the world so we're not stopping here no
- 00:38:13and our reason model is going to
- 00:38:14continue improve by accessing more tools
- 00:38:16every day yeah we're very excited to
- 00:38:18share any of the upcoming results with
- 00:38:20you all yeah the thing that keeps us
- 00:38:22going is basically being able to give G
- 00:38:24free to you and then seeing the usage go
- 00:38:26up seeing everybody enjoy gr that's what
- 00:38:30really gets us up in the morning yeah
- 00:38:34yeah thanks for tuning in thanks guys
- AI
- Machine Learning
- Grok 3
- XAI
- Data Center
- Deep Search
- Reasoning
- Tech Presentation
- Innovation
- Understanding Universe