BREAKING: Tesla Now Has the Smartest AI Ever!! Elon’s Grok 3 Demo Just STUNNED Everyone!

00:38:37
https://www.youtube.com/watch?v=etZaRHF8vec

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.

Mostra di più

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

Ottenete l'accesso immediato ai riassunti gratuiti dei video di YouTube grazie all'intelligenza artificiale!
Sottotitoli
en
Scorrimento automatico:
  • 00:00:02
    all right welcome to the grock 3
  • 00:00:04
    presentation so the mission of xai and
  • 00:00:07
    Gro is to understand the universe we
  • 00:00:10
    want to understand the nature of the
  • 00:00:11
    universe so we can figure out what's
  • 00:00:12
    going on where are the aliens what's the
  • 00:00:14
    meaning of life how does the universe
  • 00:00:15
    end how did it start all these
  • 00:00:17
    fundamental questions were driven by
  • 00:00:19
    curiosity about the nature of the
  • 00:00:20
    universe and that's also what causes us
  • 00:00:23
    to be a maximally truth-seeking AI even
  • 00:00:26
    if that truth is sometimes at odds with
  • 00:00:28
    what is politically correct
  • 00:00:30
    in order to understand the nature of the
  • 00:00:32
    universe you must absolutely rigorously
  • 00:00:34
    pursue truth or you will not understand
  • 00:00:36
    the universe you'll be suffering from
  • 00:00:38
    some amount of delusion or error that is
  • 00:00:40
    our goal figure out what's going on and
  • 00:00:43
    we're very excited to present gr 3 which
  • 00:00:45
    is we think an order of magnitude more
  • 00:00:47
    capable than gr 2 in a very short period
  • 00:00:49
    of time and that's thanks to the hard
  • 00:00:52
    work of an incredible team and I'm
  • 00:00:55
    honored to work with such a great team
  • 00:00:57
    and of course we'd love to have some of
  • 00:00:58
    the smartest humans out there join us
  • 00:01:00
    team with that let's go hi everyone my
  • 00:01:03
    name is Igor lead engineering at xci I'm
  • 00:01:06
    Jimmy Paul leading research and Tony
  • 00:01:08
    working on the reasoning Team all right
  • 00:01:10
    you I don't do
  • 00:01:12
    anything I just show up occasionally
  • 00:01:15
    yeah like I mentioned Gro is the tool
  • 00:01:17
    that we're working on Gro is our AI that
  • 00:01:19
    we're building here at XI and we've been
  • 00:01:20
    working extremely hard over the last few
  • 00:01:22
    months to improve Gro as much as we can
  • 00:01:24
    so we can give it to all of you so we
  • 00:01:25
    can give all of you access to it um we
  • 00:01:27
    think it's going to be extremely useful
  • 00:01:29
    we think it's going going to be
  • 00:01:30
    interesting to talk to funny really
  • 00:01:31
    funny and we're going to explain to you
  • 00:01:33
    how we've improved Gro over the last few
  • 00:01:34
    months we've made quite a jump in in
  • 00:01:36
    capabilities yeah actually we should
  • 00:01:38
    explain maybe also what is why do we
  • 00:01:39
    call it Gro so Gro is a word from a
  • 00:01:41
    Highland novel Stranger in a Strange
  • 00:01:43
    Land and it's used by a guy who's raised
  • 00:01:46
    on Mars and the word Gro is to fully and
  • 00:01:49
    profoundly understand something that's
  • 00:01:51
    what the word Gro means fully and
  • 00:01:52
    profoundly understand something and
  • 00:01:54
    empathy is important true
  • 00:01:58
    yeah yeah if we charted xas progress in
  • 00:02:01
    the last few months has only been 17
  • 00:02:03
    months since we started kicking off our
  • 00:02:06
    very first model grock one was almost
  • 00:02:09
    like a toy by this point only 314
  • 00:02:11
    billion parameters and now if we PR the
  • 00:02:14
    progress the time on x-axis the
  • 00:02:17
    performance of favorite Benchmark
  • 00:02:18
    numbers at mlu on the y- axis were
  • 00:02:21
    literally progressing at unprecedent
  • 00:02:23
    speed across the whole field and then we
  • 00:02:26
    kick off grock 1.5 right after grock 1
  • 00:02:29
    released after November 2023 and then gr
  • 00:02:32
    2 if you look at where the all the
  • 00:02:34
    performance coming
  • 00:02:36
    from when you have a very correct
  • 00:02:38
    engineering team and all the best AI
  • 00:02:40
    taligent the only one thing we need is a
  • 00:02:44
    big intelligence comes from big cluster
  • 00:02:47
    we can reconvert the entire progress of
  • 00:02:49
    X now replacing the Benchmark and the y
  • 00:02:51
    axis to the total amount of training
  • 00:02:53
    flops that is how many gpus we can run
  • 00:02:56
    at any given time to train our large
  • 00:02:58
    language models to impress the entire
  • 00:03:01
    internet so after all human all human
  • 00:03:03
    knowledge really that's right yeah
  • 00:03:05
    internet being part of it but it's
  • 00:03:06
    really all human knowledge all
  • 00:03:08
    everything yeah the whole internet fits
  • 00:03:09
    into a USB stick at this point it's all
  • 00:03:11
    the human tokens yeah that's right yeah
  • 00:03:14
    very soon into the real world yeah so we
  • 00:03:16
    had so much trouble actually training
  • 00:03:18
    grock 2 back in the days we kickoff the
  • 00:03:20
    model around February and we thought we
  • 00:03:23
    had a large amount of chips but turned
  • 00:03:25
    out we can barely get AK training chips
  • 00:03:27
    running coherently at any given time
  • 00:03:30
    and we have so many Cooling and power
  • 00:03:33
    issues I think you were there in the
  • 00:03:35
    data center yeah it was like really more
  • 00:03:37
    like 8K tips on average at 80%
  • 00:03:40
    efficiency more like like 6,500
  • 00:03:42
    effective uh h100s training for you know
  • 00:03:46
    several months but now now we're at 100K
  • 00:03:49
    yeah that's right more than 100K that's
  • 00:03:51
    right so what's the next step right
  • 00:03:53
    after gu to so if we want to continue
  • 00:03:56
    accelerate we have to take the matter
  • 00:03:58
    into our own hands we have to solve all
  • 00:03:59
    the ings all the power issues and
  • 00:04:02
    everything yeah so in April of last year
  • 00:04:04
    Elon decided that really the only way
  • 00:04:06
    for XI to succeed for XI to build the
  • 00:04:08
    best AI out there is to build our own
  • 00:04:10
    data center we didn't have a lot of time
  • 00:04:12
    that because we wanted to give you gr
  • 00:04:13
    free as quickly as possible so really we
  • 00:04:16
    realized we have to build the data
  • 00:04:17
    center in about 4 months and turned out
  • 00:04:20
    it took us 122 days to get the first
  • 00:04:22
    100K gpus up and running and there was a
  • 00:04:24
    Monumental effort to be able to do that
  • 00:04:27
    it's we believe it's the biggest fully
  • 00:04:29
    connected h100 cluster of its kind and
  • 00:04:32
    we didn't just stop there we actually
  • 00:04:33
    decided that we need to double the size
  • 00:04:35
    of the cluster pretty much immediately
  • 00:04:37
    if we want to build uh the kind of AI
  • 00:04:39
    that we want to build so we then had
  • 00:04:42
    another phase which we haven't talked
  • 00:04:43
    about publicly yet so this is the first
  • 00:04:44
    time that we're talking about this where
  • 00:04:46
    we doubled the capacity of the data
  • 00:04:48
    center yet again and that one only took
  • 00:04:50
    us 92 days we've been able to use all of
  • 00:04:53
    these GPS use all this compute to
  • 00:04:54
    improve grock in the meantime and
  • 00:04:56
    basically today we're going to present
  • 00:04:58
    you the results of that the the fruits
  • 00:05:00
    that came from that that's yeah all the
  • 00:05:03
    path all the RADS leads to gr 3 10x more
  • 00:05:06
    compute more than 10x really yeah really
  • 00:05:08
    maybe 15x is yep compared to our
  • 00:05:11
    previous generation model and gr
  • 00:05:13
    finished the pre-training early January
  • 00:05:16
    and we start you know the model still
  • 00:05:18
    currently training actually this is a
  • 00:05:19
    little preview of our Benchmark numbers
  • 00:05:22
    so we evaluated gr 3 on three different
  • 00:05:26
    categories U General mathematical
  • 00:05:28
    reasonings on general knowledge about
  • 00:05:31
    stem and Science and then also on
  • 00:05:34
    computer science coding Amy uh American
  • 00:05:37
    Invitational math examination host it
  • 00:05:40
    once a year uh and if we evaluate the
  • 00:05:43
    model performance we can see that the gr
  • 00:05:46
    3 across the board is in a league of its
  • 00:05:48
    own even his little brother gr 3 mini is
  • 00:05:52
    reaching the froner across all the other
  • 00:05:55
    competitors you will say at this point
  • 00:05:58
    all these benchmarks you just evaluating
  • 00:06:00
    the memorization of the textbooks
  • 00:06:02
    memorization of the GitHub repost how
  • 00:06:04
    about Real Time usefulness how about we
  • 00:06:06
    actually use those models in our product
  • 00:06:08
    what we did instead is we actually
  • 00:06:11
    kicked off a blind test of our gra 3
  • 00:06:14
    Model code named Chocolate it's pretty
  • 00:06:16
    hot yeah hot chocolate and I've been
  • 00:06:18
    running on this platform called CH arena
  • 00:06:21
    for two weeks I think the entire X
  • 00:06:24
    platform at some point speculated this
  • 00:06:26
    might be the next generation of a uh AI
  • 00:06:28
    com me your way how this chat Arena
  • 00:06:31
    works is that it strip away the entire
  • 00:06:34
    product surface right it just raw
  • 00:06:35
    comparison of the engine of those AGI
  • 00:06:38
    the language models themselves and place
  • 00:06:40
    interface where the user will submit one
  • 00:06:42
    single query and you get to show two
  • 00:06:44
    responses you don't know which model
  • 00:06:46
    they come from and in then you make the
  • 00:06:47
    vote so in this blind test gr 3 an early
  • 00:06:51
    version of gr 3 already reached 1,400 no
  • 00:06:55
    other models has reached an ELO score
  • 00:06:57
    had to have comparison to all the other
  • 00:06:59
    models at this score and it's not just
  • 00:07:02
    one single category it's 1400 aggregated
  • 00:07:05
    across all the categories in chall
  • 00:07:07
    capabilities instruction following
  • 00:07:09
    coding so it's number one across the
  • 00:07:12
    board in this blind test and it's it's
  • 00:07:13
    still climbing so we actually to keep
  • 00:07:15
    updating it so it's it's 14,400 about
  • 00:07:18
    14400 in climbing yeah in fact we have a
  • 00:07:20
    version of the model that we think is
  • 00:07:21
    already much better than the one that we
  • 00:07:22
    tested here yeah we'll see how far it
  • 00:07:24
    gets uh but that's the one that we're
  • 00:07:26
    working on we talking about today yeah
  • 00:07:28
    so actually one thing if if you're if
  • 00:07:30
    you're using grock 3 you I think you may
  • 00:07:31
    notice improvements almost every day um
  • 00:07:33
    because we're we're continuously
  • 00:07:35
    improving the model so Lally even within
  • 00:07:38
    24 hours you'll see
  • 00:07:39
    improvements yep but we believe here at
  • 00:07:42
    XI getting the best pre-training model
  • 00:07:45
    is not enough that's not enough to build
  • 00:07:47
    the best AI and the best AI need to
  • 00:07:49
    think like a human you to contemplate
  • 00:07:51
    about all the possible
  • 00:07:53
    solutions self-critique verify all the
  • 00:07:56
    solutions backtrack and also think from
  • 00:07:59
    the first principle that's a very
  • 00:08:01
    important capability so we believe that
  • 00:08:04
    as we take the best pre-train model and
  • 00:08:06
    continue training with reinforcement
  • 00:08:08
    learning it will elicit the additional
  • 00:08:10
    reasoning capabilities that allows the
  • 00:08:12
    model to become so much better and scale
  • 00:08:15
    not just in the training time but
  • 00:08:17
    actually in the test time as well we
  • 00:08:19
    already found the model is extremely
  • 00:08:20
    useful internally for our own
  • 00:08:22
    engineering saving hours of time
  • 00:08:24
    hundreds of hours of coding time equ
  • 00:08:26
    you're the power user of our graic
  • 00:08:28
    reasoning model what are some use cases
  • 00:08:30
    yeah so like Jimmy said we've added
  • 00:08:31
    Advanced reasoning capabilities to Grog
  • 00:08:33
    and we've been testing them pretty
  • 00:08:34
    heavily over the last few weeks in order
  • 00:08:36
    to give you a little bit of a taste of
  • 00:08:37
    what it looks like when Gro is solving
  • 00:08:39
    heart reasoning problems so we prepared
  • 00:08:41
    two little problems for you one comes
  • 00:08:43
    from physics and one is actually a game
  • 00:08:45
    that gr is going to ride for us when it
  • 00:08:47
    comes to the physics problem what we
  • 00:08:48
    want gr to do is to plot a viable
  • 00:08:50
    trajectory to do a transfer from Earth
  • 00:08:53
    to Mars and then at a later point in
  • 00:08:55
    time a transfer back from Mars to Earth
  • 00:08:57
    and that requires some some Physics that
  • 00:08:59
    Gro will have to understand so we're
  • 00:09:01
    going to challenge Gro come up with a
  • 00:09:02
    viable trajectory calculate it and then
  • 00:09:05
    plot it for us so we can see it and yeah
  • 00:09:08
    this is totally unscripted by the way
  • 00:09:10
    this is the that's the entirety of the
  • 00:09:12
    prompt which should be clarify is that
  • 00:09:13
    there's nothing more than that yeah
  • 00:09:15
    exactly this is the gro interface and
  • 00:09:17
    we've typed in this text that you can
  • 00:09:19
    see here generate code for an animated
  • 00:09:21
    3D plot of a launch from Earth landing
  • 00:09:24
    on Mars and then back to Earth at the
  • 00:09:26
    next launch window and we've not kicked
  • 00:09:28
    off or the query and you can see Gro is
  • 00:09:29
    thinking part of grock's advanced
  • 00:09:32
    reasoning capabilities are these
  • 00:09:33
    thinking traces that you can see here
  • 00:09:35
    you can even go inside and actually read
  • 00:09:37
    what gr is thinking as it's going
  • 00:09:38
    through the problem as it's trying to
  • 00:09:39
    solve it yeah which we are doing some
  • 00:09:42
    obscuration of the thinking so that our
  • 00:09:44
    model doesn't get totally copied
  • 00:09:45
    instantly so there's more to the
  • 00:09:48
    thinking than is displayed in yeah and
  • 00:09:52
    because this is totally unscripted
  • 00:09:54
    there's actually a chance that grock
  • 00:09:55
    might made a little coding mistake and
  • 00:09:57
    it might not actually work just in case
  • 00:09:58
    we're going to launch two more instances
  • 00:10:00
    of this so if something goes wrong we
  • 00:10:02
    were able to to switch to those and show
  • 00:10:05
    you something that's presentable so
  • 00:10:07
    we're kicking off the other two as well
  • 00:10:09
    and like I said we have a second problem
  • 00:10:11
    as well and yeah actually one of the
  • 00:10:13
    favorite one of our favorite activities
  • 00:10:15
    here XI is having Grog right games for
  • 00:10:17
    us and not just any know any old game
  • 00:10:21
    any game that you might already be
  • 00:10:22
    familiar with but actually creating new
  • 00:10:23
    games on the spot and being creative
  • 00:10:25
    about it so one example that we found
  • 00:10:27
    was really fun is create a game that's a
  • 00:10:30
    mixture of the two games Tetris and B so
  • 00:10:34
    this is that maybe an important thing
  • 00:10:35
    like this obviously if you ask an AI to
  • 00:10:38
    create a game like Tetris there's there
  • 00:10:39
    are many examples of Tetris on the
  • 00:10:40
    Internet or game like J whatever there
  • 00:10:44
    it can copy it what's interesting here
  • 00:10:46
    is it achieved a creative solution
  • 00:10:49
    combining the two games that actually
  • 00:10:51
    works and and is a good game yeah that's
  • 00:10:54
    the we're seeing the beginnings of
  • 00:10:57
    creativity yeah fingers crossed that we
  • 00:11:00
    can recreate that hopefully it works
  • 00:11:01
    hope so actually because this is a bit
  • 00:11:03
    more challenging we're going to use
  • 00:11:05
    something special here which we call Big
  • 00:11:06
    Brain that's our mode in which we use
  • 00:11:09
    more computation which more reasoning of
  • 00:11:11
    our gr just to make sure that there's a
  • 00:11:13
    good chance here that it might actually
  • 00:11:14
    do it so we're also going to fire off
  • 00:11:16
    three attempts here at at solving this
  • 00:11:19
    game at creating this game that's a
  • 00:11:21
    mixture of Tetris and B yeah let's let's
  • 00:11:24
    see what go comes up like I've played
  • 00:11:25
    the game it's pretty good like it's like
  • 00:11:28
    wow okay this is something yeah um so
  • 00:11:31
    while Gro is thinking uh in the in the
  • 00:11:33
    background um we can now actually talk
  • 00:11:34
    about some concrete know how how well is
  • 00:11:36
    Gro doing across tons of different tasks
  • 00:11:38
    that we've tested on um so we'll hand it
  • 00:11:40
    over to Tony to talk about that yeah
  • 00:11:43
    okay so let's see how Grog does on those
  • 00:11:46
    interesting challenging benchmarks so
  • 00:11:48
    yeah so reasoning again refers to those
  • 00:11:50
    models that actually thinks quite for
  • 00:11:52
    quite a long time before it tries to
  • 00:11:54
    solve a problem in this case around a
  • 00:11:56
    month ago the graph 3 pre-training
  • 00:11:58
    finishes after that we work very hard to
  • 00:12:01
    put the reasoning capability into the uh
  • 00:12:03
    current graph 3 Model but again this is
  • 00:12:06
    very early days so the model is still
  • 00:12:08
    currently in training right now what
  • 00:12:09
    we're going to show to people is this
  • 00:12:12
    beta version of the gry reasoning model
  • 00:12:14
    alongside we also are training a mini
  • 00:12:16
    version of the reasoning model
  • 00:12:18
    essentially on this plot you can see the
  • 00:12:20
    gr 3 reasoning beta and then gr 3 mini
  • 00:12:22
    reasoning the grth reason mini reasoning
  • 00:12:24
    is actually a model that we train for
  • 00:12:26
    much longer time and you can see that
  • 00:12:28
    sometimes it actually perform study
  • 00:12:29
    better compared to the gr three
  • 00:12:31
    reasoning this also just means that
  • 00:12:33
    there's a huge potential for the grth
  • 00:12:35
    three reasoning because it's trained for
  • 00:12:36
    much less time all right so let's
  • 00:12:38
    actually look at what how it does on
  • 00:12:40
    those three benchmarks so Jimmy also
  • 00:12:42
    introduced already so essentially we're
  • 00:12:44
    looking at three different areas
  • 00:12:46
    mathematics science and coding and for
  • 00:12:48
    math we're picking this high school
  • 00:12:50
    competition math problem for science we
  • 00:12:52
    actually pick those PhD level science
  • 00:12:54
    questions and for coding it's also
  • 00:12:56
    actually pretty challenging it's
  • 00:12:57
    competitive coding and also some leod
  • 00:13:00
    which is some cold inter interview
  • 00:13:01
    problems that people usually get when
  • 00:13:03
    they interview for companies so on those
  • 00:13:05
    benchmarks you can see that the gr 3
  • 00:13:07
    actually perform quite well across the
  • 00:13:09
    board compared to other competitors um
  • 00:13:12
    yeah so it's pretty promising these
  • 00:13:14
    models are very smart so Tony what what
  • 00:13:16
    what are those shaded bars yeah so okay
  • 00:13:19
    so uh I'm glad you asked this question
  • 00:13:21
    so for those models because it can
  • 00:13:23
    reason it can thinks you can also ask
  • 00:13:25
    them to even think longer uh you can
  • 00:13:27
    spend more what we call test and compute
  • 00:13:31
    which means you can spend more time to
  • 00:13:33
    reason to think about a problem before
  • 00:13:35
    you spit out the answer so in this case
  • 00:13:38
    the Shaded bar here means that we just
  • 00:13:41
    ask the model to spend more time you can
  • 00:13:43
    solve the the same problem many times
  • 00:13:45
    before it it tries to conclude what is
  • 00:13:47
    the right solution and once you give
  • 00:13:49
    this compute or this kind of budget to
  • 00:13:51
    the model it turns out the model can
  • 00:13:53
    even perform better so this is
  • 00:13:55
    essentially the Shaded bar in in those
  • 00:13:57
    BX so this is really exciting right
  • 00:14:00
    because now instead of just doing one
  • 00:14:01
    chain of thoughts with AI why not do
  • 00:14:04
    multiple once yes so that's a very
  • 00:14:06
    powerful technique that allows to
  • 00:14:07
    continue scale the model capabilities
  • 00:14:09
    after training and people often ask are
  • 00:14:12
    we actually just over fitting to the
  • 00:14:14
    benchmarks yes so how about your oration
  • 00:14:16
    so yes I think yeah this is definitely a
  • 00:14:18
    question that we are asking ourselves
  • 00:14:20
    whether we are overfitting to those
  • 00:14:22
    current benchmarks luckily we have a
  • 00:14:24
    real test so about 5 days ago Amy 2025
  • 00:14:28
    just finished this is where high school
  • 00:14:30
    students compete in this particular
  • 00:14:32
    Benchmark so we got this very fresh new
  • 00:14:35
    competition and then we asked our two
  • 00:14:37
    models to compete on the same Benchmark
  • 00:14:39
    at the same exam and it turns out very
  • 00:14:41
    interestingly the grth three reasoning
  • 00:14:43
    the big one actually does better on this
  • 00:14:46
    particular new fresh exam this also
  • 00:14:48
    means that the generalization capability
  • 00:14:50
    of the big model is stronger much
  • 00:14:52
    stronger compared to smaller model if
  • 00:14:54
    you compare to the last year's exam
  • 00:14:55
    actually this is the opposite the
  • 00:14:57
    smaller model kind of learned
  • 00:14:59
    the the previous exams better yeah so
  • 00:15:02
    this this actually shows some kind of
  • 00:15:03
    true generalization from the model
  • 00:15:05
    that's right so 17 months ago our gr
  • 00:15:07
    zero and Gro one barely solves any High
  • 00:15:09
    School problems that's right and now we
  • 00:15:11
    have a kid that just already graduate
  • 00:15:13
    the gro Gro is ready to go to college is
  • 00:15:15
    that right yeah it won't be long before
  • 00:15:18
    it's simply perfect the human exams
  • 00:15:19
    won't be hard they be too easy yeah and
  • 00:15:22
    internally we actually as gret continue
  • 00:15:24
    evolves we're going to talk about what
  • 00:15:26
    we're excited about but very soon there
  • 00:15:29
    will be no more Benchmark left
  • 00:15:31
    yeah yeah one thing that's quite
  • 00:15:33
    fascinating I think is that we basically
  • 00:15:35
    only trained Rock's reasoning abilities
  • 00:15:36
    on math problems and comparative coding
  • 00:15:39
    problems right so very specialized kinds
  • 00:15:41
    of tasks but somehow it's able to work
  • 00:15:44
    on all kinds of other different tasks so
  • 00:15:46
    including creating games no lots lots
  • 00:15:48
    and lots of different things and what
  • 00:15:50
    seems to be happening is that basically
  • 00:15:51
    Gro learns this ability to detect its
  • 00:15:54
    own mistakes and its thinking correct
  • 00:15:55
    them persist on a problem try lots of
  • 00:15:57
    different variants pick pick the one
  • 00:15:59
    that's best so there are these
  • 00:16:00
    generalized generalizing abilities that
  • 00:16:02
    Gro learns from mathematics and from
  • 00:16:04
    coding which it can then use to solve
  • 00:16:06
    all kinds of other problems that's
  • 00:16:07
    pretty reality is the instantiation of
  • 00:16:09
    mathematics that's right and one thing
  • 00:16:12
    we're actually really excited about that
  • 00:16:13
    going back to our funing mission is what
  • 00:16:15
    if one day we have a computer just like
  • 00:16:17
    deep thought that utilize our entire
  • 00:16:20
    cluster just for that one very important
  • 00:16:22
    problem in the test time all the GPU
  • 00:16:24
    turned on right so I think back then we
  • 00:16:26
    were building the GPU clusters together
  • 00:16:28
    you plug
  • 00:16:29
    cables and I remember that when we turn
  • 00:16:32
    on the first initial test you can hear
  • 00:16:34
    all the GPS humming in the hallway
  • 00:16:37
    that's almost feel like spiritual yeah
  • 00:16:39
    that's actually a pretty cool uh thing
  • 00:16:40
    that we're able to do that we can go
  • 00:16:42
    into the data center and Tinker with the
  • 00:16:44
    machines there so for example we went in
  • 00:16:46
    and we unplugged a few of the cables and
  • 00:16:49
    just made sure that our training setup
  • 00:16:50
    is still running stably so that's
  • 00:16:52
    something that I think most uh AI teams
  • 00:16:55
    out there don't usually do but it's
  • 00:16:56
    actually totally unlocks like a new
  • 00:16:58
    level of reliability and what you're
  • 00:17:00
    able to do with the hardware so okay so
  • 00:17:02
    when when are we going to solve
  • 00:17:04
    remon the easiest solution is to
  • 00:17:07
    numerate over all possible strains and
  • 00:17:10
    as long you have a verifier enough
  • 00:17:11
    compute you'll be able to do it okay my
  • 00:17:14
    projection will be what's your guess
  • 00:17:16
    what is your neural n calculate my my Bo
  • 00:17:18
    prodiction so three years ago I told you
  • 00:17:20
    this I think in now two years later two
  • 00:17:23
    things going to happen we're going to
  • 00:17:24
    see machines win some medals yes touring
  • 00:17:28
    award absolutely
  • 00:17:29
    Fields metal Nobel Prize with probably
  • 00:17:32
    some expert in the loop right so the
  • 00:17:34
    expert uplifting do you mean so this
  • 00:17:35
    year or next year oh okay that's what it
  • 00:17:39
    comes down to really yeah so it looks
  • 00:17:42
    like Gro finished know all of its
  • 00:17:43
    thinking on on the two problem so let's
  • 00:17:45
    take a look at what it
  • 00:17:47
    said all right so this was the little
  • 00:17:50
    physics problem we had no we've
  • 00:17:51
    collapsed the thoughts here so they're
  • 00:17:53
    they're hidden and then we see grock's
  • 00:17:55
    answer below that so it explains it
  • 00:17:56
    wrote a python script here using M plot
  • 00:17:58
    Li then gives us all of the code so
  • 00:18:01
    let's take a quick look at the code
  • 00:18:02
    seems like it's doing reasonable things
  • 00:18:04
    here not totally of the mark solve
  • 00:18:07
    Kepler says here so maybe it's solving
  • 00:18:09
    Kepler's laws cap Kepler law numerically
  • 00:18:12
    um yeah there's really only one way to
  • 00:18:14
    find out if this thing is working I'd
  • 00:18:16
    say let's give it a try let's run the
  • 00:18:17
    code all right and we can see yeah gr is
  • 00:18:20
    animating two different planets Earth
  • 00:18:22
    and Mars here and then the green uh ball
  • 00:18:25
    is the vehicle that's transiting the
  • 00:18:27
    spacecraft that's transitioning between
  • 00:18:29
    Earth and Mars and you could see the
  • 00:18:30
    journey from Earth to Mars and looks
  • 00:18:32
    like yeah indeed the astronauts return
  • 00:18:35
    safely at the right moment in time now
  • 00:18:38
    obviously this was just generated on the
  • 00:18:39
    spots now we can't tell you if that was
  • 00:18:41
    actually correct solution so we're going
  • 00:18:42
    to take a closer look now maybe we're
  • 00:18:43
    going to call some colleagues from space
  • 00:18:45
    X ask them if if this is legit um it's
  • 00:18:48
    pretty close it's it's uh I mean there's
  • 00:18:51
    a lot of complexities in the actual
  • 00:18:53
    orbits that have to be taken into
  • 00:18:54
    account but this is pretty close to to
  • 00:18:55
    what it what it looks like awes in fact
  • 00:18:57
    I have that on my pend here got the
  • 00:19:00
    Earth Mars home and transfer on
  • 00:19:03
    it when are we going to install groc on
  • 00:19:06
    a
  • 00:19:07
    rock I suppose in two years two years
  • 00:19:12
    everything is two years away Earth and
  • 00:19:14
    Mars Transit can occurs every 26 months
  • 00:19:17
    the next we're currently in a Transit
  • 00:19:18
    window approximately the next one would
  • 00:19:20
    be November of next year roughly end of
  • 00:19:24
    next year and if all goes well SpaceX
  • 00:19:27
    will send a Starship Rockets to Mars and
  • 00:19:30
    with Optimus robots and and Gro
  • 00:19:34
    mhm yeah I'm curious about this
  • 00:19:36
    combination of Tetris and B looks like
  • 00:19:39
    the tetris as we've named it internally
  • 00:19:43
    okay we also have an output from go here
  • 00:19:45
    it says Ro python script explains that
  • 00:19:47
    it's what it's been doing if you look at
  • 00:19:49
    the code there are some constants that
  • 00:19:51
    are being defined here some colors then
  • 00:19:53
    the trinos the pieces of Tetris are
  • 00:19:56
    there obviously very hard to see and at
  • 00:19:59
    one glance if this is good so we got to
  • 00:20:00
    run this to figure out if it's working
  • 00:20:02
    let's give it a
  • 00:20:03
    try fingers crossed all right right so
  • 00:20:06
    this kind of looks like Tetris uh but
  • 00:20:08
    the the colors are a little bit off
  • 00:20:10
    right the colors are different here and
  • 00:20:12
    if you think about what's going what's
  • 00:20:14
    going on here the J has this mechanic
  • 00:20:17
    where if you get three jws in a row you
  • 00:20:19
    know then they they disappear and also
  • 00:20:22
    gravity activates right so uh what
  • 00:20:24
    happens if you get three of the colors
  • 00:20:26
    together okay so something happens so so
  • 00:20:28
    I think what SC did in this version is
  • 00:20:31
    that once you connect three at least
  • 00:20:33
    three blocks of the same color in a row
  • 00:20:35
    then gravity activates and they
  • 00:20:38
    disappear and then gravity activates and
  • 00:20:40
    all the other blocks fall down curious
  • 00:20:42
    if there's still a Tetris mechanic here
  • 00:20:44
    where if the line is full does it
  • 00:20:46
    actually clear it or what happens then
  • 00:20:49
    it's up to interpretation who knows yeah
  • 00:20:51
    I mean it'll do different variants when
  • 00:20:53
    you ask it it doesn't do the same thing
  • 00:20:54
    every time exactly we've seen a few
  • 00:20:56
    other the tetris that work very
  • 00:20:58
    differently but this one seems cool yeah
  • 00:21:01
    are we ready for game Studio at x. a yes
  • 00:21:04
    so we're launching uh an AI gaming
  • 00:21:06
    studio at xci if you're interested in
  • 00:21:08
    joining us and building AI games please
  • 00:21:10
    join XI we're launching an AI gaming
  • 00:21:12
    studio we're announcing it tonight let's
  • 00:21:15
    go epic games but right that's an actual
  • 00:21:19
    games yeah yeah all right so I think one
  • 00:21:24
    thing is super exciting for us is that
  • 00:21:26
    once you have the best pre train model
  • 00:21:29
    you have the best reason model right we
  • 00:21:31
    already see that we actually give the
  • 00:21:33
    capability for those model to think
  • 00:21:34
    harder think longer think more broad the
  • 00:21:38
    performance continue improves and we're
  • 00:21:40
    really excited about the next front here
  • 00:21:42
    that what happen if we're not only allow
  • 00:21:44
    the model to think harder but also
  • 00:21:45
    provide more tools this I call real
  • 00:21:47
    humans to solve those problems for real
  • 00:21:50
    humans we don't ask them to solve reman
  • 00:21:52
    a hypothesis just with a piece of pen
  • 00:21:54
    and paper no internet with all the basic
  • 00:21:57
    web browsing search engine and code
  • 00:22:00
    interpreters that builds the foundations
  • 00:22:03
    and the best reasoning model builds the
  • 00:22:05
    foundations for the gr agent to come
  • 00:22:08
    today we're actually introducing a new
  • 00:22:11
    product called Deep search that is the
  • 00:22:13
    first generation of our gr agents that
  • 00:22:16
    not just helping the engineers and
  • 00:22:17
    research and scientists to do coding but
  • 00:22:19
    actually help everyone to answer
  • 00:22:21
    questions that you have day today it's
  • 00:22:23
    like a Next Generation search engine
  • 00:22:25
    that really help you to understand the
  • 00:22:26
    universe you can start asking question
  • 00:22:29
    like for example hey when is the next
  • 00:22:32
    Starship launch day for example let's
  • 00:22:34
    try that get the answer on the left hand
  • 00:22:37
    side we see a high level progress bar
  • 00:22:39
    essentially the model now is going to do
  • 00:22:41
    one single search like the current rack
  • 00:22:43
    system but actually thought very deeply
  • 00:22:45
    about hey what's the user intent here
  • 00:22:47
    and what are the facts I should consider
  • 00:22:49
    at the same time and how many different
  • 00:22:51
    website I should actually go and read
  • 00:22:52
    their content right so this can really
  • 00:22:55
    save hundreds hours of everyone's Google
  • 00:22:58
    time if you want to really look into
  • 00:22:59
    certain topics and then on the right
  • 00:23:02
    hand side you can see the bullet
  • 00:23:04
    summaries of how the current model is
  • 00:23:06
    doing what websites browsing what
  • 00:23:08
    sources is verifying and often time
  • 00:23:10
    actually cross validate different
  • 00:23:11
    sources out there to make sure the
  • 00:23:13
    answer is actually correct before it's
  • 00:23:14
    output final answer and we can at the
  • 00:23:16
    same time fire up a few more queries um
  • 00:23:19
    how about you know you're a gamer right
  • 00:23:21
    uh sure yeah so how about what are some
  • 00:23:23
    of the best builds and most popular
  • 00:23:25
    builds in path Excel hardcore right
  • 00:23:27
    hardcore League you can technically just
  • 00:23:30
    look at the hardcore ladder might be a
  • 00:23:33
    fast way to figure it out yeah we'll see
  • 00:23:34
    what model
  • 00:23:36
    does um and then we can also do
  • 00:23:39
    something more fun for example how about
  • 00:23:41
    make a prediction about the March
  • 00:23:42
    Madness out there yeah so this is go fun
  • 00:23:44
    one where Warren Buffett has a billion
  • 00:23:47
    dollar vet if you can exactly match the
  • 00:23:50
    I think the the sort of the entire
  • 00:23:52
    winning tree of marsh Madness you can
  • 00:23:54
    win a billion dollars from Warren
  • 00:23:55
    Buffett it would be pretty cool if AI
  • 00:23:57
    could help you win a billion dollars
  • 00:23:59
    from
  • 00:24:00
    Buffett that seems like a pretty good
  • 00:24:02
    investment let's go yeah all right so
  • 00:24:05
    now let's fire up the query and see what
  • 00:24:07
    model does so we can actually go back to
  • 00:24:09
    our very first one how about the buffet
  • 00:24:11
    wasn't counting on this it's sry done
  • 00:24:14
    that's right okay so we got the result
  • 00:24:16
    of the first one the model thought
  • 00:24:17
    around one minute uh so okay so the key
  • 00:24:19
    Insight here the next Starship is going
  • 00:24:21
    to be on 24th or later so no earlier
  • 00:24:24
    than February
  • 00:24:25
    24th it might be sooner
  • 00:24:29
    yeah so I think we can go down scroll
  • 00:24:31
    down what what the model does so it does
  • 00:24:32
    a little research on the fight 7 what
  • 00:24:34
    happened got grounded and actually it
  • 00:24:36
    look into the FCC filing from this data
  • 00:24:39
    Collections and then actually make the
  • 00:24:42
    new conclusion that yeah if we continue
  • 00:24:43
    scroll down let's see yeah so it makes
  • 00:24:46
    the little table I think inside xai we
  • 00:24:49
    often joked about the time to the first
  • 00:24:51
    table is the only latency that matters
  • 00:24:54
    yeah so that's how to model make
  • 00:24:56
    inference and look up all the sources
  • 00:24:58
    and then we can look into the gaming one
  • 00:25:00
    so how about
  • 00:25:04
    the for this particular one we look at
  • 00:25:07
    hey the the build is
  • 00:25:10
    light with the The Infernal is but if we
  • 00:25:13
    go down the surprising fact of all the
  • 00:25:15
    other builds look into the 12 classes
  • 00:25:18
    yeah we'll see that the Min build was
  • 00:25:20
    pretty popular whenever the game first
  • 00:25:21
    came out and now the invokers of the
  • 00:25:23
    world took over invoker monk invoker for
  • 00:25:26
    sure yeah that's right yeah by the stone
  • 00:25:28
    wavers and that's really good at mapping
  • 00:25:30
    yeah and then we can see the the match
  • 00:25:33
    manness about that one one interesting
  • 00:25:35
    thing about the Deep search is that if
  • 00:25:36
    you actually go into the panel where it
  • 00:25:39
    shows what are the subtasks you can
  • 00:25:41
    actually click the bottom left and then
  • 00:25:44
    in this case you can actually scroll
  • 00:25:45
    through actually reading through the
  • 00:25:47
    mind of grock what informations does the
  • 00:25:49
    model actually think about are
  • 00:25:51
    trustworthy what are not how does they
  • 00:25:52
    actually cross validate different
  • 00:25:53
    information sources so that makes the
  • 00:25:56
    entire search experience and information
  • 00:25:57
    retrieval process a lot more transparent
  • 00:25:59
    to our
  • 00:26:01
    users and this is much more powerful
  • 00:26:03
    than any search engine out there you can
  • 00:26:06
    literally just tell it only use sources
  • 00:26:08
    from X will try to respect that yeah and
  • 00:26:10
    so it's much more steerable much more
  • 00:26:12
    intelligent than it really should save
  • 00:26:14
    you a lot of time so something that
  • 00:26:15
    might take you half an hour or an hour
  • 00:26:17
    of researching on the web or searching
  • 00:26:19
    social media you can just ask it to go
  • 00:26:21
    do that and and come back in 10 minutes
  • 00:26:23
    later it's done an hour's worth of work
  • 00:26:25
    for you that's really what it comes down
  • 00:26:26
    to exactly and maybe better than you
  • 00:26:28
    could have done it yourself yeah think
  • 00:26:30
    about you have INF of interns working
  • 00:26:32
    for you now you can just fire up all the
  • 00:26:34
    tasks and come back a minute later so
  • 00:26:36
    this is going to be interesting one so
  • 00:26:37
    March M had not happened yet so I guess
  • 00:26:40
    we have to follow up with a next live
  • 00:26:42
    stream yeah it seems like pretty good
  • 00:26:45
    the $40 might get you a billion dollars
  • 00:26:47
    $40 subscription that's right my work
  • 00:26:51
    yeah so when are the users going to have
  • 00:26:53
    their hands on gr to yeah so the the
  • 00:26:55
    good news is we've been working
  • 00:26:56
    tirelessly to actually release all of
  • 00:26:59
    these features that we've shown you the
  • 00:27:00
    Grog free base model with amazing chat
  • 00:27:02
    capabilities that's really useful that's
  • 00:27:03
    really interesting to talk to the Deep
  • 00:27:05
    search the advanced reasoning mode all
  • 00:27:07
    of these things we want to roll them out
  • 00:27:09
    to you today starting with the premium
  • 00:27:12
    plus subscribers on X so it's the first
  • 00:27:14
    group that will initially get access
  • 00:27:16
    make sure to update your X app if you
  • 00:27:18
    want to see all of the advanced
  • 00:27:19
    capabilities because we just released
  • 00:27:21
    the update now as we're talking here and
  • 00:27:23
    yeah if you're interested in getting
  • 00:27:24
    early access to go then sign up for
  • 00:27:26
    premium plus and also we're announcing
  • 00:27:28
    that we're starting a separate
  • 00:27:30
    subscription for grock that we call
  • 00:27:31
    Super Gro for those who those real grock
  • 00:27:34
    fans that want the most advanced
  • 00:27:35
    capabilities and earliest access to new
  • 00:27:38
    features so feel free to check that out
  • 00:27:40
    as well this this is for the dedicated
  • 00:27:42
    grock app and for the website ex website
  • 00:27:44
    so our our new website is called gro.com
  • 00:27:46
    yeah and you'll also find you never
  • 00:27:47
    guess yeah you never guess and you can
  • 00:27:50
    also find our grock app in the IOS app
  • 00:27:52
    store and that gives you like a more Pol
  • 00:27:55
    even even more polished uh experience
  • 00:27:56
    that's totally grock focused if you're
  • 00:27:58
    if you want to have Gro easily available
  • 00:28:00
    one Tap Away yeah the version on gro.com
  • 00:28:03
    on on a web browser is going to be the
  • 00:28:04
    most the latest and most advanced
  • 00:28:06
    version because obviously takes us a
  • 00:28:07
    while to get thing get something into an
  • 00:28:10
    app and then get it approved by the app
  • 00:28:11
    store and then it's if something's on a
  • 00:28:13
    phone format there limitations what you
  • 00:28:15
    can do so the most powerful version of
  • 00:28:16
    grock and the latest version will be the
  • 00:28:18
    web version at gro.com yeah so watch out
  • 00:28:20
    for the name grock free in the app did
  • 00:28:22
    giveaway yeah exactly that that's that's
  • 00:28:24
    the giveaway that you have groe and if
  • 00:28:26
    it says gr through then GR hasn't quite
  • 00:28:28
    arrived for yet but we're working hard
  • 00:28:30
    to roll this out today and then to even
  • 00:28:32
    more people over the the coming days
  • 00:28:34
    yeah make sure you update your phone app
  • 00:28:36
    too where you're actually going to get
  • 00:28:37
    all the tools we're showcase today with
  • 00:28:39
    the thinking mode with the Deep search
  • 00:28:42
    so yeah really looking forward to all
  • 00:28:43
    the feedbacks you have yeah and I think
  • 00:28:45
    we we should uh emphasize that this is a
  • 00:28:48
    beta meaning that it's you should expect
  • 00:28:50
    some imperfections at first but we will
  • 00:28:52
    improve it rapidly almost every day in
  • 00:28:54
    fact every day I think it'll get better
  • 00:28:56
    if you want a more polished version I'd
  • 00:28:57
    like maybe wait a week but expect
  • 00:28:59
    improvements literally every day and
  • 00:29:01
    then we're also going to be providing a
  • 00:29:03
    voice interaction so you can have
  • 00:29:05
    conversational in fact I was trying it
  • 00:29:06
    earlier today it's working pretty well
  • 00:29:08
    but not we need these a bit more polish
  • 00:29:10
    the sort of way we can just literally
  • 00:29:11
    talk to it like you're talking to a
  • 00:29:12
    person it's that's awesome it's actually
  • 00:29:15
    I think one of the best experiences of
  • 00:29:16
    gr but that's probably about a week
  • 00:29:19
    away yeah with that said well I think we
  • 00:29:23
    might have some audience questions sure
  • 00:29:25
    yeah all right let's take a look yeah
  • 00:29:28
    let's take a look the the audience from
  • 00:29:30
    the as platform yeah so the first
  • 00:29:33
    question here is when grock voice
  • 00:29:35
    assistant when is it coming out yeah as
  • 00:29:37
    as as soon as possible just like Elon
  • 00:29:39
    said just a little bit of polishing away
  • 00:29:41
    from being reled to everybody obviously
  • 00:29:44
    it's going to be released in an early
  • 00:29:45
    form and we're going to rapidly iterate
  • 00:29:47
    on it Y and the next question is like
  • 00:29:49
    when will gr 3 be in the API so this is
  • 00:29:52
    coming in uh the gr 3 API with both the
  • 00:29:56
    reasoning models and deep is coming your
  • 00:29:58
    way in the coming weeks we're actually
  • 00:30:00
    very excited about the Enterprise use
  • 00:30:01
    cases of all these additional tools that
  • 00:30:03
    now gr has access to and how the test
  • 00:30:05
    time compute and to use car to really
  • 00:30:07
    accelerate all the business use cases
  • 00:30:09
    another one is Will voice mode be native
  • 00:30:12
    or text to speech so I think that means
  • 00:30:13
    is it going to be one one model that is
  • 00:30:16
    understanding what you say and then
  • 00:30:18
    talking back to you or is it going to be
  • 00:30:19
    some system that has text to speech
  • 00:30:21
    inside of it and the good news is it's
  • 00:30:22
    going to be one model like a variant of
  • 00:30:24
    gr free that we're going to release
  • 00:30:26
    which basically understands what you're
  • 00:30:28
    say what you're saying and then uh
  • 00:30:30
    generates the audio directly from that
  • 00:30:32
    so very much like Grog free generates
  • 00:30:34
    text that model generates audio and that
  • 00:30:36
    has a bunch of advantages I was talking
  • 00:30:38
    to it earlier today and it said hi igore
  • 00:30:40
    reading my my name from probably from
  • 00:30:42
    some text that it had um and I said no
  • 00:30:44
    no my name is Igor and it remember that
  • 00:30:47
    you know so it could continue to say
  • 00:30:48
    Igor just like a human word and you
  • 00:30:51
    can't achieve that with with Tex of
  • 00:30:52
    speech yeah oh here's a question for you
  • 00:30:54
    pretty spicy um you know is Gro a boy or
  • 00:30:58
    girl and are they sing C is whatever you
  • 00:31:00
    want it to
  • 00:31:02
    be yeah yeah are they
  • 00:31:05
    single
  • 00:31:07
    yes all right the shop is open um so
  • 00:31:11
    honestly people are going to fall in
  • 00:31:12
    love with crocet since it's 1,000%
  • 00:31:15
    probable yeah MH uh the next question
  • 00:31:18
    will Gro be able to transcribe audio
  • 00:31:20
    into text yes so we'll have this
  • 00:31:22
    capability both the app and also the API
  • 00:31:25
    we found that gr should just be your
  • 00:31:26
    personal assistant looking over your
  • 00:31:27
    shoulder
  • 00:31:28
    right and follow you along the way learn
  • 00:31:30
    everything you have learned and really
  • 00:31:31
    help you to understand the world better
  • 00:31:33
    become smarter every day yeah the voice
  • 00:31:36
    metag doesn't isn't simply it's not just
  • 00:31:38
    voice text it understands tone
  • 00:31:40
    inflection pacing everything it's wild
  • 00:31:42
    it's like talking to a
  • 00:31:44
    person okay yeah so any plans for
  • 00:31:47
    conversation memory yeah absolutely
  • 00:31:50
    we're working on it right now not really
  • 00:31:54
    forg that's right um let's see what are
  • 00:31:58
    the other
  • 00:31:59
    ones so what about the the DM features
  • 00:32:04
    right so if you have personalizations
  • 00:32:06
    and if if you have remembers your
  • 00:32:08
    previous interactions yes should it be
  • 00:32:11
    one Gro or multiple different grocs
  • 00:32:13
    except to you you can have one Gro or
  • 00:32:15
    many GRS I suspect people will probably
  • 00:32:17
    have more than one yeah I want to have a
  • 00:32:20
    do Gro yeah the gro
  • 00:32:23
    dog that's right all right cool so in
  • 00:32:27
    the past open source grock one right so
  • 00:32:30
    somebody's asking is are we going to do
  • 00:32:31
    that again with gr 2 yeah I think one
  • 00:32:34
    once gr our general approach is that we
  • 00:32:36
    will open source the last version when
  • 00:32:38
    the next version is fully out like when
  • 00:32:41
    gr 3 is mature and stable which is
  • 00:32:43
    probably within a few months then we'll
  • 00:32:46
    open source gr too okay so we probably
  • 00:32:48
    have time for one last question what was
  • 00:32:50
    the most difficult part about working on
  • 00:32:52
    this project I assume gr 3 and what I
  • 00:32:55
    most excited about I think me looking
  • 00:32:57
    looking back getting the whole model
  • 00:32:59
    training on the 100K h100 coherently
  • 00:33:03
    that's almost like battling against the
  • 00:33:05
    final boss of the universe the entropy
  • 00:33:07
    because any given time you can have a
  • 00:33:09
    cosmic rate that beaming down and flip a
  • 00:33:11
    bit in your transistor and now the
  • 00:33:13
    entire gring update if it's fit Mana bit
  • 00:33:16
    the entire grading update is out of
  • 00:33:18
    whack and now you have 100,000 of those
  • 00:33:20
    and you have to orchestrate them every
  • 00:33:22
    time any at any given time any of gpus
  • 00:33:24
    can go down yeah it's worth breaking
  • 00:33:27
    down like how were we able to get the
  • 00:33:29
    world's most powerful training cluster
  • 00:33:31
    operational within 122 days because when
  • 00:33:34
    we started off we actually weren't
  • 00:33:35
    intending to do a data center ourselves
  • 00:33:37
    we were going to just we went to the
  • 00:33:39
    data center providers and said how long
  • 00:33:40
    would it take to have 100,000 gpus
  • 00:33:43
    operating coherently in a single
  • 00:33:45
    location and we got time frames from 18
  • 00:33:47
    to 24 months so like 18 to 24 months
  • 00:33:50
    that means losing as a certainty so the
  • 00:33:52
    only option was to do it ourselves so
  • 00:33:55
    then if you break down the problem I
  • 00:33:56
    guess I'm doing like reasoning here like
  • 00:33:59
    makes you think um one single chain
  • 00:34:01
    though exactly we needed a building we
  • 00:34:03
    can't build a building so we must use an
  • 00:34:04
    existing building so we looked for for
  • 00:34:07
    basically for factories that had been
  • 00:34:09
    were that had been abandoned but the
  • 00:34:11
    factory was in good shape like a company
  • 00:34:13
    had gone bankrupt to something so we
  • 00:34:14
    found an electrox Factory in memph in
  • 00:34:16
    Memphis that's why it's in Memphis home
  • 00:34:18
    of Alvis and also one of the oldest I
  • 00:34:20
    think it was the capital of ancient
  • 00:34:21
    Egypt and it was actually very nice
  • 00:34:24
    Factory that I know for whatever reason
  • 00:34:26
    that electrox had left and uh that that
  • 00:34:29
    gave us shelter for the computers uh
  • 00:34:32
    then we needed power the we needed um at
  • 00:34:35
    least 120 megawatt at first but the
  • 00:34:37
    building only had 15 megawatts and
  • 00:34:39
    ultimately for 200,000 me 200,000 gpus
  • 00:34:41
    we needed a qu gwatt so we um initially
  • 00:34:45
    uh leased uh a whole bunch of um
  • 00:34:47
    generators so we have generators on one
  • 00:34:49
    side of the building just one trailer
  • 00:34:51
    after trail trailer of generators until
  • 00:34:53
    we can get the utility power to to come
  • 00:34:55
    in um and then but then we also need
  • 00:34:57
    Cooling so on the other side of the
  • 00:34:58
    building it was just trailer after
  • 00:34:59
    trailer of of cooling so we leased about
  • 00:35:01
    a quarter of the mobile cooling capacity
  • 00:35:03
    of the United States uh on the one other
  • 00:35:05
    side of the building um then we needed
  • 00:35:07
    to get the gpus all installed and
  • 00:35:09
    they're all liquid cooled so in order to
  • 00:35:11
    achieve the density necessary this is a
  • 00:35:13
    liquid cooled system so we had to get
  • 00:35:14
    all the plumbing for the liquid cooling
  • 00:35:16
    nobody had ever done a liquid cooling uh
  • 00:35:18
    data center at scale so this was a
  • 00:35:22
    incredibly dedicated effort by a very
  • 00:35:23
    talented team to achieve that outcome
  • 00:35:25
    now may think now now it's going to work
  • 00:35:27
    nope the the issue is that the power
  • 00:35:29
    fluctuations for GPU cluster are
  • 00:35:32
    dramatic so it's it's like a this giant
  • 00:35:34
    Symphony that is taking place imagine
  • 00:35:36
    having a symphony with 100,000 or
  • 00:35:40
    200,000 participants in the in the
  • 00:35:42
    symphony and the whole Orchestra will go
  • 00:35:44
    quiet and loud you know 100 milliseconds
  • 00:35:47
    and so this caused massive power
  • 00:35:48
    fluctuations so then uh which then
  • 00:35:51
    caused the generators to lose their
  • 00:35:52
    minds and they they weren't expecting
  • 00:35:54
    this to buffer the power we then used
  • 00:35:56
    Tesla Mega packs to smooth out the power
  • 00:36:00
    so the mega packs had to be reprogrammed
  • 00:36:03
    so with xai working with Tesla we
  • 00:36:05
    reprogrammed the MEAP packs to be able
  • 00:36:07
    to deal with these dramatic power fluctu
  • 00:36:10
    fluctuations to smooth out the power so
  • 00:36:12
    the computers could actually run
  • 00:36:13
    properly and that that worked was quite
  • 00:36:16
    tricky and and then but even at that
  • 00:36:19
    point you still have to make the
  • 00:36:20
    computers all communicate effectively so
  • 00:36:22
    all the networking had to be solved and
  • 00:36:24
    debugging a zillian network cables a
  • 00:36:27
    bugging nickel at 4: in the morning we
  • 00:36:30
    sold it like roughly 4:20 a.m. yes than
  • 00:36:34
    was figured out like there's some there
  • 00:36:36
    were a whole bunch of issues one there
  • 00:36:37
    was like a bios mismatch bios was not
  • 00:36:40
    set up correctly yeah we had d r LS PCI
  • 00:36:45
    outputs between two different machines
  • 00:36:47
    one that was working yeah one that was
  • 00:36:49
    not working yeah many other things yeah
  • 00:36:51
    exactly this would go on for a long time
  • 00:36:52
    if we actually listened to all the
  • 00:36:53
    things but know it's like like it's not
  • 00:36:54
    oh we just magically made it happen you
  • 00:36:56
    had to break down the problem just like
  • 00:36:57
    gr does for reasoning uh into the
  • 00:36:59
    constituent elements and then solve each
  • 00:37:00
    of the constituent elements in order to
  • 00:37:03
    achieve uh a a coherent training cluster
  • 00:37:06
    in a period of time that is a small
  • 00:37:08
    fraction of what anyone else was could
  • 00:37:09
    do it
  • 00:37:10
    in and then once the training cluster
  • 00:37:12
    was up and running and we could use it
  • 00:37:14
    now we had to make sure that it actually
  • 00:37:15
    stays healthy throughout which is its
  • 00:37:16
    own giant Challenge and then we had to
  • 00:37:19
    get every single detail of the training
  • 00:37:20
    right in order to get a gr Free level
  • 00:37:23
    model which is actually really hard we
  • 00:37:25
    don't know if there are any other models
  • 00:37:26
    out there that have gr's capabilities
  • 00:37:28
    but whoever trains a model better than
  • 00:37:30
    gr has to be extremely good at the the
  • 00:37:32
    science of deep learning at every aspect
  • 00:37:33
    of the engineering so it's not so easy
  • 00:37:36
    to pull this off and this is now going
  • 00:37:37
    to be the last cluster we build and last
  • 00:37:39
    Model we train oh yeah we've already
  • 00:37:41
    started work on the next
  • 00:37:43
    cluster which will
  • 00:37:45
    be yeah about five times the power so
  • 00:37:47
    instead of a quarter gaw roughly 1.2
  • 00:37:51
    gaw what's the Back to the Future
  • 00:37:54
    War what's the power you does like the
  • 00:37:57
    Back to the Future car yeah don't anyway
  • 00:38:00
    the Back to the Future power car it's
  • 00:38:02
    it's like roughly in that order I think
  • 00:38:03
    and these will be the sort of the gv200
  • 00:38:06
    SL300 cluster it it once again it will
  • 00:38:08
    be the most powerful train cluster in
  • 00:38:10
    the world so we're not stopping here no
  • 00:38:13
    and our reason model is going to
  • 00:38:14
    continue improve by accessing more tools
  • 00:38:16
    every day yeah we're very excited to
  • 00:38:18
    share any of the upcoming results with
  • 00:38:20
    you all yeah the thing that keeps us
  • 00:38:22
    going is basically being able to give G
  • 00:38:24
    free to you and then seeing the usage go
  • 00:38:26
    up seeing everybody enjoy gr that's what
  • 00:38:30
    really gets us up in the morning yeah
  • 00:38:34
    yeah thanks for tuning in thanks guys
Tag
  • AI
  • Machine Learning
  • Grok 3
  • XAI
  • Data Center
  • Deep Search
  • Reasoning
  • Tech Presentation
  • Innovation
  • Understanding Universe