Latest / Elon Musk Thinking / Latest Interview of Elon Talking About AI Grok 3!!!

Transcript

Latest Interview of Elon Talking About AI Grok 3!!! · Elon Musk Thinking

0:00
Speaker 1: Welcome to the GROC three presentation.
0:03
Speaker 2: The mission of Xai and Grock is to understand the universe.
0:08
Speaker 2: We want to understand the nature of the universe so
0:10
Speaker 2: we can figure out what's going on. Where are the aliens,
0:12
Speaker 2: what's the meaning of life, how does the universe end?
0:14
Speaker 1: How did it start? All these fundamental questions.
0:17
Speaker 2: We're driven by curiosity about the nature of the universe,
0:20
Speaker 2: and that's also what causes us to be a maximally
0:24
Speaker 2: truth seeking AI, even if that truth is sometimes at.
0:28
Speaker 1: Odds with what is politically correct.
0:31
Speaker 2: In order to understand the nature of the universe, you
0:33
Speaker 2: must absolutely rigorously pursue truth or you will not understand
0:37
Speaker 2: the universe. You'll be suffering from some amount of delusion
0:41
Speaker 2: or error. So that is our goal, figure out what's
0:44
Speaker 2: going on. And we're very excited to present GROC three,
0:47
Speaker 2: which is we think, in order of magnitude, more capable
0:51
Speaker 2: than GROC two in a very short period of time.
0:54
Speaker 1: And that's thanks to the hard work of an incredible team.
0:58
Speaker 2: My monitor to work with such a great team, and
1:00
Speaker 2: of course we'd love to have so the smartest humans
1:03
Speaker 2: out there join our team.
1:05
Speaker 1: So let's go, Hi Ron.
1:07
Speaker 3: My name is Igor lead Engineering at xi I GBPA
1:11
Speaker 3: leading research. I'm Tony working on the reasoning team, all right, Ele,
1:15
Speaker 3: And I don't do anything. I just show up occasionally.
1:19
Speaker 1: Yeah.
1:20
Speaker 3: So, like I mentioned, GROK is the tool that we're
1:23
Speaker 3: working on. Grock is our AI that we're building here
1:25
Speaker 3: at XAI, and we've been working extremely hard over the
1:27
Speaker 3: last few months to improve GROC as much as we
1:29
Speaker 3: can so we can give all of you access to it.
1:32
Speaker 3: We think it's going to be extremely useful. Do we
1:34
Speaker 3: think it's going to be interesting to talk to a
1:36
Speaker 3: funny really really funny and we're going to explain to
1:39
Speaker 3: you how we've improved GROC over the last few months.
1:41
Speaker 3: We've made quite a jump in capabilities. Actually we should
1:44
Speaker 3: explain maybe also what is why do we call it groc?
1:46
Speaker 2: So grog is a word from a Heinland novel Stranger
1:49
Speaker 2: in a Strange Land, and it's used by a guy
1:52
Speaker 2: who was raised on Mars. And the word grock is
1:55
Speaker 2: to sort of fully and profoundly understand something. That's what
1:59
Speaker 2: the word grog means, fully and profoundly understand something.
2:02
Speaker 1: And empathy is important. True.
2:04
Speaker 4: So yeah, if we charged xs progress in the last
2:08
Speaker 4: few months, it's only been seventeen months since we started
2:12
Speaker 4: kicking off our very first model. GROCK one was almost
2:17
Speaker 4: like a toy by this point on the two hundred
2:19
Speaker 4: and fourteen billion parameters. And now if we're prout the
2:22
Speaker 4: progress the time on XXI is the performance of favorite
2:27
Speaker 4: benchmark numbers at MLU on the yaxis, we're literally progressing
2:32
Speaker 4: at the unprecedent speed across the whole field. And then
2:36
Speaker 4: we kick off GROG one point five right after GUARG
2:38
Speaker 4: one released after November twenty twenty three, and then GROG two.
2:43
Speaker 1: So if you look at where all the performance coming from.
2:46
Speaker 4: We have a very correct engineering team and all the
2:49
Speaker 4: best AII talent. The only one thing we need is
2:53
Speaker 4: a big intelligence comes from big cluster, so we can
2:57
Speaker 4: reconvert the entire progress and makes AI now replacing the
3:01
Speaker 4: benchmarket of waxes to the total amount of training flops.
3:04
Speaker 4: That is how many GPS we can run at any
3:07
Speaker 4: given time to trail all large language models to compress
3:11
Speaker 4: the entire Internet.
3:12
Speaker 1: So after old human knowledge. Really that's right.
3:16
Speaker 2: Yeah, the Internet is being part of it, but it's
3:17
Speaker 2: really all human knowledge everything.
3:19
Speaker 4: Yeah, the whole Internet fits into USB stick at this
3:22
Speaker 4: point it's like all the human tokens very soon into
3:24
Speaker 4: the real world. We had so much trouble actually training
3:27
Speaker 4: Grock two back in the days.
3:29
Speaker 1: We kickoff the model around February.
3:31
Speaker 4: And we thought we had a large amount of chips,
3:34
Speaker 4: but turned out we can barely get AK training chips
3:37
Speaker 4: running coherently at any given time, and we had so
3:42
Speaker 4: many cooling and power issues.
3:45
Speaker 1: I think you were there in the data center.
3:47
Speaker 2: Yeah, it was like really sort of more like AK
3:50
Speaker 2: chips on average at eighty percent efficiency, more like like
3:53
Speaker 2: sixty five hundred effective h one hundreds training for near
3:58
Speaker 2: several months, but you know we're the one hundred k.
4:00
Speaker 1: So yeah, that's right, that's right. So what's the next step?
4:06
Speaker 4: Right, so after Groark too, So if we all continue accelerate,
4:10
Speaker 4: we have to take the matter into our own hands.
4:12
Speaker 4: We have to solve all the coolings, all the power
4:15
Speaker 4: issues and everything.
4:16
Speaker 3: So it so on April of last year, Elan decided
4:19
Speaker 3: that really the only way for XAI to succeed, for
4:22
Speaker 3: XI to build the best AI out there, is to
4:25
Speaker 3: build our own data center. So really we realize we
4:28
Speaker 3: have to build the data center in about four months.
4:31
Speaker 3: It turned out it took US one hundred and twenty
4:33
Speaker 3: two days to get the first one hundred KGPUS up
4:36
Speaker 3: and running, and that was a monumental effort to be
4:38
Speaker 3: able to do that. We believe it's the biggest fully
4:41
Speaker 3: connected h one hundred cluster of its kind. We actually
4:45
Speaker 3: decided that we need to double the size of the
4:47
Speaker 3: cluster pretty much immediately if we want to build the
4:50
Speaker 3: kind of way that we want to build.
4:52
Speaker 1: So we then had another phase.
4:55
Speaker 3: Which we haven't talked about publicly, so this is the
4:57
Speaker 3: first time that we're talking about this, where we doubled
5:00
Speaker 3: the capacity of the data center yet again. And that
5:03
Speaker 3: not only took us ninety two days. So we've been
5:06
Speaker 3: able to use all of diseributse, use all of this
5:08
Speaker 3: compute to improve GROG in the meantime. And basically today
5:11
Speaker 3: we're going to present the results of that, the fruits
5:14
Speaker 3: that came from that.
5:16
Speaker 4: That's yeah, So all the paths, all the roads leads
5:20
Speaker 4: to GROD three ten x more compute, more than ten
5:23
Speaker 4: x really maybe fifteen x compared to our private generation model,
5:28
Speaker 4: and GROCT finished the pre trading early January and we started,
5:32
Speaker 4: you know, the model is still currently trading. Actually, so
5:35
Speaker 4: this is a little preview of our benchmine numbers. So
5:39
Speaker 4: we evaluate the GROCK three on you know, three different
5:43
Speaker 4: categories on general mathematical reasoning, on general knowledge about STEM
5:49
Speaker 4: and science, and then also on computer science coding. So
5:55
Speaker 4: amy American Invitational Math Examination hosted you know a year.
6:00
Speaker 4: If we evaluated model performance, we can see that the
6:04
Speaker 4: GROD three across the board is in the league of
6:07
Speaker 4: its own, even as little brother Groctor Mini is reaching
6:12
Speaker 4: the frontier across.
6:14
Speaker 1: All the other competitors.
6:15
Speaker 4: You would say, well, at this point, all these benchmarks
6:18
Speaker 4: you're just evaluating, you know, the memorization of the textbooks,
6:22
Speaker 4: memorization of the GitHub repos, how about real time usefulness?
6:26
Speaker 4: How about we actually use those models in our product.
6:29
Speaker 4: So what we did instead is we actually kicked off
6:33
Speaker 4: a blind test of our garacy model code named Chocolate,
6:38
Speaker 4: Pretty Hot Chocolate. I've been running on this platform called
6:42
Speaker 4: chap Arena for two weeks. I think the entire x
6:46
Speaker 4: platform at some point speculated this might be the next
6:49
Speaker 4: generation of AI coming away. So how this chap arena
6:54
Speaker 4: works is that it's stripped away the entire product service, right,
6:58
Speaker 4: It just raw comparison of the engine of those Asia's
7:01
Speaker 4: the language models themselves and place interface.
7:04
Speaker 1: Where the user will submit one single querry and you.
7:07
Speaker 4: Get to show two responses, you don't know which model
7:10
Speaker 4: they come from, and India you make the vote. So
7:14
Speaker 4: in this blind test, GROX three, an early version of
7:17
Speaker 4: GROX three already reached like fourteen hundred. No other models
7:22
Speaker 4: had reached an ELO score had to have comparison to
7:25
Speaker 4: all the other models at this score.
7:27
Speaker 1: And it's not just one single category.
7:30
Speaker 4: It's fourteen hundred aggregated across all the categories in chapbock capabilities,
7:35
Speaker 4: in struction, following coding, So it's number one across the
7:39
Speaker 4: board in this.
7:40
Speaker 2: Blind test, and it's still climbing, so we actually keep
7:43
Speaker 2: to keep updating it. So it's forty fourteen hundred, but
7:45
Speaker 2: fourteen hundred and climbing.
7:47
Speaker 3: Yeah, And in fact, we have a version of the
7:48
Speaker 3: model that we think is already much better than the
7:50
Speaker 3: one that we tested here. Yeah, we'll see, I guess,
7:55
Speaker 3: but that's the one that we're working on talking about today.
7:59
Speaker 1: Yeah.
7:59
Speaker 2: So actually, thing, if you're using BARG three, you I
8:02
Speaker 2: think you may notice improvements almost every day because we're
8:06
Speaker 2: continuously improving the model. So literally even within twenty four hours,
8:10
Speaker 2: you'll see improvements.
8:11
Speaker 1: Yep.
8:13
Speaker 4: But we believe here at the XAI, getting the best
8:16
Speaker 4: pre training model is not enough. That's not enough to
8:19
Speaker 4: build the best AI and the best A. I need
8:22
Speaker 4: to think like a human to contemplate about all the
8:25
Speaker 4: possible solutions, self critique, verify all the solutions, backtrack, and
8:31
Speaker 4: also think from the first principle.
8:33
Speaker 1: That's a very important capability.
8:35
Speaker 4: So we believe that as we take the best PREA
8:38
Speaker 4: training model and continue training with reinforcement learning, it will
8:43
Speaker 4: enlicit the additional reasoning capabilities that allows the model just
8:47
Speaker 4: becomes so much better and scale, not just in the
8:50
Speaker 4: training time, but actually in the test time as well.
8:53
Speaker 4: So we already found the model is extremely useful internally,
8:56
Speaker 4: saving one hundreds of hours of coding time. So you
8:59
Speaker 4: are the user of our these model, what does the
9:03
Speaker 4: use cases yea.
9:03
Speaker 3: So, like Jimmy said, we've added advanced reasoning capabilities to
9:07
Speaker 3: GROG and we've been testing them pretty heavily over the
9:09
Speaker 3: last few weeks in order to give you a little
9:11
Speaker 3: bit of a taste of what it looks like when
9:12
Speaker 3: GROG is solving hard reasoning problems. So we've prepared two
9:15
Speaker 3: little problems for you. One that comes from physics and
9:18
Speaker 3: one is actually a game that God is going to
9:20
Speaker 3: write for us. When it comes to the physics problem,
9:23
Speaker 3: you know, what we want Grog to do is to
9:25
Speaker 3: plot a viable trajectory to do a transfer from Earth
9:29
Speaker 3: to Mars and then at a later point in time
9:31
Speaker 3: a transfer back from Mars to Earth. And that requires,
9:34
Speaker 3: you know, some physics that Grog will have to understand.
9:37
Speaker 3: So we're going to challenge Grock, you know, come up
9:39
Speaker 3: with a viable trajectory, calculate it, and then plot it
9:43
Speaker 3: for us so we can see it. And yeah, this
9:46
Speaker 3: is totally unscripted, by the way, this is the Grog interface,
9:49
Speaker 3: and we've typed in this text that you can see
9:52
Speaker 3: here generate code for an animated three D plot of
9:54
Speaker 3: a launch from Earth, landing on Mars, and then back
9:58
Speaker 3: to Earth at the next launch window. And we've not
10:01
Speaker 3: kicked off or the query and you can see Grog spinking.
10:04
Speaker 3: So out of Rock's advanced reasoning capabilities are these thinking
10:08
Speaker 3: traces that you can see here. You can even go
10:10
Speaker 3: inside and actually read what Grock is thinking as it's
10:13
Speaker 3: going through the problem, as it's trying to solve it.
10:15
Speaker 2: Yeah, weld say like we are doing some obscuration of
10:18
Speaker 2: the thinking so that our model doesn't get totally copied instantly.
10:22
Speaker 2: So there's more to the thinking than is displayed.
10:27
Speaker 3: And because this is totally unscripted, there's actually a chance
10:30
Speaker 3: that Grock might make a little courting mistake and it
10:33
Speaker 3: might not actually work. So just in case, we're going
10:35
Speaker 3: to launch two more instances of this, so if something
10:38
Speaker 3: goes wrong, we were able to search to those enshow
10:41
Speaker 3: you something that's presentable, So we're kicking off the other
10:45
Speaker 3: two as well, And like I said, we have a
10:47
Speaker 3: second problem as well. Actually, one of our favorite activities
10:51
Speaker 3: here XCI is having rock right games for us, not
10:55
Speaker 3: just any now any old game, any game that you
10:57
Speaker 3: might already be familiar with, but actually creating new games
11:00
Speaker 3: on the spot and being creative about us. So one
11:04
Speaker 3: example that we found was really really fun is create
11:07
Speaker 3: a game that's a mixture of the two games Tetris
11:11
Speaker 3: and Jewels.
11:12
Speaker 2: So this is maybe an important thing like that this
11:15
Speaker 2: obviously if you if you ask an ai to create
11:18
Speaker 2: a game like Tetris, there's there are many examples of
11:20
Speaker 2: Tetris on the on the internet.
11:21
Speaker 1: The air or game like Dueled whatever, this, it can
11:25
Speaker 1: copy it.
11:26
Speaker 2: What's interesting here is it achieved a creative solution combining
11:31
Speaker 2: the two games that actually works and is a good game. Yeah,
11:36
Speaker 2: that's the it's created. We're seeing the beginnings of creativity.
11:40
Speaker 3: Fingers crossed that we can recreate that, but hopefully it works.
11:43
Speaker 3: Hopefully it's actually because this is a bit more challenging,
11:45
Speaker 3: we're going to use something special here, which we call
11:47
Speaker 3: big Brain. That's our mode in which we use more computation,
11:51
Speaker 3: which was more reasoning of our rock, just to make
11:54
Speaker 3: sure that you know, there's a good chance here that
11:56
Speaker 3: actually might actually do it. So we're also going to
11:59
Speaker 3: fire of free attempts here at solving this game, at
12:04
Speaker 3: creating this game that's a mixture of Tetris and the Jewels.
12:09
Speaker 3: Let's let's see what god comes out.
12:10
Speaker 2: I've played the game. It's pretty good. Like it's like, wow, okay,
12:14
Speaker 2: this is something.
12:16
Speaker 3: Yeah, So while Grog is thinking in the background, we
12:19
Speaker 3: can now actually talk about some concrete numbers. How well
12:23
Speaker 3: is Groug doing across tons of different tasks that we've
12:25
Speaker 3: tested on. So we'll hand it over to Tony to talk
12:28
Speaker 3: about that.
12:30
Speaker 5: Yeah, Okay, so let's see how Greg has on those
12:33
Speaker 5: interesting challenging benchmarks. So yeah, so reasoning again refers to
12:38
Speaker 5: those models that actually thinks quite for quite a long
12:41
Speaker 5: time before it tries to solve a problem. So in
12:45
Speaker 5: this case, you know, around the months ago the growth
12:48
Speaker 5: three apprecionning finishes, so after that we worked very hard
12:52
Speaker 5: to put the reasoning capability into the current Growth three model.
12:57
Speaker 5: But again this is very early days, so the model
13:00
Speaker 5: still currently in training. So right now, what we're going
13:03
Speaker 5: to show to people is this beta version of the
13:06
Speaker 5: growth three reasoning model. Alongside, we also are training a
13:09
Speaker 5: mini version of the reasoning model. So essentially on this
13:12
Speaker 5: plot you can see the growth three reasoning beta and
13:16
Speaker 5: then Growth three mini reasoning. The reasoning mini reasoning is
13:19
Speaker 5: actually a model that we train for much longer time,
13:22
Speaker 5: and you can see that sometimes you actually perform slightly
13:25
Speaker 5: better compared to the Growth three reasoning. This also just
13:28
Speaker 5: means that there's a huge potential for the Growth three
13:31
Speaker 5: reasoning because it's trained for.
13:32
Speaker 1: Much less time.
13:34
Speaker 5: So all right, so let's actually look at how it
13:37
Speaker 5: does on those three benchmarks. So Jimmy also introduced already,
13:42
Speaker 5: so essentially we're looking at three different areas mathematics, science,
13:46
Speaker 5: and coding. And for math we're picking this high school
13:49
Speaker 5: competition math problem. For science we actually picked those PhD
13:54
Speaker 5: level science questions.
13:56
Speaker 1: And for coding, it's also actually pretty challenging.
13:59
Speaker 5: It's competitive code and also some eco, which is some
14:03
Speaker 5: interview problems that people usually get when they interview for companies.
14:07
Speaker 5: So on those benchmarks, you can see that the growth
14:10
Speaker 5: three actually performed quite well across the board compared to
14:14
Speaker 5: other competitors. Yeah, so it's pretty promising. These models are
14:18
Speaker 5: very smart.
14:19
Speaker 1: So totally what what are those shaded bars? Yeah, so okay,
14:23
Speaker 1: so I'm you asked this question.
14:25
Speaker 5: So for those models, because it can reason, it can think,
14:29
Speaker 5: you can also ask them to even think longer. You
14:32
Speaker 5: can spend more what we call test and compute, which
14:36
Speaker 5: means you can spend more time to reason, to think
14:40
Speaker 5: about the problem before you spit out the answer. So
14:43
Speaker 5: in this case, the shaded bar here means that we
14:46
Speaker 5: just asked the model to spend more time. You know,
14:50
Speaker 5: it can solve the same problem many many times before
14:53
Speaker 5: it tries to conclude what is the right solution, And
14:57
Speaker 5: once you give this compute or this kind of budget
15:00
Speaker 5: to the model. It turns out the model can even
15:02
Speaker 5: perform better. So this is it's honly the shaded part
15:06
Speaker 5: in those.
15:07
Speaker 4: Plots, right, So I think this is really exciting, right
15:10
Speaker 4: because now instead of just doing one chain of thoughts
15:13
Speaker 4: with AI while not the multiplex at once. Yes, so
15:17
Speaker 4: that's a very powerful technique that allows to continue to
15:19
Speaker 4: scale the model capabilities after training. And you know, people
15:25
Speaker 4: often ask were actually just overfitt into the benchmarks?
15:28
Speaker 1: So how mogulization?
15:30
Speaker 5: So yes, I think, yeah, this is definitely a question
15:33
Speaker 5: that we are asking ourselves whether we're overfitting into those
15:36
Speaker 5: current benchmarks.
15:38
Speaker 1: Luckily, we have a real test.
15:41
Speaker 5: So about five days ago, AIM twenty twenty five just finished.
15:45
Speaker 5: This is where high school come students compete in this
15:48
Speaker 5: particular benchmark. So we got this very fresh new competition,
15:52
Speaker 5: and then we ask our two models to compete on
15:55
Speaker 5: the same benchmark, the same exam, and it turns out
15:59
Speaker 5: very interest the gross three reasoning the big one actually
16:03
Speaker 5: does better on this particular new, fresh exam. This also
16:08
Speaker 5: means that the generalization capability of the big model is stronger,
16:12
Speaker 5: much stronger compared to the smaller model. If you compare
16:15
Speaker 5: to the last year's exam. Actually this is the opposite.
16:18
Speaker 5: The smaller model kind of learns the previous exams better.
16:23
Speaker 5: So yeah, so this this actually shows some kind of
16:25
Speaker 5: true generalization from the model.
16:28
Speaker 4: Right, So, seventeen months ago our Rock zero and Rock
16:31
Speaker 4: Web baret solved any high school problems. That's right, And
16:34
Speaker 4: now we have a kid that just already graduate. The
16:37
Speaker 4: Grock is right to go to college?
16:39
Speaker 1: Is that right? I mean, it's won't belong before. It's
16:42
Speaker 1: simply perfect.
16:43
Speaker 2: The human exams won't be part it'll be too easy.
16:46
Speaker 4: Yeah, and the internally we actually as a rocket continually evolved.
16:51
Speaker 1: We're going to talk about, you know what we're.
16:52
Speaker 4: Excited about, but very soon there will be no more
16:56
Speaker 4: benchmark left.
16:57
Speaker 1: Yeah. Yeah. One thing that's quite fascinating.
16:59
Speaker 3: I think that we basically only trained Grock's reasoning abilities
17:03
Speaker 3: on math problems and competitive coding problems. It's so very
17:07
Speaker 3: very specialized kinds of tasks, but somehow it's able to
17:11
Speaker 3: work on all kinds of other different tasks, so including
17:14
Speaker 3: creating games. No, lots and lots of different things. And
17:18
Speaker 3: what seems to be happening is that basically Grog learns
17:21
Speaker 3: this ability to detect its own mistakes and it's thinking,
17:24
Speaker 3: correct them, persist on a problem, try lots of different variants,
17:27
Speaker 3: pick them one that's best. So there are is generalized
17:30
Speaker 3: generalizing abilities that learns from mathematics and from coding, which
17:35
Speaker 3: it can then use to solve all kinds of other problems.
17:37
Speaker 1: So that's yeah, that's pretty I mean, reality is the
17:39
Speaker 1: instantiation of mathematics. That's right.
17:43
Speaker 4: And one thing we're actually really excited about that going
17:45
Speaker 4: back to our faulty mission, is what if one day
17:48
Speaker 4: we have a computer just like the thought that utilize
17:52
Speaker 4: our entire.
17:53
Speaker 1: Cluster just for the one very important problem.
17:56
Speaker 4: In the test time, all the deply turned out right,
17:58
Speaker 4: So I think back then we'll building the GBU clusters together. Uh,
18:02
Speaker 4: you're applobing cables, and I remember that when we turned
18:06
Speaker 4: on the first initial test, you can hear all the
18:09
Speaker 4: GPS humming in the hallway.
18:12
Speaker 1: That's almost feel like spiritual.
18:14
Speaker 3: Yeah, that's actually a pretty cool thing that we're able
18:16
Speaker 3: to do. That we can go into the data center
18:19
Speaker 3: and tinker with the machines there. So for example, we
18:22
Speaker 3: went in and we unplucked a few of the cables
18:25
Speaker 3: and just made sure that our training setup is still running,
18:28
Speaker 3: running stably. So that's something that's you know, I think
18:31
Speaker 3: most AI you know teams out there don't usually do.
18:34
Speaker 3: But it's actually totally unlocks like a new level of
18:37
Speaker 3: reliability and what you're able to do with the hypers.
18:40
Speaker 1: So okay, so when are we going to solve remont?
18:43
Speaker 4: So the easiest solution is to enumerate over all possible
18:47
Speaker 4: strains and as all you have a verifier, en up compute,
18:51
Speaker 4: you'll be able to do it.
18:52
Speaker 1: Okay, my projection will be what your guess, what is
18:56
Speaker 1: your neural edge? Calculate?
18:58
Speaker 4: So my my boat for the so so three years
19:00
Speaker 4: ago I told you that I think not two years later,
19:05
Speaker 4: two things is gonna happen.
19:06
Speaker 1: We're gonna see machines win some battles.
19:09
Speaker 4: Yes, two ways a word, fields metal, Globel price with
19:14
Speaker 4: probably some expert in the loop.
19:16
Speaker 3: Right, so the expert uplifting. So this year or next year,
19:20
Speaker 3: that's what it comes down to you. Yeah, So it
19:22
Speaker 3: looks like GROG finished all of its thinking on the
19:24
Speaker 3: on the two problems. So let's take a look at
19:26
Speaker 3: what it said. All right, so this was the little physics.
19:29
Speaker 1: Problem we had.
19:31
Speaker 3: You know, we we've collapsed the thoughts here, so they're
19:34
Speaker 3: you know, they're hidden. And then we see Grock's answer
19:37
Speaker 3: below that, so it explains it wrought a Python script
19:40
Speaker 3: here using map plot Lip then gives us all of
19:42
Speaker 3: the code. So let's take a quick look at the code.
19:45
Speaker 3: You know, it seems like it's doing reasonable things here,
19:48
Speaker 3: not not totally of the mark, solf Capitler says here,
19:53
Speaker 3: so maybe it's solving capitalist laws, capital capitalist law and americally. Yeah,
19:58
Speaker 3: there's really only one way to find out if this
20:00
Speaker 3: thing is working.
20:01
Speaker 1: I would say, let's let's give it a try. Let's run.
20:03
Speaker 3: Let's run the code all right, and we can see, yeah,
20:07
Speaker 3: I've got animating two different planets, Earth and Mars here,
20:11
Speaker 3: and then the green ball is the vehicle that's transiting
20:16
Speaker 3: the spacecraft that's transitioning between Earth and Mars, and you
20:19
Speaker 3: could see the journey from Earth to Mars and looks like, yeah,
20:23
Speaker 3: indeed the astronauts were turned safely, you know, at the
20:26
Speaker 3: right moment in time. So obviously this was just generated
20:31
Speaker 3: on the spot, so we can't tell you if that
20:33
Speaker 3: was actually correct solution. So we're going to take a
20:36
Speaker 3: close a look. Now, maybe we're going to call some
20:37
Speaker 3: colleagues a space X ask them if if this.
20:40
Speaker 2: Is legit, that's pretty close it's it's I mean yeah,
20:46
Speaker 2: I mean, there's there's there's there's a lot of complexities
20:48
Speaker 2: in the actual orbits that have to be taking into account,
20:50
Speaker 2: but this is this is pretty close to what looks
20:54
Speaker 2: like I add that or my pendets here. This has
20:59
Speaker 2: got the Earth home and transfer on it. When where
21:02
Speaker 2: we're going to install rock on a rocket.
21:04
Speaker 1: Well, I suppose in two years three years.
21:08
Speaker 2: Everything is two years away. Well, Earth and Mars. Transit
21:14
Speaker 2: can occurs every twenty six months. The next we're currently
21:17
Speaker 2: in a transit window approximately. The next one would be
21:21
Speaker 2: November of next year, roughly the end of next year,
21:27
Speaker 2: and if all go as well, SpaceX will send starship
21:31
Speaker 2: rockets to Mars with Optimus robots and Rock.
21:37
Speaker 1: Mh.
21:38
Speaker 3: I'm curious about this combination of Tetris and the duets.
21:42
Speaker 3: Looks like the Tetris as we've named it internally. So okay,
21:48
Speaker 3: we also have an output from Rock here wrote a
21:51
Speaker 3: Python script Spence.
21:52
Speaker 1: That is what it's been doing.
21:53
Speaker 3: If you look at the code, there are some constants
21:57
Speaker 3: that are being defined here, some.
21:58
Speaker 1: Colors, and then the totrominos.
22:01
Speaker 3: The pieces of Tetris are there, obviously very hard to
22:05
Speaker 3: see and at one glance if this is good, so
22:07
Speaker 3: we gotta we gotta run this to figure out if
22:09
Speaker 3: it's working.
22:10
Speaker 1: Well, let's let's give it a try.
22:12
Speaker 3: Fingers crossed ay, Right, So this kind of looks like Tetris,
22:16
Speaker 3: but the colors are a little bit off, right, The
22:18
Speaker 3: colors are different here, and if you think about what's
22:24
Speaker 3: going what's going on here? The jewel it has this
22:27
Speaker 3: mechanic where if you get free jewels in a row,
22:30
Speaker 3: you know, then they disappear and also gravity activates. Right,
22:35
Speaker 3: So what happens if you get three of the colors together? Okay,
22:38
Speaker 3: so something happened. So I think I think what Bruck
22:43
Speaker 3: did in this version is that, you know, once you
22:47
Speaker 3: connect three at least three blocks of the same color
22:50
Speaker 3: in a row.
22:51
Speaker 1: Then.
22:53
Speaker 3: Gravity activates and they disappear, and then gravity activates and
22:57
Speaker 3: all the other blocks fall down. I'm kind of kind
23:00
Speaker 3: of curious if there's still a Tetris mechanic here where
23:04
Speaker 3: if the line is full, does it actually clear it
23:08
Speaker 3: or what happens? Then it's up to interpretation.
23:11
Speaker 2: So who knows every I mean, when you're it'll do
23:14
Speaker 2: different variants when you ask it.
23:15
Speaker 3: It doesn't do the same thing every time exactly. We've
23:18
Speaker 3: seen a few other work very differently, but that's one
23:22
Speaker 3: seems cool.
23:22
Speaker 4: So are we ready for game studio at x l AI.
23:27
Speaker 2: Yes, so we're launching an AI gaming studio at x
23:31
Speaker 2: c I. If you're interested in joining us in building
23:33
Speaker 2: AI games, Uh, please join x AI where we're launching
23:37
Speaker 2: in AI Gaming studio or announcing it tonight.
23:39
Speaker 1: Let's go. Yeah, big games, that's an actual game. Yeah,
23:45
Speaker 1: all right.
23:45
Speaker 4: So I think one thing is super exciting for us
23:50
Speaker 4: is that once you have the best patriot model, you
23:54
Speaker 4: have the best reason model.
23:55
Speaker 1: Right, so we already see that we.
23:58
Speaker 4: Actually give a capability for those model.
24:00
Speaker 1: To think harder, think longer, think more broad.
24:04
Speaker 4: The performance continuing improves, and we're really excited about the
24:08
Speaker 4: next front here that will happen if we'll not only
24:10
Speaker 4: allow the model to think harder, but also provide more tools,
24:14
Speaker 4: just like call real humans to solve those problems. For
24:17
Speaker 4: real humans, we don't ask them to solve women hypothesis
24:21
Speaker 4: just with a piece of pen and paper. The Internet,
24:25
Speaker 4: so with all the basic web browsing, search engine and
24:30
Speaker 4: coding interpreters that builds the foundations and the best reasoning
24:34
Speaker 4: model build the foundations for the grog agent to come.
24:39
Speaker 4: So today we're actually introducing a new product called deep
24:45
Speaker 4: search that is the first generation of our grock agents
24:49
Speaker 4: that not just helping the engineers and research and scientists
24:52
Speaker 4: do coding, but actually help everyone to answer questions that.
24:56
Speaker 1: You have day today.
24:57
Speaker 4: It's kind of like a next generation search engine that
25:00
Speaker 4: really help you to understand the universe. So you can
25:03
Speaker 4: start asking questions like, for example, hey, when is the
25:07
Speaker 4: next starship launch day? For example, So let's try that
25:11
Speaker 4: the answer. On the left inside we see a high
25:17
Speaker 4: level progress bar. Essentially, you know the model knowledge is
25:21
Speaker 4: going to do one single search like the current right system,
25:24
Speaker 4: but actually thought very deeply about, hey, what's the user
25:27
Speaker 4: intent here, and what are the facts that actually consider
25:30
Speaker 4: at the same time, and how many different websites actually
25:32
Speaker 4: actually go and read the accountent right, So this can
25:36
Speaker 4: save hundreds of hours of everyone's Google time if you
25:40
Speaker 4: want to really look into certain topics.
25:42
Speaker 1: And then on the right.
25:44
Speaker 4: Inside you can see the bullet of how the current
25:46
Speaker 4: model you know, is doing what websites browsing, what sources
25:51
Speaker 4: verifying and oftentimes actually cross validate different sources out there
25:56
Speaker 4: to make sure the answer is actually correct before the
25:58
Speaker 4: output final answer, we can, you know, at the same time,
26:01
Speaker 4: fire up a few more querries. How about you know,
26:04
Speaker 4: you know your gamer right, so sure? Yeah, so how
26:08
Speaker 4: about what are some of the best bills and most
26:10
Speaker 4: popular bills in uh pathl hardcore right hardcore league?
26:15
Speaker 2: You if you can technically just look at the hardcore letterer,
26:18
Speaker 2: it might be a fast.
26:19
Speaker 1: Way to figure it out.
26:20
Speaker 4: They always see what model does. And then we can
26:22
Speaker 4: also do uh you know, something more fun. For example,
26:27
Speaker 4: how about like make a prediction about the marsh madness
26:29
Speaker 4: out there.
26:30
Speaker 1: Yeah, so this is.
26:31
Speaker 2: Kind of a fun one where Warren Buffett has a
26:34
Speaker 2: billion dollar bet if you can exactly match the I
26:38
Speaker 2: think the the the sort of the entire winning tree
26:43
Speaker 2: of marsh Madness, you can win a billion dollars from
26:45
Speaker 2: Warren Buffett. So, like, it would be pretty cool if
26:48
Speaker 2: AI could help you win a billion dollars from Buffett.
26:51
Speaker 1: It seems like a pretty good investment.
26:53
Speaker 4: Let's go yeah, all right, so now let's fire up
26:57
Speaker 4: the quarry and see what mama does.
27:00
Speaker 1: So we can actually go back to our very first one.
27:03
Speaker 4: How about the it wasn't counting on this, that's right, Okay,
27:06
Speaker 4: so we got the first one and all the thought
27:09
Speaker 4: around one minute. Okay, so the key inside here. The
27:12
Speaker 4: next starship is going to be on twenty fourth or later.
27:15
Speaker 4: So no earlier than February twenty fourth, it might be sooner.
27:19
Speaker 4: So yeah, so I think we can you know, go
27:22
Speaker 4: down to go down what a model does, so it
27:24
Speaker 4: does little research flight seven what happened got grounded and actually.
27:28
Speaker 1: Look into the FCC filing you know, from data collections.
27:34
Speaker 4: Uh, and that should make a new conclusion that yeh
27:38
Speaker 4: if we continue to roll down, let's see right, Yeah,
27:44
Speaker 4: so it makes uh the you know little table. I
27:47
Speaker 4: think inside XAI we often joked about the time to
27:51
Speaker 4: the first table is the only you know latency that matters. Yeah,
27:56
Speaker 4: so that's how the model making influence. And look at
27:59
Speaker 4: all the sources and then we can look into the
28:01
Speaker 4: gaming one. So how about the break So for this
28:05
Speaker 4: particular one, we look at the buildings light and you
28:10
Speaker 4: know it's a lot better so uh what they inferl us.
28:15
Speaker 4: But if we go down, so the surprising fact of
28:18
Speaker 4: all the other builds, So look into the twelve classes,
28:22
Speaker 4: so we'll see that the medium bill was pretty popular
28:25
Speaker 4: whenever the game first came out. And now the Invokers
28:28
Speaker 4: of the World.
28:29
Speaker 1: Took over Monkety Biker for sure. Yeah that's right.
28:33
Speaker 4: Yeah, followed by the stone Weavers. Then that's really good mapping.
28:36
Speaker 4: So yeah, and then we can see uh uh the
28:42
Speaker 4: match manners.
28:42
Speaker 1: How about that?
28:43
Speaker 4: Soe One interesting thing about the dep search is that
28:47
Speaker 4: if you actually go into the panel.
28:49
Speaker 1: Where it shows, you know, what are the subtasks.
28:52
Speaker 4: You can actually click the bottom left of this right
28:56
Speaker 4: and then in this case you can actually go to
29:00
Speaker 4: actually reading to the mind of Grock. What informations does
29:04
Speaker 4: the model actually think about our truck worthy what or not?
29:06
Speaker 4: How does it actually cross all their different information sources?
29:10
Speaker 4: So that makes the entire search experience and information with
29:13
Speaker 4: Trual process a lot more transparent to our users.
29:16
Speaker 3: This is much more powerful than any search engine out there.
29:19
Speaker 3: You can literally just tell it only use sources from X.
29:22
Speaker 3: You know, we'll try to respect that, and so it's
29:25
Speaker 3: much more steerable, much more intelligent than I mean.
29:28
Speaker 1: It really should save you a lot of times.
29:29
Speaker 2: So something that might take an hour or an hour
29:31
Speaker 2: of researching on the web or searching media. You can
29:34
Speaker 2: just ask it to go do that and come back
29:37
Speaker 2: in ten minutes later.
29:37
Speaker 1: It's done. An hour's work worth of work for you.
29:40
Speaker 2: That's really what it comes down to, exactly and and
29:42
Speaker 2: maybe better than you could have done it yourself.
29:44
Speaker 1: Yeah.
29:44
Speaker 4: Think about the informount of interns working for you that
29:48
Speaker 4: you can just fire up all the tasks and come
29:50
Speaker 4: back a minute later.
29:51
Speaker 1: This is going to be interesting one.
29:52
Speaker 4: So marchmass had not happened yet, so I guess we
29:56
Speaker 4: had to follow up with a next livestream.
30:00
Speaker 2: Yeah, it seems like pretty good. Like forty dollars might
30:03
Speaker 2: get you a billion dollars forty dollars subscription, that's.
30:06
Speaker 1: Right, I mean my work. So yeah, so where are
30:10
Speaker 1: the users gonna have their heads on? Rock three yees?
30:12
Speaker 3: So the good news is we've been working tirelessly to
30:16
Speaker 3: actually release all of these features that we've shown you.
30:19
Speaker 3: The Grock free based model with amazing chat capabilities that's
30:22
Speaker 3: really useful, that's really interesting to talk to, the deep search,
30:26
Speaker 3: the advanced reasoning mode, all of these things. We want
30:29
Speaker 3: to row them out to you today, starting with the
30:32
Speaker 3: plus subscribers on x So it's the first group that
30:35
Speaker 3: will initially get access.
30:37
Speaker 1: Make sure to update your x.
30:38
Speaker 3: Up if you want to see all of the advanced capabilities,
30:41
Speaker 3: because we just released the update now as we're as
30:44
Speaker 3: we're talking here, and yeah, if you're interested in getting
30:47
Speaker 3: early access to grock, then sign up for Premium Plus.
30:50
Speaker 3: And also we're announcing that we're starting a separate subscription
30:55
Speaker 3: for Grock that we call Super Grock for those who
30:57
Speaker 3: those real rock fans. That one of the most bands
31:00
Speaker 3: capabilities and the earliest access to new futures.
31:04
Speaker 1: So feel free to check that out as well.
31:06
Speaker 2: This this is for the dedicated grock app and for
31:08
Speaker 2: the website exactly.
31:10
Speaker 3: So our new website is called grock dot com. Yeah,
31:13
Speaker 3: and you're also guess and you can also find our
31:17
Speaker 3: Brock app in the ir S app Store, and that
31:20
Speaker 3: gives you a more even more polished experience that's totally
31:24
Speaker 3: grock focused. If you're if you want to have grock
31:27
Speaker 3: not easily available one tape away.
31:29
Speaker 2: Yeah, and the version on grock dot com on you know,
31:32
Speaker 2: on a web browser is going to be the most
31:34
Speaker 2: the latest and most advanced version because obviously it takes
31:36
Speaker 2: us a while to get it get something into an
31:39
Speaker 2: app and they get it approved by the app store.
31:41
Speaker 2: So and then if there's something on a phone format
31:44
Speaker 2: is limitations where you can do so. The most powerful
31:46
Speaker 2: version of grock and the latest version will be the
31:49
Speaker 2: web version at rock dot com.
31:51
Speaker 3: Yeah, so so watch out for the name grock Free
31:54
Speaker 3: in the app giveaway exactly that that's that's the giveaway
31:57
Speaker 3: that you have grock free. And if it says grow
31:59
Speaker 3: true then at grogby hasn't quite arrived for yet, but
32:02
Speaker 3: we're working hard to brow this out today and then
32:05
Speaker 3: to even more people over the coming days.
32:08
Speaker 4: Yeah, make sure you update your phone app too, where
32:11
Speaker 4: you're going to get all the tools were showcased today
32:14
Speaker 4: with the thinking mold with the deep search. So yeah,
32:18
Speaker 4: really looking forward to all the feedbacks you have.
32:20
Speaker 2: Yeah, I think we should emphasize that this is kind
32:24
Speaker 2: of a beta like meaning that you should expect some
32:26
Speaker 2: imperfections at first, but we will improve it rapidly almost
32:31
Speaker 2: every day.
32:31
Speaker 1: In fact, every day I think it'll get better.
32:34
Speaker 2: So if you want a more polished version, i'd like
32:36
Speaker 2: maybe wait a week, but expect improvements literally every day.
32:41
Speaker 2: And then we're also going to be providing a voice
32:44
Speaker 2: you can have conversational. In fact, I was trying it
32:47
Speaker 2: earlier today. It's working pretty well, but not we need
32:49
Speaker 2: these a bit more polish, the sort of way where
32:53
Speaker 2: you can just literally talk to it like you're talking
32:54
Speaker 2: to a person. That's awesome. It's actually I think one
32:58
Speaker 2: of the best experiences of Grog. That's that's probably about
33:02
Speaker 2: a week away.
33:03
Speaker 3: So with that said, I think we might have some
33:07
Speaker 3: audience questions. Surely, all right, take a look. Yeah, let's
33:12
Speaker 3: take a look the audience from the ass platform.
33:16
Speaker 2: Yeah.
33:17
Speaker 1: Cool.
33:18
Speaker 3: So the first question here is when Grock Voice Assistant.
33:21
Speaker 3: When is it coming out as soon as possible, just
33:23
Speaker 3: like Elan said, just a little bit of publishing away
33:26
Speaker 3: from being reached to everybody. Obviously, it's going to be
33:30
Speaker 3: released in an early form and we're going to rapidliterate
33:33
Speaker 3: among that.
33:34
Speaker 4: The next question is like when will GROX three being
33:37
Speaker 4: the API? So this is coming in the GROG three
33:40
Speaker 4: API with both the reasing models and deep search.
33:44
Speaker 1: Is coming away in the coming weeks.
33:46
Speaker 4: We're actually very excited about the enterprise use cases of
33:49
Speaker 4: all these additional tools that now Grock has access to,
33:52
Speaker 4: and how the testim computed and tool use can actually
33:54
Speaker 4: really accelerate all the business use cases. Another one is
33:58
Speaker 4: will voice mode be native or a text to speech?
34:01
Speaker 4: So I think that means is it going to be
34:02
Speaker 4: one one model that is understanding what you say and
34:06
Speaker 4: then talking back to you, or is it going to
34:08
Speaker 4: be some.
34:08
Speaker 3: System that has text of speech inside of it. And
34:10
Speaker 3: the good news is it's going to be one model,
34:12
Speaker 3: like and not a variant of grock Free that we're
34:15
Speaker 3: going to release, which basically understands what you're saying what
34:18
Speaker 3: you're saying and then generates the audio directly from that.
34:22
Speaker 3: So very much like grock Free generates text, that model
34:26
Speaker 3: generates audio, and that has a bunch of advantagers. I
34:29
Speaker 3: was talking to it earlier today and it said hi, Igor, no,
34:33
Speaker 3: reading my name from probably from some texts that it had,
34:36
Speaker 3: And I said, no, my name is Igor, and it
34:39
Speaker 3: remember that, you know, so it could continue to say Igor,
34:43
Speaker 3: just like a human world, and you can't.
34:44
Speaker 1: Achieve that with Texas speech.
34:46
Speaker 4: So yeah, so here's a question for you, pretty spicy,
34:52
Speaker 4: you know, is Grog a boy or a girl?
34:56
Speaker 1: And they think Grog is whatever you wanted to be? Yah?
34:59
Speaker 2: Yeah, single, yes, all right, the shop is open. So honestly,
35:06
Speaker 2: people are going to pull in love with Croc. It's
35:08
Speaker 2: it's like probable.
35:10
Speaker 1: Uh.
35:11
Speaker 4: The next question, will Grock be able to transcribe audio
35:14
Speaker 4: into text? Yes, so we'll have this capability in both
35:17
Speaker 4: the app and also the API without that's like Groschia
35:20
Speaker 4: just be your personal assistant looking over your shoulder, right
35:23
Speaker 4: and follow you along the way, learn everything you have learned,
35:26
Speaker 4: and really help you to understand the world better.
35:28
Speaker 1: It becomes smodery every day.
35:30
Speaker 2: Yeah, I mean the voicematter Groc doesn't isn't simply it's
35:33
Speaker 2: not just voice text. It understands like tone, inflection, pacing, everything.
35:38
Speaker 1: It's it's wild. I mean, it's like token a person. Yep.
35:42
Speaker 4: So any plans for conversation memory, absolutely, we're working on
35:46
Speaker 4: it right now.
35:47
Speaker 1: That's right.
35:50
Speaker 4: Let's see without the other ones. So what about the
35:56
Speaker 4: you know, the DM features? Right, So if you have
35:58
Speaker 4: personalization said that, if you uh, you know Grock remembers
36:02
Speaker 4: your previous interactions, yes, Should it be one Groc or
36:06
Speaker 4: multiple different.
36:07
Speaker 2: Grocks, It's up to you. You can have one Grok
36:10
Speaker 2: or many Grons. I suspect people will probably have one one.
36:14
Speaker 1: Yeah, I won't have a doctor Groc. Yeah, the Grouk doc.
36:19
Speaker 1: That's right, right, cool.
36:22
Speaker 3: So in the past we've open sourced grock one, so
36:26
Speaker 3: somebody is asking us we're going to do it again?
36:29
Speaker 2: Yeah, I think when once scrut A general approach is
36:33
Speaker 2: that we will open source the last version when the
36:36
Speaker 2: next version is fully out. It's like when when GROG
36:40
Speaker 2: three is mature and stable, which is probably within a
36:45
Speaker 2: few months, then will open source GROUG two.
36:48
Speaker 4: Mm hmm, okay, so we probably have time for one
36:51
Speaker 4: last question. What was the most difficult part about working
36:56
Speaker 4: on this project? I assume Grock three and the world
37:00
Speaker 4: I'm most excited about, so I think me looking back,
37:04
Speaker 4: you know, getting the whole model training on one hundred
37:07
Speaker 4: k h one hundred coherently, that's almost a battle against
37:12
Speaker 4: the final boss of the universe, the entropy, Because at
37:15
Speaker 4: a given time, you can have a cosmic rate that
37:17
Speaker 4: beaming downe and flip a bit you know, transistor and
37:20
Speaker 4: not the entire graded update. If it's fit mentis a bit,
37:24
Speaker 4: the entire grade update is out of whack.
37:27
Speaker 1: And nine hundred thousand of those.
37:28
Speaker 4: Orchestrate them every time at any given time, and GPS
37:33
Speaker 4: can go down, and.
37:35
Speaker 2: I mean it's worth breaking down, like how were we
37:37
Speaker 2: able to get the world's most powerful training cluster operational
37:41
Speaker 2: within one hundred and twenty two days because we started
37:45
Speaker 2: off we we actually weren't intending to do a data
37:49
Speaker 2: center ourselves. We were going to just we went to
37:51
Speaker 2: the data center providers and said, how long would it
37:54
Speaker 2: take to have one hundred thousand GPUs operating coherently in
37:59
Speaker 2: a single location, And we've got time frames from eighteen
38:02
Speaker 2: to twenty four months, so like, well, eighteen twenty four months,
38:05
Speaker 2: that means losing as a certainty. So the only option
38:08
Speaker 2: was to do it ourselves. So if you break down
38:11
Speaker 2: the problem, you guest doing like reasoning here, it takes
38:15
Speaker 2: your thing. Yeah, exactly. So what we needed a building.
38:20
Speaker 2: We can't build a building, so we must use an
38:21
Speaker 2: existing building.
38:23
Speaker 1: So we looked for basically for factories that had been.
38:29
Speaker 2: That had been abandoned, but the factory was in good shape,
38:32
Speaker 2: like the company had gone bankrupt to something.
38:33
Speaker 1: So we found an electro Luxe factory in Memphis.
38:36
Speaker 2: That's why it's in Memphis, home of Elvis and also
38:40
Speaker 2: one of the oldest I think it was the capital
38:42
Speaker 2: of ancient Egypt, and it was actually very nice factory
38:48
Speaker 2: that forever for whatever reason that Electrolux had left and
38:54
Speaker 2: that that gave us shelter for the computers. Then we
38:57
Speaker 2: needed power that we needed at least one hundred and
39:02
Speaker 2: twenty megawats at first, but the building only had fifteen megawats,
39:04
Speaker 2: and ultimately for two hundred thousand megree thousand GPUs we
39:07
Speaker 2: needed a quarter gigawat, so we they initially leased a
39:13
Speaker 2: whole bunch of generators, so we have generators on one
39:16
Speaker 2: side of the building, just trailer after trailer of generators
39:20
Speaker 2: until we get the utility power to come in. But
39:23
Speaker 2: then we also need cooling, so on the other side
39:25
Speaker 2: of the building it was just trailer after trailer of cooling,
39:28
Speaker 2: So we leased about a quarter of the mobile cooling
39:30
Speaker 2: capacity of the United States on the other side of
39:33
Speaker 2: the building. Then we needed to get the GPUs all installed,
39:36
Speaker 2: and they're all liquid cooled, so in order to achieve
39:38
Speaker 2: the density necessary, this is a liquid cooled system.
39:41
Speaker 1: So we had to get all the plumbing for liquid cooling.
39:44
Speaker 2: Now, we had ever done a liquid cooling data center
39:47
Speaker 2: at scale, so this was an incredibly dedicated effort by
39:51
Speaker 2: a very talented team to achieve that outcome.
39:55
Speaker 1: I may think, now it's going to work. Nope.
39:57
Speaker 2: The issue is that the the power fluctuations for a
40:02
Speaker 2: GPU cluster are dramatic. So it's like this giant symphony
40:07
Speaker 2: that has taking place, like having a symphony with one
40:11
Speaker 2: hundred thousand or two hundred thousand participants in the symphony,
40:15
Speaker 2: and the whole orchestra will go quiet and loud in
40:19
Speaker 2: you know, one hundred milliseconds, and so this caused massive
40:22
Speaker 2: power fluctuations. So then which then caused the generators to
40:27
Speaker 2: lose their minds and they.
40:28
Speaker 1: Weren't expecting this.
40:30
Speaker 2: So to buffer the power, we then used Tesla megapas
40:35
Speaker 2: to smooth out the power.
40:38
Speaker 1: So the mega packs had to be reprogrammed.
40:41
Speaker 2: So with the XAI we were working with Teesla, we
40:44
Speaker 2: reprogrammed the megapacs to be able to deal with these
40:48
Speaker 2: dramatic power fluctuation fluctuations to smooth out the powers that
40:51
Speaker 2: the computers could actually run properly, and that worked.
40:57
Speaker 1: It quite tricky and then but.
41:00
Speaker 2: Even at that point, just left to make the computers
41:02
Speaker 2: all communicate effectively, so all the networking had to be
41:05
Speaker 2: solved and debugging Brazilian network cables a debugging Nickel at
41:13
Speaker 2: four in the morning, or we solved it like roughly
41:16
Speaker 2: four twenty am. Was well figured out, Like there's some well,
41:21
Speaker 2: there are a whole bunch of issues. One there was
41:22
Speaker 2: like a bios mismatch.
41:25
Speaker 1: Bios was not set up correctly.
41:28
Speaker 3: We HADSPCI outputs between two different machines. One that was working, yeah,
41:35
Speaker 3: one that was not working. Many many other things.
41:38
Speaker 1: I mean, yeah, exactly.
41:38
Speaker 2: This would go on for a long time if we
41:40
Speaker 2: actually listed all the things. But it's like interesting, It's
41:42
Speaker 2: not like, oh, we just magically made it happen. You
41:45
Speaker 2: had to break down the problem, just like groctas for reasoning,
41:48
Speaker 2: into the constituent elements, and then solve each of the
41:50
Speaker 2: constituent elements in order to achieve a coherent training cluster
41:54
Speaker 2: in a period of time that is a small fraction
41:57
Speaker 2: of what anyone else could do it in.
41:59
Speaker 3: And then once the train cluster was up and running
42:01
Speaker 3: and we could use it, or we had to make
42:02
Speaker 3: sure that it actually stays healthy throughout, which is his
42:05
Speaker 3: own brand challenge. And then we had to get every
42:08
Speaker 3: single detail of the training right in order to get
42:11
Speaker 3: a rookery level model, which is actually really really hard.
42:14
Speaker 3: So we don't know if there are any other models
42:17
Speaker 3: out there that have Rockery's capabilities, but whoever trains a
42:20
Speaker 3: model better than rock Crey has to be extremely good
42:22
Speaker 3: at the science of deeplining at every aspect of the engineering.
42:26
Speaker 1: So it's it's not so easy to pull this off.
42:29
Speaker 1: And this is now going to be the last cluster
42:31
Speaker 1: were build and last model we train.
42:33
Speaker 2: Oh yeah, we're We've already started work on the next cluster,
42:36
Speaker 2: which will be about five times to power, so instead
42:39
Speaker 2: of a quarter gigawad, roughly one point to giga.
42:42
Speaker 1: What what's the what's the back to the future was?
42:47
Speaker 1: What's the power?
42:49
Speaker 2: Does the back to the Future car anyway back to
42:52
Speaker 2: the future power? It's like roughly in that order, I think. So,
42:56
Speaker 2: you know, there will be the sort of the GB
42:58
Speaker 2: two hundred slash three hundred pleasure once again. It will
43:02
Speaker 2: be the most powerful training cluster from the world. So
43:04
Speaker 2: we're not like stopping here, and.
43:05
Speaker 4: Our reason model is going to continue improve by accessing
43:08
Speaker 4: more tools every day. So yeah, we're very excited to
43:12
Speaker 4: share any of that coming results with you all.
43:14
Speaker 3: Yeah, the thing that keeps us going is basically being
43:17
Speaker 3: able to give free to you and then seeing the
43:20
Speaker 3: usage go up, seeing everybody enjoy no clock.
43:23
Speaker 1: That's that's what really gets us up in the morning.
43:26
Speaker 1: So thanks for your name. Thanks guys,