00:00:00
We are in the race for super intelligence and uh Andrew Feldman is back uh and obviously CEO and founder of
00:00:08
uh Cerebras doing inference chips pioneered the space had a successful IPO. We've talked about this a couple of
00:00:16
times. We got to see each other in January at Davos. IPO happens. Uh the boys and I got to sit with you recently.
00:00:22
>> That was fun >> at liquidity. >> That was really that was really fun. >> Had a great discussion with the boys. I
00:00:27
wanted to deep dive with you about a couple of topics. The first one is the buildout of AI. We've never seen a build
00:00:37
out like this since, you know, the Great Wall of China, >> right? Who knows? >> The pyramids. I mean, it feels like the
00:00:45
amount of capital, time, and intelligent people on the planet dedicating themselves to the buildout of something.
00:00:54
Um I I can't think of anything in our lifetimes with per perhaps uh you know before our lifetimes the war effort
00:01:00
>> right >> this is a mobilization and a scale that we read about we hear about but you're
00:01:07
actually doing it you have customers who are building data centers and you're a key piece of that
00:01:12
>> all app started with an $8 domain and no VC funding and became one of the largest ad
00:01:20
platforms in the world now that same engine powers Appleven ads for e-commerce. Your ads run inside mobile
00:01:26
games, reaching over a billion people with full screen, distraction-free attention. The platform finds buyers and
00:01:33
optimizes for profit. You set the target, it does the rest. One cookware brand went from 4 million to $16
00:01:39
million, turned profitable, and is on pace for 80 million this year. Visit apploven.com/allin
00:01:45
to launch your first campaign today. >> Maybe you could just enlighten us in 2026. What is Cerebras doing and what is
00:01:55
happening with this buildout out in Texas? These are some gigantic gigantic efforts. The the size and scope of what
00:02:04
is being built, the physical size and scope. Usually when we talk about software or we talk about hardware,
00:02:11
we're talking about chips or boxes and and they don't have the same sort of physical enormity. Right.
00:02:16
>> Right. And what we're talking about now are data centers that are in the next
00:02:21
several years going to use more power than the previous 50 years on Earth took. Wow. Right. We're talking about
00:02:28
individual buildings the size of football fields that have more power coming into them than midsize cities.
00:02:36
And they're being built they're being built across the US. They're being built in Canada. They're being built
00:02:42
throughout the Nordics. are being built here in Paris and throughout France, in Europe, in the Middle East in nations
00:02:48
that sort of weren't front and center in anybody's mind previously. You know, Kazakhstan, Tajjikstan are building out
00:02:55
Georgia building out data centers of size. Armenia, everybody's sort of focused.
00:03:01
>> Every country >> um and every state obviously in America feels they need to participate in this
00:03:08
>> and the people who are buying the capacity, the open AI, anthropics, SpaceX, SpaceX AI,
00:03:16
uh the Googles, they are insatiable right now. >> Yeah. and they're building how many years out
00:03:26
when you talk to them? They were ordering chips from Cerebras before you were finished with the chips.
00:03:32
They're putting orders in ahead of time. The irony is unlike many sort of exciting times in technology. They're
00:03:40
trying to capture yesterday's demand, right? The demand is way outstripping our ability to build data centers and to
00:03:47
fill them with hardware. >> All right? And so, you know, we have a $25 billion backlog,
00:03:54
>> $25 billion backlog >> and we we are not alone in that that that OpenAI uh Anthropic, you go through
00:04:02
this list of of uh Google wants more data centers, Microsoft wants more data centers, AWS wants more data centers,
00:04:09
right? All of these players are not chasing sort of if you build it, they will come. They're chasing the demand is
00:04:16
booked, >> right? How do we keep them from leaving? Right. And and that that's extremely
00:04:21
unusual. >> It's very unusual. And now we have people who are, you know, we have a term
00:04:28
for a token maxing. >> Yeah. >> And there's a great debate. Is this actually creating value? I'm curious
00:04:36
where you stand. You know, is it even possible that this much demand could be created if value did not exist? There is
00:04:44
clearly massive value happening. >> Yeah. But there's also massive experimentation.
00:04:49
>> Oh, for sure. You know, I you know what I I I liken this to when we first started with uh AWS
00:04:58
and it was so good to get around your own IT organization, >> right? >> That you told every engineer, yeah, go
00:05:03
ahead, put on your credit card kite and sign up. >> Yeah. >> Right. And a lot of it was really useful
00:05:08
and some of it was like, God, I wish we didn't do that. >> Yeah. And so for sure there's
00:05:13
experimentation, but it doesn't mean that the net value isn't enormous. It it means some of it is going to go nowhere.
00:05:22
>> And you know, it was the same. I remember when Costco opened up in in in in the Palo Alto area in 1988. And
00:05:30
people used to shop Costco like they shop Safeway. They go down every aisle. >> Yes.
00:05:34
>> And that's a horrible way to shop Costco because you end up with four things you
00:05:38
didn't need and each was $22. >> Right. And as people got more sort of accustomed to it, you go to the back,
00:05:45
you get the chicken, >> 18 cupcakes for the kids' birthday party, bang, you are strategic. And it's
00:05:50
exactly the same. >> I think at first people opened up and said everybody as much tokens as you
00:05:56
want. >> And the in enterprises, there's no open loop. We don't give sort of any resource
00:06:03
unconstrained to people. And now we're jumping on saying, "Whoa, all right. These guys should have as much as they
00:06:09
need. They're enormously productive over here. We can use maybe an open- source model, maybe a cheaper model over here.
00:06:16
And now we're sort of running like a business and we're really seeing a certain type of person emerge who knows
00:06:25
how to deploy this technology. Systems thinking, >> yeah, >> which developers kind of have innately
00:06:32
CEOs tend to be great strategists and understand systems. Uh but this the intelligence is getting so much better
00:06:41
every step along the way that I'm watching individuals typically startup founders but also venture capitalists
00:06:48
and associates who work at my venture firm. They start playing with the tool and then the tool starts playing with
00:06:54
them. They start to go, "Oh, I haven't clearly defined what my goal is. I don't understand what a system is. I don't
00:07:03
under I've never heard about making a requirements document." And the software's like, "Do you have a
00:07:08
requirements document? What's your goal?" The AI starts telling people, "You're token maxing and you need to get
00:07:13
a little more focused here." >> One of one of my colleagues 20 years ago, a really smart smart computer
00:07:20
scientist said, "Computers really dumb. They do exactly what you tell them." >> Yeah.
00:07:26
>> And at first, prompting was like that, right? You modified your prompt a little
00:07:30
bit and it changed the answer >> dramatically. >> Dramatically. And increasingly it's
00:07:36
understanding what your intent was. >> Right. Right. And if you if you have a chance to to to play with Fable or or 56
00:07:43
from from OpenAI, increasingly what you don't have to get the prompt just right.
00:07:49
You don't have to be a prompt whisperer. Instead, you ask it and it says, "Well,
00:07:53
here are some things and and by the way, maybe you wanted the chart to to to go two ways. You wanted a line and a bar."
00:08:00
And it's like, "Well, that's exactly what I wanted. I didn't ask for it, but that is better. And and so it's it's
00:08:07
understanding intent and that's a huge leap which if we were sitting here two years ago, the idea
00:08:14
>> we would never have been able to predict in a short 24 months that it would go
00:08:19
from being a great summarizer researcher of web results, >> right, >> to actually understanding your intent
00:08:25
and then providing a solution and abstracting it all from you. >> That's right. That's right.
00:08:30
>> Which is a very weird thing. I don't know if you've played with the Hermes agent yet.
00:08:34
>> Yeah. >> Have you played with it yet? >> I mean, I asked it just this morning.
00:08:40
Uh, and I I was given a secret bit tensor project that has the new um ZAI's uh model 52
00:08:50
>> and they gave me >> GLM 52. Yeah. >> GLM 52. So, somebody in that Bit Tensor,
00:08:55
I think you understand Bit Tensor, you've heard of it, the distributed >> um crypto project. And so they have all
00:09:02
this extra capacity. I was a whisperer told me probably some capacity in China that has free energy. Okay, fine. So
00:09:09
they gave me unlimited capacity. So I started having to do some really crazy jobs where I was saying like every hour
00:09:15
I want you to tell me what the trends in the world are that nobody else has identified yet. And you can do whatever
00:09:20
you want to do that. But my goal is to be the smartest trend hunter in the world. And I watched what it was doing
00:09:26
in the background and it started debating itself >> on where it should find the things. He
00:09:30
said, "Well, we should probably go to Hacker News and Reddit." And then I was like, "Yeah, but there's also social
00:09:34
media and trends tend to manifest on Instagram." >> That's a reasoning model. You were
00:09:38
watching a reasoning model work out. >> Yeah. Isn't that interesting? I mean, that's amazing. And and
00:09:45
it was collapsed. So, as a civilian, >> right, who doesn't hit the uncolapsed moment, and if you were using chatbt 3.5
00:09:55
or you were using 4.8, whatever it was and you haven't used this new level of reasoning and inference and unlimited
00:10:03
compute essentially, >> right? >> It opened my eyes just this morning of what a world
00:10:09
>> of unlimited tokens might look like, >> right? >> Cuz unlimited tokens, I believe, means
00:10:14
unlimited reasoning. It does. What does that mean? Yeah, it's uh I mean if you run these
00:10:23
for 25 or 48 hours, you get amazing things now and what what if by using Cerebrus we were 15 times faster and
00:10:31
then you ran it for 24 hours, >> right? A and you got weeks or months worth of thinking.
00:10:38
>> Yeah. And I mean it it is uh it is extraordinary. And I I think one of the things is people like Ilia and Sam in
00:10:46
the early days were saying this was coming. >> Right. >> Right. And I think when you look back
00:10:51
you say to yourself, "Holy crap." Those guys saw it. >> Yeah. They were They could see around
00:10:57
the corner. >> That's right. And the rest of us could see like what? I'm not sure. It's when
00:11:02
we had Sam on Allin uh at one point he uh and he said, "Oh, you know, I'd love to come on at some point." I said,
00:11:09
"Sure, come on." And he he was talking about it. He said, >> you know, I said, "What's next?" He
00:11:13
said, "Reasoning." I said, "Unpack that. What does it mean?" Well, >> understanding what your intent was just
00:11:19
as you're saying and then figuring out a strategy and then maybe talking to other
00:11:24
agents and other threads about like is this the right thing to do and vetting each other's work and I'm like wow we
00:11:29
have come a long way from guess the next word >> right right fill the sentence in you
00:11:34
know summarize this PDF >> now Cerebrus is at the center of this because this reasoning is inference
00:11:41
>> this reasoning is inference and it's computationally intensive >> right >> right and So fast compute makes this
00:11:48
sort of work fast and sort of tractable. It doesn't it by taking a huge amount of time to get a good answer.
00:11:56
>> And so it's exactly the fact that that that this reasoning consumes a huge amount of tokens internally
00:12:03
>> that allows a blisteringly fast machine like ours. And I I brought one because I
00:12:09
I'm never far without you know when one costs half a billion to make you you bring it everywhere with you. Uh we we
00:12:14
were um we were tossing this back and forth at Davos. >> Uh what's the model number of this one?
00:12:21
Th >> this was in the first eight or 10. Um >> so this has a special place. >> This has a special place. I mean my wife
00:12:27
says it's like I'm a kid with a a dirt bike for his 8th birthday. It was in his bedroom at night. I I carry him with me.
00:12:32
>> I I mean when you have um you know your your next party at the house, I highly
00:12:37
recommend just a little or something or I >> I think it would be like a great fit. It
00:12:42
would be a great bit if you had some. >> That's right. Um, but what we're looking
00:12:46
at here is the ability to do that reasoning at scale. And what is Moore's law for inference and
00:12:57
for cerebrus? Do you have something internally you discuss as we're going to double this every x time
00:13:06
period? So all chips prior to us in the processor world followed Moore's law. >> Got it.
00:13:13
>> And we broke >> doubling every 18 months. >> Doubling about every 18 months. >> Got it.
00:13:18
>> Um and we crushed it with this chip >> and we've carved out a whole new trajectory.
00:13:24
And uh my view is in the next 18 months we'll be way over 2x >> interesting. And so, uh, I I think that,
00:13:34
uh, early in an architecture, you have room to to do much better than what was traditionally Morris law. Now, if you've
00:13:41
got a 20-year-old architecture like the GPU, it's much harder, >> right? >> You you have to rely on things like
00:13:47
smaller geometry, >> right? Going to the next fab node. But in a newer architecture, you have a huge
00:13:54
amount of room still to to to learn about the the work that that is being presented and make optimizations that
00:14:02
that give you huge gains. H how do you run the company? Like just being the CEO now in the age of AI, um you have $25
00:14:12
billion in demand. You have to you have to deploy at an just an incredible blistering pace. You have to hire
00:14:21
people. You have to create a roadmap. I don't mean to give you a panic attack here.
00:14:25
>> You have to keep up with somebody like OpenAI who's moving so unbelievably quickly. Yes.
00:14:30
>> Right. And they're they're competitive. You got to keep up. >> Right. >> Right. Your hardware, your software,
00:14:35
your deployments have to keep up with some of the fastest moving organizations in history.
00:14:40
>> They're demanding customers. >> They are not they're not pushovers for sure. >> Yeah. A and also potentially competitors
00:14:49
down the road. >> I look I I think there is so much demand right now that that that
00:14:56
there is no silicon that will go unused, >> right? >> But why isn't OpenAI releasing Jalapeno?
00:15:02
Why is Amazon making their own chips? You see this reoccurring trend? Is it a way to let you know to let Jensen and
00:15:12
Nvidia know, hey, we can do this, too. So we we need good pricing. Is it is it a little bit of a flex that way or is
00:15:19
that the future that they're going to be in your business? >> No, I I I think nobody likes being
00:15:24
dependent >> and I I think some of the lessons learned by the the hyperscalers of the
00:15:31
x86 world is they were dependent on Intel. And uh some of the lessons learned by uh
00:15:40
the GPU makers was they were dependent on a small number of hyperscalers. >> Yeah.
00:15:46
>> And they wanted more customers and so they set about to to help fund these neoclouds.
00:15:53
>> And so I I think mostly it's about uh an opportunity to control at least an important part of your destiny.
00:15:59
>> Got it. >> And I I think that's a very reasonable thing. I think you don't have to sort of
00:16:04
make the fastest chip. You you just can't be entirely dependent on other people's chips.
00:16:09
>> And that dependency has become a hot topic. Not sure if you caught the episodes over the last two weeks, but
00:16:14
we've been talking over the last year about open source. I've been championing that a lot just because I was early into
00:16:22
Open Claw and quickly started using Kimmy and was like, "Wait a second. I'm blowing out my claw tokens, but this
00:16:30
Kimmy, I can't tell the difference. And then we started smart routing it and suddenly this open source started to
00:16:37
figure out reasoning and the gap has >> suddenly closed this year. >> Well, I I you know, you you don't want
00:16:46
to take your your Ferrari to the grocery store, >> right? You you you there there are times
00:16:51
you want to drive your fun car. >> Yeah. >> Right. And there are times you want to
00:16:54
throw the kids in and and don't worry if they're Cheerios on the floor. >> Minivan time,
00:16:58
>> right? There's minivan time. And and and I think that as the sort of sophistication of the user grows, right,
00:17:06
you're going to have hard problems and those are going to be frontier model problems. They're going to be open AI
00:17:11
problems. They're going to be anthropic problems. Going to be Gemini problems. And behind that, they're going to be a
00:17:16
lot of ordinary problems. Right? I mean, if you think about a company, you know how much time is spent cutting things
00:17:21
out of workday and getting it in a different cell for Yeah. >> Right. Think about
00:17:27
>> the cutting and pasting economy is real. >> That's right. And and this doesn't need,
00:17:32
right, gold medal math. No. >> What what this needs is sort of rock solid open-source capabilities. Yeah.
00:17:40
And if you think about what I mean, well, we've been thinking a lot about it in DNA, but a huge amount of DNA, all
00:17:46
right, is not invention, >> right? And you you may not need sort of the most sophisticated agents for this.
00:17:56
And another card that's turned over recently is some folks uh maybe have concerns with the
00:18:06
ambition of the frontier models and maybe sharing their data data leakage and sovereignty of intelligence and
00:18:14
they're saying hey our company is going to choose maybe we're in a regulated industry finance healthcare HIPPA you
00:18:22
know FINRA all kinds of different regulations >> we need to have this on >> prem
00:18:29
domestically and we'd liken an open-source version where we have uh a little bit more control. Yeah. And I I
00:18:36
think >> are you seeing that now? >> We are seeing that for sure. And I I think OpenAI made a good call releasing
00:18:41
OSS 12B some months back. That was a good open-source model. Um but I think in the US we need more domestic open
00:18:48
source models. We need to give the world a choice, >> right? Right. If they want to run open
00:18:53
source right now, it's OSS 12B or Chinese models. >> Nvidia has some in >> Nvidia has seen the same opportunity to
00:19:01
push open- source models. I I I think giving them more power might might be sort of
00:19:07
>> Well, I was about to that was you cut me off at the past like my understanding
00:19:12
was Jensen was like, "Hey, we we don't even want to talk about these open source models we have because our
00:19:17
customers, >> right? >> We're now going to be competing with Sam, Daario, Elon, >> uh Sergey, like do we want to be in that
00:19:26
position, >> right? >> So, but we do need some more champions here and it's open source so people can
00:19:31
fork it." Um, but that puts you in a more neutral position. >> That's right. We we we run today, we run
00:19:38
GLM, we run Kimmy, we run the Quen set of models and we run OpenAI's models, the closed source ones. We run models
00:19:47
for say Galaxos Smith Klein, which they wrote and developed. Um, we run models for uh our partner uh in the UAE uh G42
00:19:56
and MBZ UAI. Yeah. Um that are are are their models that they designed. So we we have a a a wide variety.
00:20:05
>> So sovereignty is a trend. >> Sovereignty is a trend. And I think uh the the government's actions with regard
00:20:12
to fable and uh uh 56 um where they said, "Oh, whoa, let's think and then we can act."
00:20:22
>> Um I I think sort of particularly here in Europe was a bit of a wakeup call. And when you saw this going down,
00:20:30
there's a layer of partisanship in our country right now. It's pretty fervent. Daario is pretty explicitly, you know,
00:20:38
not part of this administration. They they've been very adversarial. Both sides have been have admitted that
00:20:43
they're starting to work it out now. So, it's hard, I think, for us not being in
00:20:47
the room with these parties to understand what's partisanship, what's games here. But do you believe that what
00:20:54
they released was truly dangerous for cyber warfare, for cyber attacks, and that if you were to rate Daario's
00:21:04
not communication, because he's a very effrovescent communicator, um I think is a diplomatic way to say it, um but to
00:21:14
have a scheduled rolled out release, >> right? We'll put aside the government's control of it, but do you think that is
00:21:21
is a wise thing for us to do at this point? And do you think it's there was actually a major threat there?
00:21:26
>> So what's interesting is I hadn't seen it before, >> right? And I I think you know if we just
00:21:33
step back and say is it reasonable I don't know whe whether this was the right time but at a time
00:21:41
>> that that that a model is sufficiently uh creative in its thinking that it poses a meaningful threat for the
00:21:50
government to say we'd like you to roll it out in steps. >> Yeah. >> Now this doesn't seem unreasonable to
00:21:55
me. >> Not at all. >> Right. I mean we we do this with powerful pharmaceuticals.
00:21:59
>> Right. We we we'd like I mean we're certainly not encouraging seven years of trial and the amount of paperwork and
00:22:05
all the garbage that has acred to the FDA, >> but with a powerful new technology, it
00:22:11
certainly doesn't seem unreasonable to say, "Hey guys, um let's at least do some red teaming at the government so we
00:22:20
know our defenses can block this." >> Yeah. Have we checked Have we checked >> the infrastructure of the country like
00:22:26
>> of the NSA? Have we checked the infrastructure up? Right. And can you give us two or 3 weeks to patch any
00:22:32
obvious holes that are found? This doesn't seem to be >> an unreasonable thing for the government
00:22:38
to ask. >> Right. We But we in this very polarized time >> put on top of it, well, oh my god, it's
00:22:46
President Trump doing it. And then you have to think, well, what if it was President OA AOC or President anybody in
00:22:52
between the the two extremes? >> I I think the polarization hurts a great deal. It hurts clear thinking,
00:22:58
>> right? It hurts clear thinking and and and both sides are going to do some dumb
00:23:02
things and some really smart things. >> Right. >> Right. And in fact, what I found is that
00:23:07
the people in the government are are trying really hard. >> Um >> the rank and file,
00:23:12
>> the rank and file are trying really hard and this is moving fast. And I I think
00:23:18
that that an ability to to set aside some some of the the polarization and say h how do we do this in a reasonable
00:23:26
manner? I mean we we want Dario and Sam competing like crazy. It's awesome, right? It's good for the
00:23:34
technology. It's good for it's good for entrepreneurs to see even with thousands
00:23:39
of people this is what what you can continue to achieve. Right. >> Right. This is Google in the ass made
00:23:45
them get sharper. Amazon got everybody got better because of that. We want that and we certainly don't want to become
00:23:52
sort of a region where the first thing we want to do is regulate it. >> Right. >> Right. But as it gets more powerful
00:23:59
and the industry really should do a better job of regulating itself perhaps and and it did seem like they were
00:24:05
starting that process but then the communication was lacking maybe. >> Yeah. Uh, you know, I I think not only
00:24:14
are they racing hard, but they're inventing this as they go too. >> Yeah. Right. There's not a playbook. No.
00:24:21
>> Right. They're inventing the we we say, "Oh, just put on guardrails." Well, they
00:24:24
have to design the guard rails. >> Sure. >> Right. They the guardrails have an impact. Um, you know, one of the things
00:24:30
that fast does is it makes the guard rails less painful. And so that that we we discovered that in the last six
00:24:36
weeks. Yeah. is that the very guard rails can add time and make it feel slower and so fast ships like ours can
00:24:45
really help that. But so they're racing against competition. They're racing against their own sense of greatness.
00:24:52
>> Yeah. >> Right. Which is maybe even the biggest driver here. A and I think that they're
00:24:58
in earnest trying to think about how I do the right thing. A and all of those are are are mixed in this bucket. and
00:25:07
and sometimes you're on one side rather than the other. And >> yeah, and as you're saying, this is a
00:25:13
first time, right? When we when 3.5 came out, it wasn't like when we're using chatbt 2.5, 3.5, it was taking down
00:25:20
networks, but in talking to Nicash uh from PaloAlto Networks, I asked him like, "Hey, well, how would you grade
00:25:28
this?" And he said, "We put it against our software and we found bugs we were not aware of and
00:25:34
>> it killed them." >> Yeah. He said we had to stop everything we're doing and do patches for 6 weeks,
00:25:39
>> right? And and that's when you know, right? I mean, Nick leads, you know, maybe the leading security software
00:25:46
firm, right? >> And when it finds in an hour, right, tens of critical opens, you're
00:25:54
like, whoa, this is a powerful tool >> and we need to think and and maybe you you show it to a a group first, right?
00:26:02
May maybe you I I don't know what the right thing is but >> yeah I mean red teaming and uh we've
00:26:07
always had just when you were releasing the new version of um an operating system you know when you have your
00:26:14
iPhone you can say I want to be part of the beta. >> That's right. you know, right? And
00:26:18
there's like two other baiters that you don't even get the chance to opt into as
00:26:21
consumers. And those ones are for security. Those ones are for, you know, making sure you don't lose your data or
00:26:27
data. >> That's right. Disappear or leak or >> corruption. Any any number of these
00:26:33
things. >> I think we can also know that that there will be a massive data leak.
00:26:39
>> Of course, >> we we know this. Yeah. >> Right. And it's like uh Warren Buffett
00:26:43
talked about the reinsurance industry that you know something bad's going to happen. You don't know when.
00:26:48
>> Yeah. >> But you got to save up for it, right? You put money away for reinsurance,
00:26:53
>> but there will be a tornado. There will be a massive earthquake. I mean, we we
00:26:57
know this and we can do our best to plan, but there'll be a massive breach and they there'll be and we we have to
00:27:05
steal ourselves in advance and we have to think about it and think about the right response at the time and sort of
00:27:10
prepare ourselves for a future that is in specific unknown. But in general, we're pretty sure it's something's going
00:27:16
to happen. >> Something will happen and >> yeah, it's typically a black swan, right? By definition, it's going to be
00:27:25
something we didn't consider or a question we didn't know to ask, >> right? But but even knowing that there's
00:27:30
some unknown unknowns is a useful place to start. >> Yeah. What are we not asking ourselves?
00:27:36
>> That's right. >> With reasoning, the AI is going to be able to tell us, "Hey, schmuck humans."
00:27:40
>> That's right. By the way, here's what you're not thinking about. This is now my closing sentence when I do my
00:27:46
prompting is I need you to make me a prompt that will help me do this trend scouting for an example. And then I
00:27:53
always say at the end um please check your work >> right and uh then tell me what I haven't
00:28:01
considered in terms of my goals and uh give me ask me some questions every time you run the job and that has changed
00:28:08
everything because it's like I checked my work by the way this was incorrect right
00:28:12
>> and uh I'm wondering hey uh would you like me to also do this and some of the
00:28:16
tools like perplexity do that automatically they give you your next three prompts but if you give it
00:28:20
explicit instructions My lord is it um good at that. >> So, you know, over the course of the
00:28:27
last 10 years as I was raising money, I I thought one of the smarter questions I
00:28:31
got at the end of a conversation where someone asked, "What was the smartest question you heard that that wasn't
00:28:37
covered by what I asked?" >> It's incredible. right now. Now, that's somebody who's curious and thinking and
00:28:42
humble and and trying to >> sort of use this to get a picture of of the space and and to the extent that you
00:28:50
can ask the AI that >> and and and that it can sort of broaden your your view, you know, maybe what
00:28:56
question should I have asked >> to to be an expert in this? What what what would a a PhD level uh uh
00:29:04
questioner ask of this or a gold medal math? I mean I I think those are the sort of questions that you you you know
00:29:12
you don't even know how to ask >> which you know if you start thinking about AGI and super intelligence
00:29:20
um you know they're just definitions but they're important definitions I think to
00:29:25
kind of keep in mind because they're way points. >> That's right. And AGI I think I suspect
00:29:32
you'll agree with me that we've hit it. we just haven't exactly deployed it fully. We we we have artificial general
00:29:40
intelligence now. It feels like when we're talking about these reasoning moments and you know the the the ability
00:29:46
for it to be as smart as any human but >> let's talk about >> by any definition we had 20 years ago
00:29:51
we've hit it. >> Yes. >> Right. I mean if you think about oh there's a touring test blew it away.
00:29:56
Yes. >> I mean, you think about that that any period of time sort of 10, 15, 20, 30,
00:30:01
40, 50 years ago, we we we any definition we would have previously put forward, right?
00:30:07
>> We've blown past it. And so, which goes back to our previous point of like, do
00:30:10
we know the questions to ask 20 years ago, science fiction authors, you know, had their say and we answered all their
00:30:18
questions, >> right? >> If they were to look at this today, they'd be like, "Well, I'm out of
00:30:23
>> I'm out of questions. >> I'm out of questions. Sorry." And that's where sort of the the sort of listening
00:30:29
to people who we who sound sometimes like they're on the fringe, >> right? When when when Ilia was talking
00:30:36
eight or 10 years ago about the need for safety and then and you're like, "What?"
00:30:41
And dead right. >> Yeah. >> Right. When when when Elon was talking about building rockets and driving the
00:30:47
cost to to to near zero of of of a launch vehicle, you're like, "What?" And there it is. And now you can see and
00:30:55
that's I think that's why it's really fun to be a technologist now. >> Well, and with these tools specifically,
00:31:02
you know, we're talking about building all these tools and then the tools are starting to build themselves in this
00:31:08
recursive loop. >> That's right. >> We're kind of just starting to see people apply loops. Uh, in fact, loop
00:31:17
maxing became when I was doing my trend, when I did my trend thing, it kept picking up loop looping and it kept
00:31:23
picking up the maxing stuff and it created a buzz word for me, loop maxing, right? And then it
00:31:29
>> magically people started talking about loop maxing and I was like, "Wow, this
00:31:32
is really weird." It anticipated that this would other humans would come up with this word. But talk a little bit
00:31:37
about recursive and then the road to super intelligence. And do you have a way Andrew that you think about super
00:31:43
intelligence and what it will mean for humanity and how we will define it and how we'll experience
00:31:51
it? Yeah. >> I I I think let's begin on on on loop maxing or sort of recursive learning. I
00:31:57
I I think um I think what what what Sam and Ilia and then later Daario and and and Dana
00:32:07
Dennis saw um six years ago or five years ago was that um powerful recursive gains
00:32:19
are are exponential, >> right? you get better, you do it again. And if you continue to get gain, the the
00:32:26
the the slope of that curve is so steep. >> Yeah. >> And that um we're just beginning to see
00:32:34
that now. >> You ask it a question, you learn from the results, you ask it to do it again,
00:32:39
it the results get better and more information is added. Your answer gets better. You ask it to do again, it
00:32:45
covers more material. And the these sort of loops are producing sort of not a little bit better answers but vastly
00:32:53
better answers. >> Yeah. A >> and that is enormously powerful because we don't quite know where it ends,
00:33:00
>> right? >> You keep throwing compute at it. I mean, how much better does the answer get?
00:33:04
>> You know, we we run out of tokens or our budget or or or but but holy cow. I mean, when does the exponential stop or
00:33:12
does the answer keep going up and up and up to the right? >> Yeah. And that's sort of an enormously
00:33:17
interesting intellectual question right now. >> Yeah. Like when do we run out of
00:33:23
problems to solve and >> Well, that's right. And and when are are the the problems
00:33:29
no longer sort of intellectual problems and they're now people problems? >> Yeah.
00:33:35
>> Right. How to organize people to to get done what the AI asked for. >> Right. Right. I mean, as you know, in
00:33:41
running your company, a lot of your problems aren't hard intellectual problems. They're people working
00:33:46
together problems. >> Yeah. >> Right. And you >> motivation. >> Motivation. You spend a lot of time as a
00:33:51
leader spraying WD40 on your team. >> Right. >> Right. It just so so friction is reduced
00:33:57
and >> um how do we learn about those from AI? >> How do we get behavioral insight from
00:34:06
from AI? And and I think that's some of the things the world models are going to
00:34:09
bring us as they begin to watch human behavior. >> Yeah. We didn't even get to that. This
00:34:14
is going to be for another interview. But um when these things jump off the screens,
00:34:20
>> right, and they're in the real world and the recursiveness starts, not trying to
00:34:25
solve math problems and right, you know, humanity's most difficult ones, but hey,
00:34:30
you know, there's an incredible world out here and here's the palace of Versailles, right? You're just like now
00:34:35
we're like make me a new version of Salesforce and we're like hey you know what I'd like a palace of Versailles
00:34:41
I've got a hundred acres somewhere out Texas or Nevada I'll just send a thousand optimistes out there make me
00:34:48
the Palace of Versailles >> right >> sounds fantastical but the Palace of Versailles would seem fantastical to
00:34:54
people who lived a thousand years before it >> and it was fantastical I think to the
00:34:58
people who built it. >> Yeah. >> Right. Even to the builders, I I think they were awed
00:35:04
>> at it as they built it. >> Yeah. They're compounding. They're compounding recursive learning.
00:35:08
>> That's right. A >> and generations. We talked about you had a really such a great insight of in
00:35:13
building this place, you had generations of masons. >> Yeah. I I think in in in all these large
00:35:19
projects um often there were families uh who were specialists who you and uh you
00:35:26
you apprenticed under your father, your uncle and when you had a project that took 50 or 70 or 100 years, you might
00:35:32
have three or four generations of the same family, right? The same stonemason family working on the same structure
00:35:39
>> and passing on passing on the learnings, >> new innovations, >> right? which is what we've modeled with
00:35:47
this new >> That's right. >> models and and what you're building in the infrastructure.
00:35:52
>> It's pretty incredible when you think about it. >> Especially when we're sitting here and
00:35:57
the and that's what I mean I I think the the problem with human learning is um it
00:36:05
often moves it at the pace of a generation. uh and like uh elephants and other large mammals, we don't have
00:36:13
generations but every 15 or 20 years. And if you want to move really quickly across generations,
00:36:20
you want them happening more like Drosophili, like fruitfly. You want two a day.
00:36:24
>> Yeah. >> Right. Then and you see that in genetics, that's why we study them in
00:36:28
genetics because learning encoded in the DNA, you can study over thousands of generations.
00:36:34
>> And I I think that what we're getting is that equivalent in AI. We're getting
00:36:38
sort of learning so quickly over the equivalent of thousands of generations. >> Yeah. Darwin would be
00:36:46
>> in awe of this uh pace evolution. >> That's exactly right. You think about it
00:36:52
as there was I remember when I was getting my psychology degree and they were teaching us about paradigms and I
00:36:58
was like trying to understand how the paradigms shifted and the professor said to me uh Jason uh what you have to
00:37:06
understand is paradigms don't die. >> They don't. >> People do. >> That's right.
00:37:11
>> And that's how Freud >> he and Thomas Freud and Skinner and Young like it took them dying.
00:37:19
>> That's right. It took generation question it and that was 20 years sometimes 40 years as their students
00:37:28
maintained positions of leadership until someone said maybe we could do it differently and I I think what you're
00:37:34
seeing is this iteration is a shortening of the the the intergeneration gap and the learning is so fast.
00:37:44
It's uh always so great to talk to you because uh one it's just intellectually um uh so your your approach to it is so
00:37:54
intellectually rigorous but also um with so much poom in the world. I I feel so good that you're such an optimist about
00:38:02
this technology and you're building it >> with such thoughtfulness and it I think
00:38:06
for people who are hearing these horror stories about AI and job loss and everything, they need to understand
00:38:12
there are people like yourself who are building this in an incredibly thoughtful way and this is going to be a
00:38:17
net benefit for humanity that just is unimaginable. Yeah, >> we have a shot with this technology. So
00:38:26
not our children nor anyone they know dies of cancer. >> All right. I mean say it like that.
00:38:32
There will be some dislocation in the economy. Sure. There will be there was dislocation when cars came and and and
00:38:38
it was a bad deal to to to be a guy who shed horses, right? Or built carriages. >> But you got to also against that, you
00:38:45
know, make your tea of the cons and the pros. >> Yeah. >> Right. There's a shot that our children,
00:38:51
none of them nor their people they love will die cancer. And that that's one >> thing that we can work on wi with this
00:38:58
technology and we will have great purchase on >> and and I think you begin listing those
00:39:03
and then it's a more thoughtful discussion. >> Yeah. Unlimited energy, unlimited
00:39:08
calories, unlimited knowledge, unlimited education, unlimited housing >> and how we do it. We imagine imagine
00:39:15
sort of we know how to teach children and we don't do it right. Aristotle was a tutor to Alexander the Great. Socrates
00:39:22
was his tutor. We know that if you give a child a tutor and the tutor modifies the teaching for the child, they learn
00:39:29
better. >> That's not how we do teaching classes. Factory farming. >> That's right. We we teach to some sort
00:39:34
of mid-level. Imagine if we beat built agents that that taught children for their way of learning.
00:39:42
>> Right. >> Right. And here's a way we we've been doing it the same way for a thousand
00:39:45
years. And during that entire time, we knew how to do it better and we chose not to.
00:39:49
>> And and here's a way we can do it. Put that on the pro side. And so as long as
00:39:53
we're sort of thoughtfully and fairly writing the good and the bad, I I think it'll come out.
00:39:58
>> You got to get out there, Andrew, and keep communicating your version of the world because some people see around the
00:40:04
corner and they get a little nervous and >> Okay, fair enough. But I think the ledger, as you describe it,
00:40:11
>> is heavily weighted towards abundance. >> I think it'll create abundance for sure.
00:40:17
>> Massive abundance. >> Andrew, pleasure. Always a pleasure to see you. I'll see you in six months for
00:40:21
our checkup. That'll be great. >> Industries, capital, and intelligence are converging into a single
00:40:28
interconnected system, and the infrastructure behind it needs to evolve just as quickly. NASDAQ was built for
00:40:33
this moment, powering more than 135 marketplaces and regulators globally and connecting capital to companies shaping
00:40:40
the future. As the innovation economy accelerates, connectivity becomes the critical asset. NASDAQ is the leading
00:40:45
technology platform that makes it possible and scalable. Learn more at nasdaq.com.
00:40:52
Robin Rombach is uh the co-founder and CEO of Black Forest Labs. You are uh based in Germany in Black Forest which
00:41:01
is a city >> in Germany. >> A mountain range actually. >> A mountain range. >> Yes.
00:41:06
>> Where you grew up. >> Where I grew up. Yes. >> And uh you are working on open- source
00:41:11
image and video models. You worked at stable diffusion for a little bit. >> That's correct.
00:41:16
>> Cut your teeth on that. and uh you're known for the open source model flux and
00:41:22
maybe also for some closed source models. Tell us about the business of Black Forest Labs. What is the business
00:41:29
and what is the goal? >> 100%. Uh one uh quick addition, we are based in uh the Black Forest. It's a
00:41:36
town called Frybook and in San Francisco. We have like >> in San Francisco of course. Yeah.
00:41:40
>> Um we >> you're splitting your time or >> uh I'm splitting my time to a certain
00:41:45
degree. Um we um yeah started a company 2 years ago. Um me and my co-owners as you said like we've worked on stable
00:41:54
diffusion in the past. Before that we invented like an algorithm called latent diffusion which is basically like the
00:42:00
fundamental algorithm behind all of like generative models that are being deployed for image generation, video
00:42:06
generation, even like physical AI now. >> Yeah. Um it basically makes use of this
00:42:10
principle that uh you can compress natural data such as images, such as video, such as audio into a much more
00:42:18
like efficient representation and then train a transformer model on that. And um I mean this is the stuff why like you
00:42:26
know like JPEG uh MP3 and all of that works. And we basically translated that into like a neural uh algorithm uh a few
00:42:35
years ago when we were still like PhD students in uh in Munich actually. And uh then build on like on top of that we
00:42:42
build stable diffusion and then um on top of that uh yeah uh the generative models that we are
00:42:50
developing today and of course like the technology has advanced um but we are now tackling I would say models that are
00:42:59
really made for understanding like the whole world around us multimodal visual models uh pre-trained on images audio
00:43:08
data at the same time and we are now like entering a new paradigm which is combining that with uh something that's
00:43:15
called action prediction such that you can actually use the same model to make images to make videos to make audio and
00:43:22
to predict actions which means you can ultimately deploy it on a robot in the real world.
00:43:27
>> Wow. So from the image to the video, the audio and then eventually the real world
00:43:34
with robotics and a real world model because if you can make the image you and you can train the model that means
00:43:43
by default you understand the world. In order to make a video of the world you have to understand the world. Yeah.
00:43:50
>> And the objects in >> I I think that's Yeah. I think that's like a a really good like way to think
00:43:55
about it. Yeah, it's like it's like an intuitive way uh to interact with the world, right? Like I I
00:44:02
would say there's like these like complimentary forms of intelligence ultimately. There's like intuitive
00:44:06
intelligence and then there's like a deep reasoning layer. Now ultimately you need for like a kind of like complete
00:44:12
form you need both um and you need them to interact and I think like we've been approaching it more from like the
00:44:18
intuitive side um images is like a very natural way to approach this whole field
00:44:23
because it's not as computationally intensive as let's say video right but now yeah I think like we're combining it
00:44:30
it's converging into like a a multimodal model and yeah we see like exactly like
00:44:36
pre-training on videos gives like implicit understanding of the physics of interactions with the real world and
00:44:42
then you can get stuff like action prediction like robotics out of the same model. And with these models and the
00:44:49
training there kind of um been a limitation in creating videos and creating images
00:44:57
where the criticism of generative AI is it's a bit of a slot machine. I give a prompt, it gives me something back. But
00:45:06
how did it come up with that the training data, but you know, maybe I want a different style? Maybe I want u a
00:45:15
different color. Maybe I want a different uh you know, aesthetic. >> Yep. >> Has that how does that problem get
00:45:24
solved? And do you actually understand what's happening when the image is being made under the
00:45:31
hood? Yeah. Yeah, I I think like ultimately it's about like exposing as many like manipulation layers as
00:45:40
possible to like I don't know like a user or developer that builds on top of this model, right? And um I think like
00:45:48
we've seen that in the past with like in the past uh image models they basically
00:45:53
started from simple text to image systems, right? Then they've expanded into a text plus image to image systems
00:45:59
which means you could suddenly take an image like a real image or generated image um and iterate on that based on a
00:46:06
text font like edit it um modify it right and then this expanded into taking multiple images um and a text font and
00:46:15
combining them in a in a semantic way and producing new content. And the same principle now applies to video. And I
00:46:22
think now it becomes actually even more interesting when like all of these like modalities are actually combined inputs
00:46:27
and outputs of the same model. >> So let's talk about video. There's an announcement that you're
00:46:33
working with the greatest director of all time or living director Martin Scorsesei. We'll talk about that in a
00:46:39
second. Yeah. >> Fantastic. Uh but in a movie uh this promise of being able to make a movie in
00:46:47
which the camera angle uh the sound could be something that a Martin Scorsesei would
00:46:57
be proud to release to his fans. How close are we? And maybe tell us a little bit about this partnership, the
00:47:05
technology being able to make an actual movie like Good Fellas or a scene from Good Fellas, uh, versus where it is
00:47:14
today where you can make interesting five or 10 second clips and then maybe people struggle making 10 of them and
00:47:22
then they use some post- editing software to put them together, but you immediately understand this is not that.
00:47:30
It's not a movie. It's AI slop. It's cluji. It's doesn't pass the uncanny valley. Well, I think it's important and
00:47:41
that's at least like the view that we have is that um these AI models, they are a medium, right? They we don't want
00:47:48
to set like any way of how they are supposed to be used. We don't want to tell anyone especially not someone like
00:47:53
Martin Kazi how how is he supposed to use these models? Like he is one of the like greatest filmmakers ever. Uh it was
00:47:59
insane sitting in the same room with him multiple times and actually him seeing like exploring our models like as like
00:48:05
one of the like core researchers behind it was like just an insane feeling. Right. And at the same time I'm also
00:48:11
like a big fan. Um >> so you sat in a room with Marty Scarces and showed him your tools.
00:48:16
>> Exactly. Yeah. >> And what was his reaction? What what did he key off of? What was the thing that
00:48:23
he found most inspiring or interesting? Okay. Um I think it was really this idea
00:48:28
of like he has clearly um a a vision in his head of like a scene or a scenery where like maybe a new movie um will be
00:48:40
shot and he's trying to explore that and kind of like um we we we basically looked at the scenery of like a village
00:48:47
in Eastern Europe somewhere and he was describing it. we saw some outputs, we iterated on the outputs. Um, and
00:48:54
ultimately I think and that's what he said in the end is like the like getting like the mental picture of something out
00:49:00
of your head and communicating it in a visual way by making like these images um or the series of images um is
00:49:09
something yeah that just makes it like easier to communicate and con convey like an idea of like what is actually in
00:49:14
your head and I think that's like one of the like very interesting and powerful ways to use this technology and I think
00:49:20
ultimately >> is to get the inspiration to get the vision out of his head onto an image.
00:49:27
>> Yeah. I mean like language ultimately is like a little bit of like a lossy um
00:49:32
communication medium, right? >> Yeah. >> Um it's also interpreted in different ways, but then visual information is so
00:49:38
rich. So rich uh like an image or video, there's so much signal in it and it's just like another way of communicating.
00:49:45
And I think that's like one of the beautiful things that this technology ultimately enables. And I think like to
00:49:50
your question of making like full movies with I don't know like a video uh generation
00:49:57
model for example, I'm not sure if that is like the ultimate goal. Maybe it's like interesting to plug this into like
00:50:04
some kind of aic workflow and make like a very long video. And I think that's really cool to explore, but I think
00:50:09
ultimately like the real interesting use cases they come when you have like a human in the loop who iterates and uses
00:50:15
it as a medium. And I think this is this is at least like a perspective that that
00:50:19
I take um that makes it interesting and that this is most often when the most interesting outputs um arrive or are
00:50:26
actually being made. >> The brainstorming production level is so obviously a huge win for
00:50:32
>> you can paralyze your brainstorming basically. >> Yeah. And yeah, I like that. Paralyze
00:50:37
your brainstorming and and they have an analogy for this. They do storyboards and some of the great directors Ridley
00:50:43
Scott of Aliens and Gladiator was known for making his own. I also believe Spielberg was also like to sketch
00:50:50
Raiders of the Lost Arc and some of these. George Lucas was known for collaborating with many amazing artists.
00:50:57
Um even making miniatures and making storyboards for the Star Wars franchise. He had those people on full-time helping
00:51:04
him with that. So that's the obvious place to start. But if we look at startups, startups uh always want to try
00:51:12
to figure out how to do something cheaply. And people used to make a launch video for their startup for, you
00:51:19
know, $100,000, $250,000. So they take their $10 million venture raise and spend $250,000 on a launch video. I've
00:51:27
seen with a lot of the startups I'm investing in now, they'll just spend a week or two working with um you know, a
00:51:34
director to make a launch video. You've probably seen this trend. Yeah. And I'm sure people use Flux and some of your
00:51:40
models for this. Uh have you seen this? >> Yeah, of course. Yeah. >> Yeah. What's your take on that? Because
00:51:46
that feels like the early stage of storytelling. You're trying to communicate a product or service in a
00:51:53
fun, engaging, punchy 30-second, 90second way. Yeah. >> I mean, like again, like I think we
00:52:01
support these like exploration based on these tools, right? And I think ultimately it's great to see like all
00:52:07
different kind of like >> I don't know like launch videos products being built on top of like the same kind
00:52:13
of like base model or the same technology. Um and I think that's what's making it so interesting and also so
00:52:20
powerful. >> Yeah. And what else are people using the technology for? I understand there's a
00:52:26
Bitcoin movie coming out. instead of using a green screen in this Bitcoin movie. Um, I was talking to Galado, you
00:52:33
know, the woman who played the actress who played Wonder Woman. >> Oh, Galado, of course. She I was talking
00:52:38
to her at an event and she was telling me um it was the Breakthrough Prize uh Yuri Milner's event and she was telling
00:52:46
me she just did a Bitcoin movie and they did it on a sound stage without green screens
00:52:51
>> but all the actors just worked in like a sound stage >> and then all of the scenery behind them
00:52:57
was being done by Generative AI. That's a real movie. That's a $30 million budget movie. She said it would have
00:53:03
cost $150 million if they had to build sets and the film would have never been green lit. Are you starting to see
00:53:10
people use that in production? Not just in the back end and the ideation phase, but actually in production yet with your
00:53:18
tools. >> Yeah. >> Um, yeah, we see some use cases like that in production. I think like
00:53:23
high-end film production is kind of like the one of the like most demanding use cases. And I I think I'm glad that it's
00:53:28
being explored, but I also >> really want to um like it's I think it's important to see
00:53:34
that this technology is like on a trajectory and it's improving. It's improving rapidly. I don't know if I
00:53:39
look back at like where we started like a few years ago when I was doing my PhD in this field like the only thing that
00:53:45
you could do was like images of 64x 64 pixels. Now you can do like multi-inut videos, right, at like a high
00:53:53
resolution, but it's like it's not going to stop there. It's going to it's continue to improve and I think like
00:53:58
then it's going to unlock like even more of these like high-end use cases. But I
00:54:01
think the main thing >> before we get to that, yeah, >> hard to predict. I think hard to predict
00:54:06
and I think ultimately >> couple of years >> ultimately I think you still want to
00:54:11
have like the tool that enables like this human in the loop kind of >> of course. Yeah.
00:54:15
>> Um production workflow, right? But I think when I look at multimodal generative models um as a
00:54:25
whole, I think what really excites me is you can use the same kind of um AI model
00:54:33
to make a movie and deploy that as a brain on a robot. Ah, and I think this is like this is so interesting. Um and I
00:54:42
don't know like there's like some thoughts around trying that in the digital world, right? which would
00:54:48
be for example computer use remains to be seen if that is actually something that works or not but I think like the
00:54:54
technology is so powerful and so versatile and it's just just moving into that and all the all the talk around
00:55:02
like world models world action models all of that it's basically all the same and I think that's what's making it so
00:55:08
interesting and what I find like most exciting >> so do you believe that the technology
00:55:13
will be used to analyze or primarily to analyze is real world like here's a video of somebody you know making a
00:55:22
sandwich. Now we have the robot study it and make the sandwich or do you think there'll be a lot of synthetic data made
00:55:30
that then the robots will just study the synthetic or they're going to just in some way innately know based on all this
00:55:36
massive amounts of training data. Um I think it's a combination of prediction right and p prediction and is a way of
00:55:47
you can think about it as simulation as generation it's um predicting actions which is you have to understand the
00:55:54
input the visual inputs in order to actually predict a reasonable next action um and it's about perception it's
00:56:01
like you can only do that if you understand if you perceive the content you then you can only I don't know like
00:56:07
transform it into a new piece of content or predict an action or describe what you actually see in that um scene. And
00:56:13
the and the combination of all of that is I think um yeah is I think what's what's driving
00:56:20
it. There's not a single one of them. It's a combination of these thoughts. >> And what's the best way to get that
00:56:25
training data? Do you need to have people put on glasses, get a firsterson perspective, have them put on gloves so
00:56:33
you have that, >> you know, fidelity of understanding, hey, this glass is moving. I'm pouring
00:56:39
this glass. I'm putting ice into it, you know, and and here's how that's works and the splashing and the condensation
00:56:45
water so I can pick it up and not drop it because it's wet on the outside. Or is it going to be just, hey, take the
00:56:52
corpus of YouTube videos and the robots know exactly what to do because they'll find a thousand videos of people pouring
00:56:59
drinks. >> I mean, ultimately, I think you would want to go to a place where you could
00:57:04
like prompt a robot in context, right? as you can do with like a language model basically just tell it hey go and I
00:57:11
don't know pick up this glass with the I don't know orange juice or whatever it is yeah exactly
00:57:17
>> we're not there yet um but I think this this is like one of the goals and I think like how these models are deployed
00:57:23
currently is there's like a lot of like different hardware different robots that
00:57:28
are running in factories um that all have like some different kind of action representation that you need to kind of
00:57:35
tune the models towards right So in practice um what you do is you have like all this like visual understanding in
00:57:42
the models um and then you need only a very little bit of like a few hours of um fine-tuning data to adjust the model
00:57:51
on that specific task and I think the goal would be to kind of move away from that uh towards like as much in context
00:57:58
as possible but it is a little bit of a research problem. Uh, I I think that >> open source is kind of having a moment
00:58:06
right now. We've been discussing it on the podcast a whole bunch recently and people are also talking about
00:58:12
sovereignty. You have companies that own incredible IP libraries. I mentioned Star Wars before. Disney owns an
00:58:19
incredible library. What should your advice, what would your advice be to a company like Disney? Should they take
00:58:25
your open source software, train their own models or work with you to train their own models to control it and then
00:58:32
hey this is our IP? They've already made a point of working with chat GBT and saying hey you you can and cannot use
00:58:39
certain characters. In fact, OpenAI had a relationship with them that's for Sora
00:58:43
that's no longer happening but they officially licensed on the output some characters. So how do you think about
00:58:49
those major IP holders? What's your advice to them? Are you in discussions with them? We know about the Martin
00:58:54
Scorsesei or tour deal, but how do you think about content libraries? >> I think it is um look, I think like the
00:59:02
most interesting use cases of this like if you think about like content creation
00:59:06
um is in generating something making something that hasn't been there before, right? Like that that's a fundamental
00:59:13
like interesting aspect of this technology. And then I think like yeah when it comes to IP what we implement
00:59:19
for example on like our public facing tools is you cannot generate certain IP with these models right and I think
00:59:25
that's something that is a a sensible approach and then yes we do work with certain IP uh holders to develop models
00:59:35
together with them um some of them based on our open source models some of them based on like our more powerful
00:59:40
proprietary models but I think that is like a very like attractive >> value there what do you think that will
00:59:47
look like for consumers in another couple of years? What would potentially happen when you open up Disney Plus?
00:59:53
>> I mean, that's a good question. I'm not I'm I'm not in Disney, right? So, it's
00:59:56
up up to them to decide that. But, I think we want to enable them to build all kinds of stuff that they that they
01:00:01
that they envision. And I think we can support them. We can support like other companies in that space to
01:00:07
>> I don't know integrate the technology in the best possible way. I think like one
01:00:10
of the very interesting angles of it is that it is like it's becoming much faster. It's becoming more interactive.
01:00:15
I can imagine like a whole bunch of like very interesting interactive content creation tools that you could host on
01:00:21
Disney Plus Plus or elsewhere. >> I think the most interesting thing I've seen in this regard is uh fan films,
01:00:28
>> right? So there's a category before generative AI fanfiction. People would write their own Star Wars story.
01:00:36
Then there came fan films where people would dress up as Jedi Knights and record their own films. And George Lucas
01:00:44
said, "As long as you're not doing it commercially, you're not selling it, I give you permission to go make Jedi
01:00:50
movies." And they even released how to, you know, how-tos on how to make a lightsaber or, you know, sound files of
01:00:57
like how to make the lightsaber sound. Now, people are taking the stories that haven't been told from the Star Wars
01:01:06
universe and they're recreating them using AI. And for the fans, they're becoming quite
01:01:13
popular on YouTube. Uh Star Wars Stories Untold is, I think, the biggest one. It's getting millions of views per video
01:01:20
already. And I think that's really the future is letting the customer base pay a licensing fee or pay a fee, uh maybe
01:01:29
rent software or maybe based on the output and let them be creative with the characters, let them make their own
01:01:35
stories, and you could be in a unique position to empower that. >> No, 100%. I think like if you find like
01:01:40
a model that works for like the um IP owners uh but then also can enable like these super like creative customization
01:01:47
use cases I think that's great. Yeah. I mean like like I mean like for myself like I when I read a book or whatever
01:01:53
like watched the movie I had like so many like ideas how it could be done differently or this could have happened
01:01:58
right this is like so nice that you can actually enable people to visualize these ideas.
01:02:02
>> Yeah it's uh going to be incredible continued success with it. You have an office in San Francisco. You're hiring
01:02:09
people? Yeah, >> we do. Yeah, we just >> we raised a bunch of money. >> We raised a bunch of money. Uh we just
01:02:14
crossed 100 people. We're hiring in Germany and in San Francisco. >> Fantastic. Who are you looking for?
01:02:19
What's the right type of person, the right type of scale? >> Yeah. Um on the one hand, we're always
01:02:24
looking for researchers who have experience in large scale model training. Um experience in
01:02:31
diffusion model training, flow matching training. We're looking for engineers who want to be working with the
01:02:37
customers to you know develop these like customized um physical AI solutions or for example with like a IP owner like
01:02:46
develop these models jointly with them. Um we are looking for engineers who have
01:02:50
experience in just like large scale compute infra managing that um and making sure that the training runs runs
01:02:58
smoothly that we maximize our MFU and all that um and we're looking for people who have interest in
01:03:08
you know like uh getting the technology out there >> in the hands of people >> the the the for deployment of this
01:03:14
there's just so many great ideas and so many great partners for you I think you're going to with the open source
01:03:19
specifically. You know, it seems like the corporates really want to have some additional level of control, but they
01:03:26
also need the frontier models or your proprietary ones for some of those refined features. So, I think you have a
01:03:31
very bright >> future. All right. Continued success. Thank you so much for doing the show. Pleasure.
01:03:36
I'm going all in. I'm going all in.