Open Source Wins, AGI Is Here, and Scorsese's AI Toolkit with CEOs of Cerebras & Black Forest Labs

cerebrus doing in friendships pioneered the space had a successful IPO we've talked about this a couple times we got to see each other in January at Davos IPO happens the boys and I got to sit with you recently that was fun at liquidity that was really had a great discussion with the boys but I wanted to deep dive with you about a couple of topics the first one is the build out

of AI we've never seen a build out like this since you know the great wall of China right

who knows it the pyramids right I mean it feels like the amount of capital time and intelligent people on the planet dedicating themselves to the build out of something I can't think of anything in our lifetimes with perhaps you know before our lifetimes the war effort right this is a mobilization in a scale that we read about we hear about but you're actually doing it

“you have customers who are building data centers and you're a key piece of that's”

app love and started with an eight dollar domain and no VC funding and became one of the largest ad platforms in the world now that same engine powers app love and ads for e-commerce

your ads run inside mobile games reaching over a billion people with full screen distraction free

attention the platform finds buyers and optimizes for profit you set the target it does the rest one cookware brand went from 4 million to 16 million dollars turned profitable and is on pace for 80 million this year visit app love and calm slash all in to launch your first campaign today maybe you could just enlighten us in 2026 what is cerebrus doing and what is happening with this build out out in Texas these are some gigantic gigantic efforts the the size and scope

“of what is being built the physical size and scope usually when we talk about software”

or we talk about hardware we talk about chips and boxes and they don't have the same sort of physical enormity right and what we're talking about now are data centers that are in the next several years going to use more power than the previous 50 years on earth look wow right we're talking about individual buildings the size of football fields that have more power coming into them than mid-sized cities and they're being built they're being built across the US they're being built in

Canada they're being built throughout the Nordics they're being built here in Paris and throughout France and Europe in the Middle East in nations that sort of weren't front center and anybody's mind previously you know cause of Stan to Jesus Stan or build out Georgia building out

“data centers of size or media everybody sort of focused every country and every state obviously in”

America feels they need to participate in this and the people who are buying the capacity the opening eyes and anthropics SpaceX SpaceX AI the Google's they are insatiable right now yeah and they're building how many years out when you talk to them they were ordering chips from cerebus before you were finished with the chips they're putting orders in ahead of time the irony is unlike many sort of exciting times and technology they're trying to capture yesterday's demand

right the demand is way out stripping our ability to build data centers and to fill them with hardware

all right and so you know we have a 25 billion dollar backlog 25 billion dollar backlog

and we're not alone in that that open AI anthropic you go through this list of of Google wants more data centers Microsoft wants more data centers AWS wants more data centers right all of these players are not chasing sort of if you build that they will come they're chasing the demand is booked right how do we keep them from leaving right and and that that's extremely unusual it's very unusual and now we have people who are you know we have a term for a token

maxing yeah and there's a great debate is this actually creating value I'm curious where you stand you know is it even possible that this much demand could be created if value did not exist there is clearly massive value happening yeah but there's also massive experimentation oh for sure you know what I like in this to when we first started with AWS and it was so good to get around your own IT organization right it's told every engineer you had go ahead put on your

God I wish we didn't do that yeah and so for sure there's experimentation but it doesn't mean that the net value isn't enormous it means some of it is going to go nowhere and you know it was the same I remember when Costco opened up in in in in the Palo Alto area in 1988 and people used to shop Costco like they shop safely that go down every aisle yes that's a horrible way to shop Costco because we ended up with four things you didn't need in each with $20 right and as people got

more sort of accustomed to it you go to the back to get the chickens 18 cupcakes for the kids

“birthday party bang you were strategic and it's exactly the same I think at first people opened up”

and said everybody as much tokens as you want and the in enterprises there's no open loop we don't give sort of any resource unconstrained to people and now we're jumping on a saying whoa all right these guys should have as much as they need they're enormously productive over here we can use maybe an open source model maybe a cheaper model over here and now we're sort of running a like a business and we're really seeing a certain type of person emerge who knows how to deploy

this technology systems thinking yeah which developers kind of have in eightly CEOs tend to be great strategists and understand systems but the intelligence is getting so much better every step along the way that I'm watching individuals typically startup founders but also venture capitalists and associates who work at my venture firm they start playing with the tool and then the tool starts playing with them right 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 under I've never heard about making a required

document and the software's like do you have a requirements document what's your goal the AI

“starts telling people you're talking maxing and you need to get a little more focused here one of”

one of my colleagues twenty years ago a really smart smart computer scientist said computers really dumb they do exactly what you tell them yeah and at first prompting was like that right you modified your prompt a little bit and it changed the answer dramatically dramatically and increasingly it's understanding what your intent was right right and if you if you have a chance to to to play with fable or or five six from from open AI increasingly what you you don't

have to get the prompt just right you don't have to be a prompt whisper instead you ask it and it says well here are some things and and by the way maybe you wanted the chart to to go two ways you wanted a line in a bar and it's like well that's exactly what I wanted I didn't ask for it but that is better and so it's it's understanding intent and that's a huge leap which if we were

sitting here two years ago with idea we would never have been able to predict in a short 24

months that we go from being a great summarizer researcher of web results right to actually understanding your intent and then providing a solution and abstracting it all from you that's right which is a very weird thing I don't know if you've played with the Hermes agent yet yeah have you played with it yet I mean I asked it just this morning and I was given a secret bit tensor project that has the new Z-A-I model five two and they gave me G-L-M five two G-L-M five two

“so somebody in that bit tensor I think you understand bit tensor you've heard of it the distributed”

crypto project and so they have all this extra capacity I was a whisperer told me probably some capacity in China that is free energy okay fine so they gave me a limited capacity so I started having to do some really crazy jobs when I was saying like every hour I want you to tell me what the trends in the world are that nobody else has identified yet and you can do whatever 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 in the background and it started debating itself but where it should find the things is that we should probably go to hacker news and read it and then it was like yeah but there's also social media and trends tend to manifest on Instagram it's a reasoning model you were watching a reasoning model work out yeah it's not interesting I mean

that's amazing and it was collapsed so as a civilian right who doesn't hit the uncollapsed moment

and if you were using chatchipit three point five or you were using four point eight whatever it was and you haven't used this new level of reasoning and inference and unlimited compute essentially

right because unlimited tokens I believe means unlimited reasoning it does what does that mean

yeah it's I mean if you run these for 25 or 48 hours you get amazing things now

and what if by using three bursts we were 15 times faster and then you ran it for 24 hours right and you got weeks of months worth of thinking yeah and I mean it is

“it is extraordinary and I think one of the things is people like Ilia and Sam in the early days”

were saying this was coming right and I think when you look back you said yourself holy crap those guys they knew saw it yeah they were they could see around the corner that's right and the bill was right I'm not sure it when we had Sam on all in at one point he and he said you know I love to come on at some point I said sure come on and he was talking about it he said you know I said what's next he said reasoning there's an unpacked out what does it mean so

well understanding what your intent was just as you're saying and then figuring out a strategy and then maybe talking to other agents and other threads about like is this the right thing to do inventing each other's work and I'm like wow we have come a long way from guests the next word right fill the sentence in you know that that summarizes pdf now cerebruses at the center of this because this reasoning is inference this reasoning is inference and it's computationally

intensive right right and so fast compute makes this sort of work the fast and sort of tractable it doesn't cripple it by taking a huge amount of time to get a good answer and so it's exactly the fact that that that this reasoning consumes a huge amount of tokens internally

that allows a blisteringly fast machine like ours and I I brought one because I have never

far without you know when one costs half a billion to make you bring it everywhere with you and we were we were tossing this back and forth at Davos yeah what's the model number of this one this was in the first eight or ten got it so this has a special place this has a special place I mean my wife says it's like I'm a kid with a dirt bike for his eighth birthday within his bedroom at night I carry him with me I mean when you have you know your your next

“party at the house I highly recommend just a little more dirt I think it'd be like a great bet”

it would be a great bet if you had something that's great but what we're looking at here is the ability to do that reasoning at scale and what is Moore's law for inference and for cerebrist do you have something internally you discuss as have we're gonna double this every time period so all chips prior to us in the process or world followed Moore's law gone and we broke you doubling every 18 days doubling about every 18 months done and we crushed it

with this chip and we've carved out a whole new trajectory and my view is in the next 18 months will be way over to X interesting and so I think that early in an architecture you have room to do much better than what was traditionally Moore's law now if you've got a 20 year old

“architecture like the GPU is much harder right you you have to rely on things like smaller geometry”

right going to the next fab node but in a newer architecture you have a huge amount of room still to to learn about the work that there's being presented and make optimizations that give you huge

gains how do you run the company like just being the CEO now in the age of AI you have 25 billion

dollars in demand you have to you have to deploy at an just an incredible blister in case you have to hire people you have to create a roadmap I don't mean to give you a panic attack here you have to keep up with somebody like open AI who's moving so unbelievably quickly yes right and they're they're competitive you got to keep up right right you are hardware you're software you're deployments have to keep up with some of the fastest moving organizations in history they're demanding customers

they are not they're not pushovers for sharing yeah and also potentially competitors down the road look I think there is so much demand right now that there is no silicon that will go on used yeah why is an opening I releasing jalapeno why is Amazon making their own chips you see

too so we need good pricing is it a little bit of a flex that way or is that the future that they're

“gonna be in your business no I I think nobody likes being dependent and I I think some of the lessons”

learned by the the hyperscalers of the x86 world is they were dependent on Intel and some of the lessons learned by the GPU makers was they were dependent on a small number of hyperscalers yeah and they wanted more customers and so they set about to to help fund these new class and so I think mostly it's about an opportunity to control at least an important part of your destiny don't and I think that's a very reasonable thing I think you don't have to sort of make the fastest

chip you just can't be entirely dependent on other people's chips and that dependency is become a hot topic um not sure if you caught the episodes over the last two weeks but we've been talking over the last year about open source I've been championing that a lot just because I was early into open

claw and quickly started using kimi and was like wait a second I'm blowing out my claw tokens

but this kimi I can't tell the difference and then we started smart routing it and suddenly this open source started to figure out reasoning and the gap as well suddenly closed this year but I you know you you don't want to take your your Ferrari to the grocery store right yeah they're time

“if you want to drive your fun car yeah right and they're time you want to throw the kids in and don't”

worry if they're Cheerios on the floor mini-venta right there's mini-venta and I think that as the sort of sophistication of the user grows right you're going to have hard problems and those are

going to be front to your model problems they're going to be open AI problems they're going to be

anthropic problems maybe you have an I-promise and behind that they're going to be a lot of ordinary problems right I mean if you think about a company you know how much time is spent cutting things out of workday and getting it in a different cell for yeah all right think about the cutting in pasting economy is real that's right and this doesn't need right gold metal math what what this needs is sort of rock solid open source capabilities yeah and if you think about what I mean

when we've been thinking a lot about it in GNA but a huge amount of GNA all right is not invention right and you made on need sort of the most sophisticated agents for this and another card that's turned over recently is some folks maybe half concerns with the ambition of the front to your models and maybe sharing their data lead data leakage and sovereignty of intelligence and they're saying hey our company is going to choose maybe we're in a regulated industry finance

health care hip up you know fedra all kinds of different regulations we need to have this on prem and we want to have domestically and we'd like an open source version where we have a

“a little bit more control yeah I think are you seeing that now you are seeing that for sure and I”

I think open area made a good call releasing OSS 120B some months back that's a good open source model but I think in the US we need more domestic open source models need to give the world a choice right if they want to run open source right now it's OSS 120B or Chinese models and video has some it in in video has seen the same all opportunity yes push open source models I I think giving them more power might might be sort of well I was about to that was you you cut me out with the past

like it my understanding was Jensen was like hey we we don't even want to talk about these open source models we have because our customers right we're now going to be competing with Sam Dario Elon Sergei like do we want to be in that position right so but we do need some more champions here and it's open source so people can fork it but that puts you in a more neutral position that's right we we we run today we run GLM we run kimi we run the quense set of models and we run

open AI models the closed source ones we run models for say GLXOS Smith Klein which they wrote and developed we run models for our partner in the UAE G42 and MBC UAI that are our their models that they designed so we we have a a wide variety so sovereignty is a trend a sovereignty is a trend

hello whoa let's think and then we can act I think sort of particularly here in Europe with a bit of a wake up call and when you saw this going down there's a layer of partisanship in our country right now it's pretty fervent Dario is pretty explicitly you know not part of this administration they've been very adversarial both sides have been have admitted that they're

“starting to work it out now so it's hard I think for us not being in the room with these parties”

to understand what's partisanship what's gamesmanship here but do you believe that what they released was truly dangerous for cyber warfare for cyber attacks and that if you were to rate Dario's not communication because he's a very effervescent communicator I think is a diplomatic way to say it but to have a scheduled rolled out release right we'll put aside the governments control of it but do you think that is is a wise thing for us to do at this point and do you think it's there was

actually a major threat there so what's interesting is I hadn't seen it before right and I I think if we just step back and say is it reasonable I don't know whether this was the right time but at a time that that that a model is sufficiently creative in its thinking that it poses a meaningful threat for the government say we'd like you to roll it out in steps yeah

this doesn't seem unreasonable to me not at all right I mean we we do this with powerful

pharmaceuticals right we we'd like I mean certainly not encouraging seven years of trial and the amount of paperwork and all the garbage that has accrued to the FDA but with the powerful new technology it certainly doesn't seem unreasonable to say hey guys let's at least do some red teaming at the government yeah we know our defenses can block this yeah have we checked have we checked the infrastructure of the country like of the NSA have we checked the infrastructure of right and can

you give us two or three weeks to patch in the obvious holes that are found right this doesn't seem to be an unreasonable thing for the government to ask right we but we in this very polarised time

“put on top of it well oh my god it's president Trump doing it and then you have to think well”

what if it was president always a O.C. or president anybody in between the two extremes I think

the polarization hurts a great deal it hurts clear thinking right it it hurts clear thinking and and both sides are going to do some dumb things and some really smart things right right and in fact what I found is that the people in the government are trying really hard the ranking the rank and file are trying really hard and this is moving fast and I I think that that an ability to set aside some some of the polarization and say how do we do this in a reasonable

manner I mean we we want Dario and Sam competing like crazy we want the awesome to watch awesome yeah right it's good for the technology it's good for it it's good for entrepreneurs to see even with thousands of people this is what what you can continue to achieve right right this can draw me ask them get sharper and everything got better everybody got better because of that we want that and we certainly don't want to become sort of a region where the first and we want to

is regulate it right but as it gets more powerful and the industry really should do a better

job of regulating itself perhaps and and it did seem like they were starting that process but then

“the communication was lacking maybe I'd so you know I I think not only are they race and hard”

but they're inventing this as they go to yeah right there's not a playbook no right they're inventing the we we say I just put on guardrails look they have to design the guardrails short right they the guardrails have an impact you know one of the things that fast does is it makes the guardrails less painful and so that that is we you know we discovered that in the last six weeks yeah it is that the very guardrails can add time and make it feel slower and so fast chips like ours

right it can really help that but so they're racing against competition they're racing against their own sense of greatness yeah right which is maybe even the biggest driver here and I think they're earnest trying to think about I do the right thing and and all of those are are mixed in this bucket and and sometimes you're on one side rather than the other yeah and as you're

chatchip to 2.5 3.5 it was taking down networks but in talking to Nikash from Palo Alto networks

“I asked him like hey well how would you grade this and he said oh we put it against our software”

and we found bugs we were not aware of yes and he killed him yeah he said we had to stop everything we're doing and do patches for six weeks right and and that's when you know right I mean Nikash leads you know maybe the leading sure security software firm right and when it finds

in an hour right tens of critical opens you're like whoa there's a powerful tool yeah I

need to think and maybe you you show it to a group first right maybe you I don't know what the right thing is but I mean red teaming and we've always had just when you were releasing the new version of an operating system you know when you have your iPhone you can say I want to be part of the beta that's right you know right and there's like two other banners that you don't even get the headchance to opt into as consumers and right those ones are for security those ones are

for you know making sure you don't lose your data or data that's right disappear or leak or

“corruption yeah any any number of these things I think we can also know that there will be a massive”

data leak of course we we know this yeah right and it's like a worn buffock talked about the reinsurance industry that you know something bad is going to happen you don't know when yeah but you got to save up for it right you put money away for the reinsurance but they will be a tornado they will be a massive earthquake I mean we we know this and we can do our best to plan but they'll be a massive bridge and they'll they'll be in we we have to steal ourselves in advance

and we have to think about it think about the right response at the time and sort of prepare ourselves for a future that is in specific unknown but in general we're pretty sure it's getting something's going to happen something will happen right yeah and yeah it's typically a blocks one right that's right by definition it's going to be something we didn't consider or a question we didn't know to ask right but but even knowing that there's some unknown unknowns it is a useful place to start

yeah but we're not asking ourselves that's right with reasoning the AI is going to be able to tell us hey schmuck humans that's right while I hear what you're not thinking about this is now my closing sentence when I do my prompting is I need you to make me a prompt that will help me do this trend

scouting for an example and then I always say at the end please check your work right and

then tell me what I haven't considered in terms of my goals and give me ask me some questions every time you run the job and that has changed everything because it's like I checked my work by the way this was incorrect right and I'm wondering hey would you like me to also do this and some of the tools like perplexity do that automatically give you your next three prompts right but if you give it explicit instructions my lord is it good at that so you know over the course last 10 years

as I was raising money I thought one of the smarter questions I got at the end of a conversation where someone asked what was the smartest question you heard that it wasn't covered by what I

asked incredible right now that's somebody who's curious and how I'm thinking and humble and trying

to sort of use this to get a picture of of the space and to the extent that you can ask the AI that and and then it can sort of broaden your view you know maybe what questions should I have asked yeah to be an expert in this what what would a a PhD level a question or ask of this or a gold medal math I mean I I think those are sort of questions that you you know you don't even know how to ask which you know if you start thinking about AGI and super intelligence

“you know they're just definitions but they're important definitions I think to kind of keep in mind”

because their waypoints that's right and AGI I think I suspect you'll agree with me that we've hit it we just have an exactly deployed it fully we we have artificial general intelligence now it feels like when we're talking about these reasoning moments and you know the the ability for it to be as smart as any human but let's talk about any definition we had 20 years ago we've hit it yes right I mean if you think about all those deterring tests blew it away

yeah I mean you think about that that any period of time sort of 10 15 20 30 40 50 years ago we we any definition we would have previously put forward right we've blown past it and so which goes back to our previous point of like do we know the questions asked that's right 20 years ago

if they were to look at this today they'd be like well I'm out of not a question sorry

that's where sort of the sort of listening to people who we who sound sometimes like they're on the fringe right when when Ilya was talking eight or ten years ago about the need for safety and then and you're like what did right yeah right when when Elon was talking about building rockets and driving the cost to to near zero of of a launch vehicle you know what there it is and now you

“can see and that that's I think that's why it's really fun to be a technology's now”

well and with these tools specifically you know we're talking about building all these tools and then the tools are starting to build themselves in this recursive loop that's right we're kind of just starting to see people apply loops in fact loop maxing became when I was doing my trend but I did my trend then it kept picking up loop looping and it kept picking up the maxing stuff and it created a buzzword for me loop maxing right and then it magically people

started talking about loop maxing and I was like wow this is really weird it anticipated that this would other humans would come up with this word but talk a little bit about recursive and then the road to super intelligence and do you have a way under that you think about super intelligence and what it will mean for humanity and how we will define it and how we'll experience

“it yeah I I think let's begin on on loop maxing or sort of recursive learning I I think”

I think what what Samanelia and then later Dario and and Danna Danna's saw

six years ago or five years ago was that powerful recursive games are exponential

right you get better you do it again and if you continue to get game the slope of that curve is so steep yeah and that we're just beginning to see that now yeah I was going to question you learn from the results you ask it to do it again it that results get better and more information added your answer gets better you ask it to do again it covers more material and the sort of loops are producing sort of not a little bit better answers but vastly better answers yeah and that is

enormously powerful because we don't quite know where it ends right you keep throwing compute at it I mean how much better does the answer get you know we we run out of tokens or our budget or or but but holy cow I mean when does the exponential stop or does the answer keep going up and up and up to the right yeah and that that's sort of an enormously interesting intellectual question right now yeah like when do we run out of problems to solve and well that's right

and and when are the the problems the longer sort of intellectual problems and they're now people problems yeah right how to organize people to get done with the I asked for right I mean as you know and running your company a lot of your problems aren't hard intellectual problems there people working together problems yeah right and motivation motivation you spend a lot of times a leader spraying WD40 on your team right just so friction is reduced and how do we learn

“about those from AI right how do we get behavioral insight from AI and and I think that's”

some of the things the world models we're going to bring us as they begin to watch human behavior yeah we didn't even get to that this was going to be for another interview but when these things jump off the screens right and they're in the real world and the recursiveness starts not trying to solve math problems right you know humanity's most difficult ones but

hey you know there's an incredible world out here and here's the palace of Versailles right

you're just like now 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 I've got 100 acres somewhere right Texas or Nevada I'll just send a thousand optimists is out there right make me the palaces of Versailles right sounds fantastical but the palaces of Versailles would seem fantastical to people who lived a thousand years before it and it was fantastic I think to the people who built it yeah even to the builders

I think they were odd you know added as they built it you know they're compounding they're compounding recursive learning that's right and generations we talked about you had a really such a great insight of in building this place you had generations of mason's yeah I

you and you you were apprentice on your father your uncle and when you had a project that took their 50 or 70 or 100 years you might have three or four generations of the same family right the same stone mason family working on the same structure and passing all the classes on the learnings new innovations right which is what we've modeled with this new

that's right models and and what you're building in the infrastructure it's pretty incredible

when you think about it especially when we're sitting here and but the place I that and that that's what I mean I I think the the problem with human learning is it often moves it at the pace of a generation uh and like uh elephants and other large mammals we don't have generations but every

“15 or 20 years and if you want to move really quickly across generations you want them happening”

more like Drosophila like food fly you want to a day yeah right then you see that in genetics that's why we study them in genetics because learning encoded in the DNA you can study over thousands of generations and I I think what we're getting is that equivalent in AI we're getting sort of learning so quickly over the equivalent of thousands of generations yeah Darwin would be in awe of this pace of that version and that's exactly right you think about it as there's remember when I was

getting my psychology degree and they were teaching us about paradigms and I was like I tried to understand how the paradigms shifted and the pressure set to me Jason which after understand is paradigms don't die they don't people do that's right and that's how the Freud he was good that's right and Skinner and young like it took them dying that's right for the next generation the new generation and that was 20 years yeah sometimes 40 years right as their students

maintained positions the leadership until someone said maybe we could do it definitely and I think what you're seeing is this iteration is a shortening of the the intergeneration gap

and the learning is so fast it's always so great to talk to you because

one it's just intellectually so you're approach to it is so intellectually rigorous but also with so much P.D. in the world I feel so good that you're such an optimist about the

“technology and you're building it with such thoughtfulness and I think for people who are”

hearing these horror stories about AI and job laws and everything they need to understand there are people like yourself who are building this in an incredibly awful way and this is going to be a net benefit for humanity that just is unimaginable yeah we have a shot with this technology so not our children or anyone they know dies of cancer right I mean say it like that there will be some dislocation in the economy sure there will be

there was dislocation when cars came and it was a bad deal to to be a guy who's shoot horses right or build carriages yeah but you got to also against that you know make your tea the cons and the pros yeah right there's a shot that are children none of them nor their people they level die cancer and that that's one that's funny why we can work on with this technology

“and we will have great purchase on and I think you begin listing those and then it's a more thoughtful”

discussion yeah unlimited energy that's right calories unlimited knowledge unlimited education unlimited housing and how we do it we imagine imagine sort of we know how to teach children and we don't do it right Aristotle was a tutor tell us enter the great soccer teams was his tutor we know if you give a child a tutor and the tutor modifies the teaching for the child yeah they learn better that's not how we do teaching class no factory farming that's right we we teach to some sort of

middle-level imagine if we built agents that that taught children for their way of learning right right and here's the way we we've been doing it the same way for a thousand years and during that entire time we knew how to do it better and we just not to yeah and here's the way we can do it put that on the pro side and so as long as we're sort of thoughtfully and fairly yeah writing the good and the bad I think it'll come out of it out there and keep communicating

your version of the world because some people see around the corner and they get a little nervous and okay fair enough but I think the ledger as you describe it is heavily weighted towards abundance

I think and create abundance yeah massive and pleasure always I'll see you in the next month

converging into a single interconnected system and the infrastructure behind it needs to evolve just as

“quickly NASDAQ was built for this moment powering more than 135 marketplaces and regulators”

globally and connecting capital to company shaping the future as the innovation economy accelerates

connectivity becomes the critical asset NASDAQ is the leading technology platform that makes it possible

and scalable learn more at NASDAQ.com Robin Rombock is the co-founder and CEO Blackfars labs you are based in Germany in Blackfars which is a city in Germany a mountain range actually a mountain range yes where you grew up well I grew up yes and you are working on open source image and video models you worked at stable diffusion for a little bit that's correct cut your teeth on that and you're known for the open source model flux and maybe also for

some close source models tell us about the business of Blackfars labs what is the business and

“what is the goal 100% one quick addition we have based in the Blackfars so it's a”

town called Frybook and in San Francisco we have a lot of sandwiches go first yeah you're splitting

your time or I'm splitting my time to a certain degree we started a company two years ago me and my co-founders as you said like we've worked on stable diffusion in the past before that we invented like an algorithm called latent diffusion it's basically like the fundamental algorithm behind all of like generative models that are being deployed for image generation video generation even like physically I know yeah it basically makes use of this principle that you

can't compress natural data such as images such as video such as audio into much more like efficient representation and then train a transformer model on that and I mean this is the

stuff where like you know like jpec MP3 and all of that works and we basically translated that

into like a neural algorithm a few years ago when we were still like PhD students in in Munich actually and then build like on top of that we build stable diffusion and then on top of that yeah the generative models that we are developing today and of course like the technology has advanced but we are now tackling I would say models that are really made for understanding like a whole world around us. Fronty model visual models are pre-trained on images videos audio data at the same

time and we are now like entering a new paradigm which is combining that with something it's called action prediction such that you can actually use the same model to make images to make videos to make audio and to predict actions which means you can ultimately deploy it on a robot and the real was. Wow so from the image to the video the audio and then eventually the real world with robotics and a real world model because if you can make the image you and you

can train the model that means by default you understand the world in order to make a video of the

“world you have to understand the world yeah and the objects I think that's you know I think that's”

like a really good like way to think about it yeah it's like it's like an intuitive way to interact with the world right like I would say there's like these complementary forms of intelligence ultimately there's like intuitive intelligence and then there's like a deep reasoning layer now ultimately you need for like a kind of like complete form you need both and you need them to interact and I think like we've been approaching it more from like the intuitive

side images is like a very natural way to approach this whole field because it's not as computationally intensive let's say video right but now yeah I think like we're combining it it's converging into like a multi-modal model and yeah we see like exactly like pre-training on videos because like implicit understanding of the physics of interactions with the real world and then you can get stuff like action prediction like robotics out of the same model and with these models and the training

they're kind of been a limitation in creating videos and creating images 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 how did it come up with that the training data but you know maybe I want a different style maybe I want a different color maybe I want a different you know aesthetic yep it has that how does that problem get solved and do you actually understand what's happening

as many like manipulation layers as possible to like I don't know like a user or developer

“that builds on top of this model right and I think like we've seen that in the past with”

like in the past image models they basically started from simple text to image systems

right then they've expanded into text plus image to image systems which means you could suddenly take an image like a real image or generated image and iterate on that based on a text from like edited modify it right and then this expanded into taking multiple images and it takes from then combining them in a in a semantic way and producing new content and the same principle now applies to video and I think now it becomes actually even more interesting when like all

of these like modalities I actually combined inputs and outputs of the same model so let's talk about video there's an announcement that you're working with the greatest director of all time

or living director Martin Scorsese we'll talk about that in a second yeah so does he but

in a movie this promise of being able to make a movie in which the camera angle the sound could be something that a Martin Scorsese would be proud to release to his fans how close are we and maybe tell us a little bit about this partnership the technology being able to make an actual movie like good fellas or a scene from good fellas versus where it is today where you can make interesting five or ten second clips and then maybe how people struggle

making ten of them and then they use some post-editing software to put them together but you immediately understand this is not that it's not a movie it's AI swap it's Kluji it's

“doesn't pass the uncanny valley well I think it's a point and that's at least like the”

view that we have is that these AI models they are immediate right today we don't want to set like any way of how they are supposed to be used we don't want to tell anyone especially not someone like Martin Scorsese how the host is supposed to use his model like he is one of the like greatest filmmakers ever it was insane sitting in the same room with the multiple times and actually him seeing like exploring our models like as like one of the like all researchers behind it was like

just and insane feeling right and at the same time I'm also like a big fan so you sat in a room in Marty Scorsese and showed him your tools exactly yeah and what was his reaction what what did he key off of what was the thing that he found most inspiring or interesting I think it was really this idea of like yes clearly a vision in his head of like a scene or a scenery where like maybe a new movie will be shot and he's trying to explore that and kind of like we we basically

looked at the scenery or like a village in eastern Europe somewhere and he was describing it

“we saw some outputs we iterated on the outputs and ultimately I think and that's what he said and”

yeah and just like the like getting like the mental picture of something out of your head and communicating it in a visual way by making like these images or the series of images is something it just makes it like easier to communicate and convey like an idea of like what is actually in your

head and I think that's like one of the like very interesting and powerful ways to use the

technology and I think ultimately is to get the inspiration to get the vision out of his head onto an image yeah I mean like language ultimately is like a little bit of like a lossy communication medium right yeah it's also interpreted in different ways but then visual information is so rich so rich like an image or video there's so much signal in it and it's just like another way of communicating and I think that's like one of the beautiful things that

the technology ultimately enables and I think like the question of making like full movies with I don't know like a video generation model for example I'm not sure if that is like the ultimate goal maybe it's like interesting to like this into like some kind of a genetic work so and make like a very long video and I think that's really cool to explore but I think ultimately like the real interesting use cases they come when you have like a human in the loop

who iterates and uses it as a medium and I think this is this is at least like a perspective that I take that makes it interesting and that this is most often when the most interesting outputs arrive for actually being made. The brainstorming production level is so obviously a huge

and they have an analogy for this they do storyboards and some of the great directors really Scott of aliens and gladiator was known for making his own. I also believe Spielberg was also like to sketch Raiders on the Lost Ark and some of these George Lucas was known for collaborating with many amazing artists even making miniatures and making storyboards for the Star Wars franchise he had those people on full time helping him with that so that's the obvious place to start

but if we look at startups startups always want to try to figure out how to do something cheaply

and people used to make a launch video for their startup for you know $100,000 to $150,000 so they take their $10 million venture raise and spend $250,000 on a launch video. I've seen with a lot of the startups I'm investing in now they'll just spend a week or two working with director to make a launch video you've probably seen this trend yeah and I'm sure people use flux in some of your models for this have you seen this yeah of course yeah and what

what's your take on that because that feels like the early stage of storytelling you're trying to communicate a product or service in a fun engaging punchy 30 second 90 second way yeah I mean like again like I think we support this like exploration based on these tours right and I think like ultimately it's great to see like all different kind of like I don't know like launch videos products being built on top of like the same kind of like base model or the same

“technology and I think that's what's making it so interesting and also so powerful yeah”

what else are people using the technology for I understand there's a Bitcoin movie coming out instead of using a green screen in this Bitcoin movie I was talking to Gal Gadot you know the woman who play the actress who played Wonder Woman yeah yeah of course she I was talking to an event and she was telling me it was the breakthrough prize you're a millners event and she was telling me she just did a Bitcoin movie and they did it on a soundstage without green screens but all the

actors just worked in like a soundstage and then all of the scenery behind them was being gotten by generative AI that's a real movie that's a $30 million budget movie she said it would

have cost 150 million if they had to build sets and the film would have never been greenlit are

you starting to see people use that in production not just in the backend and the ideation phase but actually in production yet with your tools yeah we see some use cases like that in production I think like high end film production is kind of like the one of the like most demanding use cases yeah I think I'm glad that it's being explored but I also yeah really want to

“like it's I think it's important to see that this technology is like on a trajectory and it's”

improving it's improving rapidly I don't know if I look back at like where we started like a few years ago when I was doing my PhD in this feels like the only thing that you could do was like images of 64 by 64 pixels how you can do like multi-minute videos right like a high resolution but it's like it's not gonna stop there it's gonna it's continued to improve and I think like then it's going to unlock like even more of these like high end use cases but I think

the main thing where we get to that yeah how to predict I think how to predict and I think ultimately a couple of years ultimately I think you still want to have like the tool that enables like this human in the loop kind of of course yeah a production works alright but I think when I look at multi-modal generative models as a whole I think what really excites me is you can use the same kind of AI model to make a movie and deploy that as a brain on a robot

and I think this is like this is so interesting and I don't know like there's like some thoughts around trying that in the digital world right this which would be for example computer use remains to be seen if that is actually something that works or not but I think like the technology

is so powerful and so versatile and it's just just moving into that in all the all the

talk around like world models world action models all of that it's it's basically all the same

“and I think that's what's making it so interesting and what I find like most upsetting so”

do you believe that the technology will be used to analyze or primarily to analyze real world like here's a video of somebody you know making a 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 that then the robots will just

of training data I think it's a combination of prediction right and prediction and is a way of you can think about it as simulation as generation it's predicting actions which is you have to understand the input the visual inputs in order to actually predict a reasonable next action and it's about perception that's like you can only do that if you understand if you perceive the content you then you can only I don't know like transform it into new pieces of content or

predict an action or describe what you actually see in that see then the and the combination of all of that is I think yeah is I think what's what's driving it there's not a single one of them

“right it's a combination of these tasks and what's the best way to get that training data do you need to”

have people put on glasses get a first person perspective have them put on gloves so you have

that you know fidelity of understanding hey this glass is moving I'm pouring this glass I'm putting ice into it you know and here's how that's works and the splashing and the condensation 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 corpus of youtube videos and the robots know exactly what to do because they'll find a thousand videos of people pouring drinks I mean ultimately I think you would want to go to a place where

you could like prompt a rewards in context right as you can do with like a language model basically just tell that they go and I don't know pick up this glass with the I don't know why I don't want shoes yeah exactly we're not there yet but I think this this is like one of the goals and I think like how these models are deployed currently is there's like a lot of like different hardware different robots that are running in factories that all have like some different kind of action

“representation that you need to kind of tune the models towards right so in practice what you do is”

you have like all this like visual understanding in the models and then you need only a very little bit of like a few hours of fine tuning data to adjust the model on that specific task and I think the goal would be to kind of move away from that towards like as much in context as possible but it is a little bit of a research problem I think that that source has kind of having a moment right now we've been discussing it on the podcast a whole bunch recently and people are also talking

about sovereignty you have companies that own incredible IP libraries I mentioned Star Wars before

Disney owns an incredible library what should your advice what would your advice be to a company like Disney should they take your open source software train their own models or work with you to train their own models to control it and then hey this is RIP they've already made a point of working with chat GPT and saying hey you you can and cannot use certain characters in fact open air had a relationship with them that's for Sora that's no longer happening but they officially

licensed on the output some characters so how do you think about those major IP holders what you're advice to them are you in discussions with them we know about the markets we're sazy

“or tour deal but how do you think about content libraries I think it is I think like the most”

interesting use cases of just like if you think about like content creation is in generating something making something that hasn't been there before right think that that's a fundamental like interesting aspect of the technology and then I think like yeah when it comes to IP what we implement for example on like our public facing tools is you cannot generate certain IP with these models right and I think that's something that is a sensible approach and then yes we do work with

certain IP holders to develop models together with them some of them based on our open source

models some of them based on like our more powerful proprietary models but I think that is like a

very like attractiveness what would you think that will look like for consumers in another couple of years what would potentially happen when you open up Disney plus I mean that's a good question I'm not I'm not in Disney right so it's up to them to decide that but I think we want to enable them to build all kinds of stuff that they that they that they envision and I think we can support them we can support like other companies in that space too I don't know integrate the

technology in a best possible way I think like one of the very interesting angles of it is that it is like it's becoming much faster it's becoming more interactive I can envision like a whole bunch of like very interesting interactive content creation tool that you could host on Disney plus plus or elsewhere I think the most interesting thing I've seen in this regard is fan films so there's a category before a generative AI fan fiction people would write their own

and record their own films and George Lucas said as long as you're not doing it

commercially you're not selling it I give you permission to go make Jedi movies and they even released how to you know how to choose on how to make a lightsaber or you know sound files of like how to make the lightsaber sound now people are taking the stories that haven't been told from the Star Wars universe and they're recreating them using AI and for the fans they're becoming quite

“popular on YouTube Star Wars stories untold is I think the biggest one it's getting millions of”

views per video already and I think that's really the future is letting the customer base

pay a licensing fee or pay a fee maybe rent software or maybe base on the output and let them be creative with the characters let them make their own stories and you could be in a unique position to empower that well on a person I think like if you find like a model that works for like the IP owners but then also kind of enable like the super like creative

“customization skills I think that's great yeah I mean like I mean like well for myself like I”

when I read a book or whatever like watch the movie I'd like so many like I'd use how it could be done differently or this could have happened right now yeah this is like this is so nice that you can actually enable people to visualize these ideas yeah it's going to be incredible continued success with it you have an office in San Francisco you're hiring people yeah we do yeah we raised a bunch of money we raised a bunch of money we just crossed on with people

we're hiring in Germany and then San Francisco fantastic who are you looking for what's the

right type of person the right type of skill yeah on the one hand we are always looking for

researchers who have experienced in large scale model training experience in diffusion model training slow matching training we're looking for engineers who want to be working with the customers to you know develop these like customized physically eye solutions or for example with like IPO and I like develop these models jointly with them we are looking for engineers who have experience in just like large scale compute in front managing devs and making sure that the

training runs runs smoothly that we maximize our and if you and all that and we're looking for people who have interest in you know like getting the technology out there yeah in the hands of people the the the for deployment of this there's just so many great ideas and so many great

“partners for you I think you're going to with the open swore specifically you know it seems like”

the corporates really want to have some additional level of control but they also need the front your models or your proprietary ones for some of those refined features so I think you have a very bright on this and yeah all right continue success thank you so much for doing this show thank you so much pleasure I'm doing all of you I'm doing all of you you you you you you you you you you you you you you you you you you