Why one AI model won't rule the world: Fireworks AI CEO

Do you have a big concern about how fast AI is moving?

Ithink it is a concern. We still have a lot to do to bring the rest of the world along and this global movement right now is quite imbalanced. How the Silicon Valley think the world is and how the rest of the world think the world is, is still quite different. Um, so.

Lynn, welcome to the show. It's so great to uh see you. I've been following all the really great work you guys have been doing over at Fireworks AI. Take us a little bit inside of the company and what you're working on.

Oh, thanks for having me. Um, it's great that um we have seen this huge momentum across the industry in our um business, how our business going is a reflection of how the industry is going.Where uh we heavily focus on investing in specialized intelligence. A specialized intelligence is a strategy, deeply believe in the expertise each, every single company is bringing to the world, and that expertise reflected in their unique product taste, judgment and design. And obviously, that unique part is being reflecting the private data they own as a proprietary IP.And we build towards enable all of them to convert their proprietary IP into their own intelligence, customized for their business. And that's where we're seeing a huge movement across the industry for various different critical business essential reasons across like startups.Digital natives, enterprises, they're all doingthat.

What inspired you to, to co-found the company?

Ithink we are extremely impact driven group of people. By the way, we started a company with 7 co-founders. That's a very big funding team. And we all see the same vision that AI is becoming um the most critical pillar moving the economy in the future. And we have been building AI from infrastructure level to modeling level.In the past 10 years. And uh we have seen how to move hyperscales, multiple hyperscalers through technology, through product, through business.And we think this can be amplified by moving the whole entire industry in terms of impact. That's why we decided to start, start fireworks.

You know, you just recently raised, I believe, $1.5 billion I mean, $17.5 billion evaluation. Congrats to you and the team. You know, I know, I'm sure it was a big milestone for you all. You know, what, what was next? You know, when they, when they valued your company in $17.5 billion what were those investors thinking could come next from a platform like yours or a company like yours?

We're just at the beginning of this revolution.And we think model is the new form of database.In the past, databases are storing other container holding uh mission critical data for the powering the product decision making, powering a lot of kind of internal business transformation.And now a lot of that has been absorbed inside the model itself. And that's where kind of we work with so many company, they are not satisfied by using a general purpose model, where it absorbed the data for solving common tasks. They're not satisfied by that because their mission is to solve a unique problem.in a special way, exceptionally well. And that knowledge, and that his judgment cannot be captured by a general purpose intelligence model, and they really want to hone in to um to build their knowledge into this special model they own. And as critical as intelligence is their critical supply chain,That's a part for the sake of control of the IP and the control of cost business operation is essential, is essential because we heard so much concern about one, having product market fit, but scaling to bankruptcy, by losing, completely losing control of their cogs business operation. 2,is having the product market fit, but just becoming a marketing engine for the rest of the industry to copy and catch up because it's very easy to use a coding agent to replicate anyone's product. It's very easy toTo build a similar kind of product on top of the same black box API. So then you think the mode, the IP has been significantly shift from all the product hold a deep part, very thick proprietary um differentiation. And because probably in the past, it takes multiple quarters or even multiple years to get to kind of the well oil stage with a very small talent pool and all the intelligence were contained in the product.And now the the dimension changed because, because of coding agent, um, one person with brilliant idea has no idea how to code whatsoever, can create implement the product ideas and scaled quickly.At the same time, uh, they can just build on top of um the most intelligence model to get this going, but that is not the end of the game, because that doesn't mean it's a doable business because ideas are easy to create, easy to implement, easy to realize, easy to copy. So now, the companies who really want to build a durable business had to think about long term. What is a long term IP? The long term IP actually lies deeply in the customer engagement.Where they deeply understand the customer feedback, then the next phase of, of important features to build, the next phase of important value to deliver, and that flying wheel is actually the mode. So that flying wheel usually manifests itself in uh in the product data. From

an outsider looking in, we hear the words AI agents, and we think they're just really super smart already. But how smart are they really?

They're as smart as your data.So, um, so first of all, here's the industry trend macro, right, it doesn't matter whether closed or open model, the model quality has passed a threshold.It passed the threshold, it can solve many, many day to day tasks, especially as knowledge workers, and knowledge worker by definition, are deeply specialized. So look at our past history, right? So we, we start to like, we as a community evolve and build a vibrant economy in various different form and shape because we specialize. And and now our special.ization goes into a deep bushy tree. And that's where the knowledge worker space is going through deep disruption, deep disruption as in, I can see orders of magnitude, productivity boost. And the other interesting thing is the knowledge worker specialization may get reversed a little bit. I don't believe it will get fully reversed back into one person will be able to do everything.I do not believe in that ideology. But uh at the macro level, the degree of specialization may get reversed a little bit, because we, uh, I can see, you know, uh very smart people or a knowledge worker with the same knowledge base intelligence and the training, they probably can solve a little bit more problems all at once.So, for example, I see the collapsing of a product engineer and the researchers, and many companies, they build AI native product. I start to see either researcher playing a product manager or product engineering role, or product engineer, product manager start to play a researcher role, they start to kind of also tune models, customize models, and that is a collapsing of decision making.Uh, because intelligence has a lot to do with how you assess, how to evaluate, how to define the outcome, and defining outcome of a product is a product manager's role, or product engineer's role. And, uh, but it it is a necessity towards be able to train a model and hill climb towards a goal.So all these functions, roles, goals are collapsing, and then I start to see, you know, one type of people start to play multiple roles, but with that said, it is far stretched out to say, assume one person, one very smart.% were trained, it will play all possible roles in the world. I think it's just too much a jump. And I do not believe in that, you know, it's going to be realized in the

future. What role does knowledge workers, what, what role will they play?In this AI future?

I think they will play a significant part because, um, first of all, knowledge worker, they are very cutting edge, they are very adaptive to new technologies. Um, we have so many companies, they build on top of our various different extremely creative application built on top of us, either theKnowledge workers themselves start to innovate and adopt all sorts of agents and build their own agent even based on the data, the knowledge base they have, and they start to even like train models as domain experts, where they do not have deep AIana start to kind of pick up the, the tools and shovels and do things themselves. At the same time,Um, they also, uh, realized the value of thinking ahead and of kind of building a much vibrant business business using new technology, not limiting to, um, to the old way of doing things. I I view them as the most cutting edge and most creative group of people.

How far along do you think, you know, companies are in their, their AI transformations?

I think we're just at the beginning and to me, I think this technology will be kind of the most defining technology for us as a whole entire community.

You think we're headed towardsYou know, large companies being really run by 50 humans and there's thousands of agents. Is that where all this is headed, do you

think? I don't think we're gonna.Moving to a community operating like robots. I think what defines us.Homo sapien is personal touch, is the curiosity, is theCrazy pursuit of curiosity and adventure that defines who weare.

Do you have a big concern about how fast AI is moving?

I think it is a concern.It is concerned because we don't know what we don't know. The velocity of how air moving.It's not anything we have seen before.I think that the closest analogy is the Industrial revolution, but it took a long time, it took a few decades, it took a few decades.Um, for, for the community to kind of expand, um, consolidate, managing all kinds of fallouts, and the new form, adapting to new form of economy, uh, in, in that process, there are ups and downs. And there's enough timeline to absorb all of that. What we're seeing is very fast moving. But I will also see in reality, the pace of adoption of AI.Um, is heavily concentrated. There's a sharp differentiation across um Silicon Valley, and the far end of the world, right? So, uh, the, the, the difference is very sharp, actually. Um, I think we still have a lot to do to bring the rest of the world along and, and, and, and move them forward. And I think it's a global movement.Um, and this global movement right now is quite imbalanced. How the Silicon Valley think the world is and how the rest of the world think the world is, is still quite different. Um,so

is there a way, is there a way to build, like how do you build trust in this technology?

I think a lot of this trust is,Because it's new, the trust building part is is new, and people are naturally.skeptical about new things. Uh, the new things like especially around what are the implication. And I remember in the old days when TV comes, right, so we're like, this is what we're we're gonna stare at this box all the time and get information and the kind of the pictures are moving. This is crazy. Uh and it's,It's gonna, it's going to poison our brain or something, right? So, and now it's, you know, we we move way beyond that, right? We we're staring at different kind of screen, maybe too much. But, but it shape inform and shape our digital life, right? So, um, and we adapt to it as human beings. So I have, first of all, I have a huge confidence in how adaptive we are.I also have a lot of confidence in the checks and balance of how, because of diversified perspective and how we can have productive debate um about as a community, how we manage this new technology, and how we learn together, and how um we make the right balanced decision and regulation around this. Uh, so it's going to be by and large,A big community based exercise to move the needle forward. Uh, I think, right now, I'm happy that government start to kind of get more engaged to get into kind of get more educated.Um, about this technology, about the, you know, how it's going to move the economy, how it's going to have impact to our day to day life, the same as how the industry is moving on AI we're at the beginning, in terms of, um, you know, the policies, the regulation, we're also at the very infant state step of um defining that, um.But I have a lot of confidence as many challenges we are facing in the past, um, the COVID, the pandemic, it's kind of very deeply shake shaking our community, um, as well as many other things in the past, we adapt and will prevail, will prevail this one as well.

As you, you know, think about growing fireworks, AI, what's the biggest challenge you're upagainst?

I think we're fundamentally part of a big puzzle solving exercise.The puzzle solving exercise is the whole entire supply chain, the whole entire economy, the financial,Practicesis being challenged to handle a super linear demand.All AI.Uh, the system we have built in the past.is not designed for this pace of growth. So, efficiency is a new focus because if we can driveOne order, 2 orders of magnitude of efficiency from energy to the value of productivity or intelligence generated on top, then we can really address this huge imbalance between the energy constraint and other supply constraint to the demand.And and the power of the economy in a unique way. So, that's where we deeply rooted in specialized intelligence again, and we believe that's a solution to address this huge imbalance between supply and demand.

You know, you're really talking about uh uh something that we hear, we've heard a lot this year from Nvidia founder Xu Jianxi Wong, you know, where the demand is so strong and the supply is so constrained. Do you see any moment on the, the near-term horizon where thatBecomes better balanced or or demand for as long as you could see for all things AI is likely toremain strong.

I don't think the supply constraint is going to change any time because uh it goes deeply into the whole entire system and changing a system is going to be hard, but I'm very confident there are a lot of creative solutions uh emerge along the way.Because the incentive of changing that reward of changing that is so strong, but it would take some time, where we add value in the middle is to drive productivity efficiency to extreme. So here I'll give you a few examples, right? We talk a lot about token economy, but token economy is not as simple as token price. Token price is meaningless.Because at the end, ultimately, we are measuring to solve a particular task of a knowledge worker, for example, how much energy is going to consume.And uh we want to minimize the energy as little as possible. Um, and, and here, uh, it matters if a model is very verbose, and a verbose model is going to consume a lot of energy. Um, if a model is generating accurate, precise results, so, uh, the agent just need to run once and get a really ideal outcome.Versus the model and the provider create a less accurate result, an agent has to tweak all sorts of things and go rounds and rounds to get a great result and then by that, it, uh, it drain 10 times more energy. So the quality matters. Quality is efficiency, quality is token economics.The verboseness is token economics and along the way, there are a lot of smart um ar um infrastructure design to cash uh a lot more aggressively to avoid recomputation.To be able to rally um scattered resource across the globe instead of always going to a very chunky uh data center is a way to drive token. So there's so many different ways and what we are investing heavily in our R&D is to combine all this technology and all different way of perspective of reducing the energy.For solving a very complex task and drive the maximization of that difference.

You know, are people just going to have to get comfortable with having a data center somewhere near them? Actually,

I think we're going to reverse direction. So before, because the previous era of mobile first, right before AI first is a mobile first, um, and that's where application.aggressively move from desktop mainframe to a cloud to mobile devices. It makes sense because then they can access it anywhere anytime and that drives a lot of edge compute, as in the latency really matters. Um and the closer the competition is to your mobile, um, you know, your tower is going to yield really good um end user experience.But when it comes to G AI, it's by and large, obviously, there's multimodal, where images, videos is going to be network bank with hungry, but by and large, I think a lot of application lives on tax base, right? So the LLMs in tax, in comparison is tiny, it is a very small. So in essence, tax is very energy efficient, but it's also a network, networking bandwidth efficient.Um, so, so then the delay, if you kind of, hey, um, a model living um in,Australia, um, it's kind of generate intelligent tokens of a task where the user is in the US that is not too bad, actually, um, because usually people tolerate some kind of delay. Hey, I want to get a response in 2 seconds. And a 100 millisecond delay is nothing in comparison.Uh, so, but what does it bring is it bringing access to energy globally, right, instead of just constraint on the regional places, obviously, the rest of the um trust related question we need to solve is, hey, now,Kind of your data may go to, um, you know, Australia, for example, from US and then do we, do we trust that, right? So, I think more and more enterprises or end user are building comfort to, you know, access much more efficient andToken generation of value add from thatpoint.

Before I let you go, if we're having this conversation a year from now, what does your business look like? Well, have you raised more capital? Are you part of a larger company? Are you public? Like what's the big vision here?

I think we're definitely aspire to um to become a public company at some point.And we definitely see in the, in the next one year, we are going to continue our journey um of this super linear uh demand creation and um and empowering a lot more creative companies across startups. They are leading the charge of building amazing product applications to digital native and enter.Enprise transformation into efforts. We have a lot of conviction to help all this company to shine and continue to be leaders uh in their segment, and many leaders creating new segment. That is amazing movement that we want to continue to support them.

Lynn, good to see you. Thanks for making time for me. I know you're super busy. Look forward to continuing to follow your journey. Uh, we'll talk to you soon.

Thank you for having me.

All

right,

that's it for the latest episode of Power Players.