B (2:40)

First of all, fun fact about fireworks. We started with seven co founders and very big funding team. After three and a half years we are all working in the company pushing really cutting edge technologies. We know each other from Meta. That's the time when we joined Meta is finishing mobile first transition, moving its application from the Facebook, the messenger from desktop to mobile. It was a huge bet at that time. If you remember, that's the time I was having my first iPhone. My first app on iPhone was a flashlight. So that's how kind of early it was.

A (3:22)

It's a long time ago.

B (3:23)

It was a long time ago and that was a really successful transition. And interesting thing is it significantly pushed product engagement to the next level. And founder generated a lot of data, interesting data, and that data became the fuel of AI. We all joined during that time and we were bootstrapping AI infrastructure from ground up. When there's no AI hardware, everything's running on cpu, we were running tiny ML algorithms on cpu. There's no AI software, there's no AI team. So it has been fun, almost like working as a startup Founded by Meta. Inside Meta, we build AI infrastructure from ground up, powering both massive training and the massive inference at the same time. We also built Pytorch. Pytorch is now the dominating AI framework and took us many, many years to get to today's stage. Especially now when it comes to gen AI, almost all models are written in Pytorch and deployed in production In Pytorch. We started company because we saw after a few years working at Meta, the whole entire industry is also moving to AI first. Many other companies reach out to Pytorch team asking us for help because they want to do AI first transition. But there's no AI hardware, there's no AI software, there's no AI team and we know exactly how to solve that problem for the industry and that's why we get started.

A (4:54)

So explain the product, who is the customer that you're selling into right now and what does the product do for them?

B (5:00)

We sell to so called AI practitioners. So the definition is very interesting. It started as the startup founders, the CTOs, the machine learning engineering team, and now the definition of AI practitioner is actually much broader. So that's kind of really fun. Part of development, especially starting this year as we heavily focus on developing tuned models and fast and cost efficient inference of those tuned models to get to frontier level quality, speed and cost. The quality becomes so good we expand our ideal customer profile beyond the tech savvy, tech focused people to many other professionals. For example, our head of fairness use Fireworks to massage her spreadsheet and do fairness forecasting and planning. Our legal team.

A (6:03)

So who does it compete with? Is it just for like running a frontier model and I would just replace Claude or I would replace my usage of Perplexity or Groq or OpenAI with this as a rank and file employee at an enterprise company or is it for, you know, more on the developer side and APIs?

B (6:23)

There's multiple reasons. First of all, when you hit the product market fit and to scale your product to massive scale, millions of consumers or billions of consumers, then you need to become real time in your interactiveness, especially in the agentic world. And second is to scale quickly to all your customers without bankruptcy, without running to bankruptcy. That's literally a problem today. So we help them control their cost, but also get to real time while getting to frontier quality. So that's kind of where we operate, is we focus.

A (6:59)

So if somebody like Uber or DoorDash, some of your companies, if they want to test and start using Frontier models, well, they're using frontier models from the large language model companies, the proprietary ones. But if they want to do open source ones, if they want to try Quinn, if they want to try Kimi, if they want to try Gamma, they would use Fireworks.

B (7:18)

Yeah. There are also many other great open model like Nemotron. Yeah, and Mistral and many others.

A (7:24)

And what's their motivation? Is it to save money? Is it to not give their data to OpenAI and Sam Altman, which obviously has some reputation issues at the current moment that are quite acute. And I keep hearing from enterprises like, ooh, we're very concerned about our privacy, we're very concerned about proprietary data going into say an OpenAI and then they would use that for reinforcement learning. Obviously they say they won't, but people still have that concern.

B (7:49)

Yeah, it's kind of deeper than, you know, just not trusting a company. So I think let's take a look at the world's data.

C (7:59)

Right.

B (8:00)

Think about intelligence as a reflection of the collection of data. The data that goes into foundation models are the public Internet and the labeling companies label data. That's the data at essence into the foundation models. It is actually a very small fraction of wars data, less than 5%, majority of wars data, more than 95% are the private data locked inside application, locked inside enterprises. And those private data will never get shared with venture labs to train a foundation model because those private data are our individual company's ip.

A (8:43)

Yeah.

B (8:44)

So then if you look at that, it's very interesting. Majority of the intelligence reflecting the private data were not being activated represented in the foundation model. So we want to work on the next phase of frontier intelligence which is activating the private data.

A (9:01)

Got it.

B (9:01)

Customize and tune the model and to make it smarter.

A (9:04)

And do they do that customization themselves or are they relying on you to send, you know, for deployed, you know, enterprise developers into their enterprise and like work with them on fine tuning? And do they have their own instance then on their own hardware, your hardware. How does that go down at early

B (9:21)

phase of a new technology adoption? It goes through waves. The first adopters are all hackers, power users. So they want to control everything because they are capable of diving to very deep part of the technical stack. So we have the lowest level of abstraction for them to control everything. For example, Cursor recently released Composer 2 and that's their own tuned model on an open model and they use our lowest level abstraction because they are very capable of driving all kind of tweaking and adjustment. And the next level of objection is we package a lot of defaults but we still give meaningful parameters for a machine learning team who are comfortable in massaging the training process and move forward. So it's kind of media level and the highest level are the application developers. They do not have deep AI expertise where they can easily tune by many things being automated. So those three level abstractions are the platform interaction we provide and we expect adoption will go over, over time more and more towards the upper level.

A (10:34)

Let me introduce our second guest today and then we'll get into all the news of the week. Demi Goa is here and she is the co founder and CEO of Pika Agents for Creative. They launched back in, they launched AI selves back in February of 2026. This was a persistent digital twin that would learn my voice, my style, my personality for better or worse. And then you've evolved a bit since then. Tell us a little bit about what you're building Demi, and who is the customer and why are they buying it.

C (11:06)

So we started a company around like 2, 11 more 2 or 20 years ago. So you know something that's I'm always very excited about. You know, personally I'm always engineer. So I've been starting coding for like since elementary school. But at the same time Like, I also, you know, really wish I can be an artist and really, you know, and the goal for the company, like how do we help more people to be able to stuff or to enable more creativity from like more regular people. We started a company by building, you know, a web tool for, to help people to create videos. And we explored different direction and then we realized like for the web tool, it's still mostly for prosumers and then it's still very hard for a lot. Our team members cannot even use the web tool to create videos. It's really hard to prom and to edit and blah blah. And we explore like different interfaces, we explore wobble and such. And then we found out that actually maybe the best interface and the most accessible interface for people to actually create stuff, whether it's create videos or even create animation, it's through a humanized agent. More like a humanized interface where you don't actually need to like, need to learn all the complicated UI or all the editing skill or all the prompting skill. You just do it just like you're talking to a human, which is like, more like a humanized agent. So that's why we recently pivoted to focus on using a more humanized agent for people to be able to create stuff, to create videos. Create. Whether it's like social media videos or films or even potentially create, you know, like design, poster design or, or like short film, short drama or like vlog or whatever.

A (13:00)

So a creative person, they're an influencer on Instagram, they want to start creating versions of themselves or content. It knows their Persona and it just makes it. Do we have a video here? I understand we might have a, a demo video and we could show it and then you could talk over it.

C (13:14)

I want you to clarify a little bit. It doesn't have to be, I know there are the like the exist out maybe a little confusing. It doesn't have to be like a digital twin or it doesn't have the user influencer. It's really just anyone who need to create stuff, create videos, create multimedia artifacts. So yeah, we can play the video. So this is an example of like, you know, and yeah, so yeah, the motivation we talk about, which is really like, you know, what is the best interface for people to, to have more, more accessible way for them to create, create stuff. Right? To, to create videos, create image, what, whatever creation it is, we, we realized actually the best interface we try web, try mobile, the best interface is actually through a humanized agent, just like you're talking to or like it's like a creative assistant. Right. So we're like telling the agent what to do. The agent will brief you, telling a human what what you want and they'll brief for you. You can even like it's more human. It' that you can even like video call the agent. Like a humanized agent which will even like screenshot what it's working on. It's like really conversational. Really just like you're working with a human and yeah, so that's so you

A (14:31)

can create an image. I can see like some of the Personas have different animation styles. One of them is a little anime, one of them is a little Disney Pixar. You create that avatar, you create that representation of either yourself or just any Persona. And then it goes on a zoom call or books a meeting and is like a customer support agent, a sales agent or is it for entertainment? Or you're just going to let the users decide what they use these agents for?

C (14:55)

Yeah, our primary focus is about creation. That's kind of where we started. So it's really into me. It's more like evolution of the interface of how people should be able to create stuff. So the primary use case too about creating various blogs. We're creating like short film, we're creating like, like dancing video or creating singing video or creating short drama or creating ads like, like you know, maybe even beyond image videos or maybe like even like any marketing materials like like poster design or, or slides or whatever. But it's, it's really just for we, we just realized that the best interface for people to create stuff, it's not actually a web tool. It's actually through maybe not even like a commonly, like a commonly perceived agent, but really through like a humanized agent. Like you feel like you're talking to a person. You can even video call the person. You can tell them what to do. You can screenshot what you're working on. They can also screenshot what you're working on. And through that, just like you're like hiring another human.

A (16:04)

All right, so let's get through our docket. The first story we have here is a paper that came out. This is the AI layout trap paper and here's your summary. If AI displaces human workers faster than the economy can reabsorb them, it risks eroding the very consumer demand firms depend on. We show that knowing this is not enough for firms to stop it. In a competitive task based model, demand externalities trap rational firms in an automation arms race, displacing workers well beyond what is collectively optimal. The resulting loss harms both workers and firm owners. More competition and better AI amplify the excess. Wage adjustments and free entry cannot eliminate it. Neither can capital income taxes, worker equity participation, universal basic income, et cetera, et cetera. So essentially a prisoner's dilemma. If we don't cooperate, then everybody loses. And the proposed fix these UPenn and Boston University academics came up with was a robo tax charge companies for the demand they're destroying. And so I guess they call this a Pigovian automation tack. A proposed policy designated to address the negative externalities of labor displacement caused by artificial. Artificial intelligence. Artificial intelligence and automation. And of course there's been a lot of debate about this issue because block cut half their employees. People say, you know, Dorsey, Jack Dorsey said it was about AI. Other people said it was just a convenient excuse to do it. Doing more with less and that we had maybe over hiring. Historically in the tech industry, both of these things might actually be true. And CFOs are probably saying cuts will be nine times bigger than what's being reported. That's from a Fortune take HBR's take. Companies are firing based on AI's potential, not what you can actually do yet. So Lynn, what are you seeing through your customers and how they're deploying? Is it creating more jobs and are people or is your belief that we're going to see more people displaced or we're going to see more people hired because people are going to start more companies, people are going to find more problems to solve and they'll be inspired by the technology. This is the debate of 2026, clearly.

B (18:28)

Absolutely, yes. So one thing that's clear to us is we have seen a boom of ideation to production never have been faster in the past. A prototype that can reach into many hands of the people to test product market fit. It will take multiple quarters and I'll just take multiple days. And that empower. A lot of people have great design taste and the great ideas they can access so many tools for them to realize their dream quickly without a large team, without having the organizational skill set to hire people to organize them to deliver tasks together. Because now you can orchestrate a flip agent to solve those problems collaboratively. So I think the creativity is off the chart right now. And we do see a lot of a ton of startup with brilliant ideas. So the most fun operating in our space is we are the AI index. We see a lot of great use cases, crazy ideas emerge built on top of fireworks and a lot of experimentation. And many of them just take off like, you know, escape the gravity and reach a lot of people quickly. So, so that's the fun part. With that said, our company, we only have 150 people and like you mentioned earlier, we're tens of trillion tokens a day. We probably that it's extremely high traffic. Just as a reference, based on our understanding we may be mistaken that traffic is bigger than OpenAI's API traffic. So we reached that as a reflection of the direction we're heading towards is we happily operate on customized model that is not off the shelf model inference at all. It is using private data to tune the model and then bring the best quality customized focus for your application and speed and cost to your application. So that's what we see obviously that is what we see on startup land. We have many enterprise customers, especially digital natives. Interestingly those digital natives are, they were startup decades ago, one decade, two decades ago. And they are the survivors, they are the winners winning over the competition. They have a huge amount of traffic and they have a lot of people and they're rethinking how to convert themselves into AI Native Co. Again, how to reinvent themselves. So as part of process of this reinvention, it's not just a product reinvention, it's also organizational reinvention because in order to survive this wave of heavy competition, they have to revive their velocity. And that velocity got buried through layers and layers of hierarchy in the organization. And many companies start to kind of reduce that layer, especially mid management where for example, from my experience in the past there were like hierarchy of binary tree organization. That kind of probably doesn't make any sense these days. And now we're talking about a manager not just be responsible for seven to 10 people in a layer, probably they can handle 20 to 50 people because

A (22:06)

so this is, I think such a key point Lynn, is if these tools are so great, one manager can manage two, three times as many people because all of the check ins, all of the knowledge is already surfaced by AI. And you don't have to be a warden or a babysitter of employees. Which is what, let's face it, middle management was in a lot of these organizations. You probably saw that acutely, Lynn, at Facebook. Yeah, there was a middle management layer that was coordinating and writing notes and doing meetings and doing standups in your experience. Now that's all automated, correct? By AI?

B (22:43)

Yeah. So I think by and large it can be automated because as you said, information is much more discoverable and because of that, so imagine A manager's job in the past was collecting information and relay that and make sure everyone's aligned. But these can be much more efficiently done. For example, at Fireworks, we don't do a lot of one on ones because a lot of context is shared with a group of people. If we make decision, we just quickly make a decision by checking and discuss that and done. Because we are very chatty on Slack, a lot of information is on Slack and people, we all share similar context and also we can summarize what's happening on Slack per individual's needs. So that is extremely streamlined that help us stay on high velocity decision making execution. And we're very flat organization as a startup. But interestingly I also saw many large public companies start to transition their organization structure. They're flattening out because that smooth out information flow up and down. And the reason they can do that is the information discovery. Data discovery is much easier. It's not just limiting to individuals work and people management. It's also about data scientists, for example, product analytics. Right. So how do you reach a product decision and you just understand a lot of data and in the past we have large layers and layers of people trying to do that. But if you mcq your data access and then you can write an agent, you can build an agent to be able to extract that information and and summarize it and synchronize across different fronts and get very precise analysis assessment of the health of the business.

A (24:42)

And interestingly, Lyn, I don't know if you saw the report, but Rulof, bofa and Jack did a podcast this week and they did a blog post two weeks ago, Ruloff from Sequoia and Jack Dorsey is basically saying he wants all 6,000 Block employees to report directly to him and that with AI he can manage that. That's a fantastical vision, but it's not ridiculous. If you're a hardworking CEO and you work 12 hours a day and every 30 minutes you have 24 segments a day to work. 24 segments of 20 workers. You start doing the math on that, 400 workers per segment, you know, information, you know, coming in, you could actually make it work. You could actually get through in 10 days, 15 days, every single worker's output. Demi, what's your take on, you know, the combination of new structures and then what's happening in old organizations and then maybe even how you're running your organization as an AI first organization?

C (25:44)

What I really believe the human value in the future will come from the judgment and taste or Personality or like unique identity of the human. I think it's, I actually feel like it's ultimately not even necessarily the most productive thing, but what it really value for humans. For example, The only thing that human matters is you have your own judgment because it's your company, for example, or what you're doing. And then you can just have one AI agent that really reflects your judgment, your taste, your thinking style and your decision making style or, or whatever personality. And then that agent can just either can just do everything for, execute everything for you, whether through orchestra, a lot of agents together, or through just working on its own. So I think in the future that's kind of what Jack Dorsey is doing is he's trying to really amplifying his own judgment and taste for the company. And I think in the future we'll see more than that. And the value is more about there will be more people who are going to create their own unique agent that really reflects their taste and judgment. And then there will be more freelancers

A (27:05)

in the future, many more freelancers who can just come in and be that human in the loop and maybe even work with an agent. Here's a clip of Peter from Openclaw talking about you need to have a human right now in the loop with agents. That's definitely been my firsthand experience because it's, it doesn't have taste perfectly yet, but here it is.

D (27:28)

It can create code and run all night and then you have like the ultimate slop. Because what those agents don't really do yet is have taste. They are really, they are spiky smart and they're really good at things. But if you don't navigate them well, if you don't have a vision of what you're going to build, it's going to be slop. If you don't ask the right questions, it's still going to be slop. When I start a project, I have like this very rough idea what it could be. And as I built it and as I play with it and as I dare as they feel it, my vision gets more clear and I get like, I try out things, some things don't work. And I evolve my idea into what it will become. And that's like my next prompt depends on what I see and feel and think about the current state of the project. Yeah, but if you, if you try to put everything into respect up front, you miss this kind of like human machine loop. And then, and then I don't know how something good can come out without having having feelings in the loop almost like like taste.

A (28:36)

Right. So Demi, I guess that is a critically important thing for when you're running your company and people are creating agents, doing everything up front seems like the right idea. Hey, I want to create this website, I want to create this piece of content. I need it to be entertaining, I want it to have these themes, I want it to accomplish these tasks. But you can only kind of front load so much in the prompt or so much in the skill and it's really about I think iterating on the skill over and over and over again and pointing the agent in the right direction so that it does the task for you. Is that your experience as well? Demi?

C (29:13)

Yeah, the reason we transition from like a web interface or prompting or mobile to agent is because we realize in the future what matter is you train your agent to have your own taste basically. So by using your agent or like iterating with your agent by giving feedback agent, you're gradually making your agent to have your own unique taste and your agent can do a lot more things than you. So it's almost like the goal is not about like okay you're creating individual website but actually about like you're using your agent maybe whether it's just telling the agent what the taste should be or just the throughout like using your agent to create website you're like training the agent to understand your taste and then your agent will in the future when your agent have your taste it can just create infinite websites. Right. So I think what matters is really just to that's why we transition from the just like prompt interface to more like agent interface.

A (30:07)

Lyn this seems to be I think a hard fought lesson. I thought setting up my open claw and then saying I want to solve this specific test on reporting was a one time skill. I create it, I run the cron job, I never have to touch it again. It turns out that's wrong. You have to be a bit more interactive and a lot of times I'm finding these agents drift from what I told them explicitly to do and because they're so sycophantic or they're very inconsistent, I don't find they have the consistency. So maybe some thoughts on when these will anticipate a little bit better and when they'll be more consistent. Lynn, in your mind.

B (30:46)

So I, I, I have a, I have a lot of conversation with my daughter, she's in high school about AI. So it's very interesting. I'm surprised. I feel like her generation will be AI native. They grow up With AI around them, all the tools, they were just kind of be very deeply embedded in using those tools natively. But I'm surprised. Her reaction is they're like Gen Zs are cynical about AI because they don't feel it's authentic, they don't feel it's creative. They feel those are all the repetitive, mundane things if you use AI is not being thought highly of. For example, they have this school magazine and if you use it to generate pictures they like, it's better if you draw it yourself and kind of shows authenticity. So I feel like AI is getting really, really good actually. If you look at general images, it's actually really good because it's trained, using the human intelligence is able to simulate that. But I feel like we as a human, as part of our soul, we need the creativity as a satisfaction fundamentally. I'm wondering if there's a future that we can integrate deeply. Always have human in the loop to be part of the soul of this new world we're building. I think that would be fascinating. And I like the demo from Demi the New Pika agents. I feel like if I can embed my part of the creativity into these humanoids agent that can represent me and give a little bit surprise here and there because there's a little bit of impromptu of how we react and so on and. And give it that. It would be really fun.

A (33:00)

Yeah, yeah, yeah, go ahead.

C (33:04)

Oh yeah. Really a great thing about like the creativity and also the human in the loop perspective. And that's kind of what we are really like leaning towards. Because I really think right now the reason people. People are really treating. I think we need to change our mental model about what AI is. I think a lot of people in Silicon Valley are really treating agents as a tool and as a productivity tool and to people choose who they work with, not only for productivity, but also for, you know, they just enjoy working with this person. So that's why we really feel like the mental model of agent should not be a tool, but actually just be another human. And then you should really treat like having your own agent as having your own child. So it's a constant feedback and iteration. So humans should always be in a loop. It's not like you one click, everything done. It's more about, okay, you're raising your child, you're like constantly teaching it and gradually you will grow up and you can do things for you. And it's more this iteration process with children. And it's also this emotional attachment beyond productivity.

A (34:09)

Yeah. And if we just Recap this previous segment. Lyn the take on how organizations are changing. Well, even if people get laid off and organizations become smaller and flatter, there's been this Cambrian explosion, I guess, which happened hundreds of millions of years ago where life just suddenly emerged and in a very violent many people competing because of oxygen and this perfect ecological soup that occurred. I think that's kind of what's happening now with startups or just individuals being able to say, hey, I could just make a startup in a weekend, test it next week. As you were pointing out, it used to be a two or three quarter journey to get your product out and tested and you know, beta testers and then, you know, close beta and then open beta, etc. Now you're talking about doing that in three or four days, potentially with just one or two people. That means many more ideas. And you know, we could see instead of 10, 20,000 startups getting funded every year, you could maybe have 100, 200,000 or a million or 2 million created. And the startup then doesn't have to be venture backable. Lyn. It could just be enough to pay somebody's salary or maybe even half of their Facebook salary. If they were making $300,000 at Facebook they might be completely happy to make 150,000. But live at a ski resort half the year and live by the beach half a year. Check out a bit. Yellen.

B (35:42)

Yeah, so that's what I mean is I feel ultimately I see if this continue to play out. I'm actually pretty optimistic about the future is we human will focus on the most creative part because that's how we evolve over thousands, so many like tens of thousands of years is we find creative way to form new structure, new invention and new technology, new economics all the time. And every time there's some wave powering it for us to leap forward. But we always resort back to being creative. We never stand still and circling on the same spot. I feel like the AI is actually if we do it right, it's going to power the next level of creativity. I couldn't imagine what will come out of it. We may be able to do space exploration much faster. We may be able to reach the planet light years away much sooner. I think our limitations, our imagination is our limitation. So I'm very optimistic towards that direction.

A (37:05)

But continually.

B (37:08)

Yeah. In the face of doing that, even step by step, before we were like hobbyist is hobbyist. They tinker. And just as a hobby. Right. But now we are like hobbyist could be the next phase of inventors because a weekend project could really hit something fundamental, fundamentally deep and hit our biggest pain point because everyone is now empowered to create, to think, to imagine with their tools. So that's the part I really believe is the combination of human creativity. I do not believe that I'll be replaced. And our brain is only a few percentage activated. Also at the same time, there's a lot more we can derive out of it. So we actually, as a platform, well, frontier inference platform. As a platform, we pay a lot more attention to hobbyists these days because we give them the tools for them to test, invent, experiment. And once they hit something interesting, they can quickly scale and they do not need to worry about scaling because scaling is a complex system problem, especially scaling your deployment of your tune model. First of all, how you tune the model, it involves a large amount of GPU fleet or a lot of kind of data tinkling, a lot of parameters to set and experiment. And once we want to scale, you want to scale globally across many regions with low latency, high reliability, robustness and a lot of people using it and so on.

A (38:57)

How are those open source models doing on a, you know, how many months behind in human years? Not dog years or AI years, just actual human years. Because everybody knows dog years are seven years to a human year. I would say a month is seven months each month is a year in AI time right now. So are they six months behind or three months behind? Nine months hot behind. Opus 4.6, whatever.

B (39:29)

Yeah, so it's a very interesting race that's happening right now. But I want to put them against each other. I just, to us, we, we saw like every two weeks or sometime every week there's a new model released. Open, close, doesn't matter. And they're oscillating with each other on the leaderboard. Whenever things are oscillating, it's a clear signal they're converging. So last year, I think one big sticker shock is last year's Deep Seq v3 release. And that's the first time open model got very close to Frontier Labs model. And since then, Frontier Labs continue to leap forward and then openmodel catching up. So it's kind of. But overall we see the quality gaps start to converge. And that's great news for application who has a lot of data that means they tune. If they can activate their private data, they can actually forward and really get to the next level of intelligence, specific design for the application. And we see since last year we start to see a lot of adoption of this customization especially led by more frontier thinkers, more application. They are pushing the boundaries so they're

A (40:47)

getting closer, but still frontier models are ahead. So how far ahead I guess is the question I'm asking. And still people. So I get it's oscillating. There is definitely convergence, but one group seems to be leading, that's the frontier proprietary models. How far behind are the open source models in your estimation? And then depending on how many months or years you think they're behind. What is the roadmap to this convergence?

B (41:15)

For complex tasks, for example, most complicated, if you want to generate the most intelligent agent to build a distributed system, P2P system for eventual consistency or whatever, that is really hard. So I think frontier model is absolutely leading there. There are probably six month ahead, six months, one year ahead. But for many day to day tasks, for example managing spreadsheet, managing calendar, having a router or classifier of doing some kind of routing logic or doing writing improvement or doing some good, very good document processing and so on, there are many tasks varying with less complexity, but actually cover a lot of our day to day. They're very close, I would say even on par to some degree.

A (42:22)

Okay, so simple tasks, couple of months on par, complex tasks, six to 12 months behind. I think would be what I'm interpreting from your comments there, which I think is super helpful for folks because you do have this issue Demi, with the cost of these models and people with agents seem to very quickly and I don't know if this is your lived experience right now I would say but

B (42:46)

this is off the shelf model quality. But with your private data we have so many cases after tuning on that complex task it can be on par or even better than Frontier Labs model.

A (42:59)

Demi, what's your experience with your customers when they start embracing this technology? What is their usage profile in terms of token usage and how does it change when somebody goes from using a search engine or a researcher and just asking one question versus hey, I'm going to give you a complex task, I'm going to give you an agentic kind of existence, dare I use the word existence. But we're going to will this agent, this replicant to exist and I'm going to interact with it every day. How does that change? How does that token usage change on

C (43:35)

a multiple for sure, I think we do see that really depends on what kind of models we're using. We're trying to maybe get more user, more customization because it will really value a lot based on to Ling's point, if it's more open source, much cheaper versus the Frontier Lab. It's more expensive. It could be up to 10k. We have users who have 10k per month or something.

A (43:59)

120k a year to empower their agent. And do they make back 120k a year or are they just. They're so. Their business is so great they don't mind losing 120 to be on the cutting edge.

C (44:12)

I think the reason is we should really not compete like really like treat. Obviously they're the agent. When the model is better, where it's like open the source, we're like with tuning what Ling says you will probably be cheaper over time, but. But also just generally I feel like we should not really treat agent as a tool. That. Okay, you cannot pay like 10k per month for a tool. But if you really treat agent as a human. So whether it's your child, your race or. Or your employee, right. Your assistant, you're like, you're. It's like, it's like either you hire a junior creative assistant or you use for agent or like, you know, you. Either you raise your own pet or you use agent. So it's like that should be the comparison. People spend a lot, it costs a lot to have a junior creator assistant or it might cost. People are willing to pay a lot for their path. So it's not really a fair comparison of a comparison with agent was a tool. But we should really think about agent as a human or a life form. And you're keeping it, whether it's hiring or raising it in technology.

A (45:18)

There are some folks who were early on in the 70s and 80s who very much looked at what are the hobbyists doing. Because the hobbyists quickly become, if they figure something out, the entrepreneurs. And then they would look for the toys, the tinkering and the toys. And those would become the tools and the services in the future. And so a lot of what we're working on right now feels like toys. It feels like tinkering, it feels like a hobbyist. And creatives and agents certainly fills that kind of description. But then they become actual real tools and services in the future. And it's a great way to either build a company or invest money in companies is to just watch what are people tinkering with, what are the tools. And that's why I became obsessed with OpenClaw the moment I saw it. That's why I became obsessed with Tao and the Bittensor subnets or even app development and content in the form of apps back during that revolution. Lyn when you were at Facebook it just was like oh, the flashlight, oh, a little flappy birds. Became a critical economy very quickly in the future. Lyn, question for you. I believe you went to school in Shanghai. Yeah, so that's the second I think city in terms of the density of large language models, Beijing is number one still in terms of where all the computer scientists are working on large language models then maybe Beijing and Shenzhen I guess because of the hardware footprint. What is the movement there like? I'm sure you have friends and family and colleagues that you went to school with. What is their perception of the race in terms of open source frontier models and just overall what this technology will do for society?

B (47:15)

My understanding is, I think because of population, consumer facing products are much more popular and much more polished. So many of the models has been heavily used to power newer consumer facing experiences. For example, I heard open cloud at China is becoming very popular and the various different company offer open cloud as hosted managed service and get a lot of traction. So that's kind of, I think just because of the population advantage and the focus is more like consumer facing product and how to bake those new way of surfacing user experiences through LM and more like multimodal genii models, not just LMS are the kind of primary focus.

A (48:17)

Interesting. And what is the general perception of AI? We just saw some studies come out. People in America obviously have a very negative view of AI. We had this New Yorker story come out last week about Sam Altman. It was quite negative. And then we had these horrible attacks on his home. A Molotov cocktail and then somebody shot it. This is absolutely terrible. But there's an incredibly negative perception of AI in America today. And then in China my understanding is it's extremely positive. People think that this is going to be amazing for society. Is that true and why in your estimation, Lynn?

B (48:56)

I do not get. I guess I'm living in the Silicon Valley bubble. I feel extreme optimism here about AI. But my observation is I feel like China is probably more advanced on the robotic side because it's day to day, it's in delivery, food delivery, all robots and people can order ice cream. It got delivered in a couple of minutes via robots. And that's kind of the limit test of how fast things are is you order ice cream, it's still frozen.

A (49:32)

Is your ice cream melting or not? It's the ice cream test.

B (49:35)

Right. So that's my sense. Obviously robotics habit depends on AI and it's just one reflection of the consumer facing focus. And yeah, again on the AI sentiment I shared earlier even from my daughter. So she told me Gen Z is cynical. I mean she also grew up in Singapore Valley. Maybe the Silicon Valley teenagers are cynical about AI and they want to preserve the authenticity of creativity from their heart as their identity. I do think that's a human need we need to address. We shouldn't lose that along the way. So I think it's more a societal homework for us to figure out together is as we co evolve with this new technology and how to focus on maximize our fundamental needs as a human to express ourselves over time. So I think that's just we're at such an early stage of this revolution and we're going to figure it out. I'm very optimistic about it.

C (50:53)

Ling suing because that's actually something that we're trying to really solve is really like what we really believe is I think Silicon Valley is really crazy about how productive the agent are and how there's like replacing work, blah blah blah. But something we really care about is like I think probably people, a lot of people are very scared that okay, the agent are replacing my job, blah blah blah. Right. And something that we really care about is like enable people to create their own unique agent like their own agent that they could customize. And agents should not feel like a tool. The cold hearted thing, the productivity thing, you should really feel like a child you race and it should really feel like. You own it and it's your own agent and it should really be working for you and it should really be something beyond just productivity and there's also the emotional connection because like you know, it's like there's like identity creation of it and it's, I think it's very important actually beyond the productivity value of the agent. Obviously every agent will be productive, has utility but like the personality or the human perspective and identity creation agent I think is also very important and I think it's almost a sense of, it's almost a form of self expression in some sense where you could imagine like okay like now Picasso's art really inspire us. But maybe a hundred years later your own agent which was unique personality and unique voice and face really inspired people like really touched people's heart. 100 years later among other corporate agents.

A (52:28)

Yeah, I don't know if you guys saw this. This was one of the Shenzhen comes out with all these incredible toys. And again back to toys becoming tools. This one is called Bubble Pal and I don't know if you've seen it but it kind of went viral Last year, I just gave the link to the team. They'll pull it up here. Essentially, you put this device, it's like a little Alexa sort of digital assistant, but you put it on any toy, and then the toy gets a Persona and it gets a personality. And then here's a quick video. We'll play with a little bit of sound, but it's definitely having a lot more playfulness. And then beauty filters and AI filters and AI chatbots and making short videos with AI becoming incredibly popular in China with consumers. So here it is. You pick any. Here's a little girl getting left out of playing soccer with her friends, and then a little girl who wants to be an artist. And you bring your toy, whatever toy your kid has. Like, I could give this to one of my daughters in there, Teddy bear. And then I just strap this on, and it creates a Persona for the toy, and it becomes like a best friend, an imaginary best friend. But it actually does talk back to you. I'm not sure if this is dystopian or just incredibly engaging. Your thoughts, Lynn?

B (53:56)

I think we all. We all, when we were little, we all have an imaginary friend somewhere. I think it's fulfilling our need of having this imaginary friend in some form. I could see my. You know, if it goes back 10, 15 years, my kids would love playing with it.

A (54:17)

Pretty interesting stuff. I think it relates a lot to what you're doing at Pica, right? Is this concept of creating some kind of relationship with digital beings?

C (54:30)

Yeah, yeah, I think it's for sure. Like, obviously, I mean, agent, it's not just a toy in a sense that, like, everyone will have agent in the future.

B (54:37)

Right.

C (54:38)

It's. We all, like, kind of feel like Agent is not only the next interface for web and mobile, but actually replacing next computer. Because Agent, like, has storage and compute and everything. Right. So everyone have agent, but it's really about, like, okay, do you want a generic agent, or do you want your own agent aging that has more human, like, and more emotion, has more emotional beyond just productivity.

A (54:57)

Yeah. And here's just to put a final end cap on it, here's the US Perceptions of AI societal impact. These are the general public versus experts. In other words, the bubble. You responded, you. You referenced, Lynn. We're living in the AI bubble. We see all these incredible things as possible. Here's us adults versus AI experts. Here's the chart. And if we look at medical care, if you're an expert, if you're in the industry, you think it's 84%. Is your percentage that AI will have a positive impact over the next 20 years. The public 44% K through 12 education elementary school 61% people in our industry think it's going to have a great impact. 24% of people who are not in the industry just us adults helping people do their jobs. This one is the one we all see in the economy. 73% and 69% of people in the industry think it'll help people do their jobs and will help the economy. But 23% in 21 for the public and for personal relationships. People do not think this is going to help personal relationships at all. 22% of AI experts 1 in 5 of us believe this will help your relationship. 7% of the public think so. We have some alignment here for the elections and for news and relationships. Pretty good consensus. It's not going to go well. But for everything else we do have this massive swing. Your thoughts, Lynn? Just generally looking at this.

B (56:33)

Yeah.

A (56:33)

And maybe how do we change the perception I guess is the bigger issue if we accept that this is the case.

B (56:40)

So for example, in medical care, right. So we have customer. We have many customers from medical care. They are doing amazing things. I now see Dr. And by default they will ask me if they got my permission to turn on recording because they can use medical scriber to automatically take notes. We power another medical company that they are doing building preventive care software as in as I going through seeing my doctor and then they quickly can pull out the software can pull out my medical history and suggest to Dr. What kind of preventive check I need to have. And then they will give me a list. I will work on that. The idea is great because putting a more preventive either exercise or tracks or exams and whatnot is going to help make us healthier and reduce also the medical bill. So they built their business out of that. But the barrier is actually software integration with hundreds of patients record systems where there's no standard how to integrate. And they want to roll out this great preventive care idea. They are going to hit the system record barrier. However, AI is here to the rescue because there's one universal interface that is screen. Screen is one universal interface. So it's very interesting, right? In the past few decades we have made the perfect interaction between human and computer HCI through various different kind of software development in order for us to kind of make the long horizon human driven workflow more automated. Then the next phase will be automate hci. So that has already started piloting through authentic creativity from people in your Chart the first role of AI expert working in medical care system. That's why I think that makes sense. They are very optimistic about the impact of AI. I just feel the more people they see the outcome of AI the more we share the stories. Share like day to day stories. What are the benefits and we can get out of the end product.

A (59:11)

I think our industry is doing a terrible job of this. We have to show all the wins for people. Yeah Demi. And maybe a little less of the fear that everybody's losing their jobs and it's the end of the world. There's so much joy and fun and cost savings and life extension that could come from this. So where do you think the industry is getting it right? Where do you think the industry needs to improve?

C (59:35)

Demi My feeling is there is like really a differentiation between like pure like maybe like productivity AI and versus like some AI that's more like diverse or like more uniquely yours in some sense which is like maybe not. Not even have to be the most productive but. But it's uniquely your own agent and maybe like more human, more creative emotion. Whatever it is. I think what Silicon Valley is really care about is like okay, is the agent productive? Is it smart enough or is it going to solve this productive issue? The enterprise work or developer developer issue. And I think something people like and then like that really the narrative really scares like regular people because it feels like okay. Like you know, like okay agents are really like AI really replacing my job. And I think something that is really missing is how do we use AI to enable people to create their own AI and to empower themselves instead of how the angle of like okay using your AI to replace empower yourself instead of AI is replaced for productivity and really like you know, really like really good. And I actually think what. What is like to your question about it is like I feel like there will be two type agent like Nika mentioned. One is more like your own agent which is more diverse. Is really for diversity. Like really really unique. Your unique taste and judgment even though it might not be the smartest judgment and taste and personality and versus another agent which is like purely for productivity and actually for productive agent. I actually think it should have no personality. It should be maybe you should not have human judgment because human are not the most productive people. I think because human has laws like human has emotions, people has like strong personality and human are has ego or whatever. Right. Like human. And I think like so there is this two kind of AI in my opinion like the to have to achieve the best productivity you should have no human component to it. And then there's the other AI which is more human and more personality. It's really for diversity. It's really important for human because there is value for human that we might not just want the most productive thing, we might want something powerless and make us feel good. So something I think we're like kind of ignoring right now.

A (62:06)

Perfect segue. Meta introduced just last week. Musespark M U S E M S L's first model purpose built to prioritize people. So this is I guess the first language model that Facebook has produced that's closed sourced. This is I guess their frontier model. Interestingly when you look at the specific examples they gave and I had a perplexity here, do a little summary of this incredibly long announcement. They gave a bunch of what I would call domestic Mr. Mom or mom focused benefits. So planning a family trip to Florida, one agent drafts the itinerary. Another compares Orlando versus the Florida Keys. A third finds kid friendly activities all in parallel. So they're not doing here's the enterprise and how we kill jobs. They're doing here's how we make whoever the homemaker is mom or dad. Here's how they're going to do. Ask health questions and take a picture of all the snacks they keep going to travel all the snacks at an airport shelf and say which one has the most protein in it or shopping mode. Get ideas of what to wear, how to style a room or what to buy for somebody. Drawing from inspiration across Meta's apps. And this is using all of the information on Instagram, WhatsApp and Facebook. Really interesting. I think very human forward approach. So your thoughts Lynn on two things. One, Meta going from open source to closed source here. Disappointing. Interesting. What are your thoughts and how good were their open source models and then two, their human first approach and what you think the public reaction to it should be.

B (64:09)

Yeah, yeah, we're hoping Meta will open source their Avocado model for sure. But I also understand like their main business is a consumer product and this next generation model need to power their consumer product. And Meta is really good at product model co design, Very, very good at that. They have been doing that for decade and I'm pretty sure this model powering WhatsApp and Messenger and all this consumer facing product is not a static model. It's going to evolve over time. It's going to be smarter and learn from Meta's private data and so on. So that's kind of the typical pattern. I think all application developers in the future, all AI native application developers should be able to adopt and today there's not such kind of tool. It's a vacuum in the space. It should be as easy as product analytics. It should be as easy as everyone's doing a B testing. And our dream is kind of give this tool to everyone's hand and it should be turned on by default. So on my side I really hope there's a lot more US based open models. We work very closely with Nvidia. Nvidia is putting a lot of focus on driving the Nemotron model. Very very bullish about that and I think they're constantly improving the model quality and there are many other providers on the USO on the Tell me about Neo Tron.

A (65:41)

It's Nemotron N E O T R O N Not a great name. Nemo is a Pixar movie or Disney movie, I'm not sure. And Tron is a Disney movie as well. Nemo, Tron, two Disney movies put together. Very strange name. But how good is it and are people using it yet?

B (66:05)

Lynn so I think it's getting better and better. So one thing about model training is you cannot jump ahead of time. You have to go through steps. You have to train from a small size model, generate synthetic data, train medium sized model, use the synthetic data, blend with whatever data you have and then generate more synthetic data. So there are process you have to go through and then during that process and with very complex mid training post training of SFT RL and then the model becomes a lot better. So imagine the process is like first you need to build a base iq like we human when we're born we come with iq that doesn't change over time. But that IQ accumulation takes time of learning basics and then once you have base IQ then we human take years to be trained as a domain expert.

C (67:04)

Right?

B (67:05)

Whether we are a dentist or heart surgeon or a lawyer across the board it's kind of we take years to get to certain specialty and that's RR training and RL training is basically narrow down the focus and kind of really specializing certain areas solving certain kind of problems really well. So that just takes time and I think all these models, as long as there is substantial amount of research effort behind that, a lot of GPUs, it's a matter of time they will get there.

A (67:43)

What do you think about the data drought Lynn? I'm curious. And then we'll go to you Demi just on where you're sourcing data to make better models.

B (67:51)

There will be a Combination of everyone sourcing from public Internet. There's no secret sauce, everyone's kind of

A (67:57)

safe and there's nothing left. Right. There's nothing like that.

B (68:00)

And again a lot of data is locked inside of vertical private applications. You just don't have access to and only those data owner has access to. And then there's labeling companies label data. Almost all companies, all foundation frontier labs use the same labeling companies. So then what could be differentiating is how they generate inside the data and the mixture, a blend of the data could be differentiation and how the and even the algorithm of training, pre training, mid training, post training are converging. I think there's. Unless I think we are overdue across the industry, across the research community to have a new model architecture transformer is way older usually there every three years there's a leap of new model architecture. I think this is probably seven years in the making, more than seven years. So there could be likely that there is a leap forward new model architecture. It will learn knowledge in completely different way then this phase of convergence model convergence we may have a breakthrough. So far we haven't seen that yet. So if we stay in current course I think, I think whoever owns data, owns unique amount of data is gonna, it's gonna bring up a new level of intelligence.

A (69:26)

Demi, your thoughts on where people are getting data from and has that reached your customers yet? Where they say hey I wanna bring some data or I need you to go find me this data or I wanna hire a data labeling data sourcing company. These dark pools of data seem to be the next frontier, seem to be the next opportunity. We have an investment in Micro One. We had Ollie on this week in AI in one of the first pilot episodes. So maybe you could talk a little bit about where you think the next data is coming from.

C (69:55)

Yeah, I actually feel like kind of what we really believe in the future is people own their like potentially maybe they own their own data in some sense which is like the one they're using agent is technically like they're like kind of feeding their data to their own agent and then they have the ownership of their own agent. Right. Because what's important in the future that differentiate different people is actually your own taste and judgment. Right. So and then you want to use AI to really amplify your judgment taste to maximize your value. So it's almost like you're trying to train your AI with your own taste which can be feedback, iteration or it can be data and to train your AI to be like to the taste whatever Judgment you have, right. That's like kind of like a data feeding process. And then it could be like, okay, I'm a CEO. That's kind of what Mark Zuckerberg or like Ray Dally or a lot of people are doing, which is like they train agent, AI, that's like agent, that's like themselves. And then they use it to talk to rejectors. You kind of also want to do, right, to use your own judgment and to really amplify by managing the company. Right. And then probably they want to own their own data and then because that's like the value accrued in the future. So people who are probably in the future when we don't even work, everyone just going to train their own agent and to inject your own data or case and judgment to your agent and then the agent will help you to do stuff. So in some sense that is also maybe in the future people are owning their private data in some sense. From that perspective.

A (71:31)

Okay, I think that's a great place to start. Two amazing guests this week on the program. I know you're both, or I'm assuming you're both hiring and growing quickly. Lynn, tell us a little bit about what positions you're hiring for and how folks can get in touch with you directly if they're a genius or where they can go to join the fireworks team.

B (71:53)

Yeah, we are into a kind of exceptionally fast growth phase. So we're hiring across the board and from product engineering to marketing, sales and GNA across the board, even recruiting team. We're hiring.

A (72:14)

So recruiters are back to work. We have this whole recruiter apocalypse where everybody thought they didn't need recruiters and now it's back.

B (72:23)

Yeah, but we love people who love using AI tools because that does bring us extra amount of productivity and we aspire to be the smallest business, biggest company in the future and it's possible. And if you are dreaming big and you want to drive the maximum amount of creativity, it doesn't matter which domain you work on. Please talk to us and we love to have a conversation with you.

A (72:51)

Awesome. Okay, Demi, who are you hiring for? Who do you need? How do they get in touch?

C (72:56)

Yeah, for sure. We're still a very small team. We're hiring, I would say designer engineers and researchers and yeah, so we're trying to really build for people to create their own unique agent that's more humanized through the multi model research we're doing. We're hiring more designer who is really interested in the vision to help people work, people to build agent that has your unique taste, not like a general agent. We are hiring like engineers who are interested in the backend problem and more agent problem and we're hiring researcher who are interested in multimodal research. Yeah.

A (73:39)

Great. Awesome. This has been another amazing episode of this week in AI. We drop every Wednesday. We record on Tuesday, we drop on Wednesday. Please go ahead and visit us this week in AI AI and sign up with your email and you'll get our our research. We're doing proprietary research on the AI space and who's getting funded and who are the next unicorns in the space. On the show you get to meet the companies that have already reached that unicorn status and have vibrant businesses in the research department that we've created this week in AI AI. Go sign up for the email and you'll find about about the next cohort of companies that are just three, four, five people and that are growing and building interesting banks. We'll see you all next time. Bye Bye.