With Model OS Founder Fernando Jia: The Doom of the “Rent Model” Era
18VC Podcast · Episode 9 · June 20, 2026
Guests: Fernando Jia
Source: https://www.18-vc.com/podcast/modelos-fernando

All right. Good morning, good afternoon, and good evening to our audiences. And today, we are very excited to welcome Fernando, the founder and CEO of Model OS, and their company is called Intelligence Cubed. And Model OS is building a survival infrastructure for AI application layer. It helps AI apps move from prompt wrappers into companies that own their own models, evaluations, routing systems, workflows, and the intelligence loop.

Fernando, thank you so much for joining us today on the episode nine of the 18VC podcast. And for those who may be unfamiliar with our show, we are a student-run podcast featuring student and young founders, operators, and investors who are betting on the youngest and smartest minds. And the host today will be me, Philip. And I met Fernando through a mutual friend's introduction, and what immediately caught my attention was what Model OS was actually doing.

They are not building another AI app, but they are trying to solve a deeper infrastructure problem as the AI world moves more from one general model to more specialized models. And people would need to build an operation layer that helps developers discover and monetize, and customize their own models. And let's move on to your story and the story of Model OS. Then, can you share with us about your path prior to founding-...

Model OS? I see from your previous experience, you've been doing investment banking-... in Citic Securities and CDEEs. And then you move on to McKinsey, then at Y Combinator China, which is also known as Miracle Plus. And then you joined, I think, as the accelerator of Solana-... and doing your startup there. Can you just walk us through how did you end up in the place where you decided to found your own startups?

Yeah. Currently, we believe AI is so popular and someone just called there is a AI bubble currently. So we're trying to do something realistic and something can really be at the production level, and then can really be used and add some real value to the technology industry, instead of just a Ponzi game or just a focus on the storytelling or make some data farming. So, yeah.

So we believe the infrastructure for machine learning models are really important nowadays. So previously, AI applications are so popular, but most of them are still some wrappers. And currently, the web coding is so popular, and everyone actually can easily make a prototype within two days hackathon. So their mode is really low and actually some model companies like OpenAI or Anthropic, they have launched their GPS with a cloud scale, and then somehow, well, absorb the AI application layer companies.

So we're trying to build this kind of infrastructure to empower the AI application layer companies to equip with deeper infra to the co-creation of their models and build their own models instead of renting models from other side. And we believe in Silicon Valley, the most important stuff is that build something as a milestone on the way of the technology growth, especially for some specific category. And we believe that currently, the renting model era is ending now, and there will be more opportunities as the co-creation to the own model. And the other parts of the vertical AI solution will be firstly, the AI model companies.

They will take most of the market size of the vertical AI applications. And also, there are some current existing business and they are using AI to empower their current existing business like Amazon, Alibaba or Expedia. So we believe the middle layer, that is the operational layer, is the most important stuff as a whole development history for AI industry. So that's why we're focused on that.

And for my previous experiences, they are helping me to be equipped with better acumen or will be more confident because we know how the game works and how the current situations are within the whole innovation entrepreneurship community. Okay, yeah. Thank you for walking us through the overall vision of what Model OS is doing and a little bit about your previous experience. So I was actually curious, you have been working in Miracle Plus before I think you start this company.

From that experience, I assume you have kind of helped them source a lot of startups or helped them evaluate-... a lot of startups. How does that experience actually help you? Or maybe not in a good way or either a bad way when you are doing your own startups, yourself, right? You're turning from a observer or evaluator position to a founder yourself.

How does that experience help? Yeah. So yeah, this kind of experience has brought really a lot of acumen experience to me. So the most important part is the ability for industrial research. The industrial research is so-... important for reasons to invest in companies. So, for example, about 10 years ago, when Elon Musk and also Microsoft invest in OpenAI.

At that time, OpenAI was just an NPO. And after such a long time, even go to 2022, when GPT-3.5 launch, Transformer and AIGC era started by OpenAI. Similarly, like Anthropic, at that time, the RPO is not so good at some storytelling and business than STEM. So at that time, they are emphasized so much on the AI security, and most of the investors, they are totally not interested in AI security.

And based on purely business method to evaluate their business model, it doesn't make sense at all. But at that time when SBF from FTX invest Anthropic, they invest more than 500 million at that time. And then, this they are almost save FTX, but unfortunately, at the end, it doesn't, because they sell it earlier by some regulatory committee. And also Cursor at the time.

So based on these experiences, I realized that most important stuff is to select the right category at the edge or the frontier of the whole technical industry. So this experience brings really a lot from industrial research and align with some our current existing research skills and also some publications or experiment from the academia. So, mix them together, it will be easily not only for annual investment to invest in companies in a right category at the agile frontier of technical industry, but also select the right category for us to start our own startup. Yeah, definitely.

I can see that maybe you have talked to many founders doing their own AI applications and- Yes... from that experience, you may realize that maybe they don't really have the modes. Maybe one day they will be just covered by a update from these large language model providers. Yes. Right? Yes. So if you're trying to be a founder and you were investor before, and when there is a idea and you realize that you have evaluates that or you have seen this kind of deck two or three years ago, you will definitely pass this idea.

Okay. So the most important stuff is not to select the right stuff specifically, but to select something or not do- Ah.... with your time or money. Yeah, that's actually a good observation there. And then at what point... Is there a specific moment, or a specific event, or a specific person that make you decided that, "I should do this startup at this time," at 2024, I suppose, right?

Yeah. Actually, we publish our idea and finalize that as the day on January 20th, 2025. So yeah, as you said, the guy, I believe, is Wen-Hung Liang, and the event is the launch of DeepSeek. Yeah. So, before that, for co-creation, for machinery models, not only for pre-training, but also for fine-tuning, the cost of our computes are really high. And the data also requires a lot for training a model.

So, DeepSeek has improved the reinforcement learning. So after that, they just use $6 million to train their model, while OpenAI are spending tens of billions cost to train their GPT. Anthropic has spent not so much, but still a larger amount comparing to DeepSeek. And also they are with similar performance at that time, maybe somehow lower at some niche scenario.

But overall, it proved the open source machinery model can be with a similar or even better performance than closed source ones. Before that, they're just a Llama, and Llama is not so intelligent. So, it's really hard for the fine-tuning because for the full fine-tuning, for data with an updates, only open source machinery model can be fully fine-tuned. So after that, we have finalized our idea to start Model OS.

Hmm, interesting. So basically, the DeepSeek's emergence kind of inspire you that actually open source model now has the similar level of capability as these closed source models. So maybe these AI application builders, they will be willing to use these open source models because they're probably cheaper to use, and they- Yes... have a similar level of capability.

And then you found that maybe fine-tuning on them is actually a good business opportunities. Yes. Before that, the AI application layer companies were some one-person company. They cannot even fine-tune their models. It cost really a lot. Even currently, OpenAI have their FDE is just like the party B companies, just like the consulting or auditing firms.

The model company send some guy to the AI application companies, and they need to pay the salary of them, and also the whole package fee for this kind of FDE works. And for OpenAI, one person annual salary will be- A million? above a million. Yeah. So this kind of cost will be paid by the AI application companies, there are really a lot, because the valuation for AI application companies are much lower than- Yeah...

infra companies. So there'll be a barrier to entry. So, yeah. But after that, the cost has been lowered down, and also we're trying to build a automatic workflow for the fine-tuning with low-code or even no-code. So that brings really a lot of value for this kind of AI application companies. The biggest opportunity would be the lower cost, and then lower the barrier for that.

Definitely. And also, we want to move on a little bit about, I remember you mentioned your view on the AI applications. You said most of them are just wrappers of APIs of large language models, and they may get... There's a extinction event of them- Yes... maybe in the future. Can you share a little bit more about this problem you're observing, and how do you help them with your product?

Yeah. This statement is not just a story. This statement, it comes from some history for AIGC era. So when we're looking back at 2023 when GPU store launch, really a lot of wrapper died at the time. And the whole primary market recover, even as later in 2024. So, yeah. And also for this year, when Cloud CoWork, CloudCo, and also CloudSkills, and also OpenCloud are so popular.

The skills has replaced really a lot of AI application companies as well. So yeah. So that is the outcome. And also theoretically, when we analyze that, there are three parts makes them died. First would be the model companies are building more for the application. The model companies currently is not only produce models, but also they are also produce in line with some data application based on their models.

They want to monetize, yeah. Yeah. So that's where GPAS, where CloudScale come from. And the second would be the API attacks. So AI application companies, they are doing a business. They are using third-party APIs and prompt tuning them, and customize them, and build a user interface and experiences, and then provide to their customers. So that'd be a really clear workflow and also business model for them.

So what causes most is the API token price. And somehow, the model companies are monopoly that. They need to select a cheaper model with low experiences or a more expensive model with higher experiences or security. But currently, web coding is so popular, so there is not so much models to build a application prototype. It's just like a two-day hackathons-...

workload. So that means that firstly, the AI application companies, they are seeing their data, their user feedback, and also their innovation to model companies to post-training the model company's model. Okay. And then they'll use them to replace them. Yeah. And also when the model get better, even though some model companies will not cover that, there will be some competitors using the same API from them.

Yeah. So it will be like a red ocean currently. So their net profits will be much lower or even not positive. So, yeah. And the third reason is that for prompt tuning, you only change the temperature and memory. But for full fine-tuning, you change the bottom of this model, everything for the data weights and updates. So they're really clear that the performance will be much more different, right?

So that means when you use more, when users more feedback and data, or your updated innovation bring in the prompt wrappers, the model will not be developed at that time. But for fine-tuning, it can be developed. So that's why we're doing that. Yeah. Yeah. And then we can move on to your product, right? I do remember, from my understanding, Model OS have several actually quite complex or comprehensive services you're providing.

I think only one side you're providing data cleaning, like fine-tuning-... and then evaluation. And then on the other side, you provide a marketplace, right? Where the model trainers or model builders can share their models on the marketplace and some other guys who want to use it to build their own apps can use from there. And I remember there was another piece that is called, you're doing a, what's it called?

Like a- A market maker or what is it called, for both the compute and also maybe data- Matchmaker... yeah, matchmaker. Yeah. And how does these three parts come together, or how do they serve each other better than if you only do one of these business? Yeah. We're trying to do a Silicon Valley-style company that the foundry is only one part at the beginning.

But in our company, the most important stuff is the product at the production level. So when we are trying to build that, when we try to plug in some supply insight for model marketplace or some specifically usage side, we realize that most of the participants in this whole ecosystem are not so production level. So, OpenRoadher, Hugging Face, some other companies, maybe also our competitors. I cannot mention so much names, but they're not at so production level.

So if they plug in their services, and just only focus on the co-creation and the fine-tuning parts, maybe our product could be affected by them's-... quality. So that's why we're trying to build our own foundry-... and then make everything at a higher quality and then easier to run. So yeah. So that's why we're trying to build more, and fortunately, we have a good team.

I interviewed everyone, myself, and so we have really good track record for people's brand. And currently we have an initiative with Carnegie Mellon, and currently we have some researcher fellows, and all of them are post-doc or PhD from Stanford, Berkeley, Harvard, MIT, Carnegie Mellon, and also Caltech. These six schools. So, and also for the output, at the first half year, we have seven interns, and two of them go to Google, and two of them go to AWS, and two of them go to TikTok, and one of them go to Microsoft.

So we're keeping a high productivity team, so that's how we can also build our own functions and other parts of the product, and then they can provide better outputs and quality to our core function. That is the fine-tuning side. Hmm. Yeah, interesting. As you just mentioned, that your talents in your team includes people from Carnegie Mellon and also, I think, Caltech, and I do remember you mentioned you have a collaboration program-...

or initiative with CMU. How does that look like, and how does that help your business on your side? Yeah. Yeah. So we have raised a lot of compute grant at the beginning, so that is a total of more than 1.5 million, and we believe that the largest moat for this kind of model foundry companies or the model supply, we need unique high-quality machine learning models.

Definitely we can index it from public supply model that is more than 1.2 million currently on our platform. But there still will be some models developed by our own team. They're called proprietor models. So we're trying to bring more talent, that is machine learning engineer or research scientists who are model developer, to build their own model and then exclusively list on our platform.

That would be our unique moat. So yeah. But fortunately, based on my own experiences as a guest lecturer in Carnegie Mellon, and also our co-founder's background as a PhD admission officer in Carnegie Mellon as well, we have good community and relationship with alumni of top universities, and especially the research lab for computer science. So yeah.

So we have developed a new moat to collaborate with them that is a fellowship and initiative, so we're sponsoring them computes, and they're developing their own models. We do not disturb their innovations and their idea, but generally we'll recruit some fellows in the broader machine learning subject area. And then they have their own jobs, they have their own commitment in academia, but they full-time develop their own models. And we're sponsoring them, and then they exclusively list their model on our platform.

So, that's how as a startup, we have to raise a lot of budget for recruit them as our employee, and then let them develop model. The outputs are similar. And we are trying to reserve more IP for them, and then we believe that can empower and incubate more talented machine learning model developer in the future. Not only for the business, but also for the social impact.

Hmm. Interesting. And I do remember you mentioned that people can list their models on your platform. And they can kind of earn money when they're- Yes... using the model. Yes. How do you achieve that? How do you build that? How do you make that possible? I do remember you mentioned a watermarking technology. Yes. And how does that work? Does it have any- any connections to the blockchain technology or whatever, because I do see you have been in the Solana accelerators there.

How was the connection there? Yeah. For Solana, as a reader and grant from Solana Foundation, there were some payment processes considering with X02 and we're at the first position for Solana X02 Hackathon last year. And we're not for Hackathon crypto, but we are using cryptography, and also some zero-knowledge tours for the recognition of the attribution for machine learning model training.

So, yeah. And currently, my GitHub link and repo are still on Solana's official Twitter account. And so how can we track this kind of attribution that... This is a huge problem. Because about three years ago, when GP Store launched, Sam just announced that there will be some revenue sharing for the creators who publish their GP apps on GP Store, but it doesn't work at the end.

So, definitely there will be some business-side problems from traditional finance mechanism in the US. It's so legacy and that is so old. So, for a co-creator, maybe they'll get their 1099. That is ridiculous for some potential creators who like this platform. So, maybe the stablecoin helps on that. Also, the tracking process are really hard. So that's how we developed the watermark mechanism.

This mechanism is watermark on the model space, and when someone fine-tune a model as a secondary or other kind of co-creators, the watermark on the model space can be easily recognized for the- similarity or citation rates. It's just like a Turnitin on chain, just when you using Turnitin to detect the similarity rates of some research papers. It can be also works as a similar method on our platform to recognize model co-creation for fine-tuning. So, we can get the similarity rates there, and then there'll be the attribution rates, and then it will reflect to a recommendation rates.

So when a tertiary creator fine-tune their own model, have their own user, and someone pay this kind of inference cost to them, and they'll be automatically route the revenue sharing based on the similarity of citation rates and then go to the original or secondary creators. That's really fair and sustainable, and then we can encourage and empower and incentivize more model developers to develop their own model and share their own models. Because these kind of models will be not just a voluntary list on Hugging Face or open source for free. But also, even though they open source that, they can get some revenue sharing-...

when their model are fine-tuned. So that's really important for some model developers. Not every one of them have so much resources or barrier to develop their model. Yeah, I can see that. And I also remember you mentioned to make a good ecosystem, I would say, for-... the AI developer community, it's super important that these people can get paid for their contributions, right?

Basically, through the system you're building, you're building an ecosystem or it's a economic system that makes it possible for more people to engage in the process and benefit from it. Yeah. And then we can move on a little bit further into your fine-tuning process there. I do remember you mentioned some people, maybe they don't need actually a fine-tuning.

They only need a RAG service. And some of them need a more deeper fine-tuning service. And how did you decide if one user only needs a RAG or if they need a more full, well-rounded fine-tuning? Yeah. We believe most of the people use RAG instead of the full fine-tuning currently are... Full fine-tuning, they're really a higher standard or entry for people to get full fine-tune.

It's really hard. Even some software developer from big tech companies, they are not always ability for fine-tuning. Only machine learning research scientists, they are able to do that. So, actually not everyone like RAG, but RAG is most of the people can do that. Yeah, it's convenient. Yeah. It's easier. Yeah. Yeah. So, that's a pinpoint we're trying to solve for the full fine-tuning.

So where there was a local or even no code experiences for the users and then it's not so hard and it's similarly simple for users for RAG. And then that'll be totally makes sense at this time. So, yeah. And also, Since we have developed these kind of functions, and our strategy is to give the users their own choice and their own flexibility to select even RAG or full fine-tuning on our platform to use.

We're not trying to assume which kind of user will use our RAG or full fine-tuning. We just provide it to them, and with better experiences, and they can select it themselves. And why we include RAG as well, beside the full fine-tuning, because we are not just doing research, we are doing a business. So, we're trying to cover more functions that is more- Frequently used, yeah...

used scenarios and then more usage on our platform. So that's why we also cover that. Hmm, interesting. And I think there was one more thing that I would like to poke in this direction is that I do remember you mentioned, for example, because you are basically helping the AI application developers accumulate their data, right? Collect their data. And when, for example, the data accumulates to a certain level, maybe if the data is too little, it's not enough for fine-tuning.

And maybe after a certain threshold, there's enough data, or make it worth it to do the fine-tuning, right? Yeah. Is there a little bit more concrete example of, for example, if I am a developer developing, I don't know, my own app-... to what extent would you recommend them to go from, for example, a RAG to a fine-tuning or vice versa? Yeah. Definitely.

It's a incremental way. So, everyone know that more data means much better performance for the fine-tuned models comparing to the original foundation model. But most of the people ignoring the most important stuff, is the selection of the foundation model. You would like to do your niche application, which kind of model are you trying to select? So firstly, we're doing a broad coverage for our models applied publicly, and then we have some niche model supplying from our own team for proprietary models.

So, the coverage and selections will be really enough and even much easier for the whole selection coverage at this time. And the second part is routing. Most of the AI application layer companies, they are not with the knowledge and experience for selecting models. There are really a lot of models. Even our platform, there are over 1.2 million. Among the whole market, maybe there are around 2 million.

So it's really hard for them to evaluate or benchmark these models. Most of them just select a generic large language model who are listed at the top, as a consensus, or the appearance of the whole industry. And maybe that is not the best one-... but they are above the average. So yeah. So the selection is really important, and it will shape most of the productivity or the performance.

And, so we have our own router. We're not just as open router based on their price and capacity. We also have our second layer and third layer. Second layer is relevance between your prompt or your context to the model context, and the third layer is based on users. It's just like a recommendation system based on users' real usage and how much they have used and then also the frequency to benchmark how popular a model will be.

And this three-layer router not only helps the daily usage for the to C side, but also for, to B side. They will be more specifically and more confidently to select their foundation model. And then, even fewer data they are burning and the performance will be overall higher than not select the good foundation model, but with more data. And so how to select the RAG or the full fine-tuning, that is incrementally, I believe.

You can just simply use a RAG and then get more user, and with RAG helping you to collect more user data and feedback, and also from this whole experience as your innovation will be updated as well. And then, you can put all these three kind of stuffs in the full fine-tuning and then, definitely you'll be coupled with a better and more data and feedback and innovation at the same time. Okay, cool. And I do see there are different competitors out there.

For example, OpenRouter was doing the routing service. For example, Hugging Face was being like the marketplace, but I don't think they charge money-... or provide revenue to either side. And there was kind of LangChain, which does, I think, a little bit of these agent frameworks, something like that. And also these large language model providers, they do also provide their own kind of fine-tuning services.

How do you position yourself among them, and how do you see, would they one day decide to expand their business and become a direct competitor to you? Yeah. There are some competitors currently, and more specifically, the ones will be Thinking Machines Lab and Pioneer AI. And we believe the first step is that to realize procreation, especially fine-tuning for AI application companies, the modern innovation currently.

And there are only a few companies focused on that, and we believe that the most important step is production level. How fast can we deliver our product? And talented people, enough computer resources, and also good strategy and a good schedule for the product development. That is what we're focused on, and we believe our kind of demo we're in currently for the fine-tuning are fairly ahead of our competitors.

It can be really used, and we'll demo that around. And also for our foundational services like router or workflow or benchmark functions or agenting controlled execution parts, we have already with more than 2.5 million monthly active user. So, how fast you deliver your product is really important for this kind of specific niche competitors. And for others, like OpenRouter, believe they will be successful as a primary market.

The stability is the most important step for them, and by the way, you are trying to use them, you'll see the product is not so high priority for them. And for Hugging Face, they have a bad IPO, bad launch for their stocks, and currently, it caps people's imagination for the open source economy because their market cap is much lower. Yeah. And also for other competitor, we believe Civitai does really well.

But they are only focused on computer vision. But there will be some potential from them. But they are focused on a more niche part, and we believe, I believe they can do better and they have some potential, and let's keep an eye on that. And we also have some model supplying from their model, like the tool from them that have the open source there.

Hmm. Interesting. And then I would like to move the topic a little bit towards how do you guys actually make money? So I do remember the routing services, you actually are kind of free to use-... for the 2C customers, right? And I think your major revenue comes from, I would say, 2B business. Yes. Can you explain a little bit more which part of business are you currently making money, and which of them are probably more like a marketing funnel where you attract more users to onboard your system?

Yeah. We do have some incentives of free trial opportunities for some 2C user or not so deep layer of the user currently. So we have made some money from other parts. So we have four layers currently for our revenue engine. The first will be workspace subscription. If you're a small startup for applications or you're a one-person company or you're just an indie developer, you can have our workspace subscription, and this will include some basic limits of the dataset, models, evaluations, workflows, and also the deployment.

And the second part is a usage-based execution. It will charge per run, per model call, per workflow, per evaluation, or per fine-tuning job. For some professional developer, maybe they will not like to subscribe, and then we can go with a usage-based run. And also when someone exceeds the limits of their subscription coverage, they can also pay more for the usage-based execution.

And the third part will be enterprise runtime. It's just like some FDE jobs, like what OpenAI doing now. But OpenAI is still need someone as a party B companies to technical or model deployment consultant to the party A companies, and then they're doing that. But we're more providing some solutions, try to be more automatic to do that. Firstly, save the cost for the client.

Secondly, the client will be better control their product, and we're trying to deliver local or even local experiences. So, our people will not help them to do that. Instead, they can do them using their own their own team, their own innovation, and they can have a better control for the privacy for their data. And the fourth part will be the model economy layer.

So that'll be the attribution, lineage, benchmark APIs, more of the distribution and revenue sharing. Or we have the matchmaking mechanism, just as you said before, and that'll be the data and compute supplier after the scaling law broken, they need more demand, and the model developer and user need cheaper computer data for training process or inference process. And then we can charge some matchmaking fee there. And also where we can also sell our benchmark KPIs and also the full data set subscription for the benchmarks.

And also for the revenue sharing, we have our watermarking mechanism. And then, not only we are sharing with some co-creators, but we'll also charge some service fee on that for the revenue sharing. So that is a full revenue engine for us. And for some free trial or saving users, actually, they can be the first batch as a kickstart our users for our client as well.

So even though we're not charging them too much, but our client can charge them more, and our client will be really happy with that, and we can get more usage, and broader use cases from our client. Hmm. Interesting. And I would like to go back a little bit because I think, can you share with us a little bit more what type of data that, for example, a AI application collects that is actually useful for fine-tuning, and what kind of data that may be just not helpful?

Yeah. If you just purchase some data from third-party company, like Scale AI, right? Actually, the data maybe come from some uncompliant ways-... or from some data brokers, maybe from India and China or other kind of parts with lower quality. But what Scale AI do is that package them, and even some of that are not compliant, they have the best legal team in the whole AI industry.

And when the whole case ended, that cost two or three years, and the data are not so frontier or as a edge position, and there is no value or there is no need to continue the whole case, and then they succeed, and then they sell to some client companies. It's really hard to use that to train models. You cost really a lot, not only from your party, but also your human resources to put some efforts in, and you need to clean, label the data, and maybe this kind of data can only train some generic learning model. They're not so targeted.

They are not screened so much for you. Yeah. But for the most important data we're trying to bring is that, when some end user are use this kind of model, they'll pay, no matter by subscription or by usage. They need to pay that. So, since they need to pay that, they need some reimbursement, like a promotion or discount for that. And so, they will be also the co-creator for this model, even though they're not training, they're not bringing innovation, but they're committing their data for the model training process.

So, yeah. So that would be really easy, not only for the cost, to lower down their cost, but also for their self-commitment, for the whole AI industry development. And then, it will be easy to collect data, let them disclose or commit the data from their own dialogue or the conversation. And then, this kind of data will be more targeted, since some kind of model or niche models, and this kind of data supplying will need a lowered efforts for the labeling screening process.

And so that save really a lot of time and the budget from the AI application companies to bring this data in for the fine-tuning process. Yeah. And talking all about this business and technology, then we want to give a little bit more human touch of your team and yourself. So I do recognize that your team is consist of very strong, heavily academic background talents.

How's your role there? How do you make sure that this team of maybe more research-driven people, they are focusing on the right product, not only on the technology? Because sometimes they do have good technology, but if the product didn't generate revenue, that doesn't make sense, right? And how do you balance that? Yeah. Actually, the productivity of the talent or knowledge come primarily from their background, even in academia or in some big tech companies.

It depends on the age actually. So even though some of them are PhD, we are trying to find some graduates as an average student, 22, and directly pursue a PhD degree and graduate before 26. So, yeah. So at this kind of age, we believe the productivity more depends on the family. So if they have some children, no matter which kind of position or what kind of background they are, we need to be more work balanced to these kind of employees because they have their family.

Yeah. We cannot occupy so much time from the business. We're trying to be nice to our employees, but try to recruit some talent earlier. They are more energetic. There are not so much stuff from their families, and then, they'll be more productive. So, and also, currently for computer science industry, especially for machine learning, there are not so much gap from the academia and industry.

Mostly the Facebook Intelligence, like Word Labs, like LeQuinn's labs, and also, Andrew Ng, Google Brain- Yeah... and also Coursera, all of them come from academia. Yeah. And founded by researchers for machine learning. So, this kind of point are not so much we're careful and worried about. What we worry about it's we're trying to bring more, not physically young, but at least that they are young and they're men.

So then- Okay... they'll be more energetic. Yeah, definitely. And, then a little bit more about yourself. If you are to describe yourself in one word or a little bit more, rather than founder or builder or entrepreneur, what would you say? Yeah. Maybe a change maker. Hmm, interesting. We're trying to change really a lot of stuff currently. And, currently the whole economy globally are not so good, and most of the industry are red ocean currently.

So only change bring more opportunities for us. So, that's why we're trying to be the change maker. Hmm. That sounds super cool. And, we are kind of going towards the end of the episode, and I think I have two more questions for you. The first one is, looking into the future of Model OS, what are the next two or three milestones in the maybe one or two years' time?

Yeah. So, definitely the first will be the closure of this round fundraiser. And also, currently, our prediction for the secondary market is that tomorrow, that would be the largest IPO-... for human beings, for SpaceX. I also get some shares. I don't know if the audience get really lot from them as well. And it will bring really a lot of liquidity from the secondary market or even from the primary market.

And, we believe that, at the end of this month, there will be the midterms election, and also, the IPO for Anthropic and OpenAI. So after these three kind of giant companies get listed, there'll definitely be some locking period, but not so much, like half year or one year. The capitalists are trying to selling more from the secondary market, and maybe there'll be this some broken of the AI bubble. So, for two to three years, the most important stuff is that, get along with the trend of that-...

not only on technology and product, but also for the finance side. And then, we're trying to make our success for all of our partners, our investors, our team members, and also our friends as well. So, we're trying to be really fast for the listing process on the secondary market. And then, we believe that the absolute finance, the absolute money can incubate absolute edge technology.

So, that's how we're trying to be at the technology side and the finance side. Hmm. That's interesting. And last questions. For other either student founders or early age founders who are maybe feeling still uncertain about either if they should start their own startups or about their future, what do you want them to hear? Yeah. My advice may be more so different than other guys.

Although I have met some dropout students previously, but currently, I'm not recommending them to drop out. I have observed really a lot of cases for these kind of dropout students. They have good start, but it's really hard for them scale the whole experience for their capacity to pursue some knowledge or skills deeper. And also their experience for their life are limited if they are a dropout.

And some of them cannot grow even after that. Even when times grow thin, they are growing older. But there are similar ways to help them to earlier get involved into entrepreneurship and innovation community, like the investor or the entrepreneur. That is complete their academic career earlier. Like graduating in 2.5 years. Like pursue a PhD directly and graduate in three years.

That is possible if you are trying to bring more efforts in. Or there is nothing you can do, right? If there is something you can do, select more courses every semester or every quarter and be as a full GPA and make everything goes faster. We have some talented guys, as our fellow, they just go to college at 16 and also finish their undergraduate studies in just 2.5 years, and then three additional years for their PhD.

So that is possible, and you can calculate when you graduate, you may be just 22 as a PhD. So, as in you are with good capital, not for money. Maybe also for money, but not only for money. And also, your position. That's capital for your position, and the whole ecosystem will be so different. At that time, if you are clear with your idea, then a better and higher start point.

Yeah, I think you are kind of quite different in this position because I do meet a lot of people who are kind of questioning the current education system, saying that maybe students cannot learn that much from the school. Maybe they should just drop out earlier or whatever. But I think you still believe in the value of going through the system, and- Actually, not. I not believe the current education system is best solution.

I also believe this is so legacy and so traditional and waste some time, but there is no alternative. These kind of challengers, they cannot provide a good alternative. Not everyone is stupid. They know this kind of education system are not so up-to-date, but they are still sending their children to college. Why? Because there is no good alternative.

So the best solution to save time is not to not do the whole stuff and leave your 20s empty. But if you can still save some time, you can do your academic career faster. And then you can also save time and you can also have a good base on that. Hmm. Okay. Thank you so much. I think that wraps up our episode today. And thanks again for joining us- Uh-huh...

Fernando. Uh-huh. And for our audiences, just follow us on all our channels and also you can follow Model OS on their latest movements and their breakthroughs. Yeah. Thank you. Thanks. Thanks so much, 80VC.