Vol. 13 AdalFlow: Auto-Optimizing LLM Apps with Li Yin
01:02:07 Share

The investor wants to help. So that's why those male funders, even though they're like super young, right? They are really good at talking to making their dream bigger than what it is. Their vision bigger than it is. And they get their funding. No kidding. You can't be humble when you go to pitch.

There's not a big shortage of software engineers, but there's more of a shortage of AI engineers. So I'm going to give companies, empower them, not just with Adaflow as a library, but I'm going to empower them with digital AI engineers. Welcome to another episode of From Chen Ji, where we chat with overseas Chinese entrepreneurs. Today's guest speaker is Li Yin, the author of Adaflow and the founder of Silph.ai.

Adaflow is a library to build and auto-optimize any LLM task pipeline. Before Adaflow, Li was an AI research scientist at MetaAI and a PhD dropout from UT Arlington. Hi Li, welcome to our podcast. And to get us started, can you tell us briefly about what is Adaflow and what is Self AI? Okay, that's two big questions. Adaflow?

So when everybody knows prompt engineering, right? It started in 2023 and in 2024, a lot, every single one of us are building demos. So we are mostly using non-chain, NAMA index, even though, and they are still the most popular. So that's where we started to think, okay, AI actually works.

become a task that software engineer can do, right? But gradually we found the problem. Like prompting is not easy. It's very frustrating. And a lot of stacks with the AI, it becomes very tied to each API. Imagine new models coming out every day. We want to switch

right away. So I found all those problems, I found we don't really need so many abstractions to just call APIs. Doesn't make sense. We actually need something beyond on the basic API. I just need one single message. And then I add all the advanced features that make it possible for us to adapt our workflow to any models. So you have the flexibility to switch.

That's how our library was born, like to build on something simple and something actually meaningful, like actually good for your product in the long term to give you that flexibility. And what's really important is to get out of, get rid of the manual prompting. So that's the Adaflow in the GIST. Back to

our company. So Adaflow is part of our open source effort, and it will also be our secret sauce to really for us to lay out the product experience for self-AI product. And one in particular we really care about

is to actually create more AI engineers, digital AI engineers, right? We all know Devon, Magic. What they are doing is to kind of like replace the software engineer, but they are a type of engineer we don't even need to replace because we just need more of them. That's software engineer.

So our product is going to leapfrog to the future. What is the future of engineering? It's about AI. So we want to start building about the future right now. And that's self-AI to you. Yeah, sure. So talking about property engineering, what's your take on the influence of DeepSeq that we recently saw on property engineering?

And especially, I think especially there are people saying that, you know, you don't have to like prompt it as it is for chat GBT. So I'm curious in sort of your perspective. What do you mean? Like you don't have to prompt IDTs in chat GBT? Like, for instance, people will ask chat GBT that you act as a certain role, you know, and then you generate it as a certain character. And yet that is not necessary for DPC. And how do you think that might impact the prompt engineering? Yeah.

I used DeepSeq, right? I think the thing is they still have a system prompt on their website chat. So we probably don't know what prompt they used yet unless they open sourced it. And for the open source version, I think we probably should explore more on the purely raw version to understand it better.

But in terms of how DeepSeq is going to impact prompt engineering, I think it's benefiting. The reason is right now we use GPT-4.0 as the optimizer because in auto-prompting, you actually need a very smart optimizer that can identify a problem in your whole system and propose a new prompt. So you want it to actually be really good at reasoning.

And the GPT-4, oh, they are closed source models, so you couldn't see the reasoning. It's not transparent. But with DeepSeq now, we got the same level of intelligence, reasoning, but with all this transparent output. So I think it's going to benefit prompting. And in terms of, I think prompting we still need it because reasoning model, they might be good,

But you don't always need them to output so many tokens to get answers, especially for simpler tasks that don't require much reasoning. Doesn't make sense. There was a joke, someone was like, if sick isn't, the sword's 67 hours.

Yeah, 60 hours, just tell AI engineer to create their job. It's not meaningful. Yeah, yeah, definitely. I can definitely see. Yeah, so tell us about your recent launch. Yeah, so we were collaborating the research with UT Austin and the professor's name is Atlas Wong.

So out of so many people are fans of DSPY, right? Because, and they are from RIS Stanford lab, but I think Dr. Wang is one of the first few that actually see the potential of our library, especially coming from that kind of background. So that's RIS, we were collaborating and even the whole lab about DSPY

on Dr. Wang's NAP, we are using out of flow for all kinds of different use case actually. Because anything that you want to launch, you want to make sure they have the best performance. To do that,

That's where the automatic prompting genuine come into the play. And we benchmarked with DSPY and text grade the top two libraries in auto optimization and we scored pretty much higher than all the benchmarks. And

our token is more efficient. So DSPY works by few-sharp sampling, which means they have to sample a lot of demonstrations from a whole dataset. But we take a, it's not sampling approach. It's about generating a new prompt by looking at different samples, looking at where it goes wrong. So the prompt ends up being more, ends up being shorter.

and which is better for developers. Yeah, so circling back to, you talked about the products as an AI digital engineer. Can you tell us briefly more about it? What's the notion about it? I'm thinking how much I can talk. But I think one of the biggest problems we want to solve is building LAM apps because that's what we are best at.

I know Daven has a really high churn rate, like product like Daven, because the real problem with Daven is how many... I think it's useful to solve really simple problem. But if it's a simple problem, why would you even want to waste AI to solve it? You might as well just solve it yourself. And another case is

In the editor, you can use autopilot. I think my favorite AI for coding is not agent, but it's a copilot. You just give me a little bit suggestion. Give me a little bit autocomplete so that I can write code faster. Because there's a lot of code that reaches to the level. For example, when I'm writing my library, AI is not helpful. They don't understand. It's something wasn't created, right? So...

My point is, Daven has a really high churn rate because people feel they don't need it. And another thing is, Daven couldn't, even though it is able to create a prototype, but if you think about it, any product is a very long-term process. It needs to have that long-term understanding.

to continue, improve, right? There's a new feature coming out based on the user feedback, I'll do something else. And that agent don't have that long-term consistency understanding about your whole project and it will keep growing. So imagine if you just keep using agent

If your engineer don't understand, you actually go and the bike file for machine learning team. I think if I have to be all the code agent,

I would lay out more on specific tasks, like something that not a lot of people is able to do yet. And then we focus on that small group of tasks and making sure it does so well, like prompt engineering, building AOM apps, right? There's not so many of them actually knows how to do it or how to create a data set. So it sounds...

So it sounds like your approach is basically taking the hardest and most unstructured problems and try to build a platform or a framework to enable people to build their own LLMs that solve those tasks. Yeah, it's like we're trying to empower developers, right?

in different ways. One is we give them really a powerful tool that is our library Adaflow.

But the second is there's not so many people have that knowledge to even use building ALM application because it's very close to how to build AI. You still have to think about data sets, metrics. So that's why, and there's a big shortage of AI talent, right? So not software engineer. There's not a big shortage of software engineer, but there's more of a shortage of AI engineer. So I'm going to give company, not empower them, not just with AI,

out of Flowizer library, but I'm gonna empower them with a digital AI engineer that boosts their productivity and that knows how to use this advanced auto-optimization tool, knows how to create a data set, knows the new ways to research to keep them informed. - So I would like to sort of go back to the research, the research paper you've been research published.

which introduced a framework to automatically differentiate NLM workflow and go on to the front page of hackers' news. Yes. What does it feel like that? Yeah, I think it called me...

I'm pretty surprised. I'm always obsessed with this because I know the huge potential, what it means to the whole industry. So I posted it very late, late night on Haka News. Didn't even promote it. I was just like, okay, I'll just post it. It's very simple. And I also launched it on LinkedIn, even though I have like 50k followers.

somehow it didn't perform as well as I thought I'm like okay people don't care but then I suddenly we just went to the front page of the hacker news and more and more people are coming to uh to comment on the post and they got really excited um

So I think developers can start to see that. But I think the problem is, like I said, the majority of people who are building AOM applications, they are actually software engineer by-ground, right? So they might think prompt is a parameter. So they don't have that sense, okay,

Getting rid of manual prompting is good because sometimes I'm thinking developer might, like software engineer, they might not have that incentive to carry enough because they come from a software engineer background. What they're really good at is using APIs, right? What software AI engineers are really good at is creating data, manage uncertainty, iterating experiments. But the process of LAM application is actually very,

exactly the same as an AI process. But I do find even talking to a lot of software engineers, it's just way uncomfortable, the process of AI. So they would have some fraction to actually pay. But I found employers

actually care more because it doesn't make sense for my engineer to spend a few miles to manual prompting on one single model. If tomorrow there's a new model coming out and we need to do it again, you guys need to learn how to auto-prompting. Or else I need to find, imagine if you have two engineers to choose. One can build an app, LAM app, and can auto-optimize in one hour, and the other one needs to take months to

to manual tweaking it, who would you choose? I would choose the first one for sure. Yeah, that is for sure. Yeah, so I think that kind of somehow explains this a little bit lagging about how the market is catching up about our research innovation. Yeah, I can totally understand why people are getting thrilled with your research. And can you kind of walk through this research overall, the underlying mechanism of this paper?

Yeah, yeah. So, okay, I think the biggest thing is this whole optimization comes with two parts. One is you build an auto-differentiable graph on top of your application, right? Imagine application has a lot of input output. So you're kind of collecting the input output as an edge in the graph.

to represent the information flow, the feedback. So you would use that, creating all those different edges. And then at the end, you have an evaluation function

So it's an auto-differentiating pipeline. And the second part is a gradient-driven optimizer. But what does the whole pipeline, the graph does, is to, for every single load, it uses the same mini-batch training in the neural network back propagation.

So you have the training data on a mini-batch, and you observed the error on top of that batch. Then you back-pub that error to every single load in your graph, and then you want to try to identify which prompt is causing the problem and in what way. Then the optimizer will be taking that information,

The optimizer is very interesting. It takes more. The feedback from the whole thing is one part. The optimizer also, I actually injected prompt engineering knowledge in the optimizer. So I'm like, optimizer, these are the common tricks to do prompt engineering. Now, based on the feedback, can you propose a new prompt? So that's basically how it works. It proposes new prompts.

and we would validate it in evaluation dataset, and we would pick the prompt that give us the best performance.

We also observed some interesting things. It can overfill, overfitting, because you might just, on the validation data set, you just keep pushing it to get higher, higher accuracy. But on the testing set, it might actually not improve it. So there's a point you need to, but you are able to pick which checkpoints you want to

really use. Yeah, so how did you get started? Can you share briefly about what are the early days like for Adaflow? Oh yeah, Adaflow started in April 2024. Yeah, so basically I used the long-chain NAMA index and like I was building my previous product and I do a lot of RAG and

like natural language to SQL to do the search, right? But basically what I found is the existing library is just overly complicated. It's not useful and nobody should use those

uh, framework to build. And I, and I got really tired about manual prompting because I've spent a whole month to take something and it got me really frustrated every day. I'm like, why am I doing this? I'm a researcher.

And I'm doing prompting. And I'm like, I don't enjoy the process. And it's a waste of time for me. And I decided, okay, if I have to spend months as prompting, why don't I just create a library that automatically does it? So combining of everything, I started to design an architecture that would really...

reflect what LLM application is. It is very closely. It's just a different way of learning for the AI, right? We have the model weight tooling, but now we're tuning the input in a different way. But the whole thing

you still use the exact same data set. It doesn't make sense. You use the same evaluation metrics too. Yes, that's how we got started in April 2024, and it has been a long way. Yeah, it is. Can you share your biggest challenge as a founder? Yeah, I think currently as someone working on open source project,

I want to grow. I think there's a huge potential for Adaflow to become the pie torch for ALM application. We could make money

We can monetize with collaboration, partnership deals with different model provider or working with company because there's already many big company reaching out to me. Like we actually think Adaflow can be a game changer for our product. Still actually collaborate with us.

But on another way, as a founder, I need to create a long-lasting product that actually makes money. Like that is the AI teammate, the digital AI engineer. And I listen to so many things. Well, we're still a very small team right now. We're still officially two people, two people team. It's all, it is, uh,

I think the challenging is, it is a lot. Sometimes you do have to really focus and prioritize. Right.

Yeah, so I heard that saying that as you're a startup, you need to stay laser-focused. That's by Sam Altman. And you also need to ruthlessly prioritize stuff. Yeah, so it's kind of amazing to hear about Adaflow that makes it onto the front page. And then it's just two of you guys. Is the other one your co-founder? He's a founding machine learning engineer.

Yeah. Gotcha. But now we work pretty closely in the office. So we do have a, we joined a program. It's a new program started with Bessemer VC. So the program is called Bessemer Beam. It's a bunch of deep tech founders building in the

In the garage of a new CEO, ex-trainier CEO, his name is Jeff Lawson. So we get the mentorship of Jeff, which is awesome because he actually created a billion dollar, a multi-billion dollar company. And we would be almost seeing him every single day building his car in the garage while we are working on our future products.

I think he has been very, very helpful to us and I really value his advice. Like he's training me, you got to

I know, I hear you want to be all this digital AI engineer. You just got to build it, no matter how hard it is. I'm like, okay, that makes a lot of sense. I got a little bit carried away with all this research and out of flow. I think I'm doing the community a huge favor, but sometimes I feel like maybe I'm not appreciated enough. Yeah.

I mean, what do you think is like the biggest thing that you get from the Jeff's mentorship? The biggest thing I think is to, I think one thing is to be persistent, right? And to find and talk to users from very, very early days.

I didn't realize this at first, but I think from the process of Adaflow and where we started to build a community and I'm starting to see that. And another thing is to be really close to your developer because Trainio is a developer first company, right? They provide this API for communication. So I think the biggest thing is to actually know your community

and work closely with them. So we might study to do a bunch of like hackathon just for fun or maybe a little bit like small program, like be out with me. So we would actually get a bunch of engineers together and be all together and see how people are using different products. What are the problems they're facing? What if I give you a digital AI engineer on your team to boost your hackathon? Well, that would be something very exciting.

Yeah, yeah. Team without AI engineer, maybe they could pull it off or maybe with a software engineer, they could pull it off, right? They could quickly create a data set in one day or two day and auto optimize it, get the best performance. That'd be something fun to do.

Yeah, I think Hackathon is definitely a genius idea because Hackathon is limited to like 24 hours, the classic one. And people can really measure the productivity boost in one day just by using, you know, digital AI engineer that you guys deployed. Yeah, that makes sense, right? That's a pretty good idea. I just come on this, but basically, wow, we should combine.

add a flow, hack us out, plus with our trial of our product. Yeah, yeah, definitely. So it seems like, because you mentioned monetization earlier, I wonder, did you guys bootstrap to the current days, or do you guys fundraise from early investors? So we have pretty decent angel investing and VC investment already.

We are not currently very focused on fundraising, but if we have to do a next round, our round is already half full, actually. There are people pre-committed to our next round. We might be doing it in one month or two months, depends on...

when we should be ready. But right now, I think we are trying to build a very simple MVP just to prove concept. We might be collaborating with UT Austin on this because we actually have to train a model. So we would start training our own AOM. Good luck on that. Thanks.

Yeah, I remember last month you made a LinkedIn post saying that you're actively looking for a co-founder. I wonder how's the progress going so far? So there are some interesting parties. And ever since I started collaborating with UT Austin, their lab just has so many talented people. And I think gradually I've just started to...

work with people more with other people more and probably naturally get a co-founder like a potential CTO a specializing AI like in this way in this like very organic and natural way instead of like working with someone directly talking about co-founder even though we haven't worked together yeah so we also have

My previous friend might be interested with like operational kind of roles, but we should say because there's a bunch of operation like hosting the hackathon. It'd be too much for me to do that. It probably is not the best to spend all my time. So have you put in some effort in socializing and have you found it crucial for founder?

Oh, that's a good question. I actually put an effort. I moved from Mountain View to San Francisco and it's a big, big cultural shock. I know it's just one hour away, right? But

But anyway, one interesting thing is I got six tickets. My car got towed twice in two months. And some part of my car was stolen. So now my car is functioning. Damn, I'm like moving, totally ruined my car.

But anyway, I moved to the city where I joined a community. It's called Mission Control. So they actually have like three unicorns. Almost every single one of us are pretty much founders or either working at WC.

And we do a lot of events together and they host a lot of events. I'm not saying I would go for all the events. I was like really busy with my paper, but I did way more events than I was before. And I feel this is, I'm surrounded by a community and I'm going to a workplace, like a co-working place. There's a lot of other founders. So yeah, I think I'm totally socialized and more, not just the only thing.

Oh yes, your LinkedIn is very impressive, like 50,000 subscriptions.

So what have you gained through the process of socializing and also your LinkedIn? I think, well, that's a really good question, right? Like for LinkedIn, I think the year of 2024 is about really me realizing the power of social media. That's the LinkedIn, right? I had the biggest growth of year, even though I wasn't like fully a content creator. I'm not a full-time content creator.

But what I found is it really helps me build a community. And I got reached out by a lot of senior managers or direct label or even vice president of very top big companies. And

So it creates a lot of opportunity. And most of our angel and investor inbound, they reach out to me on LinkedIn. Like, I've been following you. I really like what you guys are doing. Right? So LinkedIn is a game changer in terms of how I'm thinking about

they're still building a startup nowadays. It's really about distribution channel is really important. I know product is very, very important to you, but if you have that distribution channel, it will save you lots of money in the future and you will get your product into your user much faster. And the

Come to 2024, I moved to San Francisco. 2025, right? I moved to San Francisco. I'm surrounded by a community of hustlers, founders that are young, energetic. And the biggest thing I learned from them is one person, his name is Connor, introduced me to a book called Peach Any Day. So obviously...

If you want to raise their fund, it's not just to go ask investors for their money. You have to get investors to come to you. And there's a bunch of things you need to do, you know. And it's very game-changing. It's like mind-blowing, exploding for me. And...

I really appreciate it. I started to know this process. I started to, I mean, I can't say too much, but you guys should read the book, pitch anything. It's a great book. And speaking of team building, is there any sort of ideal team structure, especially in the age of generative AI? Yeah, that's a good question. I was actually trying to make a post like, watch out for Chargiviti and Janir. Yeah.

I think, yeah, go ahead. Can you tell your library on that? CharGPD engineer, other engineer. I think especially bad for remote. I hear so many people, like to me, they fire their team that working remote because they found, I think CharGPD has changed remote work, right? Imagine if you're doing remote, then you don't really know what they're doing.

And then you look at their work. A lot of them, if you don't really seriously check their code and verify, the code might seemingly look like impressive. It might be chargeability generated, right? It might look impressive or working, make sense. But when you deep, deep down, there's just so many problems in that code. And you end up redo the code

everything. So no matter how much money you're saving on that remote engineer, but the trouble, they actually create more troubles to you. So I think a big, big thing nowadays for the founder, especially for founders who don't have a lot of budget to create a team, especially if they're in USA, they got to be really watching out. If the engineer you get is a deal,

If it's good, it's actually good for your team or is it actually purely like team value? You can totally end up losing money, losing time to train them while they're giving you something like totally not working. So I think this is one of the biggest thing at this age with the LLM

you have to tell them, be really clear how your company would want to use LLM to boost the productivity. For example, I banned my hour engineer to use LLM to write documentation because it's always like very verbose, very verbose, right? Like nobody wants to read this. If my engineer don't even want to read it themselves, why would I assume other developer want to read it? So we just banned all people to do that. Right.

Speaking from your own experience, you've mentioned about the potential risk of building teams. So how did you deal with, like, how did you choose your teammates? I learned from my mistakes. Whatever I tell you, I got the problem, yes. And

And now I think when I, now when I want to build a team, especially at this early stage, they have to be outside. They have to be here with me and showing up, right? So this is a condition that I'm deliberately going to go unless this person is impressive and has a great track record and is actually, and don't give them trust right away, right? Wait a little bit and check their work

So that would be what I'm trying to do. And in terms of, I think even some ultimate, they will ask you the question, if there were two engineers, one is more passionate about what you're doing and the other is more skilled, but less passionate, which one would you pick? Often people pick the first. I think even for us, that is the same too.

I think you just want someone so committed that can move fast with you. If the engineer is just not committed, I think that's also a big problem. And if they are passionate, I think there's things that they would work their way up, right? They would proactively ask your feedback, how I improve,

I really want to build this thing so much better, but tell me how I can do this if my skill is not there yet. So they would actually learn really fast and eventually end up being a good asset for the company. Yeah, the same thought process came up over and over again through all of our previous guest speakers who also hire the one with the most passion because they will move faster and catch up faster. And then they will stay with you. Yeah.

Yeah, exactly. And they won't use chargeability to generate code that much. Yeah, they actually want to understand everything because they are so driven to learn and improve it, you know.

Yeah. Okay. I know you talked a lot about, you know, the concept of building in public that you sharing your own stories on LinkedIn and other channels. Is there like any specific interesting story you want to share that, you know, benefits you get from building in public? I think there's a lot of building public. My biggest take is to find, it's a good process to find your audience.

And it actually gives me more motivation and it helps me create more momentum. Because every day you can see something going up. Your followers going up, your reviews going up, your GitHub report starts going up, right? And more people is reaching out to you. It gives you this

you're creating something, you're making an impact. So I think ever since I started to build in public, it has greatly shaped my mindset about being an engineer. So I think a lot of engineers, when they first started their career,

they they're like a very technical driven like let me build let me build this right it's just like fascinating let me get it more complicated let me get it deeper everybody's like making this more confident like why why do you need to do this does user feel there's actually an impact did you create more stars or more followers what is the outcome so i started to i'm like

I started to really go backward and be way impact-driven. If this thing doesn't have much impact, I will try to minimize it. But if there's something, for example, like ALM AutoD for the research paper, that's something I find

I strongly believe that is the future, right? Even though sometimes the community actually needs a little bit of time to catch up on the things. But anyway, I think the social building public really has made me to impact the driving. I was sometimes joking. If I could...

because if my goal is to become a billionaire why would I have to if I just say it out loud people can make me a billionaire that's the best solution why would I have to go all the way to beyond so much right impact driving I don't care what is the solution or whatever if someone give me a billion I'm like okay problem already solved so

So I think there's too many ways to solve a problem. You got to start with the least effort one. Yes. Building in public just changed my mindset about the less effort you put, the better. The more impact you create in the shortest time, the better.

Yeah. And also training. Yeah. A lot of the thing is if you tell people what you need, it really helps you. You never hurt. Yeah. Because whoever wants to help you, help you. And those who don't want to help you stay silent. Right. So you don't get like, either way, it's like a win. Yeah. Yeah. There's no downside. Yeah. I just want to quickly comment on the thing that you said, like people tend to build things complex. I think Steve Jobs has a good saying that you need to know your audience and

once you know your audience, stay obsessive with your audience and then you know what they want and then you build something amazingly simple and simply amazing. Yeah, yeah, that's a good one. Yeah, amazingly simple and simply amazing. I'm a strong believer of Steve Jobs. I read his biography so many years ago, like more than 10 years ago. And that's where I started to

Even from that day, I was starting to know simplicity is the real thing. It's not the complexity. I think I started to have that very early on as an engineer.

But I know a lot of engineers not. I started to delete so many code my previous engineer write in my code base. This is too complicated. Nobody wants to read this. It don't even make sense. It's not even doing what it does. So I'm just starting to cut down because even for our library, I think less code is better than more code. And the more structured code is better than just dumping a bunch of spaghetti code and nobody can read and understand and maintain.

So I have a very, very high standard in terms of the quality, the simplicity of code. It also goes to our life, our product, I think. Yeah. So switching gear a little bit, I know from our last conversation, you talked about you being like the only female founder in the like the Beam program. Why do you think, yeah, why do you think female founders are rare, even in Silicon Valley?

Yeah, I think just starting a company, most of the time people have to do it when you're pretty young, right? And men, female are 7% or 10% or 15% maximum in seeing as a software engineer, right? We have more of a biology clock.

I think that's totally another thing to stop female from actually taking the funding. And another thing might be risk-taking. I think being a founder is actually very, very uncomfortable.

And especially we are way uncomfortable because most people next to us, they are totally men, right? They are fundamentally something different about how men and female, women, we actually interact with each other. The dynamic can, it's not very balanced. So we have to really tough up

like to stand up for ourselves and start to change our mindset about to show their strengths because I think females are very, very humble. I found my female other founders, female founder friends, they would actually undersell their product. So many other male founders, they 10 times oversell what they are creating. Like a feature having launched, they would just say, we support this, we have this, right? And they started to sell.

And they say a lot of big words, you know, sometimes might not totally be true, but we tend to be very honest and tend to be more humble. And I made it actually so much harder when you compare us.

with so many others that don't do that. I think just naturally because it feels it's such a much, such a harder process for us. So a lot of people probably can potentially give up early too. The part you mentioned about like overconfident among men, I think that's something very true. I've read a journal that basically took statistics of researchers who are male and female and how they submit their journals. And it

And the male researcher overall, a lot more confident or even a bit too overconfident when submitting their journals. But that's what the investor want to hear. So that's why those male funder, even though they're like super young, right? They are really good at talking and making their dream bigger than what it is. Their vision bigger than it is. And they get their funding. No kidding. You can't be humble when you go to pitch.

Yeah. Right. That's, that's true. I mean, I've also have, because I've read a research of a sociology research who was talking about female functions in Silicon Valley. And the researcher was suggesting that because

especially during university time period, sort of the social connections is sort of more tightly, there's a stronger social connection between the male founders during universities. And when they're building their own founders, they have easier access to investors or other founders, whereas female are somewhat isolated. Oh, 100%. That's a really good point that I probably missed that it's good catch.

You felt very, very much the same? Yeah, it's very much the same because we don't bond with men the same way men bond with each other. And the majority of people here, either engineer or whoever they are, or investor, 90% decision maker of the VCs are actually male. And a lot of VC, they get their money, they invest actually based on their personal engagement.

Can you believe I, there was a statistics, 90% of the deals actually goes from, we see it actively pursuing their deal, right? And those are, a lot of those things happens on people they already know

and most of them they would build their bond with men much easier when you when you bond between men and women right there's a marriage is they're married are they not married what is this weird situation are we actually totally talking about Venus or all these like something some weird stuff romantics whatever is like involved so people tend to be

it's not exactly the same process and we might also have a boundary between men and women right we can't just go your own light just drinking with you right like you might have some some man you just go and drinking and doing bunch of like men's balls and it's something naturally not a lot of girls actually interested to do got you yeah um

Yeah, I mean, like totally. I think if I remember correctly, I think Adaflow, the name is also in honor of like Ada Laplace, which is a pioneer female mathematician who recognized that machines could go beyond mere calculations. So I think like the way that you picked the name also suggested, you know, you want to become like a pioneer in empowering the next gen of AI, AI

pioneers, especially for females. Yeah, that's totally true. I think we need the longer matters. The longer matters. Imagine there's just more people, female VCs, right? We bound having girls' nights and they decided to invest in us, right? We can't go and bound with men on their boys' night. That's for sure. Yeah, yeah. That's definitely a problem. Yeah, I do think when it comes to startups,

It is totally true. And female tend to get more defensive questions when you're talking to a female. So you really have to, we have to overcome so many natural defects that, not defects, natural, just the feature about us, like we become, we are naturally more humble, right?

We have to overcome that and be really confident and show them we are actually the alpha. Because females normally don't take the alpha role, right? But if you want to be a female founder, you have to be the alpha. And you have to make people feel your presence, speak out.

and they will doubt yourself. Correct the people. If they started to doubt you, you're just going to call them out. That's another thing. When you call out people by their mistakes or bias they put in, they feel a little bit ashamed about themselves. You got your power back. So, sure. I think, Rachel, you got the next question. Can you sort of share some of your thoughts on sort of creating a sort of a more

female founder, female founder friendly sort of environment, but what sort of thing that you would like? Or let's say what? Female founders. I think, I think that we see

you should start to realize that we are just literally a little bit more humble and we might need a little bit more help, especially at the beginning, to know how we should do, right? Because we don't have a lot of examples. There's just too many, so many men. They know they have to be bold. We just don't know. We're always told to be honest, be humble, right? So I think the community has to understand that

give the female founders more trust because if they tell you they are

It's the point of five. And it probably is the same level as a lot of founder, male founder version, given you feel it's like a point 10. So you have to understand there's a big gap in the perception. But statistics don't lie. A lot of female founder lead companies, they actually create a lot of value. They are very successful. They tend to handle risk much better than their male founders.

They don't just quickly burn their cash most of the time. I know there are exceptions, but most of us are really, really good at handling risk. And we are also very nice mostly to our employees because we want to treat people nice. So I think the whole world, the whole community should encourage and understand this and support female founders.

like more and or point them to the right way. I think they will learn. Yeah, sounds good. What do you think are Adaflow's next steps? And are there any major milestones ahead? I think Adaflow, the next one, we are still doing research with UT Austin. So the next milestone would be so many PhD students would actually use Adaflow on all of their

research exploration to prove there's a lot of value in different fields, different use cases. So that's number one. And the second, I think the code, we explored, we published the paper, LAM Autodiff, but this actually allowed more to improve

And the future of us, imagine you have a whole thing, but every single component is LAM-based classifier, LAM-based chatbot. But what if we can automatically create data set for every single smaller component?

and to train a much, much cheaper machine learning model to do that. Why do you need LLM all the time? You can use LLM to set up a full pipeline to generate a dataset. There's a lot of stuff that we can further explore in terms of how to create the training dataset for LLM.

And when you should fine-tune, when you should train, and what if we can still combine model fine-tuning together with the whole thing and with the prompting end-to-end? What if we can create a synthetic data and go back to the model fine-tuning, improve the performance, and come back to put in the application to the prompting? And then now the data is higher quality, fade it back again. You know, you could...

create a very closed-loop auto-optimization just with those sensitive data with your application with the prompt optimization. So it's a lot of very exciting

things that our library could do. And I probably have to put a little bit more thought into writing down so that the community can help us taking a little bit piece of it. Right. I can definitely see how that's important. And I mean, I sort of want to ask that, like, when...

Because previously you were in academia, right? I mean, so when you transform from researcher to now a tech founder, how did you find this? Did you find it challenging and how did you solve this dynamic? Yeah, I was doing PhD at UT Arlington, right? And I dropped out because I realized the industry gave you more GPUs.

And I joined Meta AI, so it also means I'm surrounded by the best researchers. But whenever at Facebook, we do research, but it's very grounded research. We know the research is going to land in a product in two or three years. So I'm in this applied AI, right? So I actually see the whole process, the whole life cycle about

creating dataset, we work with scale AI, we work with the data labeling team, and we work with product manager to get those things ready.

So even from the work experience, I actually see the process of building a product where we separate the whole requirements into different parts and every single team take one. And we started to give a roadmap about two years to do the initial research, where the research is and how much additional research we need. So I kind of have the process of building.

the whole thing about building a very kind of research-driven product. It's like you're doing research, but, oh, it's product-driven research. I think I'm doing exactly the same thing at Adaflow, right? I didn't build it just because it's a research. It's amazing, but it's actually very practical and I actually want to use it and other people can use it. And so...

I think I made the transition probably already started at like Meta AI. But the difference is as a founder, we have to do so many more things. For example, you have to pitch your company. You have to build your public presence. You have to build your network, right? You have to build your reputation. And you really need to understand people more. You need to understand

understand how to build team. So the entrepreneur part is actually so much more than being an engineer. That's why I'm like, I have so many things to do. If I can just go out and see it in public and people help me do it, that's probably the best. If I don't have to do it and it just got done, that's the best. So we would start to, because we have so many things to do, it started to shift your mindset about how can I be impact driven? And my impact is actually drive more people

to my region so we can work together so that I wouldn't move faster than I could do because I could never move fast enough as a single person. Right. And now we're sort of moving back, moving to the sort of the last part of our podcast, which is like, and we would like you to seek some of your advice. And I think for you, like, was there any advice that you would give to young people, especially young women founders who are interested in tech and AI? Hmm.

I think my advice would start, I really like the book, The Minimum Entrepreneur.

I think the easiest way probably to build a company is to actually find your community first. Once you engage with those community, the people in your community, you really become a domain expert in that community. You understand the problem they are facing and then you just solve problem for one or two people and maybe including yourself. And you really solve it well. You maybe don't even need to build a whole product to do it.

You can just do it manually to help them. If it works, you convert it to software products and make sure before you can only help one or two people and now you can actually help 10 people or more people. So sometimes I think this is the best way

to actually get started and to actually feel you are creating something different. And obviously you have to iterate a lot. But the second piece would be you just have to do it. You just can't think too much about the risk and the fear because whenever you start to think you're wasting your time and energy on thinking something not producing and make you feel even worse. So whenever I started to

feel self-doubt. I'm like, I'll just go do something totally different. I'm like, I don't want to think about it. I think Elon Musk was doing the same thing. It was just, we don't ever spend time to doubt ourselves. There's just no, it doesn't solve a single problem. We become way, way, I think that's the mindset. We have to take this mindset, right? Like, don't doubt yourself. Don't even bother wasting the time. Just move forward. Just keep trying.

and exploring. Yeah, and Elon Musk, they cut more than 90% of the process for SpaceX. That's why they built like 20 times cheaper rockets than NASA and those other competitors. And I know there's also a good saying that if you push this further, no doubting for yourself, you don't even planning, you just do it. Yeah, that's a good point. I think you just have to

do it so far that's why I feel like posting on social media I actually like posting because every day it gives me like I'm doing it I'm doing it right yeah I think once you started to build in public you're really nice really nice you just start to do it it's action that matters you can do something build in your house yeah

Exactly. Don't wait for that big launch. Just iterate fast, launch fast, launch often. Yeah. Yeah, and actually, back to the launch, I was a little bit disappointed because we thought we might make even bigger splash. I know we made it to the front page of Hacker News, but I'm glad people might even react more enthusiastically, even better.

But that gives me this sense of like, you just have to show up every day. There's no success is over light. I think it's a lot of compounding, compounding, a lot of struggle, maybe done work, done work at the beginning. And then as you keep going this and some people finally they started to realize the value and they started catching up and started to boost your signal.

So it's really about day-to-day momentum. It's less about, oh, I have a huge, big announcement. What you should do is announce every single day, not like announce every one month or two months or three months. It's not enough. Yeah, yeah. What would you tell yourself, Patso, as a student, as an employee, and as a new founder? Oh...

I don't know. I think I did pretty well. But I think just enjoying what you do, right? I know I'm like suffering as a founder, but I also very much enjoying it. So I would encourage people to actually find, really believe in your domain and still a little bit more focused. You got to become a domain expert. I think so many people that end up

Somehow they really like learning a lot, but I actually reserve my energy on very limited things. So being focused and become the best of a particular field would be my advice to younger students.

You can stay curious. You can study to know a little bit, but you have to remember you don't have to score 50 on every single thing in your field, right? You should aim for score 90 or 100 on one or two and just score 10 or 20 on the others. Just know a little bit on the others would be enough. So last question for the day.

If you can recommend some resources that you benefit from, like info sources, what would you recommend? It can be like books, podcasts, videos, newsletters, or articles that you read before. Yeah, I think I really like reading books. The one I'm currently reading is Peach Anything.

And I'm also reading Jeff Lawson's book. It's called Ask Your Developer. So it's basically he writes down the whole process about building tenure. And I really like the book, The Minimalist Entrepreneur. And another thing I think, the saving habits of highly effective people. I think I really like that I booked her to make sure we are effective

our day to day can be really effective and how to communicate with people

Did I mention Pitch Anything? I did, right? If you're a founder, if you read Pitch Anything, I think all my problem in the first one or two years was because I didn't read this book. Okay, so it solves everything. It's mind-blowing, yes. Oh, wow. Good. What would be your last word to all of our audience who are majority young founders?

and just do it yeah of course yeah never give up and work with your user I think so many people a lot of technical founders make mistakes is to keep building keep building nobody care actually you just got to actually make people care first and build it sometimes you really have to talk to people and really talk to people yeah okay so that will be our last word thanks so much

Thank you for your time.