DN #35: Taiwan to SF, Hackathons Everywhere & Getting Caught by Anthropic (w/ Sean Chiu)

Sean:I go on the explore page in SF. There's like three or four hagathons happening every week. And every day there's like at least ten just in SF.

Niklas:Hi and a huge welcome to you, my lovely listener. So glad you're here today. You're joining me for a chat with Sean. Sean is the multi-head San Francisco builder. with a very interesting story because you moved to the US from Taiwan. Let's start there first. What is your story?

Sean:Sure. Yeah. I I was born and raised in Taiwan up till I was sixteen and two years ago I moved to the US. So I move around a lot. I've been in my high school journey I've been to three different schools out of the four years that I have. And stayed a year in LA, then went up to the Bay in San Francisco and that's when everything popped off. Went so much of events, went so a bunch of hackathons, met cool people. And now I work at a company that hosts a bunch of hackathons and we're trying to bring the SF style builder events to the rest of the world. So that's the basics about me.

Niklas:Having seen many places, what do you feel is special about San Francisco?

Sean:honstie just the sheer amount of events that happen every day. So when I when I when I went back to Taiwan, not sure if you know, there's an app called Luma and s so many events happen there, right? I go on the explore page in SF. There's like three or four hagathons happening every week. And every day there's like at least ten just in SF. But when I go home in Taiwan there's like three or four happening every week. So That's definitely one big thing 'cause people are moving a lot. There's lots of companies that gr has built these great communities and where people builders can share information and just get to know each other. I think that's what makes SF unique, right? 'Cause the just people call it the talent density is is really high in SF. Yeah. Mm-hmm. Mm-hmm.

Niklas:I know I feel also the helpfulness. when I, on the podcast, probably the most offers like for meetups are general, just showing me the city or something when I'm there. It's no other place has had that for me. I can just say it. So it feels a lot like the people, especially if you are kind of in this tech bubble, seem to be extremely helpful. A risk if there's so much around is that you just start too many things. And GitHub... says you're working on too many things actually. Did you kill something or how do you cope with this amount of top-

Sean:Yeah, I mean AI coding came out around twenty twenty two when Chat GPT got released when we were still copy pasting code. And at that time I was experimenting with a lot of things. I've I'm watching so many YouTube videos about the latest AI tech. It's kinda died down late lately, but before it w there was like an a crazy breakthrough every week. And I think that's what That's the main reason why I have so many projects in my GitHub. But eventually I was able to, you know, double down and focus on one thing at a time. And that's what got me here right now, 'cause it's not just a one time project where you're actively building on it and people actually use it.

Niklas:I think when we look at like what happened in AI, it feels crazy still. Like the speed of progress, especially the application layer seems to be progressing incredibly fast. If we go back to like end of 22, 23, it was a lot of just copying the code from the chat over. Then we had cursor where you had type complete. Now we're talking. cloud code or the equivalents from other companies that are now able to interact with like browsers or other that they have, they've gotten really good at tool use. So, so this is quite, quite a change. And I also feel that, that the challenge is shifting, like because it's gotten so easy to build stuff. you, you guys are really good at marketing or you're kind of lost in. in the world or you're doing something that is really technically difficult, but I'm wondering what's your feeling? Like what are the technically difficult topics that are remaining or that you're seeing in NSF?

Sean:Mm difficult topics. I would say at least for me, right? the work I do everyday context is a still a huge problem. We have we use these apps, LinkedIn, WhatsApp, that can't connect as an MC P to agents. And really hard for just still cop it it brings me back to the chat gbt days where we're when we're copy pasting things when it could have been an integration. So I think that's a problem that I guess we still haven't solved with AI. It's how we store these information because there's also the chats we have with people every day, right? When we have these necklaces coming out. not sure if you've seen there's the friend necklace. that was crazy with the New York ads a while ago. But we're we're working towards it with these necklaces, but it's still hard for me to, you know, just talk to AI when doing work because I still have to do some of the manual work myself.

Niklas:Yeah, this is an interesting one, sir. Voice, I've been using just a local tool on my Mac that just allows you to put like voice into any text input, right? So for me, it's a keyboard shortcut. It starts recording.

Sean:Yeah. Mm-hmm.

Niklas:So I talk to it, then I press it, keyboard shortcut again, it inputs it into any text box. It's actually really efficient. And for me, this is actually the most efficient way. of working for me as I talk to the system, but the system gives me text back. So I'm definitely faster at reading than I'm at listening. So, and I'm definitely faster at talking than I'm at typing. So this is kind

Sean:Yeah.

Niklas:of my idea workflow is I just, talk to the system and I get an output in text, like, then I talk to the system again. So this is, this is my personal personal preference. And I'm, I also think that this is kind of If you look, think about devices, that's what I believe would be the better or the ideal device still that you can talk to it, but it gives you a visual output because we can process it faster than we can process audio. Like that's kind of also like proven scientifically. So this is, this is really interesting. What, what you also say is like context is extremely important and You're looking or you're building something called a link. What is it?

Sean:Yeah. Link was a project we started quite a while ago actually, around around May this year. And what we saw was we went to these meetup events and we talked to a lot of people, but it was kind of random. You know, they say great networkers talk to everyone in the room, but for us we talked to maybe three three or four having deeper conversations and it's really hard for us to find right people to talk to. So if you're a founder, you're looking for an investor. If you're if you're a let's say an engineer, you want to meet other engineers, but you never know what people do just with their face. So what our hypothesis was was that if we can have data from the organizer from of the guest list, because people usually already fill out their LinkedIn profile or their GitHub profile or whatever information the organizer wants before the event. And we already have the data on, you know, what why people come to these events. Why not just pair them up with the right people before they walk up to people randomly.

Niklas:And what's the use case? Can you make money with it? Is it just a fun project? What is the overarching objective behind it?

Sean:Yeah. we didn't really think too much about monetization when we when we first started. So we want it started as a tool that I used myself. So sometimes if if the organizer doesn't hide the guest list, you can see a list of people who are going to the event. And we just scraped that and we used that data to find the people we want to talk to and try to match faces during the event and got some pretty good conversations out of that. But in terms of monetizations, yeah, we really haven't thought about that yet. But I suppose we could charge organizers for the the service we offer.

Niklas:Can this be big or will it stay small? Do you have any plans or do you think it's just a side project? We'll just let it sit.

Sean:Yeah. we actually stopped one month in because we realized, you know, again, this this will stay as a side project and is is pretty hard to turn into a company because right now what we have already works, even though it's a huge improvement. it's not a pain point for organizers because all they care about is just reaching the participants, right? And great networkers already know. how to talk to everyone in the room. And yeah, I I think it would r remain as a side project as I'm working on something else already now.

Niklas:When we look at building and building agents that is something else, I think you had an experience where you got banned from, I think, was it in traffic because you used your key and open-clawed or something like that. What's your experience with

Sean:Yeah.

Niklas:these agents and also how do they compare if you just use like an open code or something like that versus the other ones?

Sean:Yeah. I mean OpenClaw was the the first ever I gu first ever agent that had memory, consistent memory, can self improve and can connect to so many apps at the same time. And I was a big cloud user at the time. It was when Chat GBT's models or open AI models still haven't beaten Anthropics yet, so I bought the two hundred dollar plan and just put it on open claw, which was not allowed at the time. I set up some kind of proxy server that turned it into more of an API key than an OAuth key. And just abused the hell out of out of it, right? Because browser at the time wasn't that efficient. So it was consuming a lot of tokens and somehow Antopic found out. And then that just was shock to me, right? The funny thing is they actually refunded the the money after like I think half a month of use. So that was k kind of interesting. But I built open claud so I don't have to you know go w go around Anthropic just to use open claw. if you could use open cloud, then you know it's built into ch claude code and it's directly on the the cloud code CLI so you don't have to pretend to be cloud code when trying to use open claw with your entropy account.

Niklas:At times seems to be really moving fast in the space. So I kind of, I just remember this whole episode where I think it was, it was actually already interesting. So OpenClaw exploded. had this mold book episode, where you had the first social network for AI agents. Then this whole thing, exactly then Entropic

Sean:Mm-hmm.

Niklas:kind of took over OpenClaw and, and, OpenAI took over OpenClaw and Entropic launched their own agent framework in this

Sean:Mm-hmm. Yeah.

Niklas:space. if we go back, because it's still an interesting topic, like I think, as you said, OpenClaw was the first real agent. What were the use cases and did you really use it for something that was really productive or was it just more like playing around?

Sean:honestly I couldn't remember what I was working on, but I knew that I was using a lot of browser tasks. I think it was before Apple Fi became a big thing. So I had to use my browser to scrape the information that I wanted with with things like Instagram or Facebook that blocked, you know, bots that were doing that are on headless browsers. So I had to use my own browser and like ten running at the same time and scraping all the data I need. that was the main use case honestly. I was also experimenting a a lot with building skills. So that I think that was also around when Anthrop like maybe one month after Anthropic released skills. So people are messing around with that a lot and Because Oprah Claw has this consistent memory where it can share across sessions, I can just build on top of the skill and continuously improve it until it it's exactly what I want.

Niklas:That's actually crazy that it's not been so long, but if I look at it and also then the evolution around harnesses that is currently happening, feel like harnesses are now evolving a lot more towards what these agents were like a few months ago. I think we're now like, if you look at Claude Cote, for example, it gets more and more capable at just doing these things. And also Claude Desktop is getting... better at doing like just running Cron jobs at the end of the day, which is a lot of what OpenCore actually was built on at the bottom. But what's your feeling there? Will this somehow diverge? Will it continue to be the same as it is now?

Sean:I think there's gonna be another breakthrough like we do in the AI age where, you know, somebody figures out a much better way to do this than what we have right now, similar to Cloud Coder Codex. harnesses, I've been digging it digging into it for a few days now. I think I've been sleeping on it, but not sure if you heard of this new model called Jeb. And yeah, it seems to Have you do you know the guy from Human Layer called Dex Dexter Horoty? He was really hyping up this thing, the harness that he was building for Jeb and if that turns into a thing I'm I think I'll be pretty excited to use it and that might be the new kind of coding agent that we have that it's different from now.

Niklas:Yeah, I'm not, so I'm not

Sean:Mm-hmm.

Niklas:100 % sure about the use cases. If I understand Jeff correctly, it's kind of a classifier. It's just a lot more powerful than the old classifiers. Like if you look back in the day

Sean:Exactly. Mm-hmm.

Niklas:when I started building stuff with machine learning and neural nets, it was a lot about like image classifiers, image classification, putting them in the right buckets. And we built like simple predictors with neural nets that would predict something, predict a number or do something like that. Then we went into the general purpose transformer and the large scale. There were a few things that we actually knew about the general purpose transformer in this famous paper, Attention is All You Need. But the crazy thing that was done back then, besides the self-attention, which is actually not the most important part, retrospectively, but... was that like traditional recurrent neural networks are not very good at being trained in parallel, like on, on really large clusters. and that's kind of the transformer is just really good at being trained as an architecture to train it parallel on like really, with really large amounts of compute. on really large datasets and which, and then there was the discovery of the scaling law, which still seems to hold true. So we're kind of like that, that was a big step. now we're kind of, Jeff is kind of a step into another direction again, in a sense that we are now talking about a classifier again, but it's a lot more powerful. think that's the large scale difference. So it's a lot cheaper to operate it than an LLM. It can do certain things really well. So this is my understanding at this point in time. So I would be really interesting if you, I'm not sure if you know a lot about it, but if you wanted to use like this classifier for programming, I think it's really challenging. How would that work? Do you know that?

Sean:Yeah. So when LN first came out, it was we're moving away from typical Python scripts where it's completely d deterministic, right? w with Jev it's kind of in between because your input can be whatever you want it to be as long as you have a defined output. So it's the kind of the middle ground between what we had before AI and what we have right now. So there's I think a lot of use cases for this. I'm not sure if you use Hot code a lot, but they have this classifier auto mode where they detect if the script the agent is trying to run is safe or not. So that's you know a great example of JEV decrease costs and also just run significantly faster. I think that's the biggest use case there. But yeah, I agree with what you said. It's it's nothing crazy where it's it's not it's not a chat GBT moment, definitely not, but it opens up a n a lot of new possibilities.

Niklas:Yeah, think like also the chat GPT moment. So in retrospect, they are easy to spot, but at the point in time, they're usually not. Like when chat GPT came out, mean, OpenAI didn't predict like the success. They have said it themselves. They would also probably have named it differently. It was an attempt to see how people would use it and if they like it. And also from GPT2. to the following systems. Also, this is the evolution, think, again, always surprised the people actually working on that. So in hindsight, it's easy to say. even for Jeff, I'm not sure. I mean, it might be really big and we just don't get it, or it just disappears again, kind of, and becomes just another thing. Both alleys, I think, are... often open on these topics, especially early on. mean, the most interesting one on this is still, think, attention is all you need as a paper that was published by Google and Google kind of missed the train of what it meant. not, it's just so hard to predict like on these outlier, outlier results. So I'm really interested to see where this goes. And I'm also interested to see like what the next like step on on technology will be in AI because I can also imagine that the GPT is not really the end of the story. That might just be the next big thing after GPT, Jeff and whatever comes next.

Sean:Yeah. Yeah, for sure.

Niklas:You're also working on something called Builder Base, like, and community, think is really important. And a lot of products focus around communities these days. So what actually is this about and how strongly are communities embedded in distribution now?

Sean:Yeah. So the story of Builder Base is that when I first moved up into SF, I attended a bunch of hackathons, bought a few and randomly decided to host my own. And they reached out, and pitched this idea of a hackathon platform 'cause it was such a pain to organize one. And it had this platform that centralized everything that so I don't have to use Google Forms, Google Sheets or Google Slides for all different kinds of thing that I wanted to do. And I thought that was cool. So I joined originally to lead our ambassador program. And now it's turned into more of a community thing because we're trying to bring these asset style hackathons where three or four happen every week into cities that really need it, right? We have London, Paris, Berlin, and New York as well that have has this demand, especially with AI right now. So hackathons Before AI used to be just a bunch of very technical engineers that could build something really cool by hand very quickly in one day. And it's usually limits it to just one very s specific problem as well. But today's hackathon is all about there's this one really cool project that one AI hackathon I hosted a while ago. They were it was a local AI event. and they solved the problem of blackouts during emergency emergency situations where you know they d deploy joint drones out and send signals back to tell firefighters where to go. So it's it's crazy how people can build such a huge project in one day. And that's what that's what I think is crazy about hack funds. people can build this and at the same time build community because you're not just scanning people's LinkedIn QR codes like a regular event, right? You're actually collaborating with the person and frantically wa running up to the organizer asking questions. So I think this is a big, big, big, big part of events, especially with community building. So really excited to bring these hackathons to cities that deserve it, right? So I'm trying to bring this to Taipei as well, through an organization called AI Collective. And yeah, it's great great times right now with AI constantly evolving and people wanting to do something with it.

Niklas:If you look at like the ideal size for a hackathon, what do you think? Is it rather small, really large, something in between? And if you had to like talk about numbers, what do you think is too big? What's too small?

Sean:I mean there's there's not really a limit on what you can build in Hackathons, right? Of course there's a time limit and also your ability to build I guess you can't build an enterprise solution in one day, but there's there's there's no limit on what what you can do. We don't limit you on what kinds of projects you build. But you definitely wanna be realistic and not over I guess over complicate your project by adding all these features that you have ideas on and be the jack of all traits but master of done.

Niklas:Yeah, I think, yeah, focus is always a good idea. Like doing too much, will just lose something on the way. you look once again at San Francisco specifically, and you're now like, you're a founder in Taiwan, in Japan, in Europe, somewhere Australia, what would you think? Is it worth the plane tickets still? Can they build it from their home? Is X enough? to get all the information you need. Should they go? What is your feeling?

Sean:Yeah. It depends on what you're looking for, right? So if you're in Japan and you're in Tokyo, there's already builders there. So unless you're really looking to come to SF and settle down for a month, if you're just here for a one day hackathon, I mean if you if you have the money then sure go for it. But me personally, if it if it's a one day trip, I wouldn't recommend it. But if you're here to actually, you know, understand SF culture and try to build a startup and g get some do fundraising, then it's definitely worth it. But if you're an engineer that's looking to try out AI, then I'd say if you're it near a tech city at in your country, it's already pretty well developed enough for you to meet cool people like yourself.

Niklas:And for me also, I mean, I'm very present

Sean:Mm-hmm.

Niklas:on X, for example. I feel like that really opened up a world to me. Obviously there might be a lot of like faked stuff, but I also feel that that access to people is actually very good and also very fast. Like it maybe was a bit different, but these days I have a bit of a following and this has gotten... quite easy for me to reach out to people. get an answer and that's actually really, that's really nice. And for me, this

Sean:Yeah. Mm-hmm.

Niklas:has helped me personally a lot, which brings me back to one point. I believe it's really important to be at the right place at the right time, which is in itself really challenging. So for you, the questions would kind of be, what do you think? Like, is it more like go to events, try to find users, get some revenue, or is it more like go to... events, find some friends and then work from there and try to create some like, what, is your feeling around this?

Sean:Yeah, that's a that's a good one. I think my mindset has changed about that recently. So before it was going to these really targeted events, right? So I want to, for example, get new customers and I'm gonna go to an event that greatly resembles with what my project is about. And I'm there to slowly try to get some users for my app. But recently it I I me personally I think it's more about making friends. So getting to know people, what they're working on, and actually building a relationship and not having these trans transactional I guess interactions is really nice because you actually get to know the person and even better, like you guys don't have to just talk during the workplace as well. You can go on a hike and actually connect with the people that you meet in events. But yeah, definitely talk to a lot of people and you'll find the person that resonates with you most.

Niklas:When we look at another project of yours, it's called GetMinds AI. What is that about actually?

Sean:Yeah, yeah. It's it's a company that I joined a while ago. It's actually the first company that I joined. And that brings me back to my point too. I met the founder at a founders gathering event and we sat down, got to know each other, and I visited his office once as well. And that was the best I would say the best experience I had in SF, actually getting to know someone. And Yeah, I worked at the company for half a year now. I just left a few months ago, but I had a really great time, learned learned a lot from the founder and really got to see what's happening inside a company. little bit about mines is that they do these synthetic personas of on anyone, honestly. So at at start at the start when the product was so vague and we were trying to clone these influencers or even professors, but eventually it was doubled down to just cloning customers and understanding what they want and what attracts them in a product.

Niklas:Besides that, also had like, and I think the AI will move a lot more in the space. Now you worked on something like bioscreen. think it's something you only worked on very briefly, but I think it will become a huge topic in the future. What was that about?

Sean:Yeah. I would say biotech is definitely a underexplored thing in AI. at least from what I'm seeing, maybe I'm missing some news, but I think there's a lot of potential there with especially genetics and proteins, 'cause these prediction models can you know they understand the real world and can make pretty accurate predictions on what could be possible. with maybe a new substance that helps you sleep sleep better or something that cures cancer even. So Bioscreen was a an experiment that I did with two very experienced C S students that had minors in biology and they were the master of both both domains and I kind of joined in as the high schooler that only knew how to make web make websites. But The project was creating proteins with AI and we used a dataset that people previously trained AI on and tried to detect these malicious proteins that were kind of disguising themselves as regular proteins that they can actually destroy your your body or whatever organic material we're injecting it into and Yeah, that's one one big example on how AI can be used in bio. But it's definitely an underexplored underexplored domain for AI.

Niklas:Yeah, I feel that it will be one of the more explored ones. my feeling is like the technically harder things now move into robotics and kind of biotech and the like a lot of what drove Silicon Valley like B2B software has become easier. still think it will be a huge market. So I'm not 100. So I believe in Jevons. paradox, I believe we will see a lot more software, but I also believe it would change a lot. Like the engineering will not be as hard. It will be a lot more like the UX and sales and distribution efforts that will be leading in it. So a lot more product work, less technical work and a lot more of the coding will be automated. That's what I think will happen there. If my last question for this podcast is, if there's somebody now who's a bit late to this AI track and they want to get into it, they want to get into agents, what do you think? Where should they start? What should they do if they had 30 days?

Sean:Yeah. It's a great question 'cause I'm trying to get my sister into AI and trying to get her explore as well. And it's it's it's really hard because they have their thing that already they're focusing on. So school, for example, is her. And two months ago I would say just start watching these YouTube videos on how to use AI agents, right? But that's what interests us. So that's why we're here. But they're not they're not interested in the technicals. So if you're And if you want to get into AI and don't want to burn out, then start with something you're interested in and see how AI can help you with that. So if you're into design, then explore these image generators or vide video generators and see what you can do out of it. And then you'll start to explore things like skills and open claw, different things that we're already familiar with. But don't start with the typical general path of learning AI, right? So what is an L L and then just the textbook way of doing things is definitely not motivating. So find something that you're interested in and then go from there.

Niklas:Thank you Sean for being on the podcast and to you my lovely listeners, see you next time.