Jon Scheele

APIs have been around for quite some time. So has artificial intelligence, but received a lot of attention lately. What we can see though is that AI and APIs are intricately interwoven because a lot of how you consume AI models is through the APIs. Very pleased to welcome Aki Ran1n to discuss this.

Aki has been in the AI data science space for quite some time and has one of the co -founders of Bambu, a robo -advisor based here in Singapore. So Aki, welcome.

Aki Ranin

Yeah, thanks for having me, Jon.

Jon Scheele

So just so that people understand, we've known each other for about five years or so since the first API Days Singapore Conference in 2019. But can you tell us a little bit about how you came to Singapore, but also what excited you about the data science artificial intelligence area?

Aki Ranin

So I actually started school in South Korea when I was six years old. So my father was a diplomat for a few decades actually, and then transitioned from that line of work into, I guess you could say like private consultancy around similar like import export work in Southeast Asia. So most of my life actually, my father has been living somewhere here, first in South Korea, then in Indonesia. I was there for a couple of years as well when I was growing up. And then he moved to Vietnam and eventually settled in Malaysia for 20 years. So to some degree, I sort of have been back and forth between Asia and my birthplace of Finland my entire life. But I did do my studies in Finland, which made a lot of sense because it's free.

And pretty good quality. And I was a computer science geek pretty much from early childhood, mostly computer games initially. And then I just had this idea that I want to work with computers for my career. And so I went to the Technic University, which is now Aalto University in Finland. But I had this thought in the back of my mind that it's just a question of when do I move back to Asia. And it took actually quite a few years. So probably sort of five, six years, at least of my career, early career I spent in Finland, but it just seemed like the economic situation was really challenging. It wasn't going in a good direction. Whereas in, you know, Southeast Asia, things were really booming. So it just seemed like, well, if I'm going to sort of invest my career might as well be in a place where there's like, lots going on. Ironically, things did change to a lot better in the Nordics after I left. So sort of when I look at what happened after I left it, you know, things move forward quite a bit in a lot of ways for the Nordics and Finland, but still at the same time, you know, it's been a great time to be in Singapore. So and, you know, I guess engage with the economic growth of the wider region. So overall, I think it's been a win.

But that's sort of the thought process of coming out here. And to your question on the data science part, what I actually ended up studying in computer science, I really enjoyed theoretical computer science, but I just didn't really see a career in that path. So this would have been sort of early 2000s when machine learning existed. And so, for example, at the university, we would have like robot soccer and sort of like rudimentary AI competitions of like chess playing algorithms and those kinds of things.

But I guess it was still seen, looked at mostly as sort of gimmicky. I think the only application that existed was computer vision. And so then I decided to pursue sort of that direction and went into robotics. And for a couple of years after graduating, I was actually working with robots in sort of factory settings. And the machine learning was really just like optimization for the most part, of some light controls and then computer vision for identifying bad packaging and those kinds of things for quality control. But the reality was that it just like the sort of curve of AI progress was so slow that on the ground it felt like nothing was happening. And the reality was that the pay was really bad. We were getting paid to do this pretty advanced work and sort of electrician rates inside factories.

And it was also pretty hard work, right? You're like in factory settings, it's noisy, it's cold, it's like a harsh environment. And so it just seemed like this is sort of, if not career suicide, it just felt like there were lots easier ways to sort of make progress and money in tech. So sadly, I think to some degree, I sort of moved in different directions, went into like ERP software and eventually just like web development and eventually like UX and mobile, but then sort of after moving to Singapore, I think that's when sort of mid 2010s, things started to really heat up and get interesting with like deep learning and various types of AI and machine learning. And so then when I got into my own startup journey, I really wanted to make that like a focus area and really get back into it.

Jon Scheele

I think that's interesting because as you say, artificial intelligence has been around for a while in terms of the concepts, but it took a lot of computing. There wasn't a lot of computing capacity available and data storage available to really utilize these models. And I have a similar sort of story. My bachelor's degree is in electronic engineering and a lot of my early work was in industrial process control. I do remember doing a subject on computer graphics when I was doing my engineering degree, but you couldn't actually do much with graphics with the PCs because there just wasn't enough power and not enough data to make good use of that. So I think computer graphics as well as artificial intelligence have really been able to explode because of the availability of cloud computing, both computing capacity as well as data storage and the ability to manipulate data. So we're in a position now where we can do a lot more. So what really prompted the decision to start Bambu.

Aki Ranin

I think it was more of an opportunity in a way that, you know, I didn't have any background in finance per se, but it just sort of happened organically that through some connections, this sort of opportunity emerged within the sort of role I was doing at the time. And so it just felt to some degree like, you know, it would be very risky, but you know, it was such as potentially a unique opportunity that it made sense to give it a shot. That even if it failed quickly, I could always get another sort of corporate job. And yeah, that was a sort of thought process is just to roll the dice, see what happens.

Jon Scheele

I remember reading an article you wrote once saying you shouldn't read business books. You should read the classics. And to be fair, there are a number of business books that maybe would make good short articles, but not necessarily an entire book. But I'm curious, which of the classics do you feel has really helped you in your startup journey?

Aki Ranin

Yeah, I think there's a few things to maybe pick apart. I think generally what I feel is, is when you're doing a startup, you're sort of trying to do something outside the distribution by definition, right? Like if the thing that you're creating already exists and it's just completely working, then the only thing you could do is just to maybe out price or out execute or something like that.

But most startups are trying to create something that doesn't quite exist. And so I think it sort of encourages this out of distribution thinking. And my sense is generally that like these business books in a sense, they are sort of capturing something in the zeitgeist that like by and large, mostly the things that you would read there are things you sort of have some idea about, you've heard about and sort of it's very iterative, like, this is a useful tip that I could apply here to do this thing 10% better. But when you're doing startups, I think the mindset is more like radically trying to rethink things. And therefore, I think getting into this mindset of sort of iterative improvement, may be counterproductive to some degree, you know, I could be totally off base here. I think there's many ways to do startups, but it's just a line of thinking that I found attractive.

And so, to sort of put myself outside the distribution and I found it more inspiring to read, you know, you could say classics, but also I think just things like sci-fi, whether it's, you know, sci-fi classics or more recent works, but it's just something that gets you thinking differently about the world. And then I think also on a personal level, if there are books that have stood the test of time that they're popular after decades, centuries, or in some cases, even like millennia, I think that says something that like, those are really the timeless things that everybody should know. And I think it's a shame that a lot of people sort of, they've heard of books like the Iliad or, you know, well, for example, the Bible, that's a pretty good one. So you're like, everybody knows the Bible and they've read some passages, whether you're religious or not.

It's not about that to me, but it's pretty much like the foundational text of Western civilization. And to not sort of have any relationship to it where you sort of even know the structure, the contents of it, I think it's a bit of a shame because we're sort of implicitly behaving, you know, our morals, ethics, everything is sort of derived pretty much from the Bible. And so to not sort of navigate that at all, I think is a bit of a shame. And so that's why I think it's just this mindset of trying to sort of look for the hidden treasure, which is often, you know, right there, just sort of avoid the trend and look at sort of the long term signal is probably just the mindset of that I've been attracted to.

Jon Scheele

And there is this concept of the hero's journey, which is a thread through a lot of stories. 

Aki Ranin

That's a good book also.

Jon Scheele

And you see this in popular books like The Hobbit and The Lord of the Rings and movies like Star Wars where someone is a pretty ordinary person, they don't know that they have anything special about them, and then they get thrust into a new situation, then over through a series of challenges they grow and develop and they realize that they have something extra that they didn't realize that they had. And then they go back to where they came from and they're different now.

I guess we're all going through that sort of journey. A startup is going through that journey. You maybe start with feeling that you're going to solve one problem and then you start to realize, well, in solving that problem, I need to solve all these other sorts of problems and I need to grow in that respect. So at Bambu, your model is helping to take away the complexity of advice by encapsulating, you're doing the hard work with building the data model and then exposing it to other companies through APIs. I mean, it's really interesting sort of a model. But what you talked about at the API Days Singapore 2024 was about this challenge that we have a lot of local optimizations of many activities that are being made possible with AI and large language models, but you touched on a more integrative approach. And I think that although we have a lot of tools now for building APIs, we still have the human there who's making key decisions about how to map one field to another field. So what do you see coming to the fore here?

Aki Ranin

Well, yeah, so a bit of context that actually a year ago, the sort of AI platform that I was building at Bambu was acquired by Goldberg, which is part of the Franklin Templeton group. And so I left Bambu to pursue that. And in fact, I think it sort of addresses your point because a platform like Bambu and sort of the AI tools we were building. I think you could think of them as sort of a collection of intelligent APIs using different forms of machine learning to do sort of specific tasks in the realm of financial planning broadly. But what's interesting about large language models now in AI is that, it sort of provides the glue because previously the glue needed to be like a user interface that you'd have to click through like a lot of screens and buttons to consume and interact with the intelligence in a way. But now what we've been actually working on recently and beta testing is this idea of just giving the large language model of these APIs.

So, something I mentioned in my talk as well is, we've had the sort of concept of tool use with large language models for quite a while with the open AI ecosystem and their platform. But others like Claude have recently entered and I found it really sort of thought provoking where they said that their API is able to handle 200 different tools in parallel, which means that, if you took a decent piece of software, certainly enough to cover what we were doing at Bambu. But you could take something much larger and say like, hey, these are all of the APIs and just recreate to some degree, the sort of overall interactions that you would be having with this suite of software and then just do it, I guess now by mainly text.

But I think as soon as we then move into this sort of voice paradigm and sort of the Omni models as you will, where they can look at screens and video and audio and text at the same time. I think it's gonna be like say in the case of financial planning, I think it really moves us into a much more compelling place where we can genuinely start talking about like AI assisted financial planning as opposed to just giving you a bunch of tools to play with. And that probably applies across a lot of industries and I feel is just like around the corner and potentially say this is that year where we start moving from sort of just like small little chat bots to much more sort of embodied AI systems that have like a large body of sort of tools and automation under the hood, which makes them much more sort of feature rich than just your basic Q&A.

Jon Scheele

So when you're setting up an organization to solve a particular problem like financial planning with the capabilities of these different large language models that may be not your own service but other people's services.

What do you feel really needs to change to make this a possibility for somebody to ask a question and then have the AI agent go and find the information that you're going to need?

Aki Ranin

Yeah, I think it definitely requires quite a lot of plumbing. I mean, if you're following the agent space, I think one of the bigger recent splashes was the Devon software development agent. And it's not like most of these companies are creating new models. For the most part, they are using whatever is the state of the art model, like GPT -4, now GPT, Omni, et cetera. But it's more that they're orchestrating, you might say, instances or threads of these AIs and managing a bunch of data going in and out of these prompts. And so ultimately, in some ways, you could say it's just like really fancy prompt engineering code that they are commanding, you know, the state of the art AIs in a very specific way.

But I think that may be less necessary if you just have better AI, where you could just tell the AI sort of like either explicitly to deal with these issues or that it just sort of knows by itself through its own intelligence how to deal with this. So if it feels like, hey, you know, I've been given a fairly chunky task here. Maybe I should spin up a few local copies of various helper models that can do certain side tasks for me and sort of, you know what I mean? That now it's very difficult. And so that's why we haven't really seen this working at scale. It's not that people aren't trying, but it is just very difficult to do it robustly, accurately and reliably and also cost efficiently. Because, you know, if you said, I can give you 10 hours and a thousand dollars to complete this sort of relatively simple task that a human could do in a few hours, yeah, you could probably do it. But again, what's the point of doing that? Because a human can just do it faster. So it does need to be sort of substantially faster and better than a human for it to make commercial sense to use this. So I think we're now in this sort of gray area where some things are starting to be possible, huge amounts of research and focus into this.

Like just the other day, I saw the news that yet another French AI startup, it's called H raised a few hundred million seed round, twice the amount of Mistral, the previous benchmark. And they were strictly focused on agents. So I think people's mindset is moving from large language models, really cool, but there's only certain things we can do there. Now we have to focus on agents because that's going to really allow us to do the bigger things that we want to do.

Jon Scheele

Okay, so since the explosion of the World Wide Web during the tech boom, discoverability has been a huge goal. Even today, "Can Google find us?" is a big challenge for anyone who has a web presence. Now, if we're providing a product or some data that's available through an API, we have the same sort of challenge; and API documentation has been a challenge for some time. So if we want to make sure that one of these AI agents can find us or find our service, what are the sort of things that we should be thinking about? What sort of metadata do we need to create so that the AI can find the information? Even if we're actually already connected to an API, the AI agent may not be looking at our particular service because there are several others to choose from. So what are your thoughts on how we should be building things for that sort of consumption by the AI agents?

Aki Ranin

Yeah, first of all, as you mentioned, the internet, which is the main source of training these AIs; it is known that you can attempt to sort of blacklist or block those crawlers, whether it's OpenAI or Anthropic or other labs. So you can try to block them from basically effectively taking for free your site's contents for their training data. So that's something that's already happening. Most people are not aware or I guess don't care about it. So they just let it happen for the most part. Although you do see now that like some of the big labs like OpenAI entering deals, for example, the last one was OpenAI and Reddit so that they're effectively buying user data without really, I guess, technically user consent. And so same thing for stack overflow.

And you even saw sort of a counter reaction where people were deleting their accounts en masse in protest and then stack overflow was trying to prevent people from deleting their accounts because that's the valuable data that they're trying to sell. So I think it says a lot about sort of the web 2 paradigm as it is that actually all this content that people were creating for themselves for free.

Before the call started, we were just chatting about Substack and Medium, and all these other blogging platforms. But the reality is whether it's Google or Instagram, Facebook, et cetera, you own none of the content. The terms and conditions are very clear that the platforms own all the content. And now, as it turns out, if there's money to be made by selling your data, they're pretty happy to do that.

But if you're thinking of it the opposite way, which is like, how do I expose my data to the AI more efficiently? I was just listening to somebody else thinking about this problem. And I think it's just a question of intelligence. Like right now it's a bit tricky. So, you know, these AI systems now, like the latest models, they tend to have sort of guaranteed JSON outputs. So generally speaking, in terms of reading APIs, API documentation, or even creating the documentation, I think those are all kind of pretty doable now.

But once the intelligence gets a little bit better, I think that may be less relevant. And it could be more a question of what's the fastest way for the AI to consume information. And for example, you know, it's an interesting question. Video could actually be pretty good if you've got audio transcripts, cause it can just zoom through all of that text immediately. Whereas like very structured, like animated, complicated websites with a lot of JavaScript moving things around is probably less ideal because then the AI sort of has to wait for something to happen. Whereas if you could just sort of have all of the, whether it's text images or things directly consumable without having to, you know, sometimes you see those scrolls where you have to scroll before more content is loaded. I think all of those kinds of sort of human conveniences are probably hindrances for the AI. So if you want to make it easy for the AI, I think you almost just like, you know, the most extreme would be like, just give it access to the database directly, which could be a bad idea in certain cases.

Jon Scheele

Okay. Well, thanks very much for sharing all of that, Aki. Just sort of parting, if somebody is looking to improve, develop in their career, understand the trajectory of AI and APIs better, what sort of things would you recommend to them?

Aki Ranin

I think there's a big difference between whether your ambition is to understand where we are now. And for that, I think there's a lot of good resources out there, starting with like really good YouTube tutorials explaining how transformers work and just getting up to speed on sort of how did we get to GPT -4 type of systems. And then, whether whatever topic you're interested in, whether it's like AI safety or the algorithms or data security, there's so much content online.

But I think the tricky part is that, is it worthwhile chasing the puck or should we think about where the puck is going? And so that's, you know, Sam Altman himself keeps talking quite a lot about this, like for startups that if you're building a business model on top of GPT -4 that's a pretty risky enterprise because we know they're working on GPT -5. So it's just a matter of time before that comes out. And if your entire business model is sort of built around, for example, the limitations of GPT -4, the kinds of things we were just talking about, like, let's make better plumbing so it's easier to read APIs. Well, what if GPT -5 just like does all of that for you, then you have no business model at all. So he talks about this like steamrolling risk. And I think if you're getting into the sort of AI world from the outside, I think that's something to think about. You know, how much value is there just understanding where we are versus thinking about where things are going. And if you're more into that, then I would probably recommend maybe some authors like Nick Bostrom, Max Tegmark, you know, Life 3.0, Super intelligence, those kind of books, which are trying to paint the path forwards.

And of course, some great podcast resources. I recently looked at Dan Fagela's The Trajectory podcast and there's like 80,000 hours, which has been around for a long time, talking about like issues of like AGI, AI safety, how are we gonna deal with all the issues when we get to like really strong AI. So if again, if you sort of wanna look at like where things are heading, I would recommend the more philosophical route. Whereas if you want to look into the technicalities.

That's much easier because that sort of content is all over the internet, whether it's Coursera or all the free courses you can take online.

Jon Scheele

Thanks very much for sharing that, Aki. I really appreciate the conversation. We're going to link to the recording of your presentation from API Days Singapore 2024. I think I'll also probably add the one from 2019, because that was in 2020, I think you spoke when we were online. But that's been really, really helpful and really appreciate it.

Aki Ranin

Thanks, Jon. Glad to be with you.