"Let's say you want to do logistics. AI can help with a lot of optimisation, specifically with complex models. The question is, should you only use AI for search? No, you should use it to optimise freight and path costs, factoring in tariffs. I had to do a routing, and it turns out O1 and O3 reasoning engines are extremely good at this. I did a demo for a retail business—they gave me all their constraints, and ChatGPT generated a highly optimized path that matched exactly what their experts expected. They said, 'That's correct.'" - Bernard Leong, founder of Dorje AI and host of Analyse Asia
"What AI has done is it destroys the ERP. First, traditional ERPs like Oracle and SAP require you to conform your business processes to their logic. With generative AI, you don’t need to—you can customize everything through code generation. The new business model should allow every user in the company to use the app. With proper access controls, finance handles the accounting, and everyone else should not have access to the company’s P&L charts. Access control remains a very human responsibility." - Bernard Leong, founder of Dorje AI and host of Analyse Asia
"My wife was an ex-lawyer who left the legal field, and one of the things she was really worried about was AI. After she saw what ChatGPT can do, she told me the way to train junior lawyers is gone. It's the same with Dev houses. They’re being paid by the hour, but now that hour of work is being shrunk to five minutes, and clients aren’t willing to pay for tasks like reading a Word document because AI can do it. The essential question is how to retrain young professionals in this new paradigm—this includes doctors and professional services like consultants." - Bernard Leong, founder of Dorje AI and host of Analyse Asia
Secure your privacy with Surfshark! Enter coupon code BRAVESEA for an extra 4 months at www.surfshark.com/BRAVESEA
Bernard Leong, founder of Dorje AI and host of Analyse Asia, joins Jeremy Au to explore how AI is transforming software development, business models, and professional roles across Southeast Asia. They break down why dev houses are losing ground, how AI accelerates coding and reshapes team structures, and why traditional SaaS and education models must evolve. Bernard shares how he replaced an outsourced dev team using AI tools, the dangers of hallucinated code libraries, and his vision for a new enterprise software model powered by prompt engineering and cloud-based trust.
00:42: Traditional software development can’t keep up with AI timelines: Bernard shares how he replaced a dev house that took five months with a feature he built in 20 minutes using 50 AI prompts during a flight. This led to firing the team and redesigning the internal workflow around speed and AI tools.
06:26: Frontend moves fast with AI, but backend demands real engineering: While vibe coding speeds up prototypes, Bernard highlights backend risks like hallucinated libraries from ChatGPT. He stresses the need for strong DevOps rules, audit trails, and secure infrastructure to prevent system vulnerabilities.
09:18: Dev houses need to reskill or become obsolete: Bernard criticizes dev houses for slow JIRA-based processes and poor QA. His lean team rebuilt what took five months in just six weeks by focusing on code quality, automation, and prompt engineering. He urges retraining junior developers to stay relevant.
20:43: AI is replacing repetitive junior roles across professions: Bernard sees AI displacing junior coders, lawyers, accountants, and consultants. He shares how his ex-lawyer wife saw this coming, and cites an MIT study where only senior professionals could spot and fix AI mistakes, while juniors added little value.
23:39: Education must shift from banning AI to measuring real thinking: Bernard describes showing students how ChatGPT completes their essays in seconds. He calls for testing reasoning and prompting skills rather than memorization.
31:57: Organizations will become lean, AI-native teams: Bernard predicts companies will move from pyramids to diamond-shaped org charts. He now trials contractors and only hires those who scale with AI.
Jeremy Au (01:10)
Hey, Bernard, good to have you on the show! Yeah, so I thought this would be a fun conversation because obviously you run the Analyse Asia, it's a huge podcast across Asia. So both podcast co-hosts, but also we're both people with both founder as well as kind of like executive experience. And right now you're also building an AI company.
Bernard Leong (01:11)
Thank you for having me on the show, Jeremy.
Yes?
Dorje AI
Jeremy Au (01:32)
Things have really changed over the past year. A full disclosure, with angel invested in you at a pre-seed stage. And then now you actually executed so many of those milestones since then. So I thought it would be a nice way to also catch up on what you're seeing and experiencing as well.
Bernard Leong (01:45)
Yeah, I've sent two investor letter updates, but we can talk a lot about what we have done so far and maybe we can discuss a bit of how the current AI industry is moving and specifically maybe looking at some things like business models and go-to-market and even development as well.
Jeremy Au (02:02)
And so what's interesting is that, you know, our last podcast about a year and a bit ago and now it's today. So things have changed. So what would you say is the difference between then and now from your perspective at a high level?
Bernard Leong (02:12)
So there is a very good quote by Sam Altman. So if you want to be an AI startup today, you need to have two mindsets. So you need to project in 12 months time. If any of the foundation AI models will have your ChatGPT, you have your Claude, you have DeepSeek, now X.AI Grok, or even Gemini. In 12 months time, are you going to be happy with the improvements in the foundation model? If that is improving and your startup is going to thrive because of that,
then you're in business. But if in 12 months, any of the features that you have done has been replicated or what people call chat GPT wrappers, or you're wrapping around the foundation model and they start cloning your features, then you're in trouble. So if I were to look at the foundation models in the past 12 months, one big thing that's really happened, I think, is in the coding side, the coding development. I suspect that 95% of development houses in our market have no idea
that this is too slow for what is to come in probably the next 12 to 36 months.
Jeremy Au (03:16)
And I think that's interesting because obviously there's a huge number of software development houses, oh, I'll say, those that are a little bit more like premium, streamlined, tight, as well as those that are more like, I say, remote outsource, maybe some sort of diaspora group across Eastern Europe or Philippines, for example, or Vietnam. So are you saying like, you think this whole industry has to change?
Bernard Leong (03:40)
They have to change because the traditional software development model currently is you developed, you have an agile project management system where you actually have to ship the code and then there is what is called a QA process, a quality assessment, and then you push it in through something called a DevOps process. What has really broke open this is that if you look at Vibe coding, or what people now talk about Vibe coding product, the part where you take from the business requirements
to just to show a prototype and or pilot product has the gap has shrink a lot. So today a business product manager cannot give me any more excuse that they cannot quote this. Very interesting. I was talking to one founder where I just demonstrated for sale and cursor to the head of engineering and to the product manager. And a few weeks later, the founder came back to me and said, Hey, you know what? My product manager said, you don't need the front end developer. I can do it now.
that the shrinkage has actually shifted in for the Dev house. think they are trying to be too slow or they basically try to pitch to any startup, say, this particular feature usually takes this amount of requirement and this is the time I need it. And then they realize the startup founder is getting very impatient with them. The second part I think is more subtle, I think a lot of people don't realize, is that it's the backend piece. So this is where you quote this really beautiful product.
Except where do you put your data? So if you're vibe coding, Vercel, because they want to give you instant gratification, they put it into a big spreadsheet. And suddenly you realize that you have this CSV file within your whole system that has like 10 gigabytes. I'm joking here, but usually for something like Vercel, you probably could do only up to 50 megabytes. So it's like, why do you have a 50 megabyte CSV file? It's because there was no backend. So you are not putting into maybe a good database like a DynamoDB, et cetera.
So there is a backend mismatch. But there are companies like Devon, you probably know of, there is Cursor. So you, so this mismatch is still not resolved. And I think the backend side is the reason why you hear everybody is so excited on LinkedIn talking about, know, I vibe coded this, et cetera, et cetera. And then after a few weeks, you'll find that everyone's already depress because they really didn't know what's going on on the backend. And ends up that the product is only
a pilot or POC but cannot go to production.
Jeremy Au (06:08)
And I think what's interesting is that obviously this AI push is generating a lot of code for the front end and so so forth. That's right. So I guess is the implication, for example, like humans can stay doing backend or is the implication like AI will also come for backend and there'll be companies that also increase productivity for backend. How do you see that fail?
Bernard Leong (06:27)
So the way how I think where all these AI coding apps are now going is there are two extremes. So the first extreme is the 0 to 70%. So this is like your non-engineering background founder or person with no coding background, they get up to the product up to what is called a 70 % mark. But the 70 % mark is not really for production. So they will go around LinkedIn telling everybody, oh, how great my product is, blah, blah, blah. Then there is the remaining the 70 to the 100%.
That I think you still need professional software engineers. You need really hardcore DevOps engineers, hardcore backend engineers. then the other extreme is where you have an engineer that's already a 10x engineer that's now putting all this AI into his actual day-to-day productivity. And he becomes a 100x or 1000x engineer. So this part you don't see a lot, but there are some engineers on that. So I have actually been coding.
in my life. Okay. So I started coding since nine. I used to get paid a hundred dollars for coding Kobo programs, you know, that was like 30, 40 years ago. And I changed my programming paradigm from just traditional coding with simple databases to object oriented programming when I did my undergrad. But I still code for fun. Then to now where we talk about things like containers and all these different DevOps stuff, right. I have to relearn every time. And when I first
Open up cursor and I tried to do vibe coding myself. I also realized that it took me. So for me it's this, will actually extremely go and learn something all the way, like two, three days and then test everything that I can. And then maybe I'll hit a lot of road bumps. And after that I was like, okay, I need to do this step by step. I suddenly realized that vibe coding has basically upended all the other automations I've done in the last many cycles. Yeah. I was like, oops, what just happened? Then I started to realize, okay.
There is something here. You have to rethink how you want to do software development. So maybe I'll tell you a story from Doge AI. in order to build one of our MVPs, we contracted a dev house. So we have a set of specs. We want to get the requirements done very, very quickly. What happened was they took five months and they couldn't get one feature that was ready done for my customer for meeting. So what I did was I was actually in Malaysia where I'm trying to get them to
finish that third feature and they say, we have one more week left, can we do that? And they said, no, we can't do it. So on my way back from Malaysia, I obviously on the plane in Asia, there's no internet. So I was already experimented a lot of vibe coding. I have some sense I actually could code out very simple prototypes on my own. What I did was I took that feature and I already have a vibe coded version of the same user interface. And then I just...
in my mind, write out all the prompts in order to code up that feature. So there were about 50 prompts in my text edit file. I was typing on the plane. I landed. I went back home. I'm supposed to go back. Then my wife, as you know, asked, I said, OK, you're going to go and fetch the kid. You have only 20 minutes more. So within that 20 minutes, I actually used cursor, turned on everything, and it worked. OK, maybe with a few things, OK? And then at that instance, I just have to.
basically went to WhatsApp, talked to the dev team and said sorry, we have to terminate.
Jeremy Au (09:47)
Wow, so you're telling me that on a flight you did, you figure out the 50 prompts and you finished the whole product, 20 minutes. Yep. And then you managed to pick up your kids.
Bernard Leong (09:56)
Yes, I went to pick up the kids. This was the This was like, it was like when I saw every, that was when I saw everything work. However, I'm not going to go to LinkedIn and go and brag about it, but there are some issues that actually comes with coding it this way. So the first thing what we did was when we terminated, we did a transition from the dev house. So for all those out there, this is not a magic moment. There are a lot of hard things that actually go through. I was also just hired a product manager on my team. And what we did was we didn't
use whatever I vibe code to actually build out this thing. What we first need to set up the rules was the rule of governance or in the DevOps side. And it turns out that my product manager was actually a DevOps engineer by training. The first week when I on-boarded him, I taught him Cursor and he said whatever he was used to be working for five months on a DevOps project has now been shrunk down to two days in DevOps work. So what we did was we designed a DevOps process and my co-founder and I, we split.
He does the backend, I done the frontend. So what we did is we vibe coded and what his task was as both the product manager and also the DevOps was to make sure that he has to be strict with us not to use libraries that are not previously known. So one of the things that I think people don't realize is in vibe coding, a lot of the tools, whether it's Vesal cursor, they're using very, very...
they're pulling libraries from all over the place. In fact, there's a way to hack you now. Basically, some hackers found the hallucinated name libraries in ChatGPT and create malicious code in these libraries and push it into the Python repository. And then as you pull these files with the hallucinated names, it actually triggers an entire takeover of your app. So what we have to do, and this is where the subtlety comes in, right? The back end side, you need to actually put
all the rules in place. The problem also, which I also discovered talking to a lot of software development people that I know across, even with some of the unicorns is that what they discover is that when they do that, the DevOps engineer actually push them back, push them back a lot. And I think what is happening is that there is a softening that needs to happen. So I told the PM and said, okay, here's what you need to do. When we use all the libraries that we use, if it works, we pass.
And anything that seems to work well and allow us to be more agile, we create it and turn it into a cursor rule. Anything that is hallucination, something that is not giling very well, maybe you require some very odd libraries that those are the rules that we have to set in order not to allow those to be pushed through. After all, I Dodger AI, we're doing a financial automation system, a ERP system. So trust is extremely important. And hence the amount of error that I can get is actually must be very little.
So we spent probably two weeks to make sure that the entire backbone is properly set up. And then we allow, and then we slowly to allow that. And I can tell you the five months of work they've done, we, this is already, think month one and a half, we have already replicated 80 % of what they have. Just I code it, they code it, all of us code things. And we even tested more of the more modern development tools. Winserv, Cursor, I think there was, now we are actually trying to do Devin, but
To be telling everybody here, please do not use Devin to generate code, but use it to write your test, the QA test, the audits to check whether the app is working. It seems to be better at that. It seems that that's the best use case for the time for what people are actually doing.
Jeremy Au (13:19)
And I think what's interesting is twofold, right? One is what it means for the company and obviously what it means for DevHouses, right? Maybe to tie off the DevHouses piece, what do you think is the evolution or how should DevHouses, should they just give up? Do you think they need to become more senior and fire the junior people? What is your advice to DevHouses?
Bernard Leong (13:38)
If you fire the junior people, then there won't be enough senior people in the future. So the way you need to train junior developers is slightly different. And you need to, for the first time, really rethink how you actually do software development. There's a course I went, I actually did in seven days, it's Generative AI for Software Development. It was done by Andrew Ng's team from Deep Learning. And one of the things that really gave me the big paradigm change was actually
the way even you want to do the DevOps. How do you want, how do you smooth out that process? And I find that 95 % of the DevHouses here, are, because we actually, because, okay, I'm probably one of those few business founders that's technical. So when I actually look through the DevHouses five months work on their JIRA and their processes, I find that they were very slow. And a lot of it seems to be locked in the QA process where they claim they were using chat GPT.
I don't think so. If you have been using ChatGPT Winserv, the speed of actually even showing me the dashboards should be much, much faster. In fact, we've replicated the front end in less than a week. And actually most of the work that we are now trying to do is to make sure that all the functions are working. To also tell everyone I'm not a SaaS company, my solution is actually we orchestrate on customers cloud and we do a pay as you go model, which we are going to come to talking about the business of the...
enterprise AI where this is going and then we can move forward from there.
Jeremy Au (16:35)
So
before we go to the business model, I think you're saying that you yourself had to change how you code. You said that the dev houses also change how they raise talent and also how they code. Yes. So what do you think is the before versus after? Because for me, before, like you said, edge out, stand up meeting, you need to have a product manager define the specs. You have to lock the document to some extent. Yes. then, so what's the before versus the
Bernard Leong (17:03)
So the product manager now is the most central, also has to work with the DevOps person to set what are the rules that you can allow Vibe code to be true or Vibe code to be out. What libraries to be used, what libraries not to be used. If you have to code something that works but requires you to run around the lines, is there a faster way to do this? How do you audit or check whether there's any malicious code? I think these are the processes where there is a softening.
It's still humans that are there, not the AI software engineers that's going to take your job. In fact, surprisingly, the AI cannot take away that job. Now what I'm going to say is that if I want to take this to a logical extreme, I think 95% of the software dev house teams are going to be in trouble because I am seeing for most of the AI startup founders now is actually very, very small internal teams, three to five men teams. I just want to get one more backend engineer.
and our burn rates are extremely low. just to be quite clear about it, the burn rate for the dev house was about twice as much. Now I have strength down and we are sending the investor letters. One of my investors was asking me, you took your burn rate down by half. Are you out of your mind? I'm like, no. Then after I just opened up the laptop and show him exactly the version before and the version after and he was like, okay, I get it.
Let's say I would like to have another 15 more months of runway. So I don't want to, until I get my first customer, I'm not going to waste a single resource. I think Dev Houses always work with the mentality, hey, we can drag on the project. That's the business model, right? We code you by the hour, but sorry, your coding by the hour is gone already because the front end design end is more or less slaughtered by Versailles and Figma. Because now I can run a very simple, let's say if I can get someone to draw up.
the sort of basic thing. And then I just put a model context protocol on top of the Figma. I could just tell the Claude to say, OK, based on this process, can you help me to code this? I would like it to be in this structure. And essentially, you can get that app done very nicely. And I think this is where the DevHouses do not really understand.
Jeremy Au (19:12)
And so how do individual quarters like yourself also change because you said that you had to throw away some of your old automations in your old.
Bernard Leong (19:19)
So you have to throw away that whole thing about everything must be audited line by line. You have to be a bit more holistic about how you code. You cannot say things like, you know, usually we need to separate out the libraries, et cetera. Yes, those structures are there. You can prompt to say, I want all these structures, but you cannot say that this will take how long on there. And there is this argument from software engineer. And I appreciate that because I'm also technically trained is that
you want things to be much more robust, resilient, high availability. All these buzzwords about foundation being robust. I think the problem is going to be that a lot of these engineers need to rethink what robust means. How fast to go to deploy to production. That's because they are not thinking in the AI timing. They are thinking in the traditional software engineering timing. I do not know. I spoke to a very, very good
technologists and he's also doing exactly the same thing as I do and he's a 10x programmer better than, no probably 100x better than I do and even he was saying that the way how you do the agile piece has to change because a lot of the front end is going to move very very fast. To be quite honest the DevHouse used to have this problem of trying to do instant gratification. The answer to the question is they can actually already give us the instant gratification. So I don't understand how much quality control you really need on the front end.
The back end is actually where quality control matters, where your data really moves, whether it's secure, whether is the plug from here to here is working. I think those two states, I think what they don't realize is that the whole development cycle is now screwed up, that not everything is equal in part. And I think that that's where the mindset change really have to come about. I always recall Jeff Bezos, who I used to work for Amazon, right? They say, your margin is my opportunity.
And I almost wanted to like, well, maybe I should build a deaf house to destroy 95 % of those businesses. Yeah, sorry. it's just like, wait, wait, wait. I just started to realize this is like, now I was asking some startup founders like, hey, maybe we should all band together and set up some, some deaf house, uh, on that can do it that way. I was like, yeah, because it's, it's, it's kind of, it's also telling me that there is a new generation of Dev houses that's going to come
Jeremy Au (21:33)
Yeah, interesting.
Bernard Leong (21:34)
Yeah,
for any of your audience, please take up this idea and build it. I think I'm happy to be angel investors, I'm sure.
Jeremy Au (21:41)
Yeah, it would be happy to be used.
It's like, we don't need to fire you, we just hire you.
Bernard Leong (21:45)
We just hire you and contract. You don't know why we have the business model. I'm trying to convince one founder, ex-founder to do it at the moment. And I said, I'll really write in the chat if you do it.
Jeremy Au (21:54)
I think what's interesting is that, okay, so the code piece is an interesting piece, whereas the celebration, it feels like the benefits are going to senior coders rather than junior coders. Is that how you feel? Because I also feel that for some of the AI software that...
Bernard Leong (22:11)
Yes, exactly the same. It's also the same with legal as well. So Harvey is taking away the jobs of the junior lawyers. Well, my wife was an ex-lawyer and she left the legal field and one of the things she was really worried about was AI. And she, after she saw what ChatGPT can do, she was like telling me the way to train junior lawyers is gone. It's the same thing with the Dev houses. They are being paid by the hour for their work. Now the hour of that work is being shrunk down to five minutes and the clients are not willing to...
pay them for reading through the entire word document, maybe a 20 page word document and sieve out the lines, right? Because the AI could do that. So the essential part of that thing is how do you retrain all these young professionals in this new paradigm? This will also include doctors, professional service, like consultants. You see, developer, the lawyer, the doctor and the consultant, these are all being disrupted because the way how they would...
the business model that they use to build their clients have now been disrupted by a technical AI agent that can do exactly the same, but maybe from 60 minutes down to five minutes. that, you know, 60 divided by five, that's 12 times, right? That's 12 times productivity. Then if you are a business owner like both of us, and it's like, hey, how should we pay for this junior lawyer who's reading our contract? Should we be paying the $60 or should we be paying the $5?
Jeremy Au (23:36)
Right. And I think it's exactly fair, right? Because also I was reading this research paper by MIT about AI software with accountants. And then it was more looking at junior versus senior. Correct. And the issue was that the senior accountants, when they saw that the accounting was wrong by the AI agent, they were able to challenge the agent and improve the output. Or if the AI agent was unconfident, because it's unclear, then they were able to step in and do the work. Whereas junior people,
they were unwilling to challenge the AI as well as they were unable to fit in for the AI when it failed. So this is not adding much incremental value.
Bernard Leong (24:11)
But then it also may become that we need fewer lawyers and we only train the best of the best to lawyers. I mean, at one point, if we think about some other professions, right? I think what's happened is that we lived through the last 150 years, maybe there is in the industrial revolution, everything is factory based. Now we are moving back to the, I would say maybe middle ages where a specific job is actually a specialized customized cottage industry world that can be customized by AI.
And I don't know whether our industrial factory setting can actually accommodate this new paradigm.
Jeremy Au (24:48)
No, I think it's entirely true because I think historically, you always think about work as a pyramid, right? Which is, one person can manage seven people and seven people can manage seven people.
Bernard Leong (24:56)
Nope,
I have a quote for that now. I was making, I'll say this. A players with AI hire A players. CD and E players are replaced by AI agents. And my dream is like, do you watch solo leveling? No, no, the Netflix show is about this young guy went from a very low rank hunter all the way to S class. You can think of entrepreneurs, you know.
Jeremy Au (25:11)
You're talking about the Amazon show?
Bernard Leong (25:21)
your S class will be Sam Ottman and Elon Musk and then your E class will be people like us, know, sorry, maybe for me. And then you're going to go to the S class and then by then the whole show was that this guy became, the protagonist became an S class hunter. And in the show, one of the things he could do is he can just call up his army, magical army, and he just say, arise. And then this hunter will show, this hunter monster will show up. And so I was thinking, yeah, you know, actually AI agents are like that. I knew well, I probably would. I think the
one person unicon, the one person, one billion dollar company will actually be somewhere maybe my son's time, et cetera. Maybe somebody maybe even within our time. But I think even for me, I'll say, well, I think we still need about a hundred people. But what I would love to do is every morning I could just say arise and then, know, there's hologram would come up and the HR person, what do you need to do? Hey, I have a new employee coming in. Can you help me to prepare his contract? Everything. And then.
arise and then you tell him, is, can you help me to make sure all my CRM calls, who do I, who, who, who, which customer do I need to meet face to face on there? Yeah. So essentially you could think of your entire army is all run basically AI agents. think that is the ideal world I really want to get.
Jeremy Au (26:35)
Yeah, I I think from my perspective, would be like instead of the classic pyramid of, you know, seven times seven times seven, which is how the organization scale, I think almost organizations are going to go back to a diamond shape. you know, think obviously there'll be a CEO leadership team that's at the top. Then there'll be a lot of managers, type who are comfortable using AI, outsourcing, fractional work, and so forth. And then like you said, only at the bottom, the people who used to be all the interns are gone. It's only the best of the best, the people who are super hungries.
Bernard Leong (27:02)
No, I will still hire interns, but if they can progress a X, if they are the 100 X players, I would want to hire them. In fact, it already changes how I want to hire my teams. So what I would do now is I would rather work with contractors. If they prove their worth, then I will give them a full employee contract. And the contractor part is actually I'll try and buy situation. So we don't have this thing called a probation period. We just need to have a contract that's fair between me and the contractor.
If it works out, works out. If it doesn't, it doesn't work out. So it's better to do this than you would go through all the unnecessary parts that actually doesn't make sense, where some of this can actually be replaced by AI agents. And I think this is what society is not ready for yet.
Jeremy Au (27:43)
So, you I think when you say that society is not ready, what do think they're not ready for?
Bernard Leong (27:48)
They're not ready for the change. So part of being an entrepreneur, sometimes you have to, I teach AI, I teach AI to corporate leaders. I teach AI to government leaders. And funny, the first question always come back from all of them. I recently taught one for YPO, by the way, for one of the chapters in Singapore. even there and even everywhere, no matter where I go, I went to UAE to teach recently. I'm actually going to go to UAE again. And the first question, all the CEOs come to me is the first question.
What should my nine-year-old be doing? And then they were like, can you take your course and condense it down for my nine-year-old so that I know that they will have a job? So I think that a lot of people are realizing that once they see how these tools are being used, they're worried about their kids going into the workforce. And I think the way how universities trains, and because I sit in...
when talking to academics that we need to prepare exams for students. And then everybody's like, now they're using ChatGPT to cheat, what should we be doing? And then they are like, and there was always one comment that will come up. We need to rethink how we want to do it. Go to action. then I just told them, look, I have just redesigned my assignment for the next term. You will see my assignments are going to be different.
Jeremy Au (28:56)
I agree with you.
Okay, so wait, so let's talk about it two ways. One is obviously teaching, but how are you teaching the students differently? And then eventually I ask about what
Bernard Leong (29:13)
Okay, I tell business school students first thing in their first class, I will turn open up ChatGPT. I'll take their individual assignment and I'll throw it into ChatGPT and say write me an essay. They will see right in front of me and they will be like, gap sync. And then I say, if any of you try to do this, I'm going to fail all of you. Yeah. And then they got worried about it. And then the next thing they will be asking me, Bernard, no, can I use ChatGPT? I write out the most of the essay and get ChatGPT to improve for me. I say that's fine because you have a train of thought and ChatGPT cannot
mimic your train of thought but I have to revise that assumption now because of 03 and 04 all these new reasoning engines coming out right you can't tell right then you have to think about okay what is the best way to test them to know how they're using AI accurately so you need to set an exercises where you have to figure out how they are prompting effectively
Jeremy Au (29:49)
Yeah
I don't know. I felt like the other way to do it is like give them multiple choice in person.
Bernard Leong (30:06)
No, no, no, no, no. Yeah, that's what some academics say. Then you end up with one CEO coming to me telling me, I went for a parent teacher day and the principal tells me that we are going to ban AI in the school because we think we cannot allow them to outsource their reasoning. And he said, I actually feel like changing him out of that school because this is like banning me from using a calculator for my job.
Jeremy Au (30:32)
Yeah, so what's the...
Bernard Leong (30:34)
The answer to the question is not to ban, is to say everyone should be able to use this tool, except that the assessment tool should be different.
Jeremy Au (30:44)
OK, so the assessment tool should be something that cannot be done by the AI doing it. Correct.
Bernard Leong (30:49)
And
there are still a lot of things that cannot be done by AI if you were to take the actual evaluations. I know everybody read the front headline about how OpenAI, cloud today, 97 benchmarks for MMLU, Helm, all these names for evaluations. They always hit 96%, 97%, right? There is something called the Humanities Last Exam, which is actually taking 50 to 60 fields. And there's a lot of reasoning things involved. This is sociology, literature, history, together with the hard science, physics, chemistry.
Until today, none of them, only think ChatGPT 04 could only get up to between 20 to 30%. Humanity's last exam is pretty difficult.
Jeremy Au (31:28)
What is the humanities last?
Bernard Leong (31:29)
It's
about reasoning. It's about how you think. It's the ability to create new knowledge. If you look at AI today, it is not ready to create new knowledge. I can tell you a story. have a... Well, I did my PhD 25 years ago. There was a problem I couldn't solve. I was trying to look for n-dimensional black holes with charge solutions. I wrote out a paper and said to myself, it looks like the solution doesn't exist. Oh, because X.AI got GROK 3 and GROK
3.7, they're all written by physicists anyway. So I said, as a theoretical physicist, maybe I should try. So I've done a deep research and say, go and, so I prompted first and asked them, tell me what happened in this area for the last 20 years. So they gave me the whole context of it. So I went back to one of my old calculations in a PDF form I said, I throw back in, I said, uhm, here's what I did. Here is the program. This is the way I looked at this solution. I wanted to generalize this
using a special technique in four dimensions. Can you tell me where I went wrong? And surprisingly, only Claude 3.7, 4, and GROK 3.0 came out and said, it's a very interesting idea. No one has done that before. Just by looking at your paper, you forgot to put these three symmetries in your calculation. And then I was like, oops. I was like, OK. So obviously, in 20 years, 25 years ago, I used a software called Mathematical to do
equation algebra solving and I was like, yeah, let's check how much mathematical used to be costing $2,000. Can I just go and buy the mathematical software today? So I oh, now it's down to 99. So they also went SaaS too. So was like, okay, can pay 99 for it. Then I asked each of the foundation engineers, okay, if I were to include these symmetries and here's the equations, can you do it step by step to show me how this actually does?
So he said, yeah, sure, whatever. And then he asked him, can you quote this in Mathematical to show me exactly? And because I'm familiar with the syntax and notation of the code itself, was like, OK. But I do know they have made some mistakes. It's not there yet. He couldn't give me the reasoning. He couldn't give me the solution. He cannot give me. But if I have ChatGPT today versus 25 years, my PhD would have gone down from three years down to probably 1.5 years.
Jeremy Au (33:47)
Yeah, I think for me the tricky part and I'll just say this is that because I also do some teaching as well is that historically before ChatGPT, I remember teaching before that, actually it was quite easy to see there's a bell curve. Yeah. Because people who put a lot of effort do well. Yeah. People who don't put any effort.
Bernard Leong (34:06)
Pretty much. I think it's the same. It is the same except it's being supercharged.
Jeremy Au (34:09)
I don't think it's a bell curve because for me, I feel like the bottom half, at least for me, has all moved all the way to a hit a certain level. And then people who are doing somewhat decently okay, they also use stretch piece. So now it looks like a, what's it called? a, what is
Bernard Leong (34:24)
It's a bit like a parallel distribution.
Jeremy Au (34:26)
Yeah, yeah, where it's basically everybody is now above everyone because everybody's using ChatGPT and they all get the same input.
Bernard Leong (34:33)
But it's a bit more U-shaped because the ones from the bottom is lifted up a bit more to that. So there's a fatter tail, so it's a bit of a U-shape.
Jeremy Au (34:40)
Yeah,
so it's a bit of a U-shape. Yeah, so exactly right. So nobody fails anymore, effectively. I'm just saying, depending on how I grade. Everybody's above average. And the weakest people got the ChatGPT . The average people got the above average. And the above average people cannot compete with the AI anyway. So they're also above average. So everybody's odd.
Bernard Leong (35:00)
They got stuck there. But
your 100X on the other side, going to 1000X. Correct.
Jeremy Au (35:05)
So
the excellent people, then the really good people that still exist and they got a bit of a boost and those who are super good became as in they got pushed to the right right so their performance became much better and a little bit of a fatter tail because it's
Bernard Leong (35:20)
I had a business mentor who told me because I was telling him how I was using the tools for my daily life and then he was explaining, said that you have to be a bit more careful because the way you're thinking about every problem now is has shrunk from hours to minutes. Everyone around you has to also think like you from hours to minutes. Otherwise you're going to have a big problem communicating to your team. And then I reflected upon that and I said, okay, yeah, I think I have to be a bit careful
when I demand that kind of speed from, let's say the Dev house situation, right? I demanded that kind of speed that goes from weeks to days because I could see how I can do it from weeks to days. And I find that this is where the big disjoint piece of development is happening. And I don't think people realized it. It probably would be very gradual and then suddenly everything just tank at once. I kind of feeling is moving that way.
Jeremy Au (36:11)
And so think that's the interesting part is like, think for us who are, you know, kind of like self-driven, comfortable with AI, then obviously we drop everything that we're not good at. We become fat tail on one side because we're like very good at X or A or B or C, right? And then obviously we're not being asked to be perfect at all dimensions. I'm using it for my exercise, but we're becoming fat tail on one side.
Bernard Leong (36:31)
No, no, we don't. Yeah,
Jeremy Au (36:36)
I don't know, is this interesting because everybody's become like, I don't know, everybody's becoming like a mini superhero.
Bernard Leong (36:41)
No, it's like the solo leveling thing, right? So for me, everything is just a rise. You want to be really... ⁓ Everybody was an agent. Yeah. Okay, come on. I just need to get this SOW done. Here are all the transcripts of the conversation from the customer. Can you help me to formulate out? I talked about this feature, but I don't want him to write the entire report. I want the AI to tell me, let's go piece by piece. And I want to see, then after that I asked the reasoning engine, can you look at this whole, what the customer want is it...
Jeremy Au (36:46)
Everyone's a superhero 100 %
Bernard Leong (37:08)
correct or are we able to help the customer to improve better with their current financial process, right? If you do that, then that's a good use of ChatGPT. The bad use of ChatGPT is what I, what Amazonians used to call it or even ex Amazonian, let me, they're boiling the ocean. This is where you're this probably, you know, 10 year plus plus engineer tell you, Oh, I've used cursor. It really junk. Then I asked him how to use it. I just throw the code in and ask you to do a code review. I'm like, come on guys. This is not how it works.
First, you put the code, you ask them, look at the repository. OK, first, analyze the repository. Tell me how the structure of this code is being written. Next, can you help me to look for what kind of libraries may have security vulnerabilities? You have to be very, very step by step. I sent that to the dev house. They were not listening to me. And I was like, guys, I have actually even give you the prompts. What are you talking about? Yeah, and they're boiling the ocean because they don't think that it can replace it.
Jeremy Au (38:01)
Thanks.
Bernard Leong (38:06)
And I'm telling them, you are in big shit because if you don't do it, it's gonna be... it's gonna come and hit you from the back without you even seeing it.
Jeremy Au (38:15)
Yeah, I think that's exactly reminds me of a time when you know for my podcasting I used the outsource to a production house. Yeah, they will charge me per episode and then it'll be like the number of shots is every time you generate a short clip an additional charge and I was like guys this cost is effectively already 10x versus what we can do using
Bernard Leong (38:23)
Yeah.
Yeah, yeah,
Correct, you can use CapCut to actually help you to select the video, the reels as well right?
Jeremy Au (38:40)
It's getting shorter and shorter, and the price of venture is getting huger. Venture has also let that go as well. But I think it goes back to the question that you said earlier, which is there's a lot of people who historically were paid based on how much time they used, hourly, lawyers, consultants, dev houses.
Bernard Leong (38:56)
You really can see companies like Gamma, which is the AI presentation app, they are actually, I think they raised only seed funding and they got profitable. And this is what I was so worried about. Startup founders who still want to raise with burn rates between 25 to 50K, I'm like, why do you need so much engineering? I'm like trying to...
squeeze down as much as my unit economics costs as possible. I only want to hire when I need it. You can be really, be very refined in your capital allocation. I'm not trying to tell everybody to be counter. What I'm trying to say is that you shouldn't be spending so much if your, what you raise versus your burn rate shouldn't be so outweighed. And that's why I've seen founders are having this situation. And I'm quite worried that because we went through the 2021, 2022, people are still
having this, if I come into AI today, I should be raising this X amount of dollars for this, right? The last I seen in the market, everybody's asking for sky high valuations. Yeah. And I don't think that that is going to work.
Jeremy Au (39:58)
So when you talk about pay as you go, what do you think that shakes out to be? Because obviously there's some destruction of value for the SaaS side. ⁓
Bernard Leong (40:06)
I'll
give you a number that I've been thinking about. A lot of medium-sized businesses in the 2021-2022, they became startup unicorns and then they ended up buying Oracle for maybe... Oracle has this thing called the 90-50-10. So the first year you pay 90% discount, second year 50 % and third year 10%. So if it's a 300k worth, your first year is really great. By your third year, you are like, oh shit, what is going on, right? So what they actually wanted in
that Oracle system, there's one feature that everybody likes. It's called reconciliation. So what people don't realize that if you are a medium-sized companies, you probably have multiple subsidiaries, one holding company, and you have a lot of bank statements, invoices of different currencies, different tax rate. And turns out that Oracle has this magic pill that if you put in all these things, you press one button, it does the whole thing for you, and it's certifiable by audits and etc.
Kudos to Oracle, they have that right? Oracle and NextSuite. that's the 3K in one button. Yeah. Right. So if you do it as a pair as you go. Now in zero, there's equivalent solution, but only for five ledgers. That means you have one holding company, four ledgers. That means only can do five. It's called joint. You can do it at 2K per subsidiary up to 10. That means, but you need to have multiple zero accounts. It's extremely painful. Joint was done by some X zero
engineers and say, hey, we know how to, the premium APIs, we know how to make it such a way that we can still artificially make it work. Now, if I were to think of it, maybe I should take the 3k, price it down by half and divide it like the way how joint splits it by ledger by ledger. Per ledger is about $500 and then implement a new version of the ledger, which I'm not going to talk about here. There's some innovations in the ledger space. So through my own
personal pride as a, well, I always work on AI, but I have this undue interest in crypto. So I know a lot about ledger technologies as a result of that and say, hey, can I use a ledger that can be on real time at half the price of what Oracle is charging? That is the holy grail that I'm going after. But okay, it's easy to state, but it's a difficult problem to solve. Yeah. I think people will first run into a few problems first.
How do you ensure multiple data sources can be doing this as fast as possible? And by the way, most of those internet companies, they have problems using the Oracle ledger. It can't support the kind of 1 million transactions per day, specifically if you are a payments company, coupons company. I heard of one of some of our startup unicorns building their own reconciliation unit outside. What they did is they aggregate all the transactions into a Databricks, then they push it into the...
Oracle ledger as one unit item. But then they are like, we lost all the information. We don't know exactly how many transactions come from where. The answer to that problem, which is what I'm thinking of solving, is you do not need the Oracle ledger and the Databricks layer. You need to put that into one layer. Yeah. And that you will solve that. That is the 300k. And I'm trying to do a pay as you go to say, OK, I think the 300k and break it down. Of course, value of money, I can give a 4%.
Jeremy Au (43:12)
I agree.
Bernard Leong (43:24)
inflation and the Munich sale will be 4% of that. Roughly, that's the business model that I'm trying to slice out. This amount that you're charging. And I want to see whether I can break it down by half. Because as Amazonian, your margin is my opportunity. If I can figure out the correct unit cost of unit economics of this, that will include your compute costs. That will include the way how you string the data together.
How you execute the data. If you run on serverless, it's half the price. If you can work all these little, little pieces down to that and make this transaction done for the accountants, they don't care about what is inside. What they care about, can you, Bernard, can you solve this problem? Currently I'm paying 300K for this problem. If you are, you can take this problem down by 150K, you go and take my money. That's for sure. That's the question I'm trying to solve.
Jeremy Au (44:13)
Right.
Yeah. And I think one thing they've hinted at as well is that it's not only a function of pricing model change, but also go-to-market has to change as well for everybody who is an AI first slash AI powered SaaS company, whatever you want to call that. I don't know what this field is called now. Every SaaS company allows AI. Every AI company is SaaS, whatever you call it. But what do you think has to change for the go-to-market motion?
Bernard Leong (44:27)
I'm call it.
I think, you know, I know there's a lot of AI startup founders, it's like, I'll go direct to the market, I sell directly to the enterprise. The hard truth is that if you want to scale, it's harder. So in the way I think about go-to market, so in Amazon, when people say you work in business development, you're actually a growth hacker. So if I'm a business development head for AI ML or the head of AI ML, which was my previous role, my job is to figure out.
How do I take a business from six digit goes to eight digit in an exponential curve? And the way we think about it is, you should break your customers into three cohorts. The whales or the lighthouse accounts, the middle stream accounts, and then the small medium businesses that will self-serve. And in that cohort analysis, what you need to do is to think about scaling functions. So where most AI businesses, their scaling functions would be
I use it, everybody likes the product, pay as you go, of mouth. ask this question, how many companies can actually do a word of mouth? mean, Cursor is number one, right? But Windsurf is a distant number two, right? So if you get that velocity, you're great. You're on great track. You'll be VC funded. You'll hit the unicorn status at the shortest time. Then what do you do with the rest of that curve within that? So
it turns out that if you look at scaling functions, one good scaling function that's really surprising and very underutilized is working with the cloud providers. Microsoft has the best, what we call, ISVs, independent software vendors. What they do is they take the Microsoft software, but they package it for things like retail. They package dynamics, but they package it for as a POS. They package as a CRM system or ERP system.
on that. So Dynamics itself is an ERP system. But the problem is that it is two-mile building breaks, and the ISVs can live on top of it. So the question now is, can you be a product type ISVs that live together with the cloud? Turns out, to everyone's surprise, you know how data breaks snowflake became so big. They were
partners to AWS. In the US, the AWS account manager selling their stuff with a SPIF included. If data breaks, snowflake are unicorns going through the AWS partner system. And by the way, they are very short of vendors. I'm helping my cloud vendor to find more partners, but the point of it is that there are very few partners in our market.
Jeremy Au (47:10)
actually agree with you about that.
Bernard Leong (47:11)
Yeah, so you find the scaling function, but you need to find a scaling function like this. That's how you go faster. But a lot of people say, you'll hear some founders say, no, it's very dirty. I have to partner with them. I have to, you know, I have to deploy on their cloud. I should be building my own server somewhere. I'm like, use, do what it takes to get the customer first. Because once you are on the hyperscaler, the hyperscaler will refer you to customers with
certain problems. My problem now is not so much of trying to work out where the customers is. In fact, actually, we have already qualified some of the customers. What I'm really worried about is how many customizations I need to do and how do I scale my product so that I could have a one-click turn on. That means the moment customer X say yes, I will launch in the cloud with all your security credentials, everything, and to ensure that I don't see any of the data.
on there. That is the problem that keeps me up at night. It's a one-click. I think we got it to one-click already. But how do I know if something has not gone wrong there? What are the things to audit? What are the things to check? Which is what a lot of AI software SaaS companies are not doing, or non-SaaS companies like mine are not doing. What they are doing is AI agents do X, Y, Z, And then someday, probably not a lot of people who are not working in AI agents.
Out of a thousand runs, typically sometimes one run your AI agent decide to be lazy and then do something else. So what typically what people do now, this is called reflection design pattern. You put a second AI agent to check. And you remember the classic internet meme, you have only one job. Yeah, so you have only one job and your one job is to do this. So a lot of my thinking now is how much trust can I get from the customer by setting in every audit trail that I can have
with my AI. I know a lot of founders will say, this is going to slow me down. This is going to whatever. Because with AI, once trust is broken, you're gone. There won't be a second chance in enterprise. And I think a lot of investors in the enterprise space do not understand that the trust level has to be extremely high. Yeah.
Jeremy Au (49:23)
I guess, to tie things up here, but what would be your advice to founders who are kind of like going after... Whether in terms of being powered by AI or deploying AI, how...
Bernard Leong (49:30)
Yeah, right.
I
think the problem statement is still the most important. Shockingly, after everything is what are you asking the AI to get the job done? I think, okay, but not to be disregarding because I know a lot of AI staff founders are working very, very hard. There are times where I feel that some of these starters are doing a one feature company that in 12 months time, OpenAI is going to incorporate the feature within the chat window.
Better way to think about it is something like this. If you have a CRM like Salesforce, how do you reimagine Salesforce in the AI world? Because what currently Salesforce is doing is patching AI into the systems, which may slow it down, which may screw things up. If you have an ERP system in the same way I've been thinking about the problem, how would you reimagine the ERP look like in a cloud-based world? You need to not just reimagine from
how the AI changes it, but also how the AI changes the entire landscape of this kind of software application business. And I think this is what people are not talking about. So if you are working on freight, right? Let's say you want to do logistics because I was in logistics. AI can actually help you to do a lot of optimisation, specifically with the complex models. Now the question then is, okay, should you just only use AI to do the search? No, you should be using AI to optimise the freight costs, path costs, thanks to tariffs. Oops.
Now I have to do a routing and it turns out O1 and O3 reasoning engine is extremely good at doing this. actually did a demo to retail people for retail business who asked me to and I use O3 and they gave me all their constraints and the chat GPT was able to give a pretty optimized path exactly what they thought as experts. They themselves say, yeah, that was correct. That's correct. the question is because they're not, they have the
foreground knowledge. Using AI for them becomes a plus and a minus. So you have to, I think the understanding the space you are in is more important than just trying to use AI. It's like everything is a hammer. Every, every, every, every, everything is a nail and you need a hammer to, AI hammer to knock it. No, it's not like that. Every nail is different. You have to use different types of hammer. It may be AI, it may not be AI. Okay.
For sometimes, even in the ERP side, I was telling my guys, hey, if this mathematical calculation can be done by just us putting this equal to this times this, just use this equal to this times this, no need to charge an LLM query to a reasoning engine for 20 sets to do this times this for me. Be practical. That's what I'm going to say.
Jeremy Au (52:14)
On that note, thank you so much. I'll just summarize the two big takeaways. First of all, thanks so much for sharing about, I think, experience of AI and how it's changing. I think AI versus coding versus Vibe coding. The whole practice of coding is a practice in terms of how it changes the product manager role, the front end, the back end. Secondly, thanks so much for sharing about how it applies to software as a service in terms of how you think it changes, how people think about pricing, go to market
as well as you know, pay as you go versus regular fees. And lastly, thanks so much for sharing about how you think about AI in the context of society, like in terms of where it benefits senior employees versus junior employees, where it benefits experienced people versus fresh interns, how students are performing in class, how instructors need to think about how to re-evaluate. So on that note, thank you so much for sharing.
Bernard Leong (53:04)
Thank you, Jeremy, for inviting me.
