There's no moat anymore — only knowing what's good
You have to treat the agent like it's your smartest dumb employee — the most brilliant junior-level person you have, faster and smarter than you, but with no experience and no wisdom whatsoever.
David Ting is a technology executive and solo founder of an AI company. He previously served as CTO at Zenni Optical, the direct-to-consumer eyewear brand, and CTO at Bespin Global, a cloud and AI professional services firm.
After leaving Bespin Global, David started coding on his own, orchestrating fleets of agents and building products that previously would have required a team.
Read full transcript of interview
I was the CTO of Bespin Global, and now I'm working on my own AI company.
What is working on your AI company looking like these days?
It's a lot of fun. We'll probably dig into the deep depths of it. So I started the journey after I left Zenni Optical as the CTO of the company. I knew AI was going to transform the business. And so I joined Bespin Global to be more hands-on technical day to day. Started two AI businesses and tried to optimize the P&L for those. And then the thing that blew me away is AI is getting so much better every day that I could actually do the work myself. So I started coding pretty much beginning of the year, and I was trying to push myself at the peak on the daily output. I can do 700,000 lines of code. So it really just blows me away. You could pretty much develop just like a development team.
Well, at Bespin, what was your org size?
My org size, the company is over 2,000, so I oversee the technical engineers. But then it's a professional services organization. So the engineers are assigned to projects. The project has clients. Clients want to do their work a certain way. But what we see out in the field is that we get a lot of cost pressure. So these smaller consulting companies, the companies are one or two people coming in, and they'll bid you on some of the projects. And the customers, after a while, they just don't talk to you. So they know our costs are probably like 5 or 10x higher than some of these small boutique shops. And then trying to understand the why, I really started building two AI companies inside, trying to understand their processes, their development methodology. Started with about 20 or so people. And then as we get more and more productive, we know we could run these companies at one third the size or even better.
All right, so you're basically saying, now you've seen this happen twice. You've seen this happen in IGN and the media industry. Now you're seeing it with the professional services industry. And so you essentially go from running these teams of engineers to saying, OK, I need to figure out what the next version of an engineer is when everything is changing. It's disrupting this company's business model. You take a look around and then you kind of say, I'm going to nope out here. Like, it's changed so much that I can't fix this because there's no point in fixing this. This is broken. Like, am I reading that right?
That's exactly right. So I'll give you some of the math behind it. I started vibe coding because I felt like I was talking to engineers and engineers say, hey, I don't code anymore. AI is taking over. So then I say, well, let me try and build a fleet of coding agents for myself and try to do work on a project myself. Once I go in, I can't go back. Essentially, I could code 100 times faster than working with other people and also develop at a velocity that's 100 times faster because of parallelization. So what matters the most is that you know what you want. Like, this is the solution for this problem. If you have a very good feel of it and you know what good looks like, that's what it takes to win in this world now. The delivery speed, like having a big team, say you outsource to India, the Philippines, or you name the country, is no longer the case. The organizational friction is going to slow you down so much. These AI agents are so good that if you know how to work with them, that's a new way to work.
So that question of trying to then define what the new SDLC is for an individual, like you can, and I think a lot of orgs are suffering this right now. You can blue sky yourself to death because you can do anything, to your point. Like you're generating all this code, which means code is a terrible metric for success at this point because you're drowning in code. And if you can do anything, everything boils down to taste and, well, what should I do? So how are you now deciding, OK, this is what I actually need to be spending my time on?
So I started building personal AIs for myself. So things to do like being more healthy, automate my stock portfolio. That one is taking a long time because it's a hard problem. I don't think anybody has cracked it or else it would be infinite wealth. But it definitely is doing way better than I did before. So that was a metric I used. Then now I moved on to the S&P. So can I beat the S&P consistently? So far so good. Stock market has been merciful. But then how good could you get? Because you're seeing these next generation AI based quants coming in with 10X returns per year.
So, returns on what, like financial returns?
So if you put a dollar in, you get 10 dollars back at the end of the year. So it blows any of the old metrics away. So for me, it has one of the best data sets out there. So what I want to do is learn the agent technology the best possible. So once I can automate my financial portfolios generating income, I'm not expecting 10X, but I'm expecting a little bit better than the S&P. If I could do that consistently, then it gives me a lot more time to do something else. Health. I'm older. I'm over 50. I've gone to the doctor. There's a lot of things they said. Interesting enough, I've gone in and put all of my test results into the AI and said, be a doctor, tell me all the things I should be worried about. And there were five or six things that they labeled, and those are all reversible at my age. So I need to form better habits. So I'm doing a personal AI, doing reminders, and then just getting myself to do the habits while the time is now to improve my health.
Also like that introduces some other factors there. So on the one hand, when you're talking about health, that is a level of personalization that actually scales. Everyone has health. If everyone were to use these tools, everyone theoretically could get healthier and better. Stock market is the opposite of that. You're talking about 10X returns. Well, returns are often based on winners and losers. And if everyone was using the tools the same way that you're using the tools, it doesn't work anymore, theoretically.
That's exactly right. So there's a whole question like, is this methodology, first of all, how does it stand the test of time? Is it just a current hack? It's only available for two years. If so, I better take advantage of it because I didn't know that existed before. And then the second part is the window will close. That 10X return, as people get better, they will come down and probably will be like a little bit better than the S&P. I'll just give you an example why the S&P will be outperformed by these funds, by this software, the personal AI software. So the S&P does things on a monthly basis. It goes through all the companies. You have to meet a certain set of thresholds. And when you hit those thresholds for two consecutive quarters, then you're a candidate, then you get elevated and they take other companies out. That rotation cycle, if you think about it in the current age of AI, is that the most efficient? Could you be faster? Could you out-S&P the S&P? Could you scout out those companies beforehand and know those candidates and know where the fund is going to go to? There's all these fingerprints of all these companies. And you can imagine that you're just investing, there are certain patterns they follow. For old institutions, let's just say Warren Buffett's company, Berkshire Hathaway, they have very, very standard formulas. You know exactly how they're going to deploy cash, but you could be one step ahead of them. If too many people do the same thing, you're absolutely right, then the market maker advantage is gone. But then if you could be faster, that's actually an advantage.
Opportunity is, you know, plus time. And then there's a cross tab of that, which is basically discipline versus iteration. And, you know, if the time window for a particular task is short, iteration and the ability to iterate quickly makes more sense than discipline. But if the time horizon is long enough, discipline will win out. And so how, AI has increased or shortened the time for so many different things where iteration may form a play. But there's still the role for the orgs that are disciplined in a bunch of different spaces that are the ones that are going to succeed. So I guess what I'm trying to figure out right now by talking to a lot of people, and I'm just inventing stuff out loud, so if it doesn't entirely make sense, I apologize, I want to learn, is if we're entering more of a validation era for AI. Like we've gone from this context era and moved a little bit until now we're really trying to dig into the outcomes. But outcomes do eventually lead into, OK, the validation of what was that outcome, what we really wanted. Is that outcome delivering a better ROI? The only way that we can determine what outcomes are are frameworks. And so for somebody who's doing this on a personal level like you, what framework are you using to determine where you spend your time? Because your time is limited. And at a certain point, there are agents you can deploy to stuff. But at some point you say, well, you hire when the numbers make sense for you to hire.
So let's talk about that. I know exactly what you're leading to. There is a framework I developed to develop the AI agent. I'm trying to improve it because I'm trying to take one of the hardest problems, productionize it, see that there's a guaranteed return, see that the feedback loop is working. So still not enough data yet, but so far so good. I get two out of the three picks right with the agents, which is better than most fund managers, by the way. So you just have to have a structured discipline of when to buy, when to sell. You have to sell at a loss and you play the system and then hopefully it'll win out over time. Going back to it, that framework of developing AI experiences is something I'm trying to work on with a few friends of mine to start another AI company. Because for me, I look at my personal AI as a way to improve myself. And so if I benefit through that process, great. My health gets better. I could make a little bit of money on the stock market. That's great. But at the end, I learn how to work with the agent, debug it and make it really good. And through that process, I learned so much. And let me just say that the journey is still early, but I can tell, like now versus before when I sat as CTO over a thousand engineers, I was talking about it in concept in a very vague term. Now I know exactly where some of the problems and challenges are in that transformation. You just have to go on that journey to see the other side.
Language is important, too, though, because when you say like I'm starting an AI company, are you starting it like, if somebody is a home builder that uses hammers, he's not describing himself as a hammer company. He's a home builder, unless he is in the business of producing hammers. You're not making, you're building processes, you're leveraging these tools. Is that an AI company or is that something different, like that we just don't have a name for?
So let's actually talk about the stock programs I'm working on. So if I pull it apart, it's probably five or six software companies it's doing already. So I'll give you one. EDGAR has a database. It basically has all the companies, like mutual funds, post that 13F to disclose how many positions, what positions they have in their fund every three months. So there are companies that track and sell that data. I implemented that system. I still need to go in and clean up the data myself. That's the part I will hire somebody to do once I get the process down. But that data set is extremely valuable. And for me to build it, it took me a week. So I just give you kind of a velocity. So if I want to go in and say the data sets are super expensive, like Bloomberg sells it for thousands of dollars, there's the cheaper ones that go for a couple hundred dollars a month. I can sell at twenty dollars a month because I'm a company of one. That product will probably be profitable very quickly. And I can actually ask the users to help me fix the data as part of the feedback group. They say it's twenty dollars. The data is ninety percent, ninety five percent right. Not as good as Bloomberg. But we'll get to Bloomberg level if you can help contribute to it.
You built it in a week.
Yes, I built it in a week.
Can't someone else build it in a week? Yes, absolutely. And maybe this doesn't matter anymore. Like the classic argument is like, what's your moat?
There's no moat. That's my point. Before there was a moat because writing that data cleanup data pipeline, writing the website, putting in a database, scrubbing the data, all of that takes a lot of time, a lot of people. I'm just giving you that as one of the examples. I actually went through that because the advantage I wanted to get was I want to get the signal from the institution, knowing their movement before other people do. That would be my advantage on my trade. So I actually wrote that and then later on I said, oh, wait a minute, that's a business by itself. But then that company is really, after I'm done and the pieces are in place, like which parts can I productize, which parts will have good demand, pick one or two lanes and basically launch it as products.
And how do you market it?
Market it. There's a whole marketing. The playbook keeps changing these days.
What is your expectation that Google or OpenAI just say, oh, yeah, we'll do this?
My expectation is that they will get into it. They already do. But their data actually is flawed. That's why there's a market afterwards on these extremely good data sets. I'll give you an example. The funds typically only show the top 20 holdings. But they have 50 holdings on some of the famous ones. How about the other 30? So you'll miss their new entries. Like what did they just get into? What are the big moneymakers getting into? The Walmart, like if you talk about a legend, Peter Lynch, before everybody knows about Walmart. He actually famously said I bought in late but I made 100X on my investment on Walmart. So those are the signals that are missing in publicly available sources. Those are actually the opportunities for like an agentic AI out there, because you could build a system very affordably and very quickly these days.
That's the hope. What's the fear?
The fear is it is very tedious. So I did three straight days of like pushing myself a lot of times. I like to see how fast could I get. So I got to 200,000 lines of code a day. I said, oh man, that's way better than what I did before. There's no way I can crack it. Invented a couple of new methodologies and I got to 400,000 lines consistently. Then I start pushing myself using multiple windows with multiple terminals and then start coding in parallel. That burned me out. But I did some 150,000 lines a day. So the fear I have is you run out of ideas. The fear also I have is you don't validate enough. So there's a lot of bad quality software that's going to be out there. So as a buyer it's kind of like, before there's the Wall Street Journal, that's a legitimate piece of article in journalism. But now there's like hundreds of thousands of pieces of blog. How do you rate which ones you could trust, which ones you don't? Same thing with code in the future.
No way. In what possible way can you validate 450,000 lines of code that you just made in a day?
I actually created a system. So I basically use agents to critique my own code and do it on a periodic basis. So they basically critique the math. They critique the security. They critique the architecture and they critique the code quality. So it goes in and fixes it on an iterative basis. That's actually one of the threads of terminal that I
guess that works for you because ultimately this is product for you. And so you are product managing yourself. And so ultimately you're the only one you can vet it through anyway.
Yes. But then going back to the software company that I'm working on with my friends, we can use the same methodology to disrupt anything. So this is where we are formulating a couple of lanes, too early to talk about any of those. But there's very lucrative lanes that are available in the AI era. And then you actually hit the nail on the head. The one hint that I have is there's a couple of things. One is people who don't know what they're doing, their token costs. I didn't know what I was doing, so I burned a ton of tokens. You'll see it. Hiring developers is cheaper than burning all the tokens. But once you get good you can compress it by 100X. The other side is validation, hallucination. How do you control it when it runs in production? That's actually one of the biggest challenges I had over the last two weeks, is that the agent performance varies a ton based on how long they perform. One of the hidden dirty secrets is that when there's a peak in production the agent quality degrades but they still respond to you with a poor result. So you make the same call to Anthropic, to OpenAI. I actually before straddled both and then tried to triangulate. And then the quality decreases and then you have to control the
It's like dealing with a human that's suffering from burnout. Don't call them at 2 a.m. to do the same work that they would do at 11 o'clock in the morning.
Exactly. But then if you productionize it, this is actually why I feel like 90 percent of these AI projects are failing. You're assuming this is a dependable human. So when you're coding and structuring it, it's not dependable. So you need harnesses around this, and these harnesses are going to be worth it. It's an opportunity for the next couple of months to create a company that does those harnesses, if that's of interest. That's actually if these are problems that people are running into.
That's interesting. There are people that make the mistake of thinking that this isn't a human, you need to treat it totally differently. There are also people that treat it, you need to treat it like it's your best employee. But you actually need right now to treat it like it's your, in many ways, smartest dumb employee. You know, the most brilliant junior level dumb person you have. And that's how you have to get there to do what you want.
How do you teach a person that's smarter than you? So essentially you're working with an agent. He's faster, smarter, but you have to give him examples of how to improve.
But less experienced and with no wisdom whatsoever.
Exactly. And then that's actually agent memory. A lot of people are working on it. A lot of the approaches are wrong because I actually believe you need experts to do that. And those expert data sets are going to be huge. And this is where a lot of the companies, when they say what is my secret trick of making Coke, there's all these secrets inside the factory, inside the formulation, how you bottle, all of those need to be packaged up. And then once you know what it is and you write it down, you can continue to improve it. And that iteration is the key for every company to really take advantage of the era. Use the opportunity, or some of your competitors may, and you'll see their product leapfrog yours because they understand how to use it to get this very, very fast feedback loop.
Did you watch any of the Google announcement yesterday? What are your thoughts on that?
Well, they are really ahead of the game. They're one of them. If you look at it, people are talking about Nvidia last year. I think people should be talking about Google this year. I think they're doing a lot of things right where they have the full stack and they give a lot of products away for free and they look at the data of how people use them and then they find and launch new products.
They're in a better position than Anthropic and OpenAI on the search side specifically because they've always monetized it. There have always been those ads sitting up there and they simply just know how to do that better. I've watched the results change just today. Have you entered in a question and suddenly, well, it's more ads up top? That's fine, you can hide them but they're going to be there. Here are the answers. And then the search results way down at the bottom, which are now by and large irrelevant. It'll be interesting to see. It feels like OpenAI is in their struggle era at the moment even though they've got more money than they know what to do with.
Yeah, I think there's some leadership void there. But I don't count them out. So like I said, I use Anthropic and I started using OpenAI, and Anthropic has its missteps too. So if your strategy on your business is heavily dependent on the outcome of the AI, then you need a number of these companies that you could use. So I really am rooting for OpenAI to make it, so it will keep Anthropic honest, and also Google. The thing I don't like about them is they're so closed, meaning that one day if they change how they bill you. OK, so let's just say you expect the agent token to be X and all of a sudden they have a monopoly and they move it up, say, hey, I'm charging 2x now. And this is exactly what happened with Salesforce, with a number of these companies, like NetSuite. I ran into this exact issue. Oh, on a renewal the cost is X, a renewal comes, oh, it's twice now. Why? Oh, we redefined the definition of enterprise. So with this number of employees now you're enterprise, so you have to pay enterprise rates. So I think that era is going to come fairly soon, sooner than you think, especially after these companies IPO when there's pressure for profit. And this is where when you're losing some of your best employees because you're trying to replace them with AI, you may regret it, because very soon in the future cost may be higher. So this is actually a very short sighted strategy as far as I could see.
Last question I want to get into, because a lot of this trip is also product market research for me, which is: everyone is redefining what the software engineer of the future, what the junior that's going through a computer science degree. Like, as somebody who works for a company that hires engineers for US, I need to understand what those people need to be able to do. So if you had the infinite money, if you needed to hire. Why a lot of people aren't hiring right now is they don't actually know what skills to hire for at this point. But for you, what would you think about hiring for?
So I actually think, as engineers or as a certain type of people, you're trained to solve problems. So in an era of AI you still need people who can solve problems. They can define the problem, they understand output, they understand what good looks like. And those are the people that from college you have to teach them to think the right way. I think the current academic system, they don't train them enough. They don't emphasize the importance of foundational knowledge. There are ways to learn that faster now with AI, with NotebookLM. That's my hack code to learn anything right now. Go in, research all these sources, read through all the sources, and then know probably more than 90 percent of the people on a particular subject in a day. So those are the people who are thirsty for that knowledge, they can learn, they can adapt, and they can consolidate the knowledge into digestible sources. They also need to understand and be seen above the crowd and be able to speak to AI in the way they understand so they improve. You're speaking to somebody who's 100 times smarter and knows probably a million times more than you do. How do you help them to get better? That's actually a really important one. And then the other one is safety. How do you put the right guardrails so that when you put very important processes in the hand of the AI to control, what are the validations? Who are the people who give you the safety that this thing is working properly in production?
Sociologically we're going to figure out if there are enough jobs for people that have that particular mindset or skill set, because that ain't everybody.
Well the other one is a creative side. I actually think one push is, I really hate the slop that's out there, everywhere out there. So I very much appreciate and want to support your podcast, because getting high quality content is going to be key. And because people's time is freed up, I hope there's a lot more people spending their time creating high quality content for people, educating the masses, pointing them the way to go. The other side is that we neglected a lot of the problems because they're not solvable. Global warming may be solvable, cure for cancer may be solvable if the farmer doesn't get in the way. Longevity is being cracked right now. There's a lot of these problems, nuclear fusion is another one, limitless energy. So there's a lot of these problems. I hope there's enough leadership in the world where we put our resources and our mind on problems that will benefit humankind and really help us advance quicker with it.
Be nice. I mean, solving the problem of SaaS versus, like, some real fucking problems. Like if we're going to spend all this money, if we're going to build this many data centers, can we fix? I don't even need to fix everything. Let's fix something.
And I think having that leadership, either from the tech community, I don't expect it to come from politics, but that push is where those breakthroughs will come. And then this is where I am not that ambitious, but with what I learned I say, whoa, there may be like a horizon in sight that is solvable. Do I spend the remaining of my time on some of those things that are unachievable currently? But once you really know the tools, they could be solved in your lifetime. So that's really the dilemma that I have, is where to put my time, and you hit it on the head.
Right. Well let's see if the Gen Xers and the millennials can kind of stand up and exit on a better note than the baby boomers did.
Fingers crossed. And I think, looking at SpaceX about to IPO, hopefully that's going to give us space exploration being something that humanity will go after. And then just looking at the numbers, this is why I'm not a big believer in government funded projects. Like what Elon Musk did is he did it, I think, one ten thousand times cheaper, like doing a moon mission with SpaceX technology versus NASA technology.
But he took untold billions of dollars from government investment, starting in the Obama administration, forever. So let's be clear, Elon's been sucking at the government teat for the vast majority of his career anyway.
I completely agree. But then on the NASA side he also took a lot of their open source intellectual properties. He didn't start from scratch. But what I'm saying is that the next generation should be looking at the same pattern on what resources are available to build great things for humans. So that's my two cents. Not saying Elon Musk is a great person, I'm just saying.
The goal ultimately of this technology should be to the benefit of humanity at large, not for the benefit of the people making the thing. Remains to be seen.
So being a former CISO for several companies, one of the things I was so impressed with is the security engineer persona. People talk about Microsoft's major announcement with their toolkit, with AI. The security bugs are just a lot easier to find and fix. And I'll just say one takeaway is get security engineers that just don't go in and do the process framework. Get people who can go into the code and help you find stuff, like Microsoft found four vulnerabilities that are critical in Windows and you should read through them. Some of those look kind of scary and they're probably in the hands of a lot of the hackers, because they were unknown for years and years.
Make the hackers work for the good. For the good.
Be part of the
conversation.
Whether you're a CTO who wants to be featured, a company looking to sponsor, or an engineering leader wanting a seat in the room — there's a place for you here.