The Optimization Era: Engineering the AI-Native Revolution | Interview with Howie Xu 🎙️

Hey, Nico. Hey everyone. Howie here. I'm the chief AI and innovation officer at Gen Digital. Gen Digital, for those of you who have not heard of the company, we actually have a number of products you have probably heard of for years: Norton, Avast, LifeLock, MoneyLine, and so on. So it's a public company with lots of consumer-facing products. Our agenda is actually helping consumers to live their digital life safer, better, more productive. And a key important agenda for us right now is to bring everyone to the AI era or the agentic era in a way that's safer, more productive. Some people are worried about, 'Am I going to survive? Am I going to thrive?' So our agenda is to help you survive and thrive in the AI era. So that's my job as chief AI and innovation officer at Gen Digital. That's what I'm thinking day in and day out.

Yeah, this is a question I think about on a daily basis, I want to say in the last 10 years, believe it or not, but a lot more intensively in the last two or three years. The way I think about AI is, let's just talk about the ChatGPT moment that everyone is familiar with. About three plus years ago, OpenAI introduced ChatGPT and it caught the world by storm and really literally changed everything from that point. And we have been living through this AI revolution day in and day out because model performance and AI impact is just bigger and bigger on a daily basis. Now what does that mean to our productivity? What does that mean to us as a consumer? What does that mean to everyone? Right? Interestingly enough, the way I think about it is the technology is strong, it's great, but that doesn't mean you are going to make immediate big impact. It happens all the time. When electricity was invented, it took another 20, 30 years for people to really grasp the value out of electricity. Now this time it's not going to take 20, 30 years, but still takes time. So let's be very specific. Enterprise as employee, let's just use that kind of scenario. How do you get value out of ChatGPT-like technology? Of course, you can talk to it, ask questions, but let's face it, getting an answer from ChatGPT is not fundamentally changing the way we live, not changing the way we work. It's a very good tool, but it's not fundamentally changing things. So between that versus fundamentally shifting the way we live and work, what's the gap? Obviously technology needs to mature to a level of less hallucination, faster speed, lower cost, all that kind of things, which OpenAI, Google, Anthropic, xAI of the world are working on. I'm actually quite confident to say that technology evolution is not in the way for making the level of impact we wanted at this point. We're not there, but it's making very fast progress. But what is the other side? The other side is we need to rethink about how we work. If you still do the same thing the same way, you're not going to get value out of the technology. Let's be very specific. We all know there's a coding agent. So all the software engineers on the planet are thinking, 'Hey, I used to have a pretty good job last 10 years, five years, two years. Am I going to have the same good job because it sounds like AI can do things faster, cheaper?' Am I going to have the same job? Clearly, as a programmer, your job is going to be very different, but that doesn't mean the programmer's job is going to be gone. It's just that the way you work should be different. I was at an event of one of the leading AI coding agent, Augment, last night. Eric Schmidt was there as well. So look, AI coding agent is very powerful. However, most companies, I want to say most companies, are not getting value because in order to get real value, you have to give context, you have to give tribal knowledge, you have to give a lot of things beyond just the documentation fed into the model. If you don't, the answer you get, the outcome you get, is suboptimal. But getting that tribal knowledge stuck in everyone's head is tough because sometimes it's unnatural, sometimes it's tough. You have to do a lot of work to get there. So from that point of view, before we joined this call, I asked about the latest development of the company. You mentioned you are releasing an optimizer product. That's the name of the product. I don't know exactly what it does, but if it's helping me to collect the tribal knowledge into a digital fashion so that I can feed it into a coding agent or some other agents, it's going to help tremendously. At that point, you can say, 'Oh, when I introduce AI coding agent, my productivity is not just 10%, 20%, 30% better because I can automate some things. I actually can rethink how I am going to do software engineering in my team.' Because at that point, I can say, 'Look, I have a 10-person team, and then seven of them are virtual agents, a virtual team member, and then three of them are human. The human's job is to do X, and the virtual staff is to do Y. But we all have similar knowledge, similar tribal knowledge, or at least we have a way to pass the knowledge from one another. Then you can plan things in a very different way.' At that point, we are talking about 2x, 3x productivity improvement. Without the knowledge collection, without the tribal knowledge, without those data context, it's hard for me to imagine AI is going to fundamentally change the way we work as an example.

Well, that's the million-dollar question. As simple as documentation of the process or the product, or PRD. But it's more than that. A lot of that is even implied. I give you an example. What is context? Maybe structure it needs to be a context because people can infer something. Just like when you and I, if you were my product manager and I'm your engineer, we work on something and we want to look up something. When Joe or John sends us something, we look up in the directory which department he works for, and that gives us some context. So the agent should have that too. It's everything and anything. We humans do not necessarily think, 'Oh, we need to pass this document or this information to the agent,' but guess what, everything matters.

Yeah, it's a tough one. Whether something is noise or not, if it's obvious, we wouldn't have this conversation. From my point of view, it's simple. I give you my simple answer, but I know in reality it's not simple. For instance, MIT reported 95% of enterprises have not got value out of it. To me, that's 100% noise. The reason it's 100% noise is because what's the expectation? Do you really expect 95% of enterprises to fundamentally change the way they work? Probably not anyway. So why do we even have that conversation? But did 95% or 100% of enterprises start a journey? 100% true. So to me, that report doesn't tell me anything. So how to tell between signal and noise? To me, it's very simple. Use your own eyes, feel it. Instead of reading from media, you have to have some calibration. You have done this program for a few series. People have to feel it. Nico asks a few questions. Typically his audience is able to give him the BS answer or not. After a while, they can calibrate. You have to do the same thing. I feel like it's not people's opinion that matters. I can say MIT report is BS or noise, but that's an opinion. What is the fact? The fact is people started the journey. Every company started making a lot of investment. Not a lot of companies have reaped the full benefit of AI, or not nearly true. We just started doing things a little bit different from three years ago. Not fundamentally different. Checking with ChatGPT once in a while. It's not fundamentally different. It's just a better Google in some ways. But once you start thinking, 'Oh, I get up in the morning, I need to think about the three virtual staff. What job do I give to the three virtual staff? I have two virtual, two physical real humans. What am I giving them? How am I thinking about they work with each other? How am I thinking about in the evening I go to bed, my three virtual staff is going to give me a report in the morning? And by the way, not just give me a report in the morning, I will have a fourth staff from a different department check on their outcome.' Once you start thinking that way, then your life is different, your work life is different. I think most people are not there, not you, not me, maybe not everyone, but we are moving towards that. So to say 95% of enterprises are failing is just not an interesting conversation. What's the real data?

Yeah, today is kind of too late. I'm just kidding. Look, as I kind of mentioned, three years ago we all started this journey. Almost everyone in our business at least started using ChatGPT here and there, maybe different degree. So we all started. But we are at a point that the technology has reached an interesting tipping point. Of course, you never know if it's today or three months ago or three months from now. With that disclaimer, we have actually reached a tipping point right now that virtual staff is a lot more reality than in the past. I think the last three years was more about getting better questions and answers from AI. I think the next three years is actually about getting better actions and coordinated actions, and treating the AI as a virtual assistant, personal butler. In my world, I'm delivering value to consumers. In your world, you are delivering value to the enterprise. So for you, your product will look more like a virtual employee, whose job is to collect tribal knowledge from people and pass it on to some other system of record. You would treat AI more like a real employee. In the morning, what do you ask the AI to do? In the afternoon, that sort of thing. In my world, it's like someone is going to watch my email, someone is going to watch my iMessage, because I may not get to iMessage during the meeting, but if it's truly important, I need to take some action. So I want a personal butler for my audience, my customers, my users.

The fact we are talking now is important because you and I, I wouldn't have given the same answer five months ago. I wouldn't. I think the technology has reached an interesting tipping point.

I think a lot of time there's a lot and then there's not a lot. Let me explain how I think about it. By the way, your founder CEO Jennifer Smith and I actually worked together at an investment firm, Greylock, one of the prestigious VC firms in Silicon Valley. We looked at a lot of startups together. That was back roughly 11 years ago. So I actually invited her to my company to do a fireside chat. We talked about the past a little bit. The reason I mention that is both Jennifer and I spend a lot of time looking into cutting edge technology and cutting edge companies, startups. As you can imagine, lots of startups didn't go anywhere. Majority of startups didn't go anywhere for obvious reasons. So from that point of view, everyone comes to Greylock saying, 'This is going to change the world,' but it turns out not the case, nine out of ten at least, if not more. So from that point of view, that's a lot. I would say the odds are that interesting technology is not going to change the world the way you think, statistically.

But however, the no part is the following. Once in a while, there is a generational technology. In my life, I came to the computer science world when PC was the thing. We got excited about new processors, Pentium this and that. The speed was the talk of the town. The PC era, internet era, cloud and mobile. I want to say that each decade or each generation there is a defining technology. Almost everything else is noise. In the 2010s, basically cloud mobile or mobile cloud was the defining thing. You can do everything else, but the most interesting part is mobile cloud. If you catch that wave, that's it. I think this decade or the coming decade is AI. I don't even know anyone would bother debating with me. You can talk about quantum computing, AR, VR, but that's not the defining technology at the moment. So I say whenever the defining technology is here, as long as you focus on that, just do it, you are going to get some value, you are going to deliver value for your audience. So I'm kind of more used to what is the high order bit of the technology of this generation. Let's be laser focused on it. My job today, my passion today, is to give consumers the value of AI so that they can survive and thrive in the AI era or agentic era. That's how I think about it. So I'm not too worried about technology coming and going. Of course there are always naysayers or non-believers of AI still today, but I don't know what to say to them. Let's not worry about the noise.

Yeah, I think this never changes generation after generation, at least in the tech world I lived through. It's really about execution. It's really about laser focus on the value. Everyone is talking about the value, everyone is talking about cloud or AI. But the focus part is not something that in reality 98.7% of the people or the companies out there have. That's a reality for one reason or another. So I would say if you have the conviction, if you laser focus and solve an unmet need, and the market is big enough, that last part is a little tough to predict, it takes a little conviction and a lot of other things have to happen, the stars have to line up for that bigger market to happen. But if you meet all of that, you're going to be taken care of. But it's easier said than done because we are humans. As much as you and I said don't forget about the noise, humans' nature is to be distracted by noise. So it's kind of a reality.

Yeah, look, unfortunately you didn't pay me for marketing so I cannot say your product. No, I'm joking. Look, it's so difficult to extract knowledge, tribal knowledge from people's heads. I can totally see your optimizer, Scribe Optimizer or Scribe other products can help. At that point, return on investment is a sort of a but I think that direction makes total sense to me at least. Because what you need to do is you can't just say, 'Hey, every human sit down for one hour every day, reflect on what's in your head, write it down, document it, pass it on to AI, and then go on work for the next seven hours.' It doesn't work that way. So you almost have to have a natural way to capture those things. Maybe Scribe is the product that is the answer. Maybe something broader, maybe something different. So I would say it's a pinpoint, it's an unmet need.

You actually used a very interesting word that I started thinking a lot recently: invisible. I was at an event organized by Augment, as I mentioned, with Eric Schmidt and a few other tech leaders. One of the big takeaways is that we already have invisible staff surrounding us. You have, I have, but are we taking full advantage of our virtual staff, the invisible staff? No, or at least not fully. I think that's part of the reason I said we reached the tipping point. Another invisible part is this knowledge. Knowledge is invisible because it's in your head, but which part of your head? When is it going to be visible? It's not something that's obvious. We need a way to extract those invisible knowledge, tribal knowledge, out naturally so that the invisible staff become visible.

Um, sorry, which angle are you thinking?

Yeah, the way... There are a lot of different ways, levels of the question I can think of, but I think the key question is really the whole thing we talked throughout the discussion: how do you leverage AI in a way fully or nearly fully? So at that point, there are different ways to measure it. For instance, one way to measure is if you are talking to an engineer, how many tokens a month or a day are you consuming? And of course, that could be a stupid thing. It's not like the more tokens you consume necessarily means the more you utilize AI. There are consequences. Let's set that aside. It is a proxy to how you interact with your AI. So my question would start from that: how many tokens do you consume? And from there, what slows you down? Why not 5x of that? If you are dealing with one agent, why not dealing with five agents at a time? If you're dealing with agents during the day, why not your agents working for you during the evening? So I would ask a lot of questions, but start from one fundamental metric and then go from there. It's almost like if you were working for Google, there's one metric for the entire company: DAU. If they see DAU increase and engagement increase, great, then they will figure out ads monetization separately. But DAU is the key thing. To me, the best metric I can think of today, and I'm pretty sure this is not the best one and I hope there is a better one soon, is tokens per employee per day or per month. This is the best metric I can think of today.

I think completely is never. On this planet, there's no such thing as completely. It's always 99.9 something like that. In that context, whether I like it or not, I'm a programmer by heart, by training. I would say unfortunately, five years from now, the programming as we knew in the last 10 years will be taken over by AI almost completely. Now that doesn't mean software engineering is taken over by AI completely, because there is always a lot of architecture decisions, design, so many decisions. Think about a Google L6 engineer, he's not coding most of the time. He's in meetings all the time. So I'm talking about the coding part. That is going to be taken over by AI completely. And coding will be a hobby, a fun part of the job, not a necessary part of the job.