Microsoft 首席产品官:如果你没有用 AI 做原型,那你的做法就是错的 | Aparna Chennapragada
Microsoft CPO: If you aren’t prototyping with AI you’re doing it wrong | Aparna Chennapragada
Opening Episode Highlights
Aparna Chennapragada: I have a cheesy Chrome extension. Literally whenever I open a new tab, it just says, how can you use AI to do what you’re going to do right now?
Stand-Up Comedy and Product Building
Lenny Rachitsky: How do you see the future of product development being different?
Aparna Chennapragada: If you’re not prototyping and building to see what you want to build, I think you’re doing it wrong. It becomes even more important to have that territorial and taste-making at the heart of it because, otherwise, you just have a Frankenstein product.
From Consumer to Enterprise Products
Lenny Rachitsky: There’s this acronym that you taught me, NLX. What is that?
Don’t Hold Back Early Adopters
Aparna Chennapragada: Natural language interface. NLX is the new UX. Often I hear a product builders say, “Oh, yeah. With AI, the model eats the products.” That doesn’t mean it’s not designed. You and I are having a conversation. It’s a podcast. I’ll have another conversation at Microsoft and that’s a meeting. Conversations also have grammars. They have structures. They have UI elements. They’re invisible. What are the new principles, new constructs in natural language as an interface?
Lenny Rachitsky: I just saw that Cursor hit 300 million ARR in two years. Interestingly, you guys were very well positioned to do really well in this AI coding tool space. You guys had Copilot, the first tool in the world at this stuff. So ahead of everyone, what happened?
Project Frontier: A Year in the Future
Aparna Chennapragada: I would say…
The Era of AI Agents
Lenny Rachitsky: Today my guest is Aparna Chennapragada. Aparna is chief product officer at Microsoft where she oversees AI product strategy for their productivity tools and their work on agents. Previously, she was chief product officer at Robinhood, vice president at Google, where she worked on Google lens, search, shopping, augmented reality, AI assistant, and a lot more. She was also a long-time engineering leader at Akamai, and on the board of eBay and Capital One.
In our conversation, we chat about how working in B2B is like being Jean-Claude Van Damme doing the splits across two moving trucks, how she’s operationalizing her team living in the future so that they’re building towards where things are going, why people still need to learn to code, why the PM role isn’t going anywhere, why NLX is the new UX, and so much more. If you enjoy this podcast, don’t forget to subscribe and follow it in your favorite podcasting app or YouTube. Also, if you become an annual subscriber of my newsletter, you get a year free of a bunch of products including Linear, Superhuman, Notion, Perplexity, and Granola. Check it out at lennysnewsletter.com and click bundle. With that, I bring you Aparna Chennapragada.
Aparna Chennapragada: Thank you, Lenny. Thanks for having me.
NLX: Natural Language Interfaces
Lenny Rachitsky: When I asked a lot of people that work with you, what I should ask you about and what I should know about you, something that came up again and again, it’s something that I think most people don’t know about you, which is that you’re big into stand-up comedy, and you take it semi-seriously. Just how serious are you about this? How much of your life is this and most importantly, how does this help you build better products?
Aparna Chennapragada: It’s hard to say I’m serious about a funny business, but I do watch and do stand-up comedy. I do open mics. I’ve done a few shows.
The Future of Product Management
Lenny Rachitsky: Wow.
Intelligence Surplus and Cognitive Updates
Aparna Chennapragada: I have one set brewing that is around AI, unsurprisingly AI and tech and Silicon Valley. It’s really interesting for me. This was an accidental discovery. I had always been an SNL fan and Discovery fan, but I went to an open mic because my son sings, and he went to the open mic for singing and he is like, “Mom, you should go do this.” And I was like, “Oh, let me go give it a try,” and I found that I enjoyed it and was good at it. To your question though, about building better products, I’d say both have PMF, I mean, product market fit, punchline market fit.
Actually, there are a couple of things that I do find really powerful and useful because in open mics or even when you’re testing these things, it’s a very tight cycle of iteration, and you get live… Open mics are the real live experiments. You put something out there, you get very clear micro feedback from users, and then you get tough feedback sometimes. And I think as product builders, that’s actually one of the great skills to have, which is you sometimes launch stuff that have a fantastic vision, but the first version is not quite there. And Reid Hoffman says this, “Hey, if you don’t launch the first version and are not embedded, you’re doing it too slow.” Just that gap in closing that, it’s good resilience.
AI Habits and Chrome Extensions
Lenny Rachitsky: Yeah. I never saw these corollaries between these two things. I didn’t realize you actually did shows, and you’re working on a set. I wasn’t going to ask you for a joke, but if you’re working on a whole thing about AI, is there something that you can share from that set?
Aparna Chennapragada: One joke I’d maybe share is people think about these AI chat products as women because you don’t know what’s going on. It’s a black box, and you don’t know what they’re thinking. There’s an entire set around that, but obviously on the flip side too, that they’re probably more like men in the sense that they hallucinate a lot. They kind of are not yet reliable.
Satya and Sundar’s Leadership Styles
Lenny Rachitsky: I’m afraid to laugh with this a little bit. This is great.
Aparna Chennapragada: And even when they don’t know the answer, they make up stuff. They’re very confident.
Counterintuitive Product Insights
Lenny Rachitsky: This is good. Where are we going to be seeing the show by the way?
Aparna Chennapragada: TBD.
The Dangers of Metrics
Lenny Rachitsky: Okay. This is great. Okay, let’s get serious again. So you worked at most of your career at a lot of consumer internet companies. You worked at Google, Robinhood, you’re on the board of eBay, you’re on the board of Capital One. Now, you’re at Microsoft. I’m curious just what is most different about working at a company like Microsoft and building product at a company like Microsoft?
Aparna Chennapragada: I think intellectually I knew that, hey, enterprise, particularly the area that I look at most at Microsoft is focused on enterprise and productivity and transforming companies through AI. And to me, I think two things really strike as very different. One, in fact, I just posted about this the other day saying, in consumer, you’re kind of like, “Oh, we have a playbook for make the product work or make the feature work and make it delightful,” but I think in the enterprise, you almost have… Every time you think you have one use case, you have really two, which is how do you make sure that the feature works well and there’s governance of the feature.
If you think about even something as simple as sharing a link to a document, you want it to be easy, frictionless, but at the same time, you want that to be secure and safe and being able to have auditability and all of those things. And often, I find that when you go from a consumer enterprise, you fall into a trap of either disregarding that and say, “Oh, we’ll just focus on one side of the house,” or overly crippling the user experience side and leaning on the other side. So I think there’s an art and science and nuance and playbook there too, so that’s one big learning for me. The other learning, and especially in the AI era for me has been about this… I think there’s a famous trailer from the 2000s on Van Damme on these two buses [inaudible 00:09:13] splits.
Three Inflection Points for Zero-to-One
Lenny Rachitsky: Like doing the splits.
Aparna Chennapragada: Yeah, doing the splits, exactly. I feel like a lot of the companies, including the tech companies, but certainly the enterprises that I talk to are in these two modes where one hand, this is the most compressed tech cycle that we’ve ever experienced. It’s in the order of weeks and months versus years and decades. If you think about mobile and cloud and internet, and there’s just so much happening, the intelligence overhang. On the other hand, there’s also humans and habits that… Productivity habits change. It’s hard to change and change management through the company is also hard. You don’t want to be rash on that. So it’s like the future is unevenly distributed but even within the companies.
Hot Seat: AI Coding Tool Landscape
Lenny Rachitsky: On the second bucket of the bus that Van Damme’s riding on of governance and adoption and changing behavior and stuff, is there something you’ve learned about how to get past that, help that along more?
The Enduring Power of Excel
Aparna Chennapragada: The thing not to do is hold back folks who are early adopters. I think that’s the other one learning. In fact, I think that’s one of the reasons why recently… I’ve been working with folks to say, “Can we have both,” which is the longer-term change management, being able to do it in a trusted way, at the same time do this program we are calling Frontier program and roll out cutting edge experimental features. We just built this world’s first deep research agent made for work, post-trained for work. And of course, it has all sorts of edges, rough edges, but if there are only adopters in an enterprise or outside, how can we put that in the hands of those folks without insisting that all of the company be completely developing different muscles?
Career Turning Points
Lenny Rachitsky: This program Frontier you’re talking about, I wanted to spend a little time on it. So what is the idea? The idea here is people are working in this futuristic environment. How does that actually work?
Aparna Chennapragada: Yeah, I think the idea is exactly this, which is I want to kind of institutionalize and operationalize my personal model of living one year in the future and say, “What does this…” Imagine a company or a setup like Frontier Consulting Group or Frontier Inc., right? And if you did live in that environment where you had all the AI tools and really advanced deep research intelligence on tap, what are the kinds of questions you’d be asking? What’s the kind of work you’d be doing? How would you change how you’re going about your work day? So that’s the premise and you’d say, “Hey, how does it change an individual?” But also down the lane, we want to think about what does a Frontier team look like. We talk a lot about Frontier labs and models. I think models layer is amazing and obviously that’s what empowers all these product building to happen, but I want to push us to think about what does a Frontier product look like? And more importantly, how does a Frontier way of working like? What does a team with three people and tons of compute and AI tools look like?
Future of Human-Agent Collaboration
Lenny Rachitsky: So how exactly does this work? There’s a team within Microsoft that’s like your job is to use all of our latest tools and build product using that. Does that work?
Aparna Chennapragada: That is the setup. We are just a few weeks into that setup, but meanwhile what we have done is we’ve actually set up a fake company and said, “Hey, if you are somebody who wants to come play with some of the cutting edge science projects and deep research agents and agents at work, come party here.”
Rapid Fire Questions
Lenny Rachitsky: Wow. And it’s only a few weeks in. Okay, so TBD how it all goes.
Aparna Chennapragada: Yeah. And again, these are micro… Let’s see. The meta point here also is that in the traditional way, we’ve kind of always thought about across the companies, across industries, really thinking about roll-outs in these macro ways. You build something and you kind of roll it out, you have a general availability for, and then you take the time. And that’s really important too because, again, we are talking about pharma companies, legal companies relying on this. So we do want to have that. But at the same time, given the compressed cycles of AI, how do we start to have people experience what’s the one year in the future?
Lenny Rachitsky: Let’s follow this thread in a few different directions. There’s how product development changes, there’s how engineering changes. There’s also just agents. I know you’re spending a lot of time in agents, feels like you’re not an AI company these days if you’re not working on agents or building an agent.
Aparna Chennapragada: Lenny, you’re doing this wrong. You didn’t use the word agents so far into the conversation.
Lenny Rachitsky: I try hard to push it out as far as I can. It’s like every conversation in San Francisco, it’s just like how long until I start talking about AI? It’s like three minutes. Average, I bet. Oh, man. Okay, so with agents, I know that you’re leading a lot of this work at Microsoft and a lot of people are wondering what the hell does this mean? What is going to change? Give us just a glimpse into how you see the world being different in a world of agents being around more.
Aparna Chennapragada: There’s a short term and there’s a long term, right? There’s a lot of hyperventilated, excited talk about the eventual future and all of that. I take a much more practical product building lens on this, and I think about these. At the end of the day, there are tools. Yes, underneath it, there’s stochastic models versus very deterministic programming models. You can tell I’m a computer scientist like the way that worldview definitely shapes how I think about this. To me, the short term is there’s an evolution. We had apps, and now I think we are firmly in the assistance era where there’s human driving the… That’s what we think of as co-pilot, right? I think the human driving in the driver’s seat but having a lot of assistance from AI.
So I think of this as then you look at the dimension of almost autonomy and delegation and intelligence. As the intelligence, for example, when deep reasoning unlock happened, of course, then you can delegate more to the agent. So I think, to me, there’s one dimension where you say, “Hey, agents are somewhat independent software processes that can kind of run tasks,” and you’re not just thinking about handholding and fine motor stuff. You’re saying, “Hey, here’s my goal. Go make this happen.” I’ll give you an example. So we are working on this researcher agent for work. And last night, I said, “Hey, I have an important meeting coming up with the leadership team. I really want to present these frameworks here and this is the roadmap here. Go back and look at all the people that are in the meeting. What are their views on this topic and come up with how I should I be thinking about the right persuasion pitch here?”
And what’s magical about this is not just that it’s saving time. Typically, we think about the, so far, AI as summarizing a document or saving time. This is like fighting synapses that I didn’t quite have and actually giving me new insights and giving me, dare I say, superpowers. So that’s a natural evolution of AI, I would say. So when I think about agents, I think about three things. One is an increasing level of autonomy and kind of independence that you can delegate higher and higher order tasks. Second thing I think of it is complexity. So it’s not just a one-shot, “Hey, create this image or do this thing or summarize the document,” it’s build me this prototype that expresses my idea of, say, an augmented reality app. It’s a complex task. And then the third thing I would say is asynchronous. It works when you are not working. I think that’s the other big thing about these things that you don’t have to sit in front of it.
Lenny Rachitsky: This answers the question of what is an agent essentially, these three bullet points. So what are the three again?
Aparna Chennapragada: When I think about agents, I think about these three things. So one, it’s autonomy like being… And it’s a spectrum, it’s not a zero-one, it’s how do I actually delegate things that it can do. Second, I think of as complexity. It’s not a one-shot, “Hey, summarize this document, generate this image, but it’s build me this prototype or help me knock this meeting out of the park.” And then the third one I think of is it’s a much more natural interaction. That doesn’t just mean chat, but it may be actually jumping on a meeting with the agent and being able to talk through all of it or point it to things that I wanted done differently. So I think all three things, the autonomy, the complexity, and the natural interaction are at least product principles that will shape really good ones, good agents.
Lenny Rachitsky: That is really helpful. Along this line of agents, there’s this acronym that you taught me as we were chatting ahead of this podcast, NLX, what is that and how does that relate to agents and why are people not thinking about this enough?
Aparna Chennapragada: Oh, that’s one of my Roman empires these days. The natural language interface. NLX is the new UX. Here’s the deal. To me, I think traditionally we’ve thought very consciously about GUI because the graphical interfaces are not something natural, and so they have had to be explicitly designed, but they’re rigid interfaces. What we have with conversational interface and natural language is it’s a much more elastic, right? That doesn’t mean it’s not designed. Often, I hear a product builders say, “Oh, yeah. With AI, the model leads to the product. So it’s just you chat with it.” You and I are having a conversation, it’s a podcast. I’ll have another conversation at Microsoft and that’s a meeting.
So conversations also have grammars, they have structures, they have UI elements, they’re invisible. And so one of the things that I see and I’m really excited about is what are the new principles, new constructs in natural language as an interface? I’ll give you a few examples. And actually a lot of startups as well as big companies are really experimenting with this stuff. One is if you think about it, prompt itself is a new construct and that’s a new UI element just like a dropdown was or a menu was. But others that are emerging, especially for agents, I think are plans. So when you give a high level goal, what we are seeing is that when the agent comes back with a plan, preferably an editable plan, that’s a new construct.
The other one that I think about a lot is showing the work, progress. You see this with different products. You see with the Copilot, you see with ChatGPT, DeepSeek, this idea of thinking aloud and it’s kind of showing the work, but how much do you do it? If it’s too verbose, it feels like I’m running some cron job and scripts, but if it’s too terse, then I don’t know if it’s going in the right path, and I don’t have the confidence yet. So there are all these new elements. So if you are a product whittler, this is a fun new space to be digging in for product design.
Lenny Rachitsky: This is really interesting because I think people chat with all these chat bots and it just feels like this is just the way it is, but you actually are designing every element of the interaction, how much to share, but how much you’re thinking, here’s my plan, what do you think. So I think this will surprise a lot of people, just realizing there’s so much that goes into just designing even these what seemingly are simple conversations.
Aparna Chennapragada: Yeah. Another good example is follow-ups, right? You could say, “Look, you asked me a question,” and then I could ask a follow-up set of things, and that’s explicitly should be designed for success. So for example, if I said, “Hey, create an image,” and it created a black and white like a clip art version of something. What are the next obvious follow-ups that it should be suggesting proactively? Now, too much and you are kind of annoying me, but too little in some sense, you’ve lost an opportunity to direct me or guide me into a happy path here.
Lenny Rachitsky: This resonates a lot with when we had Kevin Weil on the podcast, he talked about this question of just how much to show about what you’re saying. And it’s interesting that DeepSeek went the extreme of just showing everything and people liked it too. I think that was interesting.
Aparna Chennapragada: Yeah, and I think it’s a point in time thing too, Lenny, because in some sense right now these things are such black boxes. They’re almost like peeking under the hood for anything. Even if it’s verbose feels like, “Oh, I know what’s happening,” especially because the compute inference time, it’s taking long to think. So it just feels like if you just went silent, I’d be very uncomfortable, I think.
Lenny Rachitsky: I know.
Aparna Chennapragada: Exactly. So I do feel like there’s that point in time, but over time, I also feel like this is an area ripe for personalization. For example, again in human, my API would be very different from somewhere. My interface is probably different from others, and I might just want the direct, “Hey, give me the TLDR,” versus the, “Oh, so I went here and then I went there,” and I’m like…
Lenny Rachitsky: Following the start a little bit. We’re talking about just how the future is going to be different. There’s designing for these chat experiences, there’s agents, kind of zooming out to just product development in general, it feels like you’re at the forefront of a lot of the tools that are going to change the way we build products and also your teams are working with a lot of these tools that no one else has access to. So let me just ask, how do you see the future of product development being different from today most, and what do you think product builders should be preparing for doing to succeed in that future?
Aparna Chennapragada: I’ll start with one stark statement that I say internally and externally, and I am trying to live it is that in this day and age, if you’re not prototyping and building to see what you want to build, I think you’re doing it wrong. I call it the prompt sets of the new PRDs. I really insist on folks saying if you’re building new projects, new features of course come with prototypes and prompt sets. And I think the notion is not to say, “Hey, now everybody’s just a biggest version of a software engineer.” It is to say you have the fastest path to seeing and experiencing what’s in your mind to be able to communicate, right? It’s a much more high bandwidth way of communication. I think about that as a really a loop accelerator in terms of product building. That’s number one. When in doubt, as someone put it, demos before memos, right? I think that’s really number one.
I would say number two, this one is a little bit tricky I’d say, is that what I’m seeing is that the time to first demo is much shorter, but the time to a full deployment is going to take longer. So I think that there’s going to be an uneven cadence. So typically, I think there was much more of a you’ve been this thing, you take a few weeks and then you can iterate and so on. But that inner loop of prototyping and iterating and getting even user research through AI conversations, all of that gets shortened. But I think the bar for scale, therefore becomes much high. In some sense, if you look at it, there’s going to be a supply of ideas, a massive increase in supply of ideas in prototypes which is great. It raises the floor, but it raises the ceiling as well. In some sense, how do you break out in these times that you have to make sure that this is something that rises above the noise? So I would say that it’s simultaneously thinking about not chasing after every idea. I think is the second one.
I’d say the third thing is there’s a lot of conversation around full stack builders. What does the team of the future look like? A product building team. What I think about is I think that is inevitable in terms of there will be a few folks that are, especially at the prototyping early idea discovery stage that the lines of blurred, there’ll be a few taste makers at the same time. I think you can still have a lot of people experimenting. It becomes even more important to have that editorial and taste making in a Frontier, one or a few at the heart of it because otherwise you just have Frankenstein product. That definitely doesn’t change.
I have one other additional bonus thing, which is a lot of folks think about, “Oh, don’t bother studying computer science or the coding is dead,” and I just fundamentally disagree. If anything, I think we’ve always had higher and higher layers of abstraction in programming. We don’t program in assembly anymore. Most of us don’t even program in C, and then you’re kind of higher and higher layers of abstraction. So to me, they will be ways that you will tell the computer what to do, right? It’ll just be at a much higher level of abstraction, which is great. It democratizes. There’ll be an order of magnitude more software operators. Instead of Cs, maybe we’ll have SOs, but that doesn’t mean you don’t understand computer science and it’s a way of thinking and it’s a mental model. So I strongly disagree with the whole coding is dead.
Lenny Rachitsky: That’s awesome. I love that. And SO is a software operator, what is that? What that stands for?
Aparna Chennapragada: Yeah, I just made it up but yes.
Lenny Rachitsky: Okay, cool. This idea of prototyping as being kind of core to building these days, is there anything you do within Microsoft to operationalize that and make that just a thing everyone has to do? Is it just culturally do it or is it like you must show me a prototype before you show me it.
Aparna Chennapragada: Again, the future is here, unevenly distributed, even in Microsoft I would say, but there is certainly a strong cultural momentum and shift and desire to say, “Hey, let’s actually look at live demos, live prototypes, and to even communicate the ideas. And to me, I mean, it’s not always possible because obviously there are things that are deeply… If you’re trying to change something in the bowels of Excel, you probably don’t. There’s even enough depth in the product that what you need to do, and you don’t need to prototype that. But if you’re especially thinking about new things and new products, new features, absolutely.
Lenny Rachitsky: Okay, let’s talk about product management. There’s this fear that emerged as soon as all these AI coding tools came out of just like PMs are dead, we don’t need PMs. We could just build things ourselves. What are these people hanging around for? And what I found is it’s actually the opposite that now that coding is easy. Now, the question is more and more, what should we be building? Why should we be building it? Is this right? Is this the right solution? Then getting adoption for it, which is what PMs are really good at. I feel like it’s the opposite. PMs are the most important role. It’ll change too, but let me get your take. Just what do you think the future of product management looks like? Do you think it’s dead? Do you think it’s going to thrive? Do you think it’s going to change?
Aparna Chennapragada: Yes. Look, if you are a TPS report, mostly process person, and a lot of companies do get confused about product management and process and project management, I think then you do have a question of, “Hey, what is the value add here,” especially if AI can read and write 50,000 meeting notes and track things and send emails and so on, but what I do think on the flip side is the taste making and the editing function becomes really, really important. In a world where the supply of ideas, supply of prototypes becomes even more like an order of magnitude higher, you’d have to think about what is the editing function here.
So that does mean that the bar is higher for product folks, but I think there’s an interesting side effect I am observing in startups that I’m advising, companies and even within the companies that there used to be more gatekeeping I would say, in terms of like, “Oh, we should ask the product leader what they think.” And again, there is a role for that editing function, but you have to earn it now. You just don’t get it because of this title, but there’s also just unlock of latent really good ideas from smart engineers, smart user researchers, smart designers who now have this expert in their pocket to kind of round out all the other things that they’re not typically skilled at to bring forward their ideas and that’s amazing, I think.
Lenny Rachitsky: And I think that expert, it’s interesting, I’m working with an engineer on some stuff and he uses ChatGPT to even communicate to me in a more effective ways like, “Turn his pitch into something that will convince Lenny, this is a good idea.”
Aparna Chennapragada: By the way, that is actually one of my common use cases, which is the WWXD I call it. What would X do? I use it to say, “Hey, what would Satya think about this particular set of conversations or ideas that we are pitching and so on.” This is the power of, I think deep reasoning plus relevant context, right? This engineer you’re talking about has that context about you and so it’s kind of very interesting.
Lenny Rachitsky: If only everyone was as famous as Satya and had so much information out there, but I guess you can import all their emails or whatever tools exist to just understand from the conversations you’ve had with that person.
Aparna Chennapragada: Yeah. And I think this goes back to actually what you were saying too, which is I think this idea of what is the… There’s like a coil spring. There’s an intelligence overhang that I just see across the board. And I think the part of product development has to almost rewire ourselves to, I think, Tobi from Shopify calls it the reflexive AI usage. And that’s not as easy, and I’ve been thinking about why. Basically, I mean, I have a cheesy Chrome extension. Literally whenever I open a new tab, it just says, “How can you use AI to do what you’re going to do right now?” It’s very cheesy, but it kind of helps to pause and think, “Oh, what am I trying to do here?”
But the reason I find it hard, and when I talk even people who are living and breathing in this space, they find it hard is that the updating of the priors is really hard. The models couldn’t do some things one year ago. I mean, image generation was full of spellings or reasoning. You just couldn’t have deeper and smarter answers. You couldn’t do data analysis. So my impression of it from change, trying it a few months ago, that prior needs to be updated. And it’s hard to do that, right? You have to do something almost counterintuitive and against the grain to say, “No, no, ignore what you learned about what this can or cannot do.” The baby just grew up to be a 15-year-old in a month.
Lenny Rachitsky: I think that last point is so important that we’ve tried these tools over the years. And so far, it hasn’t been amazing and then all of a sudden it is, and you kind of don’t know that and you’ve given up almost and things change.
Aparna Chennapragada: I think that’s actually… If you are a product builder listening to it, that’s a really interesting arbitrage thing for you. If you can kind of cut against the grain and say, “No, I won’t have that scar tissue around.” This didn’t work a few months ago and keep setting high expectations and demand more of the AI today, I think you can unlock more.
Lenny Rachitsky: There’s a lot of alpha in doing that.
Aparna Chennapragada: That’s right.
Lenny Rachitsky:
With Coda’s collaborative all in one workspace, you get the flexibility of docs, the structure of spreadsheets, the power of applications, and the intelligence of AI all in one easy to organize tab. Like I mentioned earlier, I use Coda every single day and more than 50,000 teams trust Coda to keep them more aligned and focused. If you’re a startup team looking to increase alignment and agility, Coda can help you move from planning to execution in record time. To try it for yourself, go to coda.io/Lenny today and get six months free of the team plan for startups. That’s C-O-D-A .io/Lenny to get started for free and get six months of the team plan, coda.io/Lenny. I’m going to come back to this cheesy plugin, say more about this. So this is a plugin that just lets you put a custom message on every new tab, and you have it say, how can you use AI to do this?
Aparna Chennapragada: Yeah, it’s as cheesy as that. And it’s interesting because it works. In the last few weeks alone, I’ve been doing this experiment to say, “Hey, how much more AI pill can I get?” Both at work and in personal life to say, “When I’m trying to do anything manual, should I be demanding the AI to do this?”
Lenny Rachitsky: That’s so cool. Do you know the name of this Chrome extension by any chance otherwise?
Aparna Chennapragada: No. No. I built it.
Lenny Rachitsky: You built a Chrome extension. That’s so cool. Okay. Did you use AI to build it?
Aparna Chennapragada: Of course.
Lenny Rachitsky: Wow. Which tool did you use to do that? Some kind of Microsoft tool I imagine.
Aparna Chennapragada: Yes. No, actually, it was just like, I mean, I live in GitHub and GitHub Copilot, so I just was like, “Okay, let’s go build this Chrome extension.”
Lenny Rachitsky: Are you releasing this for the general public?
Aparna Chennapragada: No, I mean, that’s the amazing thing. It took me like 10 minutes to do this.
Lenny Rachitsky: Okay, let’s link to it. Let’s get it out there, open source this thing. Okay. You mentioned Satya, I have a question about this. So you’re one of the very few people that have worked very closely with both Satya and Sundar at Google. Let me ask you this. How do their leadership styles differ, and is there just a fun story you could share about each of them?
Aparna Chennapragada: Yeah. I do feel lucky to have a window into these two amazing leaders of this generation. I would say, I mean, again, no surprise as you’d expect from CEOs of multi-trillion dollar market cap tech companies, they are 99.99 percentile in almost every dimension you’d think of intellect, empathy, leadership, product, strategy. There are, of course, flavors of differences. I was the technical advisor for Sundar for the first… At Google and set up the office of the CEO there. And they’re, again, a matter of time and context because there’s a lot more consumer-oriented focus there. So what I did find Sundar great at it is being really calm and measured and thoughtful in terms of making sure that things have… Dealing with the complex ecosystems.
If you think about the phone ecosystem or even the search and publisher and advertiser ecosystem, it’s a very complex ecosystem. He was a master at that. He’s a master at that. And I think on Satya, I find it amazing the appetite he has for learning and fine tuning his mental models and just the zoom levels that he can operate at. The macro, the strategy, what’s the game? Also the micro, “Hey, why are we not…” Here’s a specific insight that I saw on Twitter, and you can count on the fact that he’s ahead of pretty much everybody else in terms of spotting those early things too. So it’s just been learning from the firehose as they put it.
Lenny Rachitsky: What a cool opportunity to work with two incredible folks. Okay, let’s go in a whole different direction. Let me just ask you this question that I’ve been asking people more and more. What’s the most counterintuitive lesson that you’ve learned about building products that goes against common startup wisdom, common product building wisdom.
Aparna Chennapragada: I don’t know if it’s as common as it should be, and it’s like a counterintuitive thing, but I’ve repeatedly learned that when you’re doing something new, zero-to-one, the temptation is to kind of think about… It’s like that South Park episode. Step one, think about the problem. Step two, question-
Lenny Rachitsky: Underpants. I think it’s Underpants, step one.
Aparna Chennapragada: Underpants. Exactly, right? So I do feel like there’s a temptation to rush and say to go to scale before solve. So I’ve always said to my teams solve before scale. So what that does mean is there’s a different posture and different mode when you’re trying to solve a problem versus scaling something that’s either post-product market fit or even at least in roughly in the ballpark. So to give you a couple of examples, I think when you look at the solved stage, there are wide lurches. You got to be very comfortable with the fact that you’re day one thinking about, “Hey, a plant detection tool.” And then day 15, you’re like, “Oh, actually, the tech is really good for translating foreign language.” By the way, this is not hypothetical. This is what we kind of looked at in Google Lens back then and said, “Okay, what is the intersection and so on?”
So from the outside, it looks like chaos, but actually, in the… And you should be very comfortable… Not only tolerant, I think you should have an appetite for that because the last thing you want is prematurely fix on one local hill. And then you’re climbing that in start-ups and entire product areas and companies, big companies make that mistake and three years later you’re like, “Oh, how do I get off this hill?” So I’d say that’s one big counterintuitive. When you’re trying to think about what mode you’re in, are you in the solved mode? Are you in the scale mode? One example is kind of making sure that you’re comfortable with the chaos. I think the other lesson I’ve learned is the danger of metrics. And I think again, if you have worked on Google Search or if you worked on Office products, you really have a very fine-grained sense of what are the metrics for this product?
You have the input metrics out, you have the whole shebang, but when you’re looking at something zero-to-one. If you decide on a metric two prematurely, that’s false precision first of all, right? I mean, CTR. When you have a thousand people, it doesn’t mean anything. Retention also may not mean anything. So really being very wary of this big guy, big girl of grownup metrics as I call it, right? You are looking for more qualitative, the sound of click, and what is your… The other kind of the handler uses, what is your set timer and play music? So if you look at Alexa and Siri and Google Assistant and all these things, they had a very promising broad interface. You could say anything, but I think there was one or two things that it was really good at. You could set a timer, you could play music, and you could play trivia. And so you’ve got to nail those things before you say, “Oh yeah, here you can do anything with it,” which is not a good recipe.
Lenny Rachitsky: Not so funny. That’s exactly what I use my Google Home for, so basic. I don’t do the trivia thing now maybe I got to give this shot.
Aparna Chennapragada: Got to try that. Yeah.
Lenny Rachitsky: There’s something along these lines that I’ve also seen you talk about, which is how to go zero-to-one with something, just a little framework for helping you know if this is the right time for this idea. How do you think about that?
Aparna Chennapragada: Yeah. And when you think about the solved mode, and this is again sticking with my whole living in one year in the future, I gravitated towards the zero-to-one and solved mode products completely thinking about new category of products. And what I’ve found, both the hard way I would say, is that you do want to look for at least two out of these three factors, inflection points here if you want to make a really good product. Number one is there a… Shift is a step function in the tech. That’s somewhat obvious I would say. Deep learning was one for Google lens. Back then, speech recognition was a step function for conversational search. I would say for Robinhood, the generational shift was very clearly, and the fact that phones were a primary means for you could actually have mobile app for finance that you could use. So look for that inflection. What is the tech inflection? And right now, of course, like LLMs and reasoning models are that step function, but that’s not enough.
I would say the second factor that we should look for is, what is the consumer behavior shift? So to give you an example, when we started working on Google Lens, what we said is, “Look, people were taking mostly pictures for sharing, selfies and sunsets and so on. And suddenly, when storage became free, mostly free, and everybody had phones everywhere all the time, you took pictures of everything. And then you had enough of pictures or you use the camera as the keyboard for your world, for the real world. And so how do you then say, “Oh, this consumer shift is big, and so therefore, as you go order of magnitude more photos, then you want more to come out of them and you can apply AI to that.”
And I’d say the third inflection point, particularly I would say in enterprise but also in consumer, is the business model shift. Is there an inflection natural inflection point in the business model? So any great products, if you think about all the way from search, again, the second price option and the fact that you had CPCs, same thing with SaaS and the fact that you could actually charge or monetize enterprise products in a different way. And with AI, of course the monetization is a whole different… We’ve just barely scratched the surface of whether you do seat monetization, usage like on tap, and then of course outcome-based stuff, outcome-based monetization. Hey, have you solved the problem for me and then I will pay you some fees. So all three to me are kind of like, great, but at least two out of three for a good product.
Lenny Rachitsky: So this essentially… When investors look at startups, they’re always asking, why now? Why is this the time to start this thing? And so your advice here is there’s three ways to look at it. Two of these three should be true. There should be a shift in technology, some new technology that has enabled this now recently. There’s a shift in consumer behavior, and then there’s maybe a new sort of… Or you’ve invented a new business model, any way to monetize something that it gives you an advantage over folks trying to do it today.
Aparna Chennapragada: Yep, absolutely.
Lenny Rachitsky: Awesome. You did mention Robinhood, I think in that example. That was another good example of phones-
Aparna Chennapragada: Yeah, I mean, talk about the business model of, again, not having a zero fees. And again, that combination of all of these things is what can unlock it. You can’t just say, “Oh, we’ll just have a much more better intuitive interface and hope that people will switch to it.”
Lenny Rachitsky: Okay, so speaking of zero-to-one products, I’m going to take us to a occasional segment on this podcast that I call Hot Seat Corner. And I have a question for you that is on my mind and it’s come up in a couple recent podcasts actually. So there’s these companies like Cursor, VZero, Lovable, Bolt, Replit that are the fastest growing company’s history. I just saw that Cursor hit 300 million ARR in two years. Interestingly, you guys were very well positioned to do really well in this space, this AI coding tool space. You guys had Copilot, the first tool in the world at this stuff, so ahead of everyone. You build VS Code, which is what all these companies are forking to build on. You have incredible AI infrastructure, incredible AI talent. So this could have been your market. What happened? What happened, Aparna?
Aparna Chennapragada: It’s interesting, the framing… So I’m a dead user of GitHub Copilot, and I would say, “Look, if you unpack…” I think the beauty of this is that code generation has become an amazing tool that LLMs have unlocked. So it is actually really good excitement and action that now code generation has just opened up all of these things that… We talked about the whole idea of prototyping, goes from idea to marks and idea to a clickable prototype in a few minutes. Those are the kinds of things that, of course, we should expect code generation to enable. The way I think about how we are positioned and what we do with GitHub is… So it’s a system, not just a product or a set of features.
If I think about GitHub, it’s for folks who have the repo there and you have… Of course, you have the assistance in terms of autocomplete and you can chat, but now we have the agent board. It’s one of the fastest loops that we are seeing, really strong positive feedback. So in some sense, when you have a system, what you are looking for in terms of building and designing it is not just a single product that can grow, but what is the repository? What is your context? What are the set of features that grow from your expertise? If you’re a really expert coder, you want the assistance this product needs to scale for that. If you’re a wide coder, you should still be able to do that and so on. So that I think is the way that GitHub is positioned to build on and growing honestly really well.
Lenny Rachitsky: That’s so interesting. So the core of this is everyone ends up in GitHub anyway, no matter what tool they use and that’s kind of the-
Aparna Chennapragada: Yeah. The idea again is that code generation as a tool will unlock lot more products. I mean, they’re not all competitors to the fact of… They’re not all kind of doing the same job. I think when you are… At the end of the day, you are building code for companies to run on, you need to have a system. You need to have kind of the ability, an entire Swiss Army toolkit, not just the autocomplete, not just a chat, not just like a software agent that runs and you kind of hand hold. You need all of this to work together, and that’s what the GitHub product is going after.
Lenny Rachitsky: All roads lead to GitHub. On the flip side of this question, there have been probably 5,000 startups that have tried to disrupt Excel and you guys just keep winning, so something there is working really well.
Aparna Chennapragada: That is so interesting you say that. So when I came to Microsoft, and I’m an Excel fan, so I actually had a conversation with one of the OG Excel product folks. I was like, “an, what is it about this product?” And he said a couple things that were really interesting for me that just stuck with me. One is and I said, “Hey, Excel is a proof that non-coders also have to program.” Programming is really powerful and it’s the tool that gives all of the non-coders a really powerful programming ability, and I thought that was just really striking.
And then the second thing that I found out super cool, I don’t know if you know this, but I didn’t know at least before two years ago that there are these amazing Excel championships like World Excel championships where you see folks who can do just magic. And to me, I think the insight here is also that some tools are harder to learn. Perhaps in the beginning there’s friction in terms of learning, but great to use. So it’s a very good case of, hey, the learning curve initially, the one-time learning curve might be tricky, but it is because there’s so much power and depth in the tool.
Lenny Rachitsky: That’s so interesting. I never thought of Excel as a programming language, but it makes sense and I feel like once you get used to it and this is just the way things work, you’re kind of stuck there and everything else has to basically copy that model, which is hard to be as good.
Aparna Chennapragada: Yeah. And I think the depth then the attention that the team has given, and again, that’s the compounding effect over decades of working on deep, deep signal from people who depend on it day in and day out.
Lenny Rachitsky: Okay. To kind of start to close out our conversation, I want to ask this question around your career. I find that most people have one moment in their career that changes the trajectory of their career. It could be a manager they had, it could be a project they worked on, it could be just the job they landed. What would you say is the most pivotal moment in your career that eventually led you to becoming chief product officer at Microsoft?
Aparna Chennapragada: Actually, there is one moment where it was a turning point for me. I was in Google Search, I was working on this idea that I thought should just work and it didn’t. I said, “Hey, these phones are becoming a thing. Personalization has to be important.” So I probably banged my head against the wall for a year or so trying to make personalization work. And it turns out when you have a query that you put into Google Search, the personalization didn’t matter as much. And so we disbanded the team, but then I think I started working on this product called Google Now, which was a twist on that, which said, “Hey, actually on the phone, we should be able to push content. It’s not about searching with personalization.” For example, if you have a flight coming up, we should be able to say, “Hey,” connect the dots and say, “you should leave now given the traffic and where you need to go,” and so on or if you’re deeply interested in stand-up comedy with deadpan artists, you should check out Mitch Hedberg.
These are kind of these really moments that the smartphone should be smarter. So I let that product through the initial zero-to-one phase, and that was a pivotal moment. It made me realize two things. One, I really love seeing around the corner and kind of seeing where things go and building the product rise to the occasion way more than the scaling and sustaining products. Second, it’s harsh, but being early is the same as being wrong. This is pre-LLMs, pre-deep learning a lot of the really amazing ideas in terms of next token predictor, et cetera. We’d been thinking of it but didn’t have the horsepower to go… The interface was great, the intelligence wasn’t there. And I’d say the third thing that stuck with me is I got to work with some really smart… They talk about talent density now, and I think really smart people who have gone on to do amazing things, and so it gave me a taste of what a small group of people can do.
Lenny Rachitsky: It’s such a great story because it didn’t work out in the end. Google Now kind of went away. And by the way, I super remember that product. It was very cool. I remember looking at it was very delightful and happy. And so I also have this segment on the podcast called Failure Corner, where people share a story of failure and how that helped them. And I love this as a combination of those two.
Aparna Chennapragada: Yeah. I mean, I’m not going to lie. I think it was painful when you do that because you see the vision of what can be and what is, and sometimes it’s hard limitations. Sometimes, in this case, it takes five years or 10 years to really unlock the intelligence, but sometimes it’s one or two key click stops away from the product being great and part of figuring out is knowing when you’re in what situation.
Lenny Rachitsky: How long was that period from starting out until just moving on and it’s not working?
Aparna Chennapragada: Yeah, I would say in that case, one of the good things is, again, it led the foundation of… It was one of the foundations of the Google Assistant. And of course, as the LLMs step function happened now with Gemini, it kind of works out. And I think it’s the same thing across the board, which is sometimes you want to figure out the invariance that do work that then go on to the next version of the product. And other times, you just have to start over.
Lenny Rachitsky: Is Google Now the first agent before agents? That’s what it feels like.
Aparna Chennapragada: That was certainly the idea, but it is fascinating to me that the interface, that there, we had the opposite problem. Whether you think about all the voice assistants, the interface is like we overshot and the intelligence wasn’t there. Today, I feel like there’s an opposite problem. I think these things have amazing intelligence and the interface we have largely is like the AOL Dial-Up Modem Chatbot.
Lenny Rachitsky: We’ve covered a lot of ground. Is there anything that you wanted to chat about or leave listeners with, maybe a last nugget of wisdom before we get to a very exciting lightning round?
Aparna Chennapragada: I think I would say one thing that I’m really excited about is this idea of figuring out how we as people and agents collaborate together. I think there’s some great set of products and experiences to be reimagined. That’s my other Roman empire, which is how do we actually have this co-working space where you have the humans and agents and how do you actually have an output that’s much, much more significant than what any one of us or any few of us can produce?
Lenny Rachitsky: Well, I need to hear more about this. When do you imagine a co-working space of humans and agents? What does this look like? Is this Microsoft teams or is this a physical place with little robots?
Aparna Chennapragada: Oh, I had a thought of the physical place, but I am thinking a lot about… Right now, all of these experiences are very civil player, and I do think there’s an opportunity to think about how do we… Again, I’m living one year in the future, how do we actually have collaborate with each other, but also with agents and really figure out, for example, what tasks can we delegate? What can we inspect? How do we actually have information that flows between people that agents can mediate, and so on.
Lenny Rachitsky: All right, I’m curious to see what you guys got cooking. With that, we’ve reached our very exciting lightning round. Are you ready?
Aparna Chennapragada: Let’s do it.
Lenny Rachitsky: Let’s do it. First question, what two or three books that you find yourself recommending most to other people?
Aparna Chennapragada: Oh, I have recency bias, but I’ve been reading this book called The Brief History of Intelligence, phenomenal book and like lots of underlining from me. And I think it kind of… The premises too, it looks at the evolution of intelligence like human intelligence and the brain development and connects that to what we are seeing with AI.
Lenny Rachitsky: Do you have a favorite recent movie or TV show that you’ve really enjoyed?
Aparna Chennapragada: Hacks. I’ve been watching this. It’s about a woman who’s a great standup comedian of… I think it’s set in the fact that she grew up in the ’70s and ’80s and really tried to break through in an industry that hasn’t traditionally been very friendly to women, so really fun and quirky.
Lenny Rachitsky: Do you have a favorite product that you’ve recently discovered that you really love, could be an app, could be some physical?
Aparna Chennapragada: I do use a lot of Microsoft products, GitHub Copilot being one of them, but I think the one that maybe I’ll pick is Granola, I think, is the name of the app. I found it really useful. I just gave it a spin the other day and I’m like, “Oh, this is really useful in terms of being able to, again, without being intrusive, just capture the thoughts, notes, and structure it, put some…” It felt like one of those things where, yep, the confidence of a few things like we were talking about like the transcription, real-time transcription tech has gotten really good. Voice recognition is great, and then enough of the LLM magic on top of it to make it structured and contextual.
Lenny Rachitsky: I am a huge fan of Granola. I’ll give a quick picture here. If you become an annual subscriber of my newsletter, you get a year free of Granola for your entire company.
Aparna Chennapragada: Did not know that.
Lenny Rachitsky: There we go, and then just check that out, lennysnewsletter.com, and you click the word bundle and you’ll see how to do that.
Aparna Chennapragada: Very cool.
Lenny Rachitsky: Very cool. Two more questions. Do you have a favorite life motto that you often come back to when you’re dealing with something maybe you share with folks that they find useful as well in work or in life?
Aparna Chennapragada: I have one. In fact, actually, this is my email signature for, I don’t know, for the last 20 years or so. It says the best way to predict the future is to invent it. I think it’s a quote by Alan Kay. I find it useful for two things. One is no one knows anything. When you think about all the folks who think about, “Hey, this is exactly how everything’s going to look and this is exactly the sequence,” and so on, I think there’s no substitute to experientially building it. I think the second part is if you think there’s something that should exist, go build it.
Lenny Rachitsky: I love that. Final question. We’ve talked about standup comedy a bit. Is there a favorite under the radar standup comedian that you think people should go check out?
Aparna Chennapragada: Oh, there’s a couple of them. So one, I think, there’s an Indian American or I think a British Indian standup comedian. Her name is Sindhu Vee, super smart, mom comedy, and I think the other one that… This is definitely not under the radar, but I just love his stick is Nate Bargatze. He’s just so good.
Lenny Rachitsky: Aparna, this was amazing. Two final questions. Where can folks find you online if they want to reach out maybe and follow up on anything you shared and how can listeners be useful to you?
Aparna Chennapragada: You can find me on LinkedIn and Twitter. Aparna CD is the handle. I do post stuff a lot more on LinkedIn these days, so would love to hear thoughts, comments, conversations there. I’d say one thing that would be super interesting is if any of this stuff spark conversations, particularly around this, what can a small team with a lot of AI tools do or new products that folks are really excited about, saying that they should exist, hit me up.
Lenny Rachitsky: Amazing. Aparna, thank you so much for being here.
Aparna Chennapragada: Thank you.
Lenny Rachitsky: Bye, everyone. Thank you so much for listening. If you found this valuable, you can subscribe to the show on Apple Podcasts, Spotify, or your favorite podcast app. Also, please consider giving us a rating or leaving a review as that really helps other listeners find the podcast. You can find all past episodes or learn more about the show at lennyspodcast.com. See you in the next episode.
Glossary
| English | 中文 |
|---|---|
| agent mode | agent 模式 |
| Alan Kay | Alan Kay(计算机科学家) |
| alpha | alpha(套利机会/超额收益) |
| ARR (Annual Recurring Revenue) | ARR(年度经常性收入) |
| ChatGPT | ChatGPT |
| Copilot | Copilot |
| CPO (Chief Product Officer) | 首席产品官 |
| cron job | cron 定时任务 |
| DeepSeek | DeepSeek |
| fork | fork(基于已有代码库创建分支) |
| full stack builders | 全栈构建者 |
| Granola | Granola(AI 会议笔记应用) |
| GUI (Graphical User Interface) | GUI(图形用户界面) |
| intelligence overhang | 智力过剩 |
| Jean-Claude Van Damme | 尚格·云顿 |
| Kevin Weil | Kevin Weil |
| Mitch Hedberg | Mitch Hedberg(美国冷面笑匠风格脱口秀演员) |
| Nate Bargatze | Nate Bargatze(美国单口喜剧演员) |
| next token predictor | 下一个 token 预测器 |
| NLX (Natural Language Interface) | 自然语言界面 |
| PMF (Product-Market Fit) | 产品市场契合 |
| PRD (Product Requirements Document) | PRD(产品需求文档) |
| Reid Hoffman | 里德·霍夫曼 |
| Satya Nadella | Satya |
| scar tissue | 旧伤疤(过往负面经验形成的心理障碍) |
| Sindhu Vee | Sindhu Vee(印裔英国单口喜剧演员) |
| SNL (Saturday Night Live) | 《周六夜现场》 |
| software operator | 软件操作者 |
| Sundar Pichai | Sundar |
| talent density | 人才密度 |
| Tobi (Tobi Lütke) | Tobi |
| TPS report | TPS 报告 |
| UX (User Experience) | UX(用户体验) |
| WWXD (What would X do) | WWXD(X 会怎么做) |
Reformatted by reformat_english.py
Microsoft 首席产品官:如果你没有用 AI 做原型,那你的做法就是错的 | Aparna Chennapragada
文字稿
开场精选
Aparna Chennapragada: 我有一个有点傻的 Chrome 扩展。每次我打开一个新标签页,它就显示一句话:你怎么用 AI 来做你马上要做的事?
Lenny Rachitsky: 你认为产品开发的未来会有什么不同?
Aparna Chennapragada: 如果你不做原型、不通过构建来看清自己到底想做什么,我认为你的做法就是错的。在这种时候,把领地意识和品位把控放在核心变得更加重要,否则你只会得到一个弗兰肯斯坦式的拼凑产品。
Lenny Rachitsky: 你教过我这个缩写——NLX,那是什么?
Aparna Chennapragada: 自然语言界面(Natural Language Interface)。NLX 是新的 UX。我经常听产品构建者说,“哦,有了 AI,模型吞噬了产品。“但这不意味着它不需要设计。你我正在进行一场对话,这是一档播客;我在微软还会有另一场对话,那是一场会议。对话也有语法,有结构,有 UI 元素,只是它们是无形的。那么,作为界面的自然语言,有哪些新的原则和构造?
Lenny Rachitsky: 我刚看到 Cursor 两年就达到了 3 亿 ARR。有意思的是,你们在这个 AI 编码工具领域其实处于非常有利的地位。你们有 Copilot,是这个领域世界上第一个工具,领先所有人那么多。后来发生了什么?
Aparna Chennapragada: 我想说的是……
Lenny Rachitsky: 今天的嘉宾是 Aparna Chennapragada。Aparna 是 Microsoft 的首席产品官,负责其生产力工具的 AI 产品策略以及 Agent 相关工作。此前她是 Robinhood 的首席产品官、Google 副总裁,在那里负责 Google Lens、搜索、购物、增强现实、AI 助手等多项业务。她也曾是 Akamai 的资深工程领导者,并在 eBay 和 Capital One 董事会任职。
在我们这次对话中,我们聊到了做 B2B 就像尚格·云顿在两辆行驶的卡车之间做劈叉,她如何让团队”活在未来”以朝事物发展方向构建,为什么人们仍然需要学习编程,为什么 PM 这个角色不会消失,为什么 NLX 是新的 UX,以及更多内容。
Lenny Rachitsky: Aparna,非常感谢你来这里,欢迎来到播客。
Aparna Chennapragada: 谢谢你,Lenny。感谢邀请。
单口喜剧与产品构建
Lenny Rachitsky: 当我问很多和你共事过的人,我应该问你什么、应该了解你什么,有一件事被反复提到,而且我觉得大多数人并不知道——你非常喜欢单口喜剧,而且对此还挺认真的。你到底有多认真?这在你生活中占多大比重?最重要的是,这怎么帮你做出更好的产品?
Aparna Chennapragada: 很难说我对于一个搞笑的事情有多”认真”,但我确实会看也会演单口喜剧。我参加开放麦(open mic),已经做过几场演出。
Lenny Rachitsky: 哇。
Aparna Chennapragada: 我正在酝酿一套关于 AI 的段子——不出所料,关于 AI、科技和硅谷。这对我来说很有意思,是一个偶然的发现。我一直都是 SNL(《周六夜现场》)的粉丝,也是 Discovery 的粉丝,但我去开放麦是因为我儿子唱歌,他去开放麦唱,然后他说”妈,你也应该去试试。“我说”好吧,让我试试看。“然后我发现我既喜欢这件事又擅长这件事。回到你的问题,关于构建更好的产品,我觉得两者都有 PMF,我是说产品市场契合(product-market fit),以及包袱市场契合(punchline-market fit)。
实际上有几件事我觉得非常有力量、非常有用。在开放麦或者测试这些段子的时候,这是一个非常紧凑的迭代循环,你会得到实时的……开放麦就是真正的活实验。你把东西放出去,从用户那里得到非常清晰的微观反馈,有时会得到很严厉的反馈。我认为作为产品构建者,这其实是应该具备的重要能力之一——你有时发布的东西有很棒的愿景,但第一版还不太到位。Reid Hoffman 说过,“如果你发布的第一版不让你觉得尴尬,那说明你太慢了。“去弥合那个差距,就是一种很好的韧性。
Lenny Rachitsky: 我从来没想过这两件事之间有这些共通之处。我不知道你真的上台表演过,还在准备一整套段子。我本来不打算让你讲笑话的,但既然你在准备一整套关于 AI 的内容,有没有什么可以分享的?
Aparna Chennapragada: 我也许可以分享一个笑话——人们觉得这些 AI 聊天产品像女人,因为你不知道里面在发生什么,它是个黑箱,你不知道它们在想什么。围绕这个有一整套段子,但显然反过来也成立——它们可能更像男人,因为它们经常产生幻觉,而且目前还不太可靠。
Lenny Rachitsky: 我都有点不敢笑了。这太棒了。
Aparna Chennapragada: 而且就算不知道答案,它们也会编一个出来,而且非常自信。
Lenny Rachitsky: 不错。话说我们到时候去哪里看这个演出?
Aparna Chennapragada: 待定。
从消费互联网到企业产品
Lenny Rachitsky: 好的。非常好。好吧,让我们回到正题。你职业生涯的大部分时间都在消费互联网公司工作——Google、Robinhood,你是 eBay 和 Capital One 的董事会成员,现在在 Microsoft。我很好奇,在 Microsoft 这样的公司工作、在 Microsoft 做产品,最不同的地方是什么?
Aparna Chennapragada: 我觉得从理智上我早就知道,企业市场——尤其是我在 Microsoft 最关注的领域,就是围绕企业和生产力、通过 AI 来变革企业——有两点让我觉得非常不同。第一点,其实我前几天刚发了一个帖子说,在消费领域,你大概会觉得”我们有一套打法:把产品做好、把功能做好、让它令人愉悦”,但在企业领域,你几乎……每次你觉得自己有一个用例,实际上你有两个——一个是确保功能本身运作良好,另一个是确保对这个功能有治理机制。
就拿分享一个文档链接这么简单的事情来说,你希望它方便、无摩擦,但同时又希望它安全可靠、可审计,等等。我经常发现,从消费转到企业时,人们容易掉进一个陷阱:要么完全忽视其中一面,说”我们就专注一边就行了”;要么过度牺牲用户体验,把重心全压在另一边。所以这里面也有艺术、科学和微妙之处,也有一套打法。这是我的一个大收获。另一个收获,尤其是在 AI 时代,让我感触特别深的是……我想 2000 年代有一支很著名的广告片,尚格·云顿站在两辆卡车之间做一字马劈叉。
Lenny Rachitsky: 对,做一字马劈叉。
Aparna Chennapragada: 对,做一字马劈叉,没错。我觉得很多公司,包括科技公司,当然还有我交流过的那些企业,都处于这两种状态之间:一方面,这是我们有史以来经历过的最压缩的技术周期——时间尺度是几周和几个月,而不是几年和几十年。想想移动、云计算、互联网的发展就知道了,而现在有这么多事情在同时发生,还有智能过剩。但另一方面,人也存在惯性和习惯——生产力习惯是很难改变的,整个公司的变革管理也很难。你不能操之过急。所以就像那句话说的:未来分布不均,而且即使在同一家公司内部也是如此。
Lenny Rachitsky: 关于尚格·云顿骑的那第二辆卡车——也就是治理、采用和改变行为这些方面——你有没有学到什么方法来推进这些、帮助克服障碍?
不要拖住早期采用者
Aparna Chennapragada: 一个不该做的事情就是拖住那些早期采用者。这是我的另一个收获。事实上,这也是为什么最近我一直在和同事们说:“我们能不能两者兼顾?“——既做长期的变革管理,以可信的方式来推进,同时又推出一个我们叫做 Frontier 的项目,发布最前沿的实验性功能。我们刚刚构建了世界上第一个专为工作场景打造的深度研究智能体,针对工作场景做了后训练。当然,它还有各种各样的粗糙之处、边界问题,但如果企业内部或外部有一些早期采用者,我们怎么能把这些工具交到他们手上,而不要求整个公司都立刻发展出不同的能力?
Lenny Rachitsky: 你提到的这个 Frontier 项目,我想多花点时间聊聊。具体想法是什么?就是让大家在一个未来的环境中工作?实际怎么运作的?
Frontier 项目:提前一年生活在未来
Aparna Chennapragada: 对,我觉得想法就是这样——我想把我个人”提前一年生活在未来”的模式制度化、可操作化,然后说”想象一下……”想象一家公司,或者一个像 Frontier 咨询公司、Frontier 公司这样的机构。如果你真的生活在那种环境中——拥有所有 AI 工具,还有真正先进的深度研究智能力随时可用——你会问什么样的问题?会做什么样的工作?会怎么改变自己的工作方式?这就是前提。你会说”这如何改变一个个体?“但再往远处看,我们还要思考一个 Frontier 团队是什么样的。我们经常谈论 Frontier 实验室和 Frontier 模型。我觉得模型层非常了不起,它显然是所有这些产品构建的基础,但我想推动大家去思考:一个 Frontier 产品是什么样的?更重要的是,一种 Frontier 的工作方式是什么样的?一个三个人加上大量算力和 AI 工具的团队是什么样的?
Lenny Rachitsky: 那具体怎么运作?Microsoft 内部有一个团队,任务就是使用所有最新工具来构建产品,是这样吗?
Aparna Chennapragada: 这就是基本设定。我们才刚开始几周,但与此同时我们实际上注册了一家假公司,然后说:“如果你想体验一些最前沿的科学项目和深度研究智能体、工作中的智能体,来这里玩。”
Lenny Rachitsky: 哇,才几周。好的,所以最终效果如何还有待观察。
Aparna Chennapragada: 对。而且再说一下,这些都是微小的……让我们看看。这里的元观点是:传统上,我们在跨公司、跨行业的时候,总是以宏观的方式思考推广。你构建一个东西,然后把它推广出去,做一个正式发布,然后花时间慢慢来。这确实很重要,因为我们说的是制药公司、法律公司在依赖这些东西,所以我们确实需要那种稳重的节奏。但与此同时,考虑到 AI 的压缩周期,我们怎么让大家提前体验一年之后的未来?
智能体时代
Lenny Rachitsky: 让我们沿着这个线索往几个不同方向展开。产品开发方式如何变化,工程方式如何变化,还有智能体(agents)。我知道你在智能体上花了很多时间,现在感觉如果你不是在做智能体或构建智能体,你都不好意思说自己是 AI 公司。
Aparna Chennapragada: Lenny,你这就不对了。都聊到这儿了你才第一次说到”智能体”这个词。
Lenny Rachitsky: 我尽量把它往后推。在旧金山,每一段对话就是看多久之后会开始聊 AI——平均大概三分钟吧,我打赌。天哪。好吧,关于智能体,我知道你在 Microsoft 领导了很多这方面的工作,很多人都在想这到底意味着什么?会发生什么变化?给我们透露一下,你觉得在一个智能体更加普及的世界里,世界会有什么不同?
Aparna Chennapragada: 有短期和长期两个维度。大家对未来的远景有很多过度兴奋的谈论,但我更倾向于从一个务实的、做产品的视角来看待这些问题。说到底,这些都是工具。底层当然是随机模型,而不是非常确定性的编程模型。你可以看出来我是计算机科学出身的——这种世界观确实影响了我对这些问题的思考方式。对我来说,短期来看是一种演进。我们经历了应用时代,现在我认为我们已经进入了辅助时代——人在主导……这就是我们所说的 Copilot,对吧?人在驾驶座上,但有大量 AI 辅助。
所以我认为可以沿着几乎自主性、委托程度和智能水平这个维度来观察。当智能水平提升时,比如深度推理的突破出现时,你自然可以把更多事情委托给智能体。所以对我来说,一个维度是——智能体是某种程度上独立的软件进程,可以自行执行任务,而不仅仅是需要你手把手地操控那些精细操作。你说的是,“嘿,这是我的目标,去把它搞定。“我举个例子。我们正在开发一个面向工作的研究智能体。昨晚我对它说:“我马上要和领导团队有一个重要会议,我真的很想展示这些框架,这是路线图。你回去查一下所有参会的人,了解他们在这个话题上的观点,然后帮我想想我应该如何构建最有说服力的提案。”
这件事的神奇之处不只是节省时间。过去我们通常把 AI 看作是总结文档或节省时间的工具。而这次更像是激活了我原本不具备的神经连接,真正给了我新的洞察,甚至可以说给了我超能力。所以这是 AI 自然的演进方向。当谈到智能体时,我会想到三个维度。一是不断提高的自主性——你可以把越来越高阶的任务委托出去。二是复杂性——它不只是”帮我生成这张图片”或”帮我总结这个文档”这种一次性操作,而是”帮我搭建一个原型来表达我关于某个增强现实应用的想法”——这是复杂任务。第三点是异步性——它在你不在的时候也在工作。这是这些智能体的另一个重要特点,你不需要一直坐在它面前。
Lenny Rachitsky: 这回答了”什么是智能体”这个问题,就是这三个要点。再说一遍是哪三个?
Aparna Chennapragada: 当我想到智能体,我会关注这三点。第一,自主性——它是一个光谱,不是非零即一的,而是我究竟能委托它做多少事情。第二,复杂性——不是一次性的”总结这个文档、生成这张图片”,而是”帮我搭建这个原型”或”帮我把这次会议拿下”。第三,更自然的交互。这不只是聊天,而可能是直接和智能体一起开个会,能够边讨论边指出哪些地方我想要不同的做法。所以自主性、复杂性和自然交互,这三者至少是我认为能塑造出优秀智能体的产品原则。
NLX:自然语言界面
Lenny Rachitsky: 这非常有帮助。沿着智能体这条线,有一个你在我们播客录制前聊天时教我的缩写——NLX,它是什么?它和智能体有什么关系?为什么大家对它关注得不够?
Aparna Chennapragada: 哦,这是我这段时间最着迷的话题。自然语言界面(NLX)。NLX 就是新的 UX。关键在于:过去我们对 GUI 的思考非常刻意,因为图形界面不是天然的东西,所以必须经过专门设计,但它们是刚性的界面。而对话式界面和自然语言则更有弹性。但这不意味着它不需要设计。我经常听到产品人说:“有了 AI,模型本身就决定了产品,所以你只需要跟它聊天就行了。“你我在录播客,这是一段对话;我在 Microsoft 还会有另一段对话,那是一个会议。
对话也有语法、有结构、有 UI 元素,只是它们是不可见的。所以我现在非常关注、也感到非常兴奋的是:自然语言作为界面,有哪些新的原则和新的构造?我举几个例子。实际上很多创业公司和大公司都在这方面做大量实验。第一,如果你仔细想,提示词本身就是一个新的构造,它就是一个新的 UI 元素,就像当年的下拉菜单或菜单栏一样。但正在涌现的、尤其是针对智能体的新构造,我认为是计划。当你给出一个高层目标时,我们看到当智能体返回一个计划——最好是可编辑的计划——那就是一个新的构造。
另一个我经常思考的是展示工作过程、展示进度。你在不同产品中都能看到这一点。Copilot、ChatGPT、DeepSeek 都有,就是”出声思考”的理念,展示工作过程。但展示多少?太多了,感觉像是在跑什么 cron 脚本;太少了,我又不知道它是否走在正确的方向上,缺乏信心。所以这些都是新的界面元素。如果你是做产品的,这是一个值得深入挖掘的有趣新领域。
Lenny Rachitsky: 这真的很有意思,因为大家跟这些聊天机器人对话时会觉得”就是这样嘛”,但实际上你在设计交互的每一个元素——展示多少、思考过程展示多少、“这是我的计划,你觉得怎样”。我觉得这会让很多人惊讶,意识到即便是看似简单的对话,背后都有大量的设计工作。
Aparna Chennapragada: 对。另一个很好的例子是后续跟进(follow-ups)。你问了我一个问题,我可以接着提一系列后续问题,这应该被刻意设计好。比如我对它说”生成一张图片”,结果它生成了一张黑白剪贴画风格的东西。它应该主动建议哪些明显的后续操作?太多了会让人觉得烦,太少了则白白错失了引导用户走向满意路径的机会。
Lenny Rachitsky: 这让我很有共鸣。Kevin Weil 上我们播客的时候也谈到了这个关于”展示多少思考过程”的问题。有趣的是 DeepSeek 走了极端——什么都展示出来,结果大家还挺喜欢。我觉得这很有意思。
Aparna Chennapragada: 对,而且我觉得这也是一个时间点的问题,Lenny。在某种意义上,现在这些东西还像黑箱一样,所以对任何东西人们都想偷看一眼内部。即使很啰嗦,也会觉得”哦,我知道在发生什么”。尤其是因为推理计算时间比较长,它在思考。所以如果它突然沉默了,我会非常不安。
Lenny Rachitsky: 我理解。
Aparna Chennapragada: 没错。所以我确实觉得有这样一个时间节点的问题,但随着时间推移,我也觉得这个领域非常适合个性化。比如,还是以人来做类比,我的 API 和别人的会很不一样。我的界面可能也和别人不同,我可能只想直接得到”给我个摘要”,而不是”哦,我先去了这儿,然后又去了那儿”,我就……
Lenny Rachitsky: 稍微跟一下开头。我们在谈未来会有什么不同。设计聊天体验是一方面,智能体是另一方面。把视角拉远到整体的产品开发——你似乎站在了很多即将改变我们构建产品方式的工具的最前沿,而且你的团队也在使用很多其他人还接触不到的工具。所以我想直接问:你认为未来的产品开发和今天最大的不同是什么?产品构建者现在应该做些什么准备,才能在未来取得成功?
Aparna Chennapragada: 我先说一个我在内部和外部都在说的、相当直白的观点,我也在努力践行它——在当今这个时代,如果你不是在通过原型和构建来看到自己想做什么,那你的做法就是错的。我把 prompt 集称为新的 PRD(产品需求文档)。我真的坚持要求大家在启动新项目、新功能时,必须带着原型和 prompt 集来。这个意思不是说”嘿,现在人人都变成了最高级别的软件工程师”,而是说你拥有了最快的路径,把你脑海中的东西看到、体验到,然后能够沟通出来,对吧?这是一种带宽高得多的沟通方式。我认为这是产品构建中一个真正的加速器。这是第一点。拿不准的时候,就像有人说的那样——先做演示,别写备忘录。我觉得这确实是第一要务。
第二点,这一点有点微妙——我看到的情况是,做出第一个演示的时间大大缩短了,但完整部署的时间反而会更长。所以节奏会出现不均衡。以前通常是:你锁定一个东西,花几周时间,然后迭代,如此往复。但现在,原型制作、迭代、甚至通过 AI 对话来做用户研究,这个内循环整个都缩短了。不过,达到规模化标准的门槛因此变得更高了。从某种角度看,创意和原型的供给会大量增加,这很好。它抬高了下限,但也拉高了上限。在这种情况下,如何脱颖而出?你必须确保这个东西能越过噪音被看到。所以我想说的是,同时也要注意不要去追逐每一个想法。这是第二点。
第三点,现在有很多关于”全栈构建者”的讨论。未来的团队会是什么样子?产品构建团队会是什么样?我认为这是不可避免的趋势——在原型和早期创意探索阶段,会有一些人角色界限变得模糊,同时也会出现几个真正有品味的人。我觉得你仍然可以让很多人去实验,但在前沿领域,有一个或几个具备编辑判断力和品味的人作为核心变得更加重要,否则你就只会得到一个拼凑出来的科学怪人式产品。这一点绝对没有变。
还有一个额外要说的——很多人觉得”哦,不用学计算机科学了,编程已死”——我完全不同意。如果说什么的话,我认为编程一直都在走向越来越高的抽象层。我们不再用汇编语言编程了,大多数人甚至不再用 C 语言,然后抽象层越来越高。所以对我来说,你告诉计算机做什么的方式会一直存在,只不过会在一个更高的抽象层上操作,这很好。它实现了民主化。软件操作者的数量会增加一个数量级。以前是 CS(计算机科学),以后可能变成 SO(软件操作者),但这并不意味着你不需要理解计算机科学——它是一种思维方式,一种心智模型。所以我强烈反对”编程已死”这种说法。
Lenny Rachitsky: 太好了,我很喜欢这个说法。SO 是软件操作者(software operator),这个缩写代表什么?
Aparna Chennapragada: 对,我刚编的,但没错。
Lenny Rachitsky: 好的,明白了。关于原型制作是当今构建产品的核心这个理念,你在微软内部有没有什么机制来落实这一点,让每个人都必须这样做?是文化上的引导,还是说必须”先给我看原型再说”?
Aparna Chennapragada: 同样,未来已经到来,只是分布不均——在微软内部也是如此。但确实有一股很强的文化势能和转变,大家都在说”嘿,让我们看实际演示、实际原型,用这种方式来沟通想法”。对我来说,这不是在任何情况下都可行的——显然有些东西非常底层,如果你要在 Excel 的深处改什么东西,你可能不需要。产品本身有足够的深度,你知道自己要做什么,不需要去做原型。但如果你是在思考新的东西、新产品、新功能,那就绝对需要。
产品管理的未来
Lenny Rachitsky: 好,我们来聊聊产品管理。这些 AI 编程工具一出来,就出现了一种恐惧——PM 已死,我们不需要 PM 了,我们自己就能做东西,这些人留着干嘛?但我发现实际情况恰恰相反——既然编码变得容易了,现在越来越多的问题是:我们应该做什么?为什么要做?这个方向对吗?这是正确的解决方案吗?然后推动产品的采用——这些恰恰是 PM 最擅长的事情。我觉得恰恰相反,PM 是最重要的角色。它也会变化,但先听听你的看法。你觉得产品管理的未来会怎样?你觉得它会消亡吗?会蓬勃发展吗?还是会发生变化?
Aparna Chennapragada: 是的。你看,如果你是一个主要做 TPS 报告、以流程为主的人——很多公司确实会把产品管理和流程、项目管理搞混——那你确实要面临一个问题:“这里的价值增量在哪里?“尤其是 AI 可以读写五万份会议记录、追踪事项、发邮件等等。但另一方面,我认为品味判断和编辑功能变得极其、极其重要。在一个创意供给、原型供给增加一个数量级的世界里,你必须思考这里的编辑功能是什么。
这意味着对产品人员的门槛更高了。但我观察到一个有趣的副作用——在我提供建议的初创公司、其他公司,以及公司内部——以前有更多的把门现象,就是”哦,我们应该问问产品负责人的看法”。编辑功能当然有它的作用,但现在你必须靠实力去赢得这个角色,而不是仅凭头衔就能获得。同时,那些聪明的工程师、用户研究员、设计师——他们脑子里一直有些很好的想法——现在口袋里有了一个专家,可以弥补他们不擅长的其他方面,从而把想法推进出来。我觉得这非常了不起。
Lenny Rachitsky: 我觉得这个”专家”很有意思。我现在和一个工程师合作一些东西,他用 ChatGPT 来和我更有效地沟通,比如”把这个提案改成一个能说服 Lenny 这是好主意的形式”。
Aparna Chennapragada: 顺便说一下,这其实也是我常用的一个场景,我叫它 WWXD——X 会怎么做?我会用它说,“嘿,Satya 会怎么看待我们正在推介的这组对话或想法?“这就是深度推理加上相关上下文的力量,对吧?你说的那个工程师掌握了关于你的上下文,所以这非常有趣。
Lenny Rachitsky: 要是每个人都像 Satya 那么有名、网上有那么多信息就好了。不过我想你可以导入他们所有的邮件,或者用现有的工具,从你和那个人的对话记录中去理解他们。
智力过剩与认知更新
Aparna Chennapragada: 对。我觉得这也回到了你之前说的,就是……这里有一种被压缩的弹簧的感觉。我看到各处都存在一种智力过剩(intelligence overhang)。我认为做产品开发的人几乎需要重新自我塑造,就像 Shopify 的 Tobi 所说的——反射性地使用 AI。但这并不容易,我也一直在想为什么。基本上,我搞了一个很土的 Chrome 扩展。真的,每次我打开新标签页,它就显示一句话:“你怎么能用 AI 来做你接下来要做的事?“非常土,但它确实能让你停下来想一想,“哦,我到底要做什么?”
但我发现它难用的原因——而且我跟那些在这个领域全职投入的人交流,他们也觉得难——就是更新先验认知非常困难。这些模型一年前还做不了某些事情。比如,图像生成满屏都是拼写错误,推理也做不到更深更聪明的回答。也做不了数据分析。所以我对它的印象停留在几个月前试用时的状态,而那个先验需要被更新。这很难做到,对吧?你必须做一件几乎违反直觉、逆着惯性的事,告诉自己,“不,不,忘掉你之前认为它能做什么不能做什么。“这个婴儿在一个月内就长成了十五岁的少年。
Lenny Rachitsky: 我觉得最后这点非常重要。我们这些年一直在尝试这些工具,到目前为止它一直不太好,然后突然就好了,而你几乎不知道,因为你基本上已经放弃了,但事情已经变了。
Aparna Chennapragada: 我觉得这其实……如果你是一个正在听这个播客的产品构建者,这对来说是一个很有意思的套利机会。如果你能逆着惯性,说”不,我不会被那些旧伤疤束缚”——几个月前这不好使——然后不断设定高期望、对今天的 AI 提出更多要求,我觉得你能解锁更多东西。
Lenny Rachitsky: 这样做里面有很多 alpha。
Aparna Chennapragada: 没错。
AI 使用习惯与 Chrome 扩展
Lenny Rachitsky: 我想回到你说的那个很土的插件,再多聊聊。就是一个让你在每个新标签页上放自定义消息的插件,你让它显示”你怎么能用 AI 来做这件事?”
Aparna Chennapragada: 对,就这么土。但有意思的是它确实管用。就在过去几周,我做了一个实验,想看看”我能多吃多少 AI 药片”,不管是工作还是个人生活,当我在做任何手工操作的时候,都问自己”我是不是应该让 AI 来做这件事?”
Lenny Rachitsky: 这太酷了。你知道这个 Chrome 扩展叫什么名字吗?
Aparna Chennapragada: 不知道。我自己做的。
Lenny Rachitsky: 你做了一个 Chrome 扩展。太酷了。你用 AI 做的吗?
Aparna Chennapragada: 当然。
Lenny Rachitsky: 哇。你用什么工具做的?我猜是某种微软的工具。
Aparna Chennapragada: 嗯,其实我就是,我日常就在 GitHub 和 GitHub Copilot 里工作,所以我就说”好吧,来做这个 Chrome 扩展吧”。
Lenny Rachitsky: 你打算公开发布吗?
Aparna Chennapragada: 不,我是说,这就是厉害的地方——我只花了大概十分钟就搞定了。
Lenny Rachitsky: 好,我们把链接放出来吧。让它开源。好。
Satya 与 Sundar 的领导风格
Lenny Rachitsky: 你提到了 Satya,我有一个相关问题。你是极少数同时和 Satya 以及 Google 的 Sundar 都有过密切合作的人。我来问你——他们的领导风格有什么不同?你有没有关于他们各自的有趣故事可以分享?
Aparna Chennapragada: 好。我确实觉得自己很幸运,能有机会近距离观察这两位这个时代最杰出的领导者。我想说,正如你对市值数万亿美元的科技公司 CEO 所预期的那样,他们在你想到的几乎所有维度上——智力、共情、领导力、产品、战略——都属于 99.99 百分位。当然,风格上的差异是存在的。我在 Google 时担任 Sundar 的技术顾问,帮他搭建了 CEO 办公室。当时的背景有很大不同,因为 Google 有更多的消费者导向业务。我发现 Sundar 非常擅长的一件事是,在处理复杂生态系统时保持冷静、审慎和深思熟虑。
想想手机生态,甚至搜索、出版商和广告主的生态——那是一个非常复杂的生态系统。他是这方面的大师。而 Satya,我觉得令人惊叹的是他对学习和调整心智模型的渴望,以及他能在不同的缩放级别上运作——宏观层面、战略层面,大局是什么;同时也能到微观层面,“嘿,为什么我们没有……”他在 Twitter 上看到了某个具体的洞察,而且你几乎可以确信他在发现这些早期信号方面领先于几乎所有其他人。所以用他们的话说,就是从消防水龙头里学习。
最反直觉的产品心得
Lenny Rachitsky: 能和这两位杰出的人共事,真是难得的机会。好,我们换个完全不同的方向。我想问你一个我越来越多地用来问别人的问题:关于打造产品,你学到的最反直觉的一课是什么?那种与常见的创业智慧、常见的产品构建认知相悖的教训。
Aparna Chennapragada: 我不确定这个认知是否像它应有的那样被广泛传播,它确实有点反直觉,但我反复学到的一点是:当你在做从零到一的新事物时,诱惑在于……就像那集《南方公园》。第一步,想问题。第二步,问……
Lenny Rachitsky: 内裤。应该是内裤,第一步。
Aparna Chennapragada: 内裤。没错,对吧?所以我确实觉得有一种诱惑,让人急于在解决问题之前就去追求规模化。所以我一直对团队说:先解决,再规模化。这句话的意思是,当你试图解决一个问题时,与你规模化一个已经达到产品市场契合、或者至少大致方向对的产品,你的姿态和模式是完全不同的。
举几个例子。在解决阶段,你会发现方向有很大的摇摆。你必须非常坦然地接受,第一天你还在想”嘿,做一个植物识别工具”,到了第十五天你发现”哦,实际上这个技术用来翻译外语非常好”。顺便说一下,这不是假设,这基本上就是当时我们在 Google Lens 里经历的过程——“好,交叉点在哪里”等等。
从外面看,这像是混乱,但实际上……你不仅应该容忍它,我认为你应该对这种混乱有胃口,因为最糟糕的事就是过早地锁定在一个局部小山丘上。然后你开始往上爬,创业公司、整个产品领域、大公司都会犯这种错误,三年后你才发现”天哪,我怎么从这个山丘上下来”。所以我觉得这是一大反直觉的地方。你要想清楚自己处于什么模式——是解决模式还是规模化模式。一个体现就是确保你对这种混乱感到舒适。
指标的危险
我学到的另一课是指标的危险。如果你在 Google Search 或 Office 产品团队工作过,你对这个产品的指标体系会有一套非常精细的感知——输入指标、输出指标,全套都有。但当你在做从零到一的事情时,如果你过早确定一个指标,首先是虚假的精确度,对吧?我的意思是,CTR。当你只有一千个用户时,这个数字毫无意义。留存率可能也没有意义。所以真的要非常警惕那些我称之为”大人式的大指标”。你应该更多地看定性的东西——点击的声音是什么样的,你的……另一个维度是使用场景,比如”设个计时器”和”播放音乐”。如果你看 Alexa、Siri 和 Google Assistant 这些产品,它们都有一个很有前景的宽泛界面——你什么都能说,但我觉得真正好用的就那一两件事。设计时器、播放音乐、玩问答。你得先把这几件事做到极致,然后才能说”哦,什么都能做”——那不是一个好的配方。
Lenny Rachitsky: 还真是。我的 Google Home 就只干这些,太基础了。我没有用它玩过问答,也许我得试试。
Aparna Chennapragada: 得试试。是的。
从零到一的三个拐点框架
Lenny Rachitsky: 你还谈过一个与之相关的东西,就是如何从零到一去做一件事——一个小框架,帮你判断现在是不是做这个想法的合适时机。你是怎么想的?
Aparna Chennapragada: 对。说到解决模式,这又回到了我一直说的”活在未来一年”的思路。我倾向于从零到一的解决模式产品,完全是在思考新品类的产品。我发现——应该说是在吃亏之后才学到的——如果你想做出一个真正好的产品,你至少需要以下三个拐点因素中的两个。第一,技术层面是否有一个阶梯函数式的跃迁?这个比较显而易见。深度学习对 Google Lens 就是一个拐点。再早一些,语音识别的突破是对话式搜索的拐点。对 Robinhood 来说,代际变迁很明确——手机成为主要工具,你可以真正拥有一个用于理财的移动应用。所以要寻找这种拐点:技术拐点是什么?而现在,当然,LLM 和推理模型就是那个阶梯函数。但仅仅这一点是不够的。
我要说第二个应该关注的因素是:消费者行为的变化是什么?举个例子,当我们开始做 Google Lens 的时候,我们观察到的是:人们拍照主要是为了分享——自拍、夕阳之类的。但后来,当存储变得基本免费、人人都随时随地带着手机时,你开始拍一切东西。你拍了大量的照片,或者说你把相机当成了真实世界的键盘。那么你怎么判断”这个消费者行为变化足够大”,随着照片量增长了一个数量级,人们自然想从照片中获取更多信息,你就可以把 AI 应用上去。
第三个拐点,特别是在企业端但也在消费者端,是商业模式的变革。是否存在一个天然的商业模式拐点?任何伟大的产品,从搜索开始——二价拍卖和 CPC 计费模式;SaaS 也是如此——你可以用不同的方式对企业产品收费或变现。而在 AI 时代,变现方式完全是另一回事——我们才刚刚触及表面:按席位收费、按使用量即用即付,以及基于结果的变现——“嘿,你帮我解决了这个问题,我再付你费用”。这三个拐点都具备当然最好,但至少要做到三选二,才能做出一个好产品。
Lenny Rachitsky: 所以这本质上就是……当投资者看创业公司时,他们总在问:为什么是现在?为什么现在是做这件事的时机?所以你这里的建议是,有三个维度来回答这个问题,其中两个应该成立:技术上要有变革,某种新技术使得现在可以做这件事;消费者行为上要有转变;还有就是你可能发明了一种新的商业模式——某种变现方式,让你相对于当前在做这件事的人拥有优势。
Aparna Chennapragada: 没错,完全正确。
Lenny Rachitsky: 太好了。你确实提到了 Robinhood,我想就是在那个例子里。那也是一个很好的例子,关于手机……
Aparna Chennapragada: 对,说到商业模式——零佣金。再一次,所有这些因素的组合才是解锁关键。你不能仅仅说”哦,我们会有一个更直观的界面,然后指望人们会迁移过来”。
热座环节:AI 编程工具的竞争格局
Lenny Rachitsky: 好的,说到从零到一的产品,我想带大家进入这个播客偶尔会有的一个环节,我称之为”热座角”。我有一个问题想问你,这个问题一直在我的脑海中,而且最近几期播客也出现过。现在有一些公司,比如 Cursor、VZero、Lovable、Bolt、Replit,它们是有史以来增长最快的公司。我刚看到 Cursor 在两年内达到了 3 亿美元的 ARR。有意思的是,你们其实在这个领域——AI 编程工具这个领域——占据了非常好的位置。你们有 Copilot,世界上第一个做这个的工具,领先所有人。你们构建了 VS Code,所有这些公司都是基于它 fork 出来构建产品的。你们有令人难以置信的 AI 基础设施,令人难以置信的 AI 人才。所以这本应该是你们的市场。发生了什么?Aparna,发生了什么?
Aparna Chennapragada: 这个问题的框架很有意思……我自己是 GitHub Copilot 的死忠用户,我想说,“你看,如果你拆解一下……”我觉得这件事的美妙之处在于,代码生成已经成为 LLM 解锁的一种令人惊叹的工具。所以代码生成确实带来了非常好的兴奋和行动力,它打开了所有这些可能性……我们之前谈到过原型设计的概念——从想法到标记、从想法到可点击的原型,几分钟内就能完成。当然,这些都是我们应该期待代码生成能够实现的事情。
我对我们的定位以及我们在 GitHub 上所做的事情的看法是……它是一个系统,而不仅仅是一个产品或一组功能。
GitHub 是面向那些把代码仓库放在那里的人的,你当然有自动补全方面的辅助,也可以聊天,但现在我们有了 agent 模式。这是我们看到的最快的迭代循环之一,获得了非常强烈的正面反馈。所以在某种意义上,当你拥有一个系统时,你在构建和设计它时所追求的不仅仅是一个能增长的单一产品,而是:你的代码仓库是什么?你的上下文是什么?从你的专业能力出发能生长出什么样的功能集?如果你是一个非常资深的程序员,你需要这个产品能够为你的水平提供匹配的辅助。如果你是一个广泛型开发者,你也应该能够使用它。我认为这就是 GitHub 的定位所在,而且说实话增长得非常好。
Lenny Rachitsky: 这太有意思了。所以核心是——无论大家使用什么工具,最终都会回到 GitHub,这就是……
Aparna Chennapragada: 对,同样的道理,代码生成作为一种工具会解锁更多的产品。它们并不都是竞争对手,它们并不都在做同样的事情。我认为当你在……说到底,你构建的代码是要让公司运转的,你需要一个系统。你需要一整套瑞士军刀式的工具箱,不仅仅是自动补全,不仅仅是聊天,也不仅仅是一个需要你手把手引导的软件代理。你需要所有这些东西协同工作,而这正是 GitHub 产品在追求的方向。
Lenny Rachitsky: 条条大路通 GitHub。反过来看这个问题,大概有五千家创业公司试图颠覆 Excel,但你们一直在赢,所以那方面肯定有什么东西运作得很好。
Excel 的持久生命力
Aparna Chennapragada: 你说这个太有意思了。我刚来微软的时候,我是个 Excel 粉丝,所以我跟 Excel 的一个元老级产品人员聊过。我说,“嘿,这个产品到底是怎么回事?“他说了几件让我印象深刻的事情。第一,我说,“Excel 证明了非程序员也需要编程。“编程是非常强大的能力,而 Excel 这个工具赋予了所有非程序员一种非常强大的编程能力,我觉得这一点非常发人深省。
第二件我觉得特别酷的事情——我不知道你知不知道,反正我两年前才知道——有这些令人惊叹的 Excel 锦标赛,比如世界 Excel 锦标赛,你能看到那些选手做出简直像魔法一样的操作。对我来说,这里的洞察是:有些工具学习曲线更陡。也许一开始在学习上会有摩擦,但用起来非常好。所以这是一个非常好的案例——初始的学习曲线,一次性付出的学习成本可能有点棘手,但那是因为这个工具有如此强大的能力和深度。
Lenny Rachitsky: 这太有意思了。我从没想过把 Excel 当作一种编程语言,但这说得通。而且我觉得一旦你习惯了,这就是事情运转的方式,你就被锁在那里了,其他所有东西基本上都得复制那个模式,而要做到同样好是很难的。
Aparna Chennapragada: 是的。我觉得还有这种深度——以及团队倾注的注意力——这是一种几十年来的复利效应,来自那些日复一日依赖它的人所提供的深层信号。
职业生涯的转折点
Lenny Rachitsky: 好的,作为我们对话的收尾,我想问一个关于你职业生涯的问题。我发现大多数人的职业生涯中都有一个改变轨迹的关键时刻,可能是一位上司,可能是一个参与过的项目,也可能是得到的一份工作。你认为你职业生涯中最重要的转折点是什么,最终引导你成为了微软的首席产品官?
Aparna Chennapragada: 确实有一个时刻对我来说是一个转折点。当时我在 Google Search,我在做一个我觉得应该能行的想法,但结果并没有。我说,“嘿,手机正在成为一种趋势,个性化一定很重要。“于是我大概花了一年左右的时间绞尽脑汁想让个性化发挥作用。结果发现,当你在 Google Search 中输入一个查询时,个性化其实没那么重要。于是我们解散了团队,但后来我开始做一个叫 Google Now 的产品,这是那个思路的一个转折,它的理念是,“嘿,其实在手机上,我们应该能够推送内容,问题不在于带个性化的搜索。“比如,如果你有一趟航班即将起飞,我们应该能够连接各个信息点,告诉你”考虑到交通状况和你要去的地方,你现在应该出发了”;或者如果你对冷面笑匠风格的脱口秀非常感兴趣,你应该去看看 Mitch Hedberg。
这些就是那些真正体现智能手机应该更智能的时刻。我带领那个产品走过了最初的从零到一阶段,那是一个转折点。它让我认识到两件事。第一,我真的很喜欢看到拐角处的东西,看到事情的发展方向,并构建能够应对挑战的产品,这比扩展和维护产品更让我着迷。第二,这很残酷,但过早和做错是一回事。那是在 LLM 之前、深度学习之前,很多关于下一个 token 预测器等真正了不起的想法,我们都想过,但没有足够的算力去实现……界面很棒,但智能还不够。第三件让我铭记的事情是,我有机会和一些非常聪明的人共事……现在人们谈论人才密度,我认为那些真正聪明的人后来都做出了了不起的事业,所以这让我体会到一小群人能做到什么。
Lenny Rachitsky: 这个故事真的很棒,虽然结局并不如意——Google Now 后来确实消失了。顺便说一句,我对那个产品印象深刻,真的非常酷,我还记得当时看到它时的那种惊喜和愉悦。另外,我的播客里有一个环节叫”失败角”,邀请嘉宾分享一段失败经历以及从中得到的收获。你这个故事恰好把这两个主题结合在了一起。
Aparna Chennapragada: 是的,我不会假装不痛心。当你投入其中时确实很痛苦,因为你看到了愿景和现实之间的差距,有时候限制是硬性的。像这个例子,需要五年甚至十年才能真正解锁足够的智能;但有时候,产品离优秀只差一两个关键的卡扣,而搞清楚自己身处哪种情况,本身就是难题的一部分。
Lenny Rachitsky: 从开始做到最终放弃、承认行不通,这个过程持续了多久?
Aparna Chennapragada: 在这个案例中,一件好事是它奠定了……它是 Google Assistant 的基础之一。当然,随着 LLM 在 Gemini 上实现阶跃式突破,这个思路现在算是走通了。我觉得这其实是普遍规律:有时候你需要找出那些真正有效的、不变的东西,让它们延续到产品的下一个版本;而另外一些时候,你只能从头再来。
Lenny Rachitsky: Google Now 是不是 agent 出现之前的第一个 agent?感觉就是这个定位。
Aparna Chennapragada: 当时的想法确实如此。但让我觉得有趣的是界面的问题——我们当时遇到了相反的困境。不管是哪个语音助手,界面都过度超前了,而智能还没跟上。而今天,我觉得问题反过来了——这些系统具备了惊人的智能,但我们现有的界面基本上还是 AOL 拨号上网时代的聊天机器人水平。
人与 agent 协作的未来
Lenny Rachitsky: 我们已经聊了很多话题了。你还有什么想聊的,或者想在结束前留给听众的,也许是一句最后的智慧箴言?然后我们就要进入非常精彩的快问快答环节了。
Aparna Chennapragada: 我想说的是,我目前非常兴奋的一件事,就是探索人与 agent 之间如何协作。我认为有大量出色的产品和体验值得重新构想。这是我的另一个”罗马帝国”——就是我经常忍不住去想的事情——我们如何真正打造一个人与 agent 共存的工作空间,如何让产出远远超过我们任何一个人或几个人能独立完成的水准?
Lenny Rachitsky: 我需要听你展开讲讲。你想象中人与 agent 的协作空间是什么样的?是 Microsoft Teams 那样的东西,还是说一个有小型机器人的实体空间?
Aparna Chennapragada: 哦,我也想过实体空间的可能,但我现在更多思考的是……目前所有这些体验都还是单人模式,我确实认为存在一个机会去思考——还是那句话,我活在未来一年——我们如何真正实现人与人之间的协作,同时也与 agent 协作,真正搞清楚哪些任务可以委派、哪些需要检查,信息如何通过 agent 在人之间流转,等等。
Lenny Rachitsky: 好的,我很期待看到你们正在打造的东西。接下来,我们到了非常精彩的快问快答环节。准备好了吗?
Aparna Chennapragada: 来吧。
快问快答
Lenny Rachitsky: 来吧。第一个问题:你有哪两三本书是经常推荐给别人的?
Aparna Chennapragada: 哦,我有近因偏差,不过我最近在读一本叫《A Brief History of Intelligence》(智能简史)的书,非常精彩,我画了很多下划线。它的核心前提是考察人类智能的进化、大脑的发育,并将其与我们今天在 AI 中看到的现象联系起来。
Lenny Rachitsky: 你有最近特别喜欢的电影或剧集吗?
Aparna Chennapragada: 《Hacks》。我最近一直在看。它讲的是一位非常优秀的单口喜剧演员的故事,她的背景是在七八十年代长大的,试图在一个对女性并不友好的行业中闯出一片天地。非常有趣,也很古灵精怪。
Lenny Rachitsky: 你最近有没有发现一个特别喜欢的产品?可以是应用,也可以是实物。
Aparna Chennapragada: 我确实用了很多微软的产品,GitHub Copilot 是其中之一,但我想特别提一下一个叫 Granola 的应用。我觉得它非常好用。前几天试了一下就觉得”哦,这真的很有用”——它能在不造成打扰的情况下捕捉你的想法、笔记,并把它们结构化整理……它让我感觉像是几项技术终于达到了可信度的临界点,就像我们刚才聊到的,实时转录技术已经非常好了,语音识别很出色,再加上 LLM 的魔法在上面做结构化和上下文理解。
Lenny Rachitsky: 我是 Granola 的超级粉丝。我顺便插个广告:如果你成为我 newsletter 的年度订阅者,你的整个公司可以免费获得一年的 Granola 使用权。
Aparna Chennapragada: 这个我还真不知道。
Lenny Rachitsky: 就是这个,可以去 lennysnewsletter.com,点击 “bundle” 那个词就能看到具体操作方式。
Aparna Chennapragada: 很酷。
Lenny Rachitsky: 很酷。还有两个问题。你有没有一个经常回来的人生座右铭,在工作或生活中遇到问题时会想到的,也许你也会分享给别人?
Aparna Chennapragada: 有一个。事实上,这是我大概过去二十年来一直用的邮件签名。上面写的是:“预测未来的最好方式就是去创造它。“这是 Alan Kay 的一句名言。我觉得它在两方面很有用。第一,没有人真正知道未来会怎样。当你听到那些人说”一切都会是这个样子、发展顺序就是这样”的时候,没有什么能替代亲身去构建的体验。第二,如果你认为某样东西应该存在,那就去把它造出来。
Lenny Rachitsky: 我很喜欢这句话。最后一个问题。我们之前聊到了单口喜剧。有没有一位你觉得还不够出名的单口喜剧演员,值得大家去看看的?
Aparna Chennapragada: 有几位。一位是印裔英国单口喜剧演员,叫 Sindhu Vee,非常聪明,擅长讲妈妈式喜剧。另一位……这位绝对不算”不够出名”了,但我实在太喜欢他的风格了——Nate Bargatze。他就是那么好。
Lenny Rachitsky: Aparna,今天聊得太棒了。最后两个问题:大家在网上哪里可以找到你,如果想要联系你或就今天聊到的内容跟进?以及,听众怎样能帮到你?
Aparna Chennapragada: 你可以在 LinkedIn 和 Twitter 上找到我,账号是 Aparna CD。现在我更多是在 LinkedIn 上发内容,欢迎在那里跟我分享想法、评论和交流。我想说一件特别有意思的事——如果今天聊到的任何内容引发了讨论,尤其是围绕”一个小团队借助大量 AI 工具能做出什么”,或者大家真正期待、觉得应该被做出来的新产品,请随时联系我。
Lenny Rachitsky: 太棒了。Aparna,非常感谢你来参加这次访谈。
Aparna Chennapragada: 谢谢你。
Lenny Rachitsky: 大家再见。非常感谢收听。如果你觉得这期节目有价值,可以在 Apple Podcasts、Spotify 或你喜欢的播客应用上订阅。也请考虑给我们评分或留下评价,这真的能帮助更多听众发现这个播客。你可以在 lennyspodcast.com 找到所有往期节目或了解更多关于这个节目的信息。下期再见。
术语表
| 原文 | 中文 |
|---|---|
| agent mode | agent 模式 |
| Alan Kay | Alan Kay(计算机科学家) |
| alpha | alpha(套利机会/超额收益) |
| ARR (Annual Recurring Revenue) | ARR(年度经常性收入) |
| ChatGPT | ChatGPT |
| Copilot | Copilot |
| CPO (Chief Product Officer) | 首席产品官 |
| cron job | cron 定时任务 |
| DeepSeek | DeepSeek |
| fork | fork(基于已有代码库创建分支) |
| full stack builders | 全栈构建者 |
| Granola | Granola(AI 会议笔记应用) |
| GUI (Graphical User Interface) | GUI(图形用户界面) |
| intelligence overhang | 智力过剩 |
| Jean-Claude Van Damme | 尚格·云顿 |
| Kevin Weil | Kevin Weil |
| Mitch Hedberg | Mitch Hedberg(美国冷面笑匠风格脱口秀演员) |
| Nate Bargatze | Nate Bargatze(美国单口喜剧演员) |
| next token predictor | 下一个 token 预测器 |
| NLX (Natural Language Interface) | 自然语言界面 |
| PMF (Product-Market Fit) | 产品市场契合 |
| PRD (Product Requirements Document) | PRD(产品需求文档) |
| Reid Hoffman | 里德·霍夫曼 |
| Satya Nadella | Satya |
| scar tissue | 旧伤疤(过往负面经验形成的心理障碍) |
| Sindhu Vee | Sindhu Vee(印裔英国单口喜剧演员) |
| SNL (Saturday Night Live) | 《周六夜现场》 |
| software operator | 软件操作者 |
| Sundar Pichai | Sundar |
| talent density | 人才密度 |
| Tobi (Tobi Lütke) | Tobi |
| TPS report | TPS 报告 |
| UX (User Experience) | UX(用户体验) |
| WWXD (What would X do) | WWXD(X 会怎么做) |
此文档由 AI 分片翻译(translate_long_document)