Scale AI CEO 谈 Meta 的 140 亿美元交易、将 Uber Eats 做到 800 亿美元,以及前沿实验室接下来在构建什么
Scale AI CEO on Metas 80B, & what frontier labs are building next
Has AI Delivered on Its Promise
Lenny Rachitsky: There’s been a lot of talk these days about AI not delivering on the promise that we hear, especially at enterprises.
Jason Droege: These things take 6 to 12 months to get them truly robust enough where an important process can be automated. Like with any of these major tech revolutions, headlines tell one story and then on the ground, laying broadband means you need to dig up every single road in America to lay it. Someone’s got to dig up the road or someone’s got to run the undersea cable.
Scale AI and Data Labeling
Lenny Rachitsky: Is there anything you think people don’t truly grasp or understand about where AI models are going to be in the next two, three years?
Jason Droege: The general trend right now is going from models knowing things to models doing things. The next question becomes, what can it do for me? How does the agent make decisions for you?
The Standard for Startups
Lenny Rachitsky: Let’s talk about Scale and this whole world of AI that you’re in, you essentially pioneered data labeling, trading data, creating evals for labs.
Jason Droege: 18 months ago, you would get a short story and it would say, “Is this short story better than this short story?” And now you’re at a point where one task is building an entire website by one of the world’s best web developers, or it is explaining some very nuanced topic on cancer to a model. These tasks now take hours of time and they require PhDs and professionals.
About Jason Droege
Lenny Rachitsky: I’ve talked to a bunch of people that have worked with you over the years, and I heard a lot about just how high of a bar you set for new businesses.
Jason Droege: From an entrepreneurship standpoint, it truly is about what insight do I have? Why am I so lucky to have this insight? Why in a world of a million entrepreneurs who are thinking, who are smart, who are trying everything, why am I in the position where I likely have an insight that others do not?
Starting with Scour
Lenny Rachitsky: Today, my guest is Jason Droege. Jason is the new CEO of Scale AI. This is the first interview that he’s done since taking over for Alex Wang after the Meta deal. Alex now leads the super intelligence team at Meta. Prior to Scale, Jason co-founded a company with Travis Kalanick. Before, he started Uber, worked at a couple startups. Most famously, Jason launched and led Uber Eats, which went from an idea that he and his team had to what is now a multi-billion dollar run rate business and one that basically saved Uber during the pandemic when nobody was taking rides. This interview is following a theme that I’ve been following through a bunch of interviews, which is the evolution of how AI models actually gets smarter. Along with scaling, compute and improving the actual model code, much of the improvements we’re seeing in ChatGPT and Claude and every frontier AI model is these labs hiring experts to filling gaps in their knowledge and correcting their understanding of how things work, and basically showing them what good looks like in every domain that consumers are using models.
Scale was the pioneer in this space. They created the category, and in our conversation we talk about what is happening at Scale and just how this deal with Meta worked, what experts like doctors and software engineers are specifically doing to help models get smarter, how the whole market of data labeling and evals and data training has changed from when Scale entered the market to today, and also just how long will we need humans to keep helping AI get smarter. We also get into where Jason sees models going in the next few years because they have such a unique glimpse into the future. We also talk about a ton of really unique and really important product lessons from the course of Jason’s career, including a bunch of advice on how to start a new business, both startups and within existing companies, and also a bunch of advice on hiring and leadership and so much more.
A huge thank you to Allen Penn and Stephen Chau for suggesting topics for this conversation. If you enjoy this podcast, don’t forget to subscribe and follow it in your favorite podcasting app or YouTube. It helps tremendously. And if you become an annual subscriber of my newsletter, you get a year free of 15 incredible products including Lovable, Replet, Bolt, n8n, Linear, Superhuman, Descript, Wispr Flow, Gamma, Perflexity, Warp, Granola, Magic Patterns, Raycast, ChatPRD, and Mobbin. Head on over to lennysnewsletter.com and click product pass. With that, I bring you Jason Droege.
That’s why Figma built Figma Make. With just a few prompts, you can make any idea or design into a fully functional prototype or app that anyone can iterate on and validate with customers. Figma Make is a different kind of vibe coding tool. Because it’s all in Figma, you can use your team’s existing design building blocks, making it easy to create outputs that look good and feel real and are connected to how your team builds. Stop spending so much time telling people about your product vision and instead show it to them. Make code-backed prototypes and apps fast with Figma Make. Check it out at figma.com/lenny. Jason, thank you so much for being here and welcome to the podcast.
Jason Droege: Yeah, thanks for having me. Excited to be here.
Scale AI and the AI Data Industry
Lenny Rachitsky: As I was researching your background and prepping for this podcast, I learned a really interesting fun fact about you that I don’t think a lot of people know. So Travis Kalanick, he had a startup before Uber. It was called Scour. It was a peer-to-peer file sharing app, and then I think got shut down. You were his co-founder. This was the early part of your career. I’m guessing there are hours of stories we could talk about during this experience. So let me just ask you this one question. What’s just a lesson that has stuck with you from that experience that you’ve taken with you to future places you’ve worked and built product at?
The Evolution of Data Labeling
Jason Droege: I mean, there’s so many lessons. I like to pick one. I think that the main lesson is that in business and in startups, everything’s negotiable. I think that’s the main thing. Because we were 19 at the time, 19, 20 at the time, we built this search engine in a dorm room and we were running it out of the dorm room and our first URL was scour.cs.ucla.edu. These things were not necessarily in fractions at the time, but we were just being practical. It was basically a project that we had started, and so we built the search engine and people started using it and we thought we would get in trouble, but it turned out the computer science department was excited about it even though we had basically parked a domain on their servers and we were using our own computers in the dorms to serve up this website and product.
And then, when we got into financing, the financing process was fascinating, and this is where the everything is negotiable lesson came from, which is, it was Ron Burkle and Mike Ovitz, who are the initial investors in the business. We were in LA, so we were at UCLA, so we were not quite wired into the entire Sand Hill Road scene. And as we were doing the deal, the terms kept changing on us. We thought you went and raised money and it was like, “Oh, we’ll get a few million dollars at a $5 million valuation.” This is back when that was actually a series A valuation. And then over the course of the deal, it was like, “We’re doing the deal. We’re not doing the deal. Oh, you should give us 50% of the company. Oh, you should give us 75% of the company. Oh, if you want to sign the document today, this person’s going to show up for breakfast and if you don’t sign today and give us 80% of the company, the person’s not going to show up.”
It was just completely wild, the things that we saw from day one of what can happen in business, and we thought there was a way to do things, and at a very young age we realized there is no way to do things. There is just the way that you can negotiate your way through the world, which I actually think influenced Travis heavily and then me later heavily at Uber in terms of if you can imagine it and it makes sense and you can align incentives, then it can happen. But there is no way. And to learn that at 19 or 20 years old I think was highly imprinting.
Lenny Rachitsky: That is an amazing lesson. What happened to Scour? It got shut down, I think. What happened there?
Finding and Keeping Experts
Jason Droege: Well, yeah, so basically what Scour was was it was a multimedia search engine and then peer-to-peer file sharing network. But what it was used for was finding free content. And at the time, the laws were on this were pretty ambiguous because we weren’t, mix tapes were legal, but this was like a hyperversion of that. But we were eventually sued for a quarter of a trillion dollars. So I guess if you’re going to experience something that’s potentially as life devastating as that, doing it when you’re, I think we were 21 or 22 at the time is the time to do it, but it was just this very cold splash of water about how the real world really works, because the MBAA and the RAA were the ones who sued us, the entertainment industry sued us or the associations that represent the entertainment industry, and then they settled it for $1 million.
So we’re like, “Wait, you wanted a quarter of a trillion dollars and then you settle for $1 million.” And of course they were just trying to drive us in a bankruptcy, drive us out of the market, and these are established companies. So we’re like, “If these guys don’t have a playbook to follow, they just make up numbers, then wow, how should we navigate the rest of our lives?”
Reinforcement Learning and RL Environments
Lenny Rachitsky: Let’s talk about Scale and this whole world of AI that you’re in. This is the first interview that you’re doing since taking over CEO at Scale. I’m honored to have you here to talk through this stuff. This is also the first interview you’re doing since the whole Meta deal, which is very complicated, confused a lot of people. So I’m just curious to hear the current state of Scale, what people should know. For example, what’s your relationship with Meta? What’s your relationship with Alex? What is the current state of Scale?
Data Generalizability
Jason Droege: Yeah, so Scale is a fully independent company. The transaction was Meta invested a little bit over $14 billion to get 49% of the company, non-voting stock, didn’t take a new board seat. Alex fills the board seat. So the board is the same, the governance is largely the same. There’s no preferential access to anything that Meta has. There’s no preferential relationship. I mean, we’ve had a longstanding relationship with Meta on the data side of the business for a long time and even on some business development related things to maybe working on things in government together, et cetera. And so, those might get bigger just as we’re closer now, but there’s nothing that prevents us from doing things with other parties and they have no access to anything that they wouldn’t have had otherwise. All the privacy still in place, all the data security still in place that was there before.
And in fact, only about 15 people went over in the transaction. So Scale has about 1,100 employees or so now, and we have two major businesses. Each of those businesses, each of them has hundreds of millions of revenue. So we have two unicorns inside the company today that sustains. The business has grown every month since the deal happened, which I’ve read, the reporting is not consistently reported. We haven’t talked about it, so this is part of getting the word out and we’re excited to continue to build, deliver data, and do what we did before.
Lenny Rachitsky: So the company today, independent, its own company. Alex, just to be clear, he works at Meta now. He’s no longer at Scale.
Enterprise Labeling Needs
Jason Droege: Yeah, that’s right. Excuse me, I should have talked about that more.
The Future of Human Labelers
Lenny Rachitsky: I think that’s really interesting. So basically, it was an investment. Some people left to join Meta, the company continues, you’re running the ship. Let’s talk about this whole space that you guys essentially pioneered, I don’t know best way to call it, data labeling, training data, creating evals for labs. You guys were at this before anyone even knew this was a thing. I know even Scale pivoted into this market from other things. I think there was a bunch of stuff they tried with self-driving cars and all these things, and then it’s like, “Oh shit, AI labs need this data.”
One of the main stories I’ve been hearing is, and I’ve had a bunch of CEOs from this space on the podcast, is that there’s been this big shift from the way, from what Scale had pioneered and had been doing for a long time, which is generalists, low-cost labor training. From that to now, labs mostly need experts, lawyers, doctors, engineers doing training, writing evals, things like that. I’m curious just what you’re seeing, how that’s impacting you guys, where you think things are heading, what people should know about this whole market of data training data.
Jason Droege: Yeah, totally. I think the current positioning out there from competitors is just bogus. So I’ll start with that and then maybe talk a little bit about, I’ll explain what I mean by that in a second. But I think it’s important to just give 30 seconds on what the history of Scale is and what’s the thread going back to 2016. So Alex had this insight in very early days that the important thing to models was data. And I think he was 19 or 20 years old at the time as well. And so, he’s like, “Okay, well what business would I create around this?” And the business that he created around it was, okay, let’s do labeling for autonomous vehicles, because if you label the data that they have, the cars do better. And then, that wave turned into the computer vision wave, which we have a relationship with the Department of Defense where we do labeling for them, and that was in 2020.
And then, you move forward and the models have gotten better over this period of time. And so, as models get better, they need different types of data. So we’ve constantly been adapting to the type of data that models need to be successful. And so, then the gen AI wave hit, and this went through the moon or to the moon. And so, as part of that, that industry is changing constantly too. So it is correct that when the models came out two or three years ago, I mean we remember using them, they would hallucinate all the time, they would get basic answers wrong, they didn’t know which poem was better, this poem or that poem. And that was the state of labeling a couple years ago. And things have changed quickly and we’ve changed with it. And now the state for everyone, and we’ve been at the forefront of all of this, is expert data labeling, more sophisticated tasks.
So to give you a sense of what the task was 18 months ago, I’ve been here about 13 months. So I was interviewing and I remember seeing it. You would get a short story and it would say, “Is this short story better than this short story?” And then you would edit it and be like, “Yeah, it would be better if it was this,” and you would give some preference ranking to it. It was pretty basic 18 months ago, and you had the rise of some experts, but the models were so far behind that they needed just even the basic stuff they needed. And now, you’re at a point where a task is, one task is building an entire website by one of the world’s best web developers, or it is explaining some very nuanced topic on cancer to a model. And these tasks now take hours of time and they require PhDs and professionals.
So to give you a stat to back this up, 80% of the people that we have on our expert network have a bachelor’s degree or greater, which is very contrary to some of the positioning that’s out there and some of the understanding of this industry. About 15% have a PhD that’s greater, and we have PhDs on the network earning significant amounts of money doing labeling, contributing their expertise to these models. So we’ve been doing expert data labeling ever since the models need it. I mean, this game is keeping in touch with the researchers, knowing what they need, coming up with ideas internally. In some ways, we drove this because we were seeing that the models were not sufficient in more expert ways. And so, we would go to the model builders and say, “Hey, we noticed that this is a problem. If you would like to fix it, this cadre of experts can do that for you.” So the counter positioning is out there, but I think that’s just what competitors say sometimes. It has nothing to do with reality.
The Role of Evals
Lenny Rachitsky: Okay. That was extremely interesting. So what I’m hearing is yes, there has been a big shift to labs need more expert folks involved in training, labeling, writing evals. You guys are very aware of that and have evolved with that. One of the, I don’t know, allegations I guess in the market is that it’s hard to find these experts. So all these companies have their proprietary network of experts and how they find them. Is there anything you could share about just how you guys go about that because that feels like the hardest part is finding these experts and keeping them from other companies?
Jason Droege: They are hard to find. You have to have many, many tactics. So we get, as you would expect, there’s not one way you do it. The largest way is that they refer each other because when you are enjoying what you’re doing and you are using your expertise to contribute to AI, which is pretty cool. If you’re a PhD on this pretty specific topic and you’re using a model and you’re frustrated that, oh, it doesn’t interact with me in the way that I want, this is a paid way to have an outlet for that and to make hundreds or thousands of dollars doing that. And so, a lot of times they refer each other.
We also have campus programs where we will literally go onto the campus and talk to the professors, talk to the students, ask about who would like to do this type of work. And then, of course, there’s the more traditional scaled ways of LinkedIn and places like that. But the best ones come from these grassroots and referral networks. And the only way you get that is providing a great experience to these people, because these people, they’re doing it partly for money, but they’re also doing it because they think that their contribution to the AI models is important and interesting, and in many times it solves a problem for them.
AI Model Advances and Infrastructure
Lenny Rachitsky: So something that I’ve been seeing on Twitter just this week as I was preparing for this is there’s the information headline. This came out and this mirrored something that Brendan from Workhorse said that over time the entire economy is going to move towards just reinforcement learning and everyone’s just training AI is basically the jobs that will be left. Thoughts on that? Is that where you think things are going? Is there another perspective?
Jason Droege: Reinforcement learning is very important, and I think this is a broader comment about the move to environments. There’s these things called RL environments that effectively are sandboxes for AI agents to play in to accomplish a goal so that they can learn how to accomplish that goal. We’ve been doing this for over a year. So for example, you have a Salesforce instance. How does an AI agent navigate that instance? That instance has data that it needs to recognize, it has configurations. Salesforce is a highly configurable product. It has configurations, it needs to understand how to navigate. You’re asking the agent to do a business process that needs very high reliability, and then the agent needs to know, “Hey, if I can’t accomplish what I’m going to accomplish, or I think if there’s a low accuracy of what I’m about to accomplish, how do I pop it up to a human being for feedback so I can get guidance?”
All of those things need to be trained and there’s no alchemy to it. You just have to put the AI agent in an environment that represents what a human being would be doing. And you can imagine the number of environments in the world and the number of goals within each environment is enormous. So the question is, and the research that we have done over the past year to try to be a good partner to our model builders, our model builder customers, is how generalizable is each individual task or each individual environment. So if you imagine the world of environments of software systems, configurations, data types, sizes, user counts, complexities, it’s like the permutations are endless. So what you need is you need a strategy that allows a lab to collect data that is generalizable enough across a broad spectrum of use cases so that they don’t have to collect 45 trillion combinations of what should the agent do in this particular situation.
So sometimes the work and the data is highly generalizable, and by generalizable I mean you have it accomplished in a simple way. The task might be find the meeting on my calendar for my interview with Lenny, and the agent goes and it looks through all my calendar and then it pops it out, very simple example. That needs to be generalizable to any calendar search potentially or potentially any calendar action. And the more generalizable it is, the more valuable the data is. So our job is to provide the most valuable data to model builders that accomplishes the goal of making agents as useful as possible for their end users.
Outlook for the Next Few Years
Lenny Rachitsky: I love that you’ve been sharing these examples of what this stuff is specifically that these people are doing, the data you’re providing to labs. So just to mirror back a few of the examples you’ve shared, one is an engineer building a website, sharing the code essentially with the model. And here’s how I would do it. And in that example, is it just like here’s the code or is it a recording of them building it? What is the data?
Jason Droege: It could be both. So in some cases, it’s just the website and here’s an example, and then they design it. In some cases, it needs to be annotated in such a way that’s like, I made this decision for this reason or this decision for that reason, or here’s how I would think about it. So it depends on what the model builders are trying to accomplish. And so, it can get quite nuanced in terms of what they’re trying to train on.
Is AI Overhyped
Lenny Rachitsky: Got it.
Jason Droege: So it’s not like here’s a website and then it’s created doing websites. It’s like, here’s a website, here’s why I made this decision, here’s why I didn’t make this decision, or here’s a broken website and here’s why it’s broken if they’re trying to accomplish, I don’t know, a debugging tool for a website builder or something like that.
Focusing on Product and Customers
Lenny Rachitsky: And another example you shared is a short story where it’s like, here’s one short story, here’s another I imagine generated by a model. And then it’s like, which is better, and then how would you make it better? The other example you just shared is a Salesforce agent where it’s like, Hey, book a meeting with a prospect and then teach it how that happens. I love just how concrete these are because it’s like, okay, I get it. This is the stuff that these companies do. Is there another maybe one or two examples just to give people a sense of what this data looks like?
Reverse-Engineering Restaurant Economics
Jason Droege: Absolutely. I can actually give you an example from, so we have two sides of our business. One, we supply data to model builders. We sell the data, and then the other is we actually do solutions. We sell applications and services to healthcare systems, insurance systems, et cetera. I actually think it would paint a more colorful picture if I gave you an example of one of those because it involves data, but it involves the use of data, the manipulation of data for a very, very specific goal. And so, one example there is we work with a healthcare system and health systems have lots of problems. This particular healthcare system has experts that see very rare cases on a regular basis. So you go there only if no one else can figure out your problem, and there’s a huge backlog. So there’s a productivity element to this implementation tier.
So there’s a huge backlog. They want to be able to see more patients, they want to be able to provide better care, and they want to prevent the number of revisits because they want to give the accurate diagnosis day one and what the treatment should be. Well, to do this today without the help of AI, the doctor really needs to read 200 to 300 pages of documentation and it’s rolled into one document, but in different formats. And so, if you’re a doctor, how are you going to read 200 or 300 pages of everything? So what they do is they do the best they can. They scan it, they ask a nurse to look at it, they ask maybe a more junior doctor to take a look at this case. They want to treat the patient well, obviously this is why they became a doctor. And then, they go into the room and they talk to the person and then they make a diagnosis.
Well, we basically built a tool that will read that document for them and point out the top 5 to 10 things that they should take into consideration, either allergies that might not be obvious is one example where we actually, we picked up on an allergy that a patient had that would not have been obvious from reading the document and that allergy actually would’ve had a conflict with the medication that they were going to be prescribed. And so, the AI tool basically pulled out this correlation that would’ve even been hard for a human being to do. To make this tool better and better, you get to a certain limit with the models off the shelf, and actually the people inside of this healthcare system have to do their own labeling.
So we talk about labeling for model builders, but we are starting to see the labeling move into enterprises and into governments because you can only get so far with off the shelf plus rag plus some fine-tuning based on recorded data. One thing people often miss about these systems is we assume because you hear these numbers of like, “Oh, this bank in just 200 petabytes of data a year or whatever fantastical number.” What we miss is is that the right data? Which of that data is useful to the models? And most of it is not useful. Some of it is, but a lot of what we do when we’re talking about knowledge work, when we’re talking about making judgment is human judgment based on synthesizing how would this doctor in this case or how would this banker in this case make this decision and how would they make decision in the context of their overall enterprise? And that might be different bank to bank, healthcare system to healthcare system, because of the culture, the objectives, the incentives, et cetera. And so, we’re getting to the point now where we see that digitizing judgment, human judgment, true subject matter, deep expertise is becoming a bottleneck that we’re unblocking for our customers.
Lenny Rachitsky: That’s really interesting. It’s like the spectrum went from just low skill generous labor to experts to now the specific expert at this one company who needs to do this work, this labeling.
Urgency from the Buyer’s Perspective
Jason Droege: Absolutely. I mean understanding what, there’s this broad narrative. We have two narratives. We have the AGI, everything is just going to become AGI, and then there’s the skeptics, which is like, “Hey, this is all bunk, this is a bubble, et cetera.” And of course, my view is most things are kind of like there’s truth in between and some of the extreme parts of the extreme probably correct, but the reality is is that it’s very hard to get machine critical use cases in agentic systems where agents are talking to agents to a level of accuracy that is necessary to accomplish a goal. And one of the main issues is that a one document, think about the problem of even understanding a document, a document that reads the exact same words in company A will have a different meaning and importance in company B. So how do you have a system that knows that? So this is all got to be built. So if you’re going to make good decisions.
Lenny Rachitsky: This is a good segue to this question that is always on people’s minds when they look at companies like yours and the other folks in the space is just how long do we need people to be doing this? At what point will AI be smart enough to do it themselves? I know your incentives are to say we’ll never run out of people because it’s aligned with your growth, but just how should we think about just why do we need people, I don’t know, in 10 years? How long do we need these experts telling AI things it doesn’t know?
Independent Insights and Alpha
Jason Droege: First off, the history of data labeling is a history of new beginnings. Autonomous vehicles do not need as much data labeling as they did in the past. I mean, Scale is a company that believes that data will always be important at the point at which you don’t need external data, human data in models. I think we’ve gotten to a level of advancement in the world that is almost like unfathomable because you’re effectively saying that no new human skill and no new human knowledge is important enough to put into these models. That feels like pretty far out there. And so, for a business like ours, we’re constantly looking at how do you build operations that can constantly find the new needs and then work with the contributor network we call the experts contributors to unearth that data, to unearth that information. And sometimes it’s new people, sometimes within our existing base we find that existing people have expertise that we didn’t know about that maybe wasn’t useful to a model a year ago, but now is useful.
So this is a constant progression of getting more and more data into these models. Yes, we are financially incentivized to believe that humans will always be in the loop, but that’s not just a business belief, it is a personal belief. These systems need to work for us, and if these systems work for us, then we will need to be on the loop or in the loop on any of the decisions that these systems make. As to the broader point around labor, which I think comes up around white collar apocalypse and these things that come up, I’m definitely on the more maybe practical side of this, possibly just because of my nature, possibly because I see what’s going on on the ground actually in these customers where supposedly this transformation is going to happen in the next one to two years. And I just think that it might happen. The space is moving super fast, but I don’t think it’s going to happen.
It is definitely not going to happen in the next year. The idea that it happens in the next two years I think is very far-fetched, but nothing’s impossible here. And long-term, I think that if you go back through, I don’t know, pessimist archive or whatever, these accounts that post, the radio was invented and then all of this will be eliminated. There will be change, but the change, I think humans are very good at adapting. So I think what we’re underestimating in all of the doom and gloom is we believe in human adaptability. We as a company are highly adaptable and I think the history of technology has shown that people are adaptable.
How High Should Startup Standards Be
Lenny Rachitsky: I really like that takeaway. I’m an optimist as well, so I’m always looking for reasons to be optimistic. I want to follow that thread before I get there, something very tactical I want to ask about is evals seems to be coming up a lot, especially with companies in your space. I’m still learning a lot about just what this all is, especially in your market. How much of what you or experts are providing are evals versus other types of data?
Jason Droege: A lot of it’s evals, and within enterprise customers and government customers, it’s mostly evals because somebody’s got to establish the benchmark for what good looks like. That’s the simple way to think about evals. What does good look like and do you have a comprehensive set of evals so that the system knows what good looks like? It’s as simple as that.
How Uber Eats Was Chosen
Lenny Rachitsky: So in the case maybe of the healthcare example you shared, essentially this doctor would be sitting there looking at all these reports, creating evals that are like, this is what this should be discovering in this report, in this record. Is that a way to think about it?
Jason Droege: Yeah, that’s a very big part of it, which is what does good look like?
The McDonald’s Story
Lenny Rachitsky: Awesome, okay.
Jason Droege: I have to reduce things down to simple terms.
Gross Margin Thinking
Lenny Rachitsky: It’s interesting you say good versus correct. Is that a specific term you like to use good versus just this is the correct answer.
Jason Droege: I didn’t intentionally use that word, but these are probabilistic systems and so depending upon… Yeah, so I can get into some nuance here about the right types of problems that AI is good at solving. So if you have a human process that is 10 or 20% accurate or 10 or 20% liked, AI is awesome. Because if you get to 50, 60, 70, 80% accurate, you’re in the money, you’re in the green, everybody’s happy. Now, the system then has to know, hey, for the remainder, how do I make sure that humans are involved for the remainder of the decision making? But from a net value add standpoint, the humans are pumped in that scenario.
If you have a human process, a workflow that is 98% accurate, and you expect an AI system to get you the remaining 2%, not totally there yet. And so, when I say what does good look like? A lot of the processes and a lot of the things that people are asking these systems to do and systems for us to build are making judgments on their behalf. And so, just like we would ask a human being, “Hey, what do you think we should do in this scenario?” What you’re looking for is you’re looking for the best recommendation or course of action given the current information.
Not Losing Is the Key to Winning
Lenny Rachitsky: To you, this is so obvious and to people in your market that I think a lot of people think about AI being trained on just here’s a bunch of data, check it out, learn everything you can from all of human history and all of written record. But what’s wild is basically people are sitting around teaching AI things it doesn’t know, filling gaps. That’s how AI is getting smarter now. There’s no more real data for it to feed on. It’s just like, here’s what I don’t know, or here’s what an expert found you’re wrong. I’m going to teach you this. And the fact that it scales and that’s keeping models improving is so mind-boggling.
Lessons from Selling Golf Clubs
Jason Droege: Yes. No, yeah, I agree. I mean, like with any of these major tech revolutions, the headlines tell one story and then on the ground, laying broadband means you need to dig up every single road in America to lay it. There is the, yeah, it’s as simple as that. Someone’s got to dig up the road or someone’s got to run the undersea cable. There’s always some operational chiseling that’s going on in all of these industries. I mean, if you think about how magical these models are, they’re remarkable that if you’ve been in technology long enough, it blows my mind even today that they get the punctuation right consistently. I mean, that sounds like almost daft to say at this point in the market, but if you were to go back three years and think about that from a technological standpoint, a lot of things that we think are trivial now are very sophisticated, and it’s a combination of, I mean, the real answer is it’s a combination of computational power, model improvement, and data, and all three are getting better at once.
Right Team Over Best Talent
Lenny Rachitsky: Let’s follow that thread. You’ve been at Scale for a long time, CEO for, you said, 13 months. I feel like you see a lot more about where things are heading because you work with labs on things they haven’t even announced yet. You see more than most people, and I know there’s only so much you can share about what companies are doing, but just is there anything you think people don’t truly grasp or understand about where AI models are going to be in the next two, three years?
Jason Droege: Look, there’s so much talk. I think it depends on how much X or news you consume. So I think it’s like what sort of our perspective. The general trend right now is going from models knowing things to models doing things. And we’re pushing the boundaries of knowledge, like the benchmarks that we put out and that others put out are showing that the knowledge that these models have is getting, it’s quite robust. And then, the next question becomes, well, what can it do for me? And as soon as you get into that world, that’s where the environments we were talking about start to come into play. How do you navigate a Salesforce instance? How do you navigate a healthcare system? How do you navigate even a weather app on your phone, and how does the agent make decisions for you?
We’re just getting into the beginning of that. It’ll be very interesting to see how quickly that happens. And I think that’s where a lot of the speculation has a wide variance because we’re at the beginning of it. People take different trajectories on how that’s going to improve. And so, if you take a trajectory of the most aggressive trajectory, which is like, oh, it’s actually going to be quite easy to train on these things, and then it’s just a change management exercise in the economy, which by the way, change management exercises are not to be underestimated.
There’s still people in the world without an email address. And so, the adoption curve then becomes a human and policy issue, not a technological issue. We’re not there from the technology standpoint, but I do think in the next two to three years, if I take the bait and have to make a guess is the technology will get to a point where it will push the change management and policy makers to say like, “Oh, what do we do with this because it’s getting pretty close?” That’s probably two or three years away.
AI Quick-Fire Questions
Lenny Rachitsky: There’s been a lot of talk these days about AI not delivering on the promise that we hear, especially at enterprises. There’s this MIT study that just showed that there’s all these pilots that people are excited about and then they don’t work and companies aren’t adopting these tools. There’s data showing engineers are not actually as productive with tools. It actually slows them down sometimes. You work with a ton of companies implementing all kinds of AI. What are you seeing on the ground? What kind of gains are you seeing? Do you feel like it’s overhyped, underhyped?
Jason Droege: There’s a lot of hype out there, and our job is to actually build products that work, that deliver value for our customers and figure out where the rubber hits the road. And to get a sophisticated, my healthcare example is one, we do other sophisticated workflows, claims management for insurance companies. This is a financial decision that’s happening, but it’s an automatable process. But basically what happens is the POCs get to 60 or 70% of the way there, and the human mind goes, oh, the rest is no big deal. But it’s like uptime in data centers where every nine is an order of magnitude investment in terms of reliability, backups, et cetera. One nine is basically a web server in a dorm room like we had at UCLA, and then five nines is this crazy high bar, but it just seems like a very small movement.
So you have a similar dynamic going on here where you have a bunch of people, one of the reasons why the POCs have failed, one, there’s a denominator effect because it’s so easy to do, “Hey, I spun up a project, I spun up a project, I spun up a project.” So it’s really easy for people to try. So I don’t necessarily know that the 95% number, I think is a bit of clickbait in a way. It tells the right story, but it is a little bit hyperbolic because if you take the efforts that happen in the company where they actually get a quality partner like we are, or if you do it yourself, if you have engineers who’ve worked with models before and they put in the time, and I’m talking about months, not like minutes like you see in these videos to actually get legal approval, policy approval, regulatory approval, change managements like an accuracy that everybody’s comfortable with. If you actually do that, these things take 6 to 12 months to get them truly robust enough where an important process can be automated.
So I think that’s where the hype is right that when you do it, the impact is like, whoa, I never would’ve figured that out myself, and I’m one of the most educated doctors in the world as an example. But the time to get there is just longer than what people are selling.
Thinking Fast and Slow & Bias
Lenny Rachitsky: It’s such a good point that it’s not only is it easy to try these things, it’s just like everyone’s doing it so everyone’s feeling FOMO like, “I got to try these things. I got to try all these prototyping tools, Cursor, all these things.” Just goes, “Everyone’s doing it,” and then you just rush into it and it doesn’t actually work out.
Jason Droege: Easy to learn, hard to master. That’s my summary.
Recent Favorite Movies and Shows
Lenny Rachitsky:
Whether you’re a funded tech startup looking for ways to pay contractors and earn yield on your idle cash, or an agency that needs to invoice customers and keep them current, or an e-commerce brand that needs to stay on top of cashflow and access capital, Mercury can be tailored to help your business perform at its highest level. See what over 200,000 entrepreneurs love about Mercury. Visit mercury.com to apply online in 10 minutes. Mercury is a FinTech, not a bank. Banking services provided through Mercury’s FDIC-insured partner banks. For more details, check out the show notes.
Okay, let’s move on from AI. This could be an endless discussion about AI, but you’ve got a lot more lessons to teach us. You’ve helped build Uber Eats, you’ve had a couple startups in the past. We talked about Scour for a bit. I’ve talked to a bunch of people that have worked with you over the years, and I got a lot of really interesting insights into the stuff that you’re extremely good at. So I’m just going to go through a bunch of these. One is your obsession with being close to customers, talking to customers, and I love this topic because it’s something everybody thinks they’re great at, and they feel like they completely understand how this is important, why this is important. They all feel like I’m doing this, don’t worry about… Everyone else is not doing this, but I am. Talk about just what you think maybe people miss about how this looks when you’re doing it well, and just why this is so important.
Jason Droege: I mean, I probably fall in the category of what you just described, which is maybe part of the hubris you need to start anything new. But I mean, I don’t think it’s a clean process. I think my process is I’m constantly questioning every single thing that I’m hearing at the beginning of anything. I don’t take what a customer says literally. And there’s been a lot talked about on this topic from a product management standpoint in terms of like, oh, don’t do what they say, do what they mean, and look at the real problems and underlying things. I think the way that I look at it that might be additive to the discussion is I look at the underlying incentives of the customer. And the underlying incentives of customers are not always financial. Sometimes it’s ego, sometimes it’s career growth.
If you’re selling enterprise software to someone, there’s an executive sponsor as an example, that person needs to trust that you’re going to do a good job for them. How do you get them to jump with you on this big project? Well, that’s part of the journey of not just the product, but what do they need to hear from us? What do we need to supply them? What do we need to do to actually unlock the opportunity to implement the product? So I think there’s an incentives alignment baseline. I’m a big believer that it’s cliche, but show me the incentive and I’ll show you the outcome. I think that’s absolutely true. And even when customers will tell you things, I’ll give you an example. I’ve been out of the game for a while so I can be open about it, Uber Eats.
So when we launched Uber Eats, I looked at the business in terms of being close to the customer. We actually couldn’t get a restaurant tour. I knew nothing about this industry. So at Uber, my job was to figure out what other businesses we should get into. And so, we looked at a billion businesses and Uber Eats, food delivery was the one that we thought was most interesting, which turned out to be right so good for us.
Recently Discovered Products
Lenny Rachitsky: Very right.
Life Motto
Jason Droege: And we couldn’t get a restaurant tour to help us understand their unit economics. And they’d say like, “Oh, it’d be this percentage or that percentage, or Why do you want to know?” And then we’d go to a different restaurant tour and they would explain it, but they were a little suspicious of why are these Uber guys talking to me about how much my ham costs? And so, what we did is we ordered just a bunch of food from these places, and then we got a restaurant supplier to give us a base catalog, and we just matched up how much does the ham weigh? How much does the cheese weigh? How much does the bread weigh? How many pieces of lettuce were on there? And we tried to actually just compose our own independent view of what’s the ingredients cost versus what’s the labor cost? And then, we triangulated what was our ground truth, and then what are we being told by restaurant tours, and then what is the site guys telling us about restaurant economics?
And if those things all overlapped, and we’re like, okay, we have an insight about what to do here and how does this relate to Uber Eats? Well, what we found as part of this is that roughly a restaurant pays 20 to 30% of every meal to ingredients, and they pay roughly 20 or 30% to labor, and they pay roughly 10% to real estate and a bunch of other, anyway, so goes down the chain. But the important parts is what’s the value of incrementality?
And so, we came in and we said, “We’re going to charge you 30% of the bill.” And they were like, “Oh my God, is this group on all over again? This is way too high. Oh my gosh.” And we explained the economics to them and they were like, “Okay, we’ll give it a try, but this is way too high.” And they were right, the real number, the real clearing prices aren’t 25%, but we weren’t that far off. And so, when you go to find product market fit or be close to the customers, it’s a combination of what’s the most valuable thing. Well, in a restaurant tours case, give me incremental demand. Because if you were to take a restaurant location and triple demand based on the same labor but you’re just scaling ingredients, you’ve got a 70, 80% incremental gross margin product.
Restaurant tours would hate when we would say this because it doesn’t work out exactly like that in reality. But because we had that insight, we had confidence that we could go to market with, we need to charge you this so that the delivery fee can be that. And then, if the delivery fee is that and we charge you this, then we think the consumers will adopt, and that’s what you need to get your incremental demand, and then we could pay the driver this. And so, you fit this whole puzzle together without totally satisfying, in the case of a marketplace, you’re not totally satisfying any individuals 100% of their needs. What you’re satisfying is is you’re getting a clearing rate for them to participate in the market in the case of a marketplace. So that’s one example.
Uber Eats and Goodbye
Lenny Rachitsky: Yeah. I love this example as you almost you figure out how to help them with something they don’t even fully themselves know yet. So as you think through their goals for them as if you were them, break down the economics and then here’s the solution versus, hey, what can we do for you guys?
Jason Droege: Yeah. I mean, if you walked into a restaurant, they would tell you a bunch of things. They would say, “Oh, labor schedule is an issue.” They would say, “My rent is an issue.” They would say, “All these, my ingredients prices are an issue, that’s 20 or 30%.” If you could shave off 3% of that, that would be huge. You might then take that and go, “I’m going to go build a business. It’s going to save you 10% of your ingredients costs.”
Well, but that doesn’t actually get into their head on what’s truly important day-to-day. That might be important for them on an annual basis, but on a daily basis, what are they doing? They’re looking at their numbers, they’re looking do people show up. Did I make money yesterday? Am I going to make money tomorrow? So the urgency, I think the biggest thing people miss when they’re building new products is the urgency of the buyer part of it. You can build something that provides a lot of value, but if it’s not the top thing that the customer is thinking about in their busy days, then you’re just going to have a long road to a small town.
Lenny Rachitsky: This touches on just the theme I heard a lot about, this idea of independent thinking and how much you value that, and this feels like a really good example of that. Is there anything else along those lines of just why this way of thinking is so critical?
Jason Droege: I think as a founder’s job, and I mean I stretched that term because at Uber we had all of the benefits of Uber so I wasn’t really a founder. I just started the business there. But there are some elements of founding there is you’re looking for alpha in the market. When we started our first company in ‘97, it wasn’t that cool. It might’ve been cool in Silicon Valley, but it was definitely not cool in LA. Now, it’s super cool to start a business. So as a result, everyone’s trying everything. So how do you get alpha on that market. If your research is highly influenced by what the world is saying around you, you’re not going to have an independent insight. You have to go off and do your own thing.
And this is why from an entrepreneurship standpoint, I have very strong feelings about what the approach to founding a company should be and is probably very particular to me. But it truly is about what insight do I have, because why am I so lucky to have this insight? Why in a world of a million entrepreneurs who are thinking, who are smart, who are trying everything, why am I in the position where I likely have an insight that others do not? And then, why am I the one to do it?
And the answer might be I’m in this narrow, far-flung place. The other answer might be, I am inherently a contrarian personality type, so I’m just constantly looking for the thing that’s true that people don’t believe is true, which sometimes worked. But then, the second part of that’s super important, which is why do I want to work on this problem for 5 to 10 years? And people get this wrong all the time. They go and talk to a customer and they go, “They have a problem. I’m going to go solve it.” And it’s just not a great way to start a business. You really have to have this burning desire to constantly be questioning yourself.
The other thing about independent thinking is is that you can’t fall in love with your ideas. And I do not proclaim to be the world’s greatest thinker for what it’s worth, this is what you’ve been told, but is just part of that is basically throwing away who you are, who you’ve been, all your ideas for the mission that you’re on, which is trying to accomplish something for our customer.
Lenny Rachitsky: This is great. I’m glad you went here. This touches on the other theme I heard often about you is just how high of a bar you set for new businesses. And I think this advice is useful both for founders, as you said, and also people starting companies within companies, new business lines. So you’ve talked about this a bit already, but is there anything more there, just how high that bar needs to be for it to likely work out when you’re starting something new?
Jason Droege: Look, if you want to give yourself the best chance, and this isn’t always how it works, but if you’re in my position 25 plus years in their career, if you want to give yourself the best chance, I think there’s two ways that companies end up working out. And the first way, which is probably the most important, quite frankly, is that the founder is just a force of nature over a long duration of time. Because you’re going to have to pivot, you have to have that energy to pivot. You have to go years and years and years with it being hard, and that’s probably the most important thing.
But the second most important thing is that you can easily educate yourself on what are good business models, what are bad business models, what are good markets, what are bad markets? And even if you’re this force of nature, having the knowledge, if you’re going to go into a bad market with all your energy, you should at least know, maybe ignorance is bliss because you just throw yourself into it and it just works out with time. But that’s not how I would operate, which is marketplaces are good businesses. SaaS, at least historically, we’ll see how this changes, but SaaS, historically, great businesses, recurring revenue businesses, sticky businesses, network effect businesses.
And if you look at what the top VCs invest in, yes, there is a lot of portfolio building, but there are similarities in terms of the types of business models that they believe could be worth tens of billions of dollars. And they have network effects, they have lock-in. They are more valuable at scale, a big scale than low scale. So if you just take a filter on a new business, this is what I did at Uber, which is like if you just have a filtering mechanism on a new business, it doesn’t take that long to eliminate the bad ideas. And then, of what’s left, you can pick, oh, I’m very passionate about this, even though it might have more problems than this other thing that on paper looks better. And then, you have to have passionate about it. But I think people just miss a basic understanding of what businesses even have a chance of being worth $100 billion.
Lenny Rachitsky: So you launched Uber Eats, you figured out this is the place to go and bet. As an outsider, feels obvious, of course this is going to be a massive success. Of course, food delivery, such a good idea. I know you looked at a ton of ideas in that process. Can you just talk about what you explored and why you ended up picking Uber Eats?
Jason Droege: I am definitely not the smartest person in the room when it comes to figuring these things out. And so, I keep a very, very wide aperture on ideas for as long as I can until I’m like, okay, everything is coalescing. And I think there’s a bunch of reasons why you have to keep an open aperture on considering ideas that might seem bad at the start, but you just keep digging and see if you’re right that they’re bad or you’re wrong. So just as a general philosophical principle, I’ll start there. We looked at, we did some crazy stuff. I went walking around San Francisco one day and I looked down Market Street and there was a CVS, a 7-Eleven, a CVS, a Walgreens, a 7-Eleven, and I’m like, “How many SKUs could possibly be inside one of these things that people want and couldn’t you just put that into a van and you hit the button on the van and the van comes around and you get whatever convenience items you have, and they’re convenience items, so why would that be a problem?”
And we launched that in DC. We put 10 of these trucks on the road, we put 250 SKUs in them. And I mean, crickets is an understatement of how bad it was. I mean, we couldn’t get an order to save our lives. And what we realized was that we hadn’t really done the research on what convenience stores really were. It was if you didn’t have cigarettes, you didn’t have beer, you didn’t have Slurpees, you didn’t have these things, for example, you didn’t bring people in to sell all the other things. So we didn’t know anything about retail. We were clueless. So that’s one idea. We looked at grocery, but honestly the unit economics just terrified me of all the pick packing and everything like that. I think Instacart did a remarkably good job at getting the unit economics to a good spot and probably the hardest operational problem you could tackle.
We did generalized delivery, point to point delivery, what’s now, I forget what Uber’s product is called, but Uber Direct I think it’s called, where you have something that needs to go point to point in a city. That was a flop from the beginning because the truth is is how consumers don’t really have this need, business sort of have this need, and in 2014 when we were doing this, no one had this need. But we tried 15 versions of all these things before we eventually just said, “Okay, the food delivery thing is just popping off on all signals and we can make the unit economics work. People seem to want it. It’s a super cool problem because we can enable independent restaurants with all these tools and allow them to compete with the big guys. We can take the real estate out of the equation. So you can have a real estate location that’s non-prime, but if you have prime food, then you get to compete.” So we’re like, “Oh, this is a very interesting problem and we can really help local economies.”
Lenny Rachitsky: And this ended up being, if I remember correctly, this basically saved Uber during COVID. Lyft didn’t have something like this. And how big is this business at this point? Anything you share about just how important this turned out to be for Uber?
Jason Droege: Yeah, of course. Well, we launched it in December of 2015 in Toronto and within two hours we had done 20 billion. So it was 0 to 20 billion in four and a half years, which is pretty good. Uber was very good at scaling things, but competitive market. Others did well. We beat a lot of people. Some people beat us. And then, now I think it’s pushing 80 billion, and that’s been for another four and a half years since I left. I think COVID turned it from 20, I left right before COVID, total coincidence, 20 to 50 in a year. So I mean, ride-sharing went this and food delivery just went to Pluto.
Lenny Rachitsky: What luck. Well done.
Jason Droege: Luck is part of the game. That’s the other thing that’s important to realize. Luck is part of the game, so do not begrudge people for luck. This industry is hard. All these things we’re doing are really, really hard. Luck is just part of the game.
Lenny Rachitsky: Maybe speaking that maybe not. One of your colleagues, Stephen Chau, who I am an investor in his new company, he worked with you at Uber Eats for a long time. He told me to ask you about the McDonald’s story. I imagine that was just a big milestone, a big moment enough for you guys. So why’d you decide put McDonald’s in Uber Eats and there’s apparently a story of how you won that deal.
Jason Droege: So it was interesting, and this just goes to maybe where sometimes ignorance leads you to accidentally the right answer. So we had launched Uber Eats and Uber had a global footprint and we were the only food delivery network with a global footprint excluding China. Everything at Uber needed to be launched globally. That was a very big part of the culture, et cetera, which is a lot of work and you can spread yourself too thin and cause other problems. But in this way it was good. My vision was, okay, let’s help the little guy compete with all these chains. They have these systematized food systems and food is what makes a city amazing. And no one talks about the chain restaurant that they visited in Paris. They talk about the local place that they found and let’s be part of that. That’s who we want to be.
And so, McDonald’s actually approached us and they said, “Hey, we’d love to do food delivery with you.” And I said, “No.” And they’re like, “Hold on a second. We have 80 million consumers a day. You don’t want to do this together?” I’m like, “It’s not really our vibe right now.” And so, I pushed them off for four or five months until my team is like, “You’re insane. These people are going to put marketing behind it. They really want to do this. They want to lean in.” So we actually had, because of that, I think it’s hard to correlate these things, we ended up with this exclusive relationship with them, got an insane number of customers of… Chains at this point actually weren’t really on food delivery networks because everybody was so worried about the unit economics, because they’re so sensitive to the basket size.
And my approach was like, eh, figure it out, which is a very Uber culture thing. Okay, the basket’s $17, it’s our job to make that work, reduce the radius on the delivery, figure out the economics, maybe mark up some of the food someplace. There’s always a way to figure it out. So we did it and then three months later the business just started hockey sticking again at a different level. And my team is just like, “Dude, you were so stubborn on this point,” but I think it actually ended up being in net benefit because we got a great deal with them.
Lenny Rachitsky: So the fact that you pushed him out helped you get a better deal is what I’m hearing. That’s amazing.
Jason Droege: Yeah, I think that’s the story he would be referencing. And then, the onboarding of it was crazy because we basically went global with them in six months, and at this point the business was less than two years old. So activating this, I don’t even know, an 80-year-old company that expects processes to be in place and we have two of our office managers in New York managing it. It’s just mayhem.
Lenny Rachitsky: I’m still sad In-N-Out is still not on any of these apps.
Jason Droege: Yeah, me too.
Lenny Rachitsky: I remember someone was hacking it. There’s all these ways people found a way around and they’re like, “No, no. Okay, you’re Postmates. We know we’re not going to give you any food.”
Jason Droege: Yes, love In-N-Out.
Lenny Rachitsky: You’ve touched on this idea of gross margins and margins, how obsessed you are with this. I wanted to spend a little time on here. I’ve heard just you’re obsessed with understanding gross margins before going in on anything. Most founders have no idea what they’re doing here. What have you learned about just what people should be paying attention to, what they might be forgetting when they think about just the feasibility of a business?
Jason Droege: Yeah, look, it’s one filter like many filters. There are certainly businesses that have low gross margins that are great businesses. Costco, Walmart, et cetera. Amazon talks about this all the time of there’s companies that increase prices and there’s companies at lower prices. But I would say that by and large, high gross margins combined with healthy churn curves are a very healthy sign for the business. I mean, think about it. If I were to sell you something and I can’t mark it up a lot, how much value am I adding beyond what’s in my hand? And if I’m not adding that much value, then what am I in the business of doing? And I’m in business of adding value. And it’s not quite that simple. This is just a litmus test of when someone comes to me and they go, especially in a new business, and we deal with this. I dealt with this at Uber, I’ve dealt with it everywhere.
Someone comes up with an idea and they go, “We can get into this business and I think we can charge this and it’ll get us to a 40% gross margin.” And then, my next question is start at a 60% gross margin. Why does that not work? And they go, “Oh, well, the customer…” And immediately, you short circuit to what the real problem is. Oh, the customer has an alternative. Oh, okay, well who’s the alternative? Oh, it’s some offshoring company. Well, what’s their gross margin? Oh, we don’t know. You go find out. It’s like 20% and they’ve been around for a long time and they have scaled operations. And you’re like, okay, so your gross margin is going to go from 40 to 20 quicker than you think, and you’re going to be in a world of hurt unless you do something to differentiate.
So I take gross margin is just a very coarse instrument, not a perfect instrument to think about, am I adding enough value? Am I differentiated? It’s not perfect, but it’s a very quick short circuit filter to even to see if someone’s pitching you an idea, have they thought through this dynamic? Because if the response is gross margin is super low right now, but here’s the dynamic I’m going after. And then you’re like, “Oh, okay.” And sometimes it’s like, we’ll just make it up with volume and then the gross margin will go negative for a while and you’re like, “Wait, this doesn’t work.”
Lenny Rachitsky: So what I love about this is just a lens into is my idea good enough if studying, can I keep a high gross margin? Is there a reason why people in this space haven’t been able to have a higher margin?
Jason Droege: Yeah, exactly. And like I said, it’s meant to disqualify just you’re doing these large for larger companies and everybody has ideas. And so, it’s a way to cut through. Do you understand the machine that is going to need to be in place in two or three years? You might have a 70% gross margin now because the next question is why can’t someone else do this? And if you have an answer of like, “Well, they can now, but they can’t in two years, if we run really fast.” Okay, we might have something. If they can now and they will be able to in two years, you’re going to have margin compression.
Lenny Rachitsky: Along these lines I was just listening to, I think it was the a16z podcast. Alex Rampell I think was sharing this story about Costco, how as you said, their strategy is actually to keep margins very, very low because all their revenue comes from their membership. So they have something like 50 million members paying 100 bucks a month and that’s their entire business. And so, they don’t plan and they don’t want to make money off the products.
Jason Droege: Yeah, that’s right. I mean, they’re playing a slightly different game, not an expert on Costco, have spent some time with the company, but they use price as a way to get to scale. And so, they’re basically saying if we discount, same with Walmart, we will get so much volume that we will just take the air out of the room for all of our competition. And so, then the question of, okay, so if you have a low gross margin today, in two or three years, once you land one of these centers in a market, why won’t your margins to get eroded? The answer is because we will have already absorbed all of the demand. You try to go to 8% versus 10% gross margin, which I roughly think is what their gross margin is. That’s going to be a really hard business. If you already have a habit with a customer, they have already built their weekly trips around you, you already have relationships with suppliers, you already have general managers that know how to stock inventory, that’s not a straightforward exercise. So they’re first to scale and then good luck competing with them.
Lenny Rachitsky: Okay. Just a couple more questions. One is there’s this term that I’ve heard that you often say and believe in is this idea of not losing is a precursor to winning.
Jason Droege: Yes, yes.
Lenny Rachitsky: Talk about that.
Jason Droege: Tech is a culture where portfolios are built by investors, and a lot of the narrative is controlled by investors frankly. Founders obviously participate, but this idea that you should just go for it is consensus. Just go for it. Who cares? Well, I don’t know, if it’s my life and I only have one moment to take a shot, I might want to just not just go for it. I might want to think for a little bit, and I think the best entrepreneurs, I have no data to back this up, but just these are my friend, this is my friend group. I think the best entrepreneurs and the best business owners look at the risk profile of the decisions that they’re making and they try to make asymmetrically positive decisions all along the way.
And so, oftentimes I feel like we forget about the risk of a decision, and there’s more to unpack there because I actually think taking highly risky decisions and then having it work out is a weird cultural thing too, because then how do you train people to do that? Because it’s a very hard thing to take high risk decisions and be right enough because it creates a lot of volatility. But it goes back to my comment about the most important thing in founders, which is just this ability to persevere through. Survival is just part of the game, and most people just give up before they get their timing right, before they get the right insight with the customer before they get the right product in the market. And life can change quickly in tech. You can go from being a dog to being a hero in a very short period of time, but you’re on this very, very long journey, but you have to survive for that condition to be met.
And so, then the question is is when you’re in a hype cycle, I would argue that we are right now, everyone wants to go for it and then go for it more and then go for it more and go for it more and you don’t realize, guys, all of our customers are going to be around in five years. They just want us to solve their problems. We have to be around to solve their problem for them. And so, survival is a precursor to that. So let’s not put ourselves in position that could potentially compromise the enterprise along the way. It doesn’t mean don’t take risks, but think about how you calculate it.
Lenny Rachitsky: I love how clear it is that this lesson and many of the lessons along these lines have come from just failure and things not working out and things breaking, which is the best outcome.
Jason Droege: If you ever get on the other side of a high reward, high risk decision, it is so painful because you are just cooked. You are done, and often there’s no way out.
Lenny Rachitsky: Is there a story along those lines that comes to mind or an example of that?
Jason Droege: Well, this is where it is together on why I try to be so I think you can spend a little bit of time thinking upfront to save yourself a lot of pain downstream. I had this business not worth detailing it, but after the bubble burst in 2001, I’m like, “I’m going to self-fund a business. I’m going to build a profitable business. I want to prove that I can do this.” And we had started Scour, which had all the things we talked about. And so, what I did is I’m like, I was a golfer and frankly, there was nothing to do in tech.
So I started selling golf clubs on the internet and I was making real money and I might’ve learned more from this business than any other because I started on eBay and I was 22, and I didn’t really understand that my margins would come down because anyone can do this, but I was one of the first ones to do it, so I was making a ton of money and then I built this business and then I just failed to recognize I had a lot of hubris. I was like, “Oh, if I could just buy all the used golf clubs in America, I can be the market maker for prices,” and don’t people do that?
Lenny Rachitsky: I love this ambition. That’s great.
Jason Droege: And it’s just like it’s madness to actually think about the practicality of that. And so, I just didn’t spend the time thinking and then I ended up in this business. The business was profitable, it got to a couple million of revenue, whatever, paid me a dividend for a while, but it was painful the entire way.
Lenny Rachitsky: I love the spectrum of experiences you’ve had. You’ve sold golf clubs, you’re helping achieve AGI, you could say. There’s also a whole part of your career. We haven’t talked about where you built tasers and body cams and drones and all these things. Also, peer-to-peer file sharing before anyone else. Final topic I just want to spend a little time on based on this experience is hiring and building teams, something that I know you have a really strong take on. That I’ve been hearing a lot on this podcast recently is this idea of it’s more important to build the right team than find the most optimal top talent. Talk about that, why that’s so interesting and important.
Jason Droege: As of late, I’ve developed a more nuanced view of this, which is for certain roles, you absolutely need the right experience in this current market. You see this with researchers, because the market’s moving so fast, you don’t have time to train up some people, so you actually have to go find people either who have the right relationships with customers that you want to get or you have to, who might not check other boxes but are awesome at that, might not check the classic boxes that I think you’re referencing of they’re a problem solver, they can grow with the company, they have a high trajectory, et cetera. I would say that’s 5% of the roles in the company, but very important whenever speed to market is important.
And then, for interviewing, I just interview for three things and I have to interview across all kinds of expertises, which is hard. I can’t be an expert in everything. And so, I reduce it down to just three things, which is like, are you a curious problem solver and can you articulate that verbally? Can you work across people? Are you humble enough to work across and are you a good leader? And if you just do those three things, I think you have a pretty high chance of success, at least in an organization that I’m running, because the world’s changing. So you do need people that are adaptable. So all the experience is not necessarily one-to-one relevant.
And then, the working across to your team point, this actually came up at Uber Eats. So when I was building the Uber Eats management team, I’m not sure if this was mentioned to you from that group, but whenever I would hire people, I was trying to compose almost like an organism of strengths and then minimize the conflicts. That management team for the most part outside of some of the operations side, but for the most part, that management team was the same management team from day one when we had nothing to $20 billion. And I just believed that the team, knowing each other’s strengths and weaknesses and being able to compensate for each other was more important than the classic advice you get around, “Well, that person hasn’t seen this much scale.” And you’re like, “Well, yeah, but can they learn it?” I learned it. So you do have to kind of believe in people a little bit, which is my job, not necessarily their job. And so, I mean, these are people systems. They’re not straightforward rules-based things you can apply.
Lenny Rachitsky: And I especially love this advice because there’s all this talk about what skills will matter in this world of AI doing all our jobs, and it feels like these three buckets are maybe the same thing, just are they good at solving problems? Are they good leaders? Can they collaborate well with other people?
Jason Droege: Yeah, I don’t think that the core rise of humanity, it will change, and I think that these things are pretty core to how humans have been successful for a long time.
Lenny Rachitsky: Speaking of that, I’m going to take us to a recurring segment of this podcast that I call AI Corner, where I ask folks this question, what’s some way that you’ve found a use for AI in your day-to-day life in your work that makes you more effective, get more done, get better stuff done?
Jason Droege: Honestly, when I came into Scale, so my history was in consumer and I’ve done some application level stuff with government, and this space is moving so quickly. AI is my, I use it as a tutor. As these new concepts come up, I have a lot of people in the company who can educate me on the nuances of the technicals of all of, excuse me, the technical nature of the data and the products, but they only have so much time. And honestly, there’s new concepts coming up all the time and I need to stay on top of it.
So it might sound crazy, but a large percentage of my job is not dealing with the engineering issues related to AI. I’m managing an organization, but I love understanding it. It’s one of the most enjoyable, rewarding parts of my job is to learn from all these AI researchers, but they don’t always have the time to do it, so I use it as a tutor. I turn on voice mode and talk to it on my way into work. So I think that’s probably the most impactful thing that I use it for that’s also relevant to this topic.
Lenny Rachitsky: I do exactly the same thing, especially when I’m prepping for this podcast. What exactly is this? I think about when you say this, I did an interview with the founders of Perplexity a few years ago asking about how they work at Perplexity, and the founders said that before, they were ruled, before they ask a question of anyone on the team, they have to ask AI first. And I was just like, “That’s crazy.” Now, it’s so obvious. But back then, I was like, “That’s an insane way of working. I’ve never heard of this before.” Just a sign of how ahead of the curve they were.
Jason Droege: Yeah, I think number two would be I’ll take internal documents and I’ll ask, what’s the most important thing in this document? And I’m shocked, and then I’ll read it and just double check, but I’m shocked at how good it is at just pulling out. There’s so much in organizations that is like, I don’t know what you want me to say and I don’t know what I need to know, but we each have our own agendas, and so this matching of, and so then there’s this huge broadcast problem where it’s like, of all of the information you might want to receive, what’s actually important to you? And so, I use it a lot for that too.
Lenny Rachitsky: Amazing. That’s a really good tip. I use it for legal documents, just like what do they know about what they’re trying to do here for me or against me? Jason, is there anything else you wanted to share or leave listeners with, maybe double down on a point before we get to a very exciting lightning round?
Jason Droege: Yeah, absolutely. I mean, I think the really important, the reason why I’m doing this, the reason why want to spend time here outside of wanting to be on the show for a while and being a long-term listener is, our long-time listener, excuse me, is there’s a lot of amazing work going on at Scale. The teams are working super hard, we’re delivering a ton of value for our customers. The public narrative has not represented the work that the people here are doing and the work that our customers are doing with what we’re doing for them. And I just think that deserves the respect and reward that all those people are putting in, and we’d like people to know that.
Lenny Rachitsky: I appreciate you saying all that. With that, we’ve reached our very exciting lightning round. We’ve got five questions for you. You ready?
Jason Droege: Yeah, let’s go for it.
Lenny Rachitsky: What are two or three books that you find yourself recommending most to other people?
Jason Droege: Some of this is going to sound interesting. The Selfish Gene is one of my favorite books.
Lenny Rachitsky: Love that book. I don’t know if anyone’s ever mentioned, it was one of the most influential books for me too. So sorry, keep going.
Jason Droege: Yes. I think Selfish Gene. Road Less Traveled, I’ve read more than once. I mean, it’s just one of the classic human psychology book. And then, I think in business, I think Good to Great. It’s not the read that you’re going to be most excited to enjoy on a vacation, but it’s pretty much right, and I think we should take advice from people who have analyzed these business problems before because not a lot’s changed, but we keep acting like everything’s changed.
Lenny Rachitsky: What’s crazy about that book, you look at all the companies they talk about, I haven’t read in a while, but just the whole book is about companies that last, I believe, or maybe that’s the other book, I don’t know. But anyway, all the companies that they talk about, I don’t know if they’re still around. It’s so hard for a business to last a long, long time.
Jason Droege: I would also recommend Thinking Slow and Fast, that’s the… Yes.
Lenny Rachitsky: Thinking, Fast and Slow.
Jason Droege: Thinking, Fast and Slow. Excuse me, sorry. It’s been like a decade since I read it, but just in terms of point there being human biases are very important to understand.
Lenny Rachitsky: What’s really crazy to me about that book and Kahneman in general, someone asked them just, how’s your life been impacted by learning all these biases humans have? He’s like, “Not much. I have the same biases. Knowing them doesn’t really help me avoid them.”
Jason Droege: See, I find myself checking myself. Whenever I get super convicted on something now I will be like, okay, what is the list of things that I’m inclined to do to try to catch myself? Because I think we’re most inclined to have these bad decisions impulsively, which is what I think the book is largely about. I mean, it’s a long book.
Lenny Rachitsky: So long. Oh, my God. It feels like that’s where AI can help us in the future. Just like, “Hey, Jason, are you sure this isn’t framing a fact or whatever?”
Jason Droege: Yes.
Lenny Rachitsky: Okay. Next question. Do you have a favorite recent movie or TV show that you’ve really enjoyed?
Jason Droege: Most of the movies I watch are with my kids, so I wish I had something deep and profound.
Lenny Rachitsky: No, kids content also is a very acceptable-
Jason Droege: The Formula 1 movie I thought was really good. I mean, it’s a classic action movie. I don’t think it informs anything in AI or business, but it’s good to check out from the craziness of tech once in a while.
Lenny Rachitsky: Is there a product you recently discovered that you really love? Could be an app, could be clothing, could be a kitchen gadget, anything along those lines?
Jason Droege: VO3. Not totally new, but when I was in high school, I wanted to be a screenwriter. I actually grew up in the Bay Area and everybody was an engineer, but I wanted be a screenwriter. And so, I went back and I got the first page of one of my old scripts, which not good scripts, but I got the first page. I took a picture of the script and I fed it to VO3, and I said, “Make this scene,” and it got it right.
Lenny Rachitsky: Wow.
Jason Droege: I was shocked. I was just absolutely shocked that you could just take a picture of a script. And so, now I’m thinking about that for how do I use these tools for family videos? Some of the grad tools now with making live images more active, I think are really interesting. I think they need one more step of iteration, but I think those are going to be really emotionally life-changing for people because just a little bit of movement in an image from a grandparent or a relative or whatever you haven’t seen in a while, it really does make a big emotional impact on you.
Lenny Rachitsky: I love that when you play with these tools, you probably can think about, oh, here’s the people that help train this thing. Here’s the people that helped on the problem that it had.
Jason Droege: I was actually talking to someone who was working on VO3, and I told him the script thing and he goes, “Oh, actually scripts. Yeah, no, the way the data is formatted in a script, that would actually be very good.” Because they start with set looks dark interior, this character says it in this raspy voice, and so it gives you all the instructions in the script.
Lenny Rachitsky: Oh, man, just unlocked a whole new business unit right there. Two more questions. One is do you have a favorite life motto that you often think about, find useful in work or in life?
Jason Droege: Yeah. The end is never the end. That’s my favorite internal saying, and it goes to the comments before about survival being a precursor, surviving being a precursor to thriving. You got to survive before you thrive, which is your brain tells you, and along these entrepreneurial journeys, I think this is most applicable. I mean, this is the hardest journey anyone can go on. If you go on this journey for five years, you are mentally harder than 99.9% of the population. People don’t understand the Chinese water torture of having self-doubt and having things go wrong, et cetera.
And so, more tactically, you get this when you’re working out like in a day like, “Oh, I’m too tired. I need to stop.” But the truth is is you can keep going and the world’s going to keep spinning. So I find in the moments where it’s just the hardest or you have this hard decision that seems impassable and your body, you’re having this visceral reaction to this is impassable, just to remind yourself that I’m going to wake up tomorrow. This isn’t the end. There’s another end somewhere. I just find that to unlock me to be like, okay, there might not be a perfect solution, there might be an imperfect solution, but it’s a solution so let’s just keep going.
Lenny Rachitsky: Final question. You helped create Uber Eats. I imagine you’re still a power user of Uber Eats. You have a favorite restaurant on Uber Eats that maybe people should know about, maybe that you order most from?
Jason Droege: I order a shocking amount of McDonald’s actually. Despite my original story, it’s the family treat in the house. I would say that that’s probably the top thing that we order.
Lenny Rachitsky: Oh, man, I’m worried for your health, but I love, I haven’t had McDonald’s so long. This is like, maybe I should give it another-
Jason Droege: I mean, more practically we will order mixed greens or tender greens or something like that on a day-to-day basis, but I think that the more notable, surprising thing is is that despite my initial aversion to working with a global chain, it’s a good treat once in a while. You just shouldn’t have it all the time.
Lenny Rachitsky: Jason, this was incredible. I really appreciate you making time for this. I’m really honored to be the first chat you’ve had since taking over at Scale. Where can folks find you online if they want to maybe reach out, learn more about what you’re, I don’t know, maybe join Scale. Where do you want to point people to and how can listeners be useful to you?
Jason Droege: Yeah, absolutely. I’m @jdroege, J-D-R-O-E-G-E on X. That’s probably the easiest way to follow me, keep up with things and you can shoot me a DM if you like. And so, I think that’s how you would keep in touch and, sorry, what was your other question? Sorry.
Lenny Rachitsky: If you’re hiring, I don’t know, where should people go check it out if you are, and then also just-
Jason Droege: Absolutely. Just go to scale.com, go to our careers page, and we have 250 open roles. To the point about we’re in business and we’re growing, we’re hiring a ton of people. Our data business is growing, our applications and services business is growing like crazy, and so we’re going to need a lot of people to help us on that journey.
Lenny Rachitsky: You guys just signed some insanely large contracts with the government I was reading.
Jason Droege: Two $100 million contracts.
Lenny Rachitsky: $100 million contracts.
Jason Droege: 100, yeah. We didn’t sign just one. We signed two in one month, so yes, no, our federal business is doing well. Our enterprise business is doing well. Our international government’s business is doing well. There’s a lot of demand out there.
Lenny Rachitsky: Some salespeople are getting some great commissions. Good job. Jason, thank you so much for being here.
Jason Droege: Yeah, thank you. Honor to be a guest here. Super excited to be with you, especially so early in the journey, or at least my journey here leading Scale.
Lenny Rachitsky: Appreciate it. Thanks for coming. Thanks for joining us. 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 |
|---|---|
| a16z | 保留原文(Andreessen Horowitz 风险投资公司简称) |
| agent | 智能体(agent) |
| AI Corner | AI 角 |
| Alex Rampell | 保留原文(a16z 合伙人) |
| Alex Wang | 保留原文(Scale AI 前任 CEO,现领导 Meta 超级智能团队) |
| Allen Penn | 保留原文 |
| Brendan | 保留原文(Workhorse 相关人物) |
| churn curves | 流失曲线 |
| Costco | 保留原文(美国连锁仓储式超市) |
| evals | 评测(evals) |
| frontier labs | 前沿实验室 |
| Good to Great | 《从优秀到卓越》(Jim Collins 著作) |
| hockey sticking | 曲棍球棒式增长(指突然加速的指数级增长) |
| In-N-Out | 保留原文(美国知名汉堡连锁品牌) |
| Instacart | 保留原文(美国杂货配送平台) |
| Jason Droege | 保留原文(Scale AI 新任 CEO) |
| Kahneman | 保留原文(Daniel Kahneman,诺贝尔经济学奖得主、行为经济学家) |
| Lenny Rachitsky | 保留原文(播客主持人) |
| Lyft | 保留原文(美国网约车平台) |
| McDonald’s | 麦当劳 |
| Meta | Meta(保留原文) |
| Mike Ovitz | 保留原文(投资人) |
| Perplexity | 保留原文(AI 搜索引擎公司) |
| Postmates | 保留原文(美国外卖配送平台) |
| RAG | RAG(检索增强生成) |
| RL environments | RL 环境 |
| Ron Burkle | 保留原文(投资人) |
| Salesforce | 保留原文(产品名) |
| Sand Hill Road | 沙山路(Sand Hill Road) |
| sandbox | 沙盒 |
| Scale AI | Scale AI(保留原文) |
| Scour | 保留原文(Travis Kalanick 早期创业公司) |
| SKU | SKU(库存量单位) |
| Slurpee | Slurpee 冰沙饮料(7-Eleven 品牌产品) |
| Stephen Chau | 保留原文 |
| The Information | 保留原文(科技媒体) |
| The Road Less Traveled | 《少有人走的路》(M. Scott Peck 著作) |
| The Selfish Gene | 《自私的基因》(Richard Dawkins 著作) |
| Thinking, Fast and Slow | 《思考,快与慢》(Daniel Kahneman 著作) |
| Travis Kalanick | 保留原文(Uber 及 Scour 联合创始人) |
| Uber Eats | Uber Eats(保留原文) |
| unit economics | 单位经济模型 |
| VC | 风险投资(VC) |
| VO3 | 保留原文(AI 视频生成工具) |
| Walmart | 保留原文(美国连锁零售巨头) |
| Workhorse | 保留原文(公司名) |
| X | X(社交媒体平台) |
Reformatted by reformat_english.py
Scale AI CEO 谈 Meta 的 140 亿美元交易、将 Uber Eats 做到 800 亿美元,以及前沿实验室接下来在构建什么
文字稿
AI 是否兑现了承诺
Lenny Rachitsky: 最近关于 AI 未能兑现我们所听到的承诺的讨论很多,尤其是在企业层面。
Jason Droege: 这些事情需要 6 到 12 个月才能真正达到足够的稳健程度,使得重要流程可以被自动化。就像任何一次重大科技革命一样,头条新闻讲的是一个故事,但在实际落地层面,铺设宽带意味着你需要挖开美国的每一条道路来铺设线路。总得有人去挖路,总得有人去铺设海底电缆。
Lenny Rachitsky: 关于未来两三年 AI 模型会发展到什么程度,你觉得有什么是人们还没有真正理解或意识到的?
Jason Droege: 目前的大趋势是从模型”知道事情”转向模型”做事情”。下一个问题就变成了:它能为我做什么?智能体(agent)如何替你做决策?
Scale AI 与数据标注
Lenny Rachitsky: 我们来聊聊 Scale 以及你所处的整个 AI 世界。你本质上开创了数据标注、交易数据、为实验室创建评测(evals)的先河。
Jason Droege: 18 个月前,你可能拿到一篇短篇小说,任务是判断”这篇短篇小说是否比那篇更好”。而现在,一个任务可能是由世界上最优秀的网页开发者之一构建一个完整的网站,或者是向模型解释某个非常微妙的癌症话题。这些任务现在需要耗费数小时,而且需要博士和专业人士来完成。
创业的标准
Lenny Rachitsky: 我和你多年来共事过的一些人聊过,我听到很多关于你为新业务设定的标准有多高。
Jason Droege: 从创业的角度来看,真正关键的是:我有什么洞察?为什么我如此幸运拥有这个洞察?在一个有一百万聪明且在不断尝试的创业者在思考的世界里,为什么我处于这样一个位置——我可能拥有其他人没有的洞察?
关于 Jason Droege
Lenny Rachitsky: 今天的嘉宾是 Jason Droege。Jason 是 Scale AI 的新任 CEO。这是他在 Meta 交易后接替 Alex Wang 以来首次接受采访。Alex 现在领导 Meta 的超级智能团队。在加入 Scale 之前,Jason 与 Travis Kalanick 共同创立了一家公司。在此之前,他参与了 Uber 的起步,还在几家创业公司工作过。最著名的经历是,Jason 创立并领导了 Uber Eats,从他和团队的一个想法发展到现在一个数十亿美元营收规模的业务,而且在新冠疫情期间,当没人打车时,这项业务基本上拯救了 Uber。
这次访谈延续了我一直以来的一系列访谈主题,即 AI 模型实际上是如何变得更聪明的。除了扩展计算规模和改进模型代码本身之外,我们在 ChatGPT、Claude 以及每一个前沿 AI 模型上看到的许多改进,都来自这些实验室聘请专家来填补知识的空白、纠正对事物运作方式的理解,并基本上在每个消费者使用模型的领域向模型展示什么样的表现才算好。
Scale 是这个领域的先驱,他们开创了这个品类。在我们的对话中,我们聊到了 Scale 正在发生的事情、与 Meta 的交易是如何达成的、像医生和软件工程师这样的专家具体在做什么来帮助模型变得更聪明、从 Scale 进入市场到现在,数据标注、评测和数据训练的整个市场发生了怎样的变化,以及我们还需要人类继续帮助 AI 变聪明多久。我们还谈到 Jason 认为未来几年模型会朝什么方向发展,因为 Scale 对未来有着非常独特的视角。我们还聊到了 Jason 职业生涯中大量独特而重要的产品经验,包括很多关于如何在现有公司内外开展新业务的建议,以及关于招聘和领导力等方面的诸多内容。感谢 Allen Penn 和 Stephen Chau 为这次对话建议了话题。
从 Scour 说起
Lenny Rachitsky: Jason,非常感谢你来参加节目,欢迎来到播客。
Jason Droege: 谢谢邀请,很高兴来到这里。
Lenny Rachitsky: 我在研究你的背景、为这次播客做准备的时候,了解到了一个关于你的很有意思的事情,我觉得很多人不知道。Travis Kalanick 在创办 Uber 之前还有一家创业公司,叫 Scour,是一个点对点文件共享应用,后来好像被关停了。你是他的联合创始人。这是你职业生涯早期的事情。我猜这段经历里有讲不完的故事,所以我只想问你一个问题:从那段经历中,有什么教训一直伴随着你,带到了你后来工作和做产品的地方?
Jason Droege: 说实话,有太多的教训了。我试着挑一个。我觉得最重要的教训是,在商业和创业中,一切都是可以谈判的。我觉得这是最核心的一点。因为当时我们才十九、二十岁,在宿舍里建了一个搜索引擎,就在宿舍里运营,我们最初的网址是 scour.cs.ucla.edu。这些东西在当时未必是常规操作,但我们只是很务实。基本上就是我们自己发起的一个项目,我们搭了搜索引擎,人们开始用,我们本以为会惹上麻烦,结果计算机科学系对此非常兴奋,尽管我们基本上是在他们的服务器上挂了一个域名,还在宿舍里用自己的电脑来提供这个网站和产品服务。
后来进入融资阶段时,融资过程非常令人大开眼界,“一切都是可以谈判的”这个教训就是从那时候来的。最初的投资人是 Ron Burkle 和 Mike Ovitz。我们在洛杉矶,在 UCLA,所以并没有真正融入整个 Sand Hill Road 的圈子。在我们做交易的过程中,条款不断在变。我们以为融资就是——“我们拿到几百万美元,估值五百万美元。” 那个时候这还算是一个正经的 A 轮估值。然后在交易过程中,变成”我们在做这个交易,我们不做这个交易。哦,你们应该给我们 50% 的公司股份。哦,你们应该给我们 75%。哦,如果你想今天签文件的话,这个人会来吃早餐,如果你今天不签、不给我们 80% 的股份,这个人就不会出现。”
这完全是疯狂的,我们从第一天起就看到了商业世界中可能发生的种种事情。我们原以为做事有一种固定的方式,但在很年轻的时候我们就意识到,做事没有固定的方式,只有你能通过谈判在这个世界中争取到的方式。我实际上认为这对 Travis 影响很大,后来对我在 Uber 也影响很大——如果你能想象到一件事,它合理,你能对齐各方激励,那它就可以发生。但没有什么”标准方式”。在十九、二十岁就学到这一点,我觉得影响非常深远。
Lenny Rachitsky: 这是一个非常棒的教训。Scour 后来怎么样了?好像被关停了?当时发生了什么?
Jason Droege: 对,基本上 Scour 是一个多媒体搜索引擎,后来变成了点对点文件共享网络。但人们用它来寻找免费内容。当时相关的法律还比较模糊,因为混音带是合法的,但这相当于一个超级加强版的混音带。最终我们被告索赔四分之一万亿美元。所以我想,如果要经历可能毁掉人生的事情,大概在二十一、二十二岁的时候经历是最好的时机。但这确实像一盆冷水,让我们看到了现实世界真正是如何运转的。MBAA 和 RAA 起诉了我们,娱乐行业起诉了我们,准确说是代表娱乐行业的协会起诉了我们。然后他们以一百万美元和解了。
我们就想,“等等,你们要四分之一万亿美元,然后以一百万美元和解?” 当然,他们只是想把我们逼到破产,把我们赶出市场,而且这些都是老牌公司。所以我们想,“如果这些人连可遵循的剧本都没有,只是随口编数字,那我们该怎么规划接下来的人生?“
Scale AI 与 AI 数据行业
Lenny Rachitsky: 我们来聊聊 Scale 和你所在的整个 AI 领域吧。这是你接任 Scale CEO 以来第一次接受采访,我很荣幸你能来这里聊这些。这也是你自 Meta 交易以来第一次接受采访,那笔交易非常复杂,让很多人感到困惑。所以我很好奇想听听 Scale 的现状,大家应该了解些什么。比如,你们和 Meta 的关系是什么?你和 Alex 的关系是什么?Scale 现在的状态是怎样的?
Jason Droege: 好,Scale 是一家完全独立的公司。这笔交易是 Meta 投资了略超过 140 亿美元,获得公司 49% 的非投票股份,没有新增董事会席位。Alex 填补了那个董事会席位。所以董事会没变,治理结构基本没变。Meta 对任何东西都没有优先获取权,没有优先合作关系。我的意思是,我们和 Meta 在数据业务方面有长期的合作关系,甚至在一些商务拓展相关的事情上也有合作,比如一起在政府项目上开展工作等等。这些合作可能会因为我们现在关系更近而扩大,但没有任何东西阻止我们和其他方合作,他们也没有获得以前不曾有过的任何东西的访问权。所有的隐私保护仍然有效,所有的数据安全仍然有效,跟以前一样。
事实上,交易中只有大约 15 人转去了 Meta。Scale 现在约有 1100 名员工,我们有两大业务。每项业务都有数亿美元的收入。所以我们现在公司内部有两个独角兽级别的业务。自交易发生以来,业务每个月都在增长——我看过一些报道,说法不太一致。我们之前没有公开谈过,所以这也是我们开始对外传递信息的一部分。我们很高兴继续建设、交付数据,做我们一直在做的事。
Lenny Rachitsky: 所以公司现在是一家独立的、自己的公司。Alex,明确一下,他现在在 Meta 工作,不在 Scale 了。
Jason Droege: 对,没错。抱歉,我应该多谈谈这一点。
Lenny Rachitsky: 我觉得这很有意思。所以基本上,这是一笔投资。一些人离开加入了 Meta,公司继续运转,你在掌舵。我们来聊聊你们基本上开创的这个领域——我不知道最好的叫法是什么——数据标注、训练数据、为实验室创建评测(evals)。你们在任何人知道这是回事之前就在做了。我知道甚至 Scale 也是从其他方向转型进入这个市场的,好像之前尝试了很多自动驾驶汽车之类的东西,然后发现,“哦,AI 实验室需要这些数据。”
我一直在听到的一个主要说法是——我的播客上也有过这个领域的一批 CEO——就是行业发生了很大转变:从 Scale 曾经开创并长期从事的模式,也就是用通用的、低成本劳动力进行训练,转向了现在实验室主要需要专家——律师、医生、工程师来做训练、写评测(evals)之类的工作。我很好奇你看到了什么,这对你们有什么影响,你觉得事情在朝什么方向发展,大家应该了解关于训练数据这个市场的什么信息。
数据标注行业的演变
Jason Droege: 完全同意。我认为目前竞争对手在市场上的定位说法完全站不住脚。我就先从这一点说起,然后稍微解释一下我的意思。但我觉得先花 30 秒钟回顾一下 Scale 的历史以及追溯到 2016 年的主线很重要。Alex 在很早期就有了一个洞察:对模型来说,最重要的东西是数据。他当时大概也就 19 或 20 岁。于是他想:“好吧,那我要围绕这个创建什么业务呢?“他创建的业务是:做自动驾驶汽车的标注,因为如果你给它们的数据做标注,汽车的表现就会更好。然后那波浪潮演变成了计算机视觉的浪潮,我们与美国国防部有合作关系,为他们做标注,那是在 2020 年。
再往后,模型在这段时间里不断变好。随着模型变得更好,它们需要不同类型的数据。所以我们一直在不断适应模型成功所需要的数据类型。然后,生成式 AI 的浪潮来了,这一波简直一飞冲天。作为这个过程的一部分,这个行业本身也在不断变化。所以,两三年前模型刚出来的时候,确实——我们都记得用它们的体验——它们经常产生幻觉,基本的问题都答错,不知道这首诗和那首诗哪首更好。这就是几年前标注工作的状态。情况变化很快,我们也随之而变。现在对所有人来说——而我们一直走在这所有变化的最前沿——就是专家数据标注,更复杂的任务。
给你一个概念,18 个月前的任务是什么样的。我来这里大概 13 个月了,所以我在面试的时候亲眼见过。你会拿到一篇短篇小说,上面写着”这篇短篇小说是不是比那篇更好?“然后你编辑一下,说”嗯,如果改成这样会更好”,再给一个偏好排序。18 个月前就是这种相当基础的工作,虽然已经有一些专家参与了,但当时模型差得远,它们连最基本的东西都需要。而现在,你到了这样一个阶段:一个任务可能是让世界上最顶尖的网页开发者之一搭建一个完整的网站,或者向模型解释某个非常微妙的癌症话题。这些任务现在需要耗费数小时,需要博士和专业人士来完成。
给你一个数据来佐证:我们专家网络上有 80% 的人拥有学士或更高学位,这与市场上一些定位说法和人们对这个行业的理解截然相反。大约 15% 拥有博士及以上学位,我们的网络上有博士通过做标注赚取可观的收入,把他们的专业知识贡献给这些模型。所以自从模型需要专家以来,我们一直就在做专家数据标注。这个游戏的核心就是与研究人员保持联系,了解他们需要什么,在内部提出想法。在某些方面,是我们推动了这一转变,因为我们看到模型在更专业的领域存在不足。于是我们会去找模型构建者,说:“嘿,我们注意到这是个问题。如果你想修复它,这批专家可以帮你。“所以那些反向定位的说法确实存在,但我认为那只不过是竞争对手有时会说的话,与实际情况毫无关系。
Lenny Rachitsky: 好的,这非常有趣。所以我的理解是,确实发生了很大转变——实验室需要更多专家参与训练、标注、写评测(evals)。你们对此非常清楚,并且一直在随之演变。市场上有一个——我不知道该不该叫指控——说法是这些专家很难找到。所有这些公司都有自己的专有专家网络,各有各的寻找方式。关于你们是如何做到的,你能分享些什么吗?因为感觉最难的部分就是找到这些专家,并防止他们被其他公司挖走。
如何找到并留住专家
Jason Droege: 他们确实很难找。你必须采取很多很多策略。正如你所预期的,我们获得他们的方式不是单一的。最主要的方式是他们互相推荐,因为当你享受你正在做的事情,并且你正在用自己的专业知识为 AI 做贡献——这其实挺酷的。如果你是某个相当特定领域的博士,你在使用某个模型时会感到沮丧——它不能按照你想要的方式与你交互——这是一种有偿的途径来解决这个问题,而且还能赚到几百或几千美元。所以他们经常会互相推荐。
我们也有校园项目,会直接走进校园,和教授交谈,和学生交谈,询问谁愿意做这类工作。当然还有更传统的规模化方式,比如 LinkedIn 等平台。但最优质的来源还是这些草根和推荐网络。而获得这些的唯一方式就是为这些人提供良好的体验,因为他们做这件事部分是为了钱,但也因为他们认为自己对 AI 模型的贡献是重要的、有趣的,而且很多时候这也解决了他们自己的问题。
强化学习与 RL 环境
Lenny Rachitsky: 就在本周,我在 Twitter 上看到一些东西,正好是我在准备这次访谈的时候——有一篇 The Information 的报道标题。这和 Workhorse 的 Brendan 说的观点很相似:随着时间的推移,整个经济将转向强化学习,所有人都在训练 AI,基本上这就是将来剩下的工作。你对这个怎么看?你觉得事情会朝那个方向发展吗?有没有不同的视角?
Jason Droege: 强化学习非常重要,我认为这实际上是关于向环境迁移的一个更广泛的评论。有一种叫做 RL 环境(RL environments)的东西,本质上是供 AI 智能体(agent)在其中玩耍以完成目标的沙盒,这样它们就能学习如何完成那个目标。我们已经做了一年多了。举个例子,你有一个 Salesforce 实例。AI 智能体(agent)如何在这个实例中导航?这个实例中有它需要识别的数据,有配置。Salesforce 是一个高度可配置的产品,它有各种配置,智能体(agent)需要理解如何导航。你要求智能体(agent)执行一个需要非常高可靠性的业务流程,然后智能体(agent)还需要知道:“嘿,如果我无法完成我想要完成的事情,或者我认为我即将完成的事情准确率很低,我该怎么把它交给人类来获取反馈和指导?”
所有这些都需要训练,而且其中没有什么炼金术。你只需要把 AI 智能体(agent)放到一个代表人类所做工作的环境中。你可以想象世界上环境的数量,以及每个环境中目标的数量,是巨大的。所以问题是——我们在过去一年里所做的研究,为了成为我们模型构建者客户的好伙伴——就是每一个单独的任务或每一个单独的环境的可泛化性有多高。如果你想象一下软件系统的环境世界——配置、数据类型、规模、用户数量、复杂度——排列组合是无穷无尽的。所以你需要的是一种策略,让实验室能够收集到足够泛化的数据,覆盖足够广泛的使用场景,这样他们就不必为了每种特定情况下去收集 45 万亿种”智能体(agent)在这种情况下应该怎么做”的组合。
数据的可泛化性
Jason Droege: 所以有时候工作和数据是高度可泛化的,所谓可泛化,我的意思是用简单的方式完成任务。比如任务是”找到我日历上和 Lenny 面试的那个会议”,然后智能体(agent)就去翻遍我的整个日历,把它找出来,一个非常简单的例子。这种能力需要能泛化到任何日历搜索,甚至任何日历操作。可泛化性越高,数据就越有价值。所以我们的工作就是为模型构建者提供最有价值的数据,以实现让智能体(agent)对终端用户尽可能有用这一目标。
Lenny Rachitsky: 我很喜欢你一直在分享这些具体的例子,让大家知道这些人到底在做什么,以及你们提供给实验室的数据是什么样子的。回顾一下你分享过的几个例子,一个是工程师搭建网站,本质上就是把代码分享给模型,告诉它”我会这样做”。在这个例子中,只是”这是代码”,还是说有一段他们搭建过程的录制?数据到底是什么形式的?
Jason Droege: 两种都有可能。在某些情况下,就是一个网站,这是一个示例,然后他们进行设计。在另一些情况下,需要进行标注,比如”我出于这个原因做了这个决定”,或者”我出于那个原因做了那个决定”,或者”我是这样思考这个问题的”。所以这取决于模型构建者想要实现什么目标。在训练目标方面,情况可能会变得相当细致。
Lenny Rachitsky: 明白了。
Jason Droege: 所以不是说”这是一个网站”,然后就学会了创建网站。而是”这是一个网站,这是我做这个决定的原因,这是我没有做那个决定的原因”,或者”这是一个出问题的网站,这是它出问题的原因”——如果他们想训练的是一个网站构建器的调试工具之类的。
Lenny Rachitsky: 你分享的另一个例子是一个短篇小说——这是一篇短故事,那是另一篇(我猜是模型生成的),然后判断哪篇更好,以及如何改进它。你刚才分享的另一个例子是 Salesforce 智能体(agent),比如”嘿,和一个潜在客户约个会议”,然后教它这个流程怎么完成。我很喜欢这些具体的例子,因为它让人一看就明白了——好的,我懂了,这些公司做的就是这些事。还有没有一两个例子,让大家进一步感受一下这些数据长什么样?
企业内部的标注需求
Jason Droege: 当然可以。我可以给你一个例子——我们的业务其实分为两面。一面是向模型构建者供应数据,我们把数据卖给他们;另一面是我们也做解决方案,向医疗系统、保险系统等销售应用和服务。我觉得如果给你一个后者的例子,画面会更丰富,因为它涉及数据,但更涉及数据的使用——为了一个非常具体的目标对数据进行操控和处理。举个例子,我们与一个医疗系统合作。医疗系统有很多问题,这个特定的医疗系统有专家定期诊治非常罕见的病例。你去那里,前提是其他人都无法诊断你的问题,所以积压量非常大。因此,这种落地实施层面也有一个生产力要素。积压量很大,他们希望能看更多患者,提供更好的护理,并减少复诊次数——因为他们想在第一天就给出准确诊断和治疗方案。然而,在今天没有 AI 辅助的情况下,医生实际上需要阅读 200 到 300 页的文档,而且这些文档汇成一个文件,但格式各不相同。如果你是一位医生,你怎么可能读得完 200 或 300 页的所有内容?所以他们的做法是尽力而为——扫描一下,让护士看一下,也许再让一位资历较浅的医生看一下这个病例。他们当然想好好治疗患者,这也是他们当医生的原因。然后他们走进诊室,和患者交谈,做出诊断。
我们基本上构建了一个工具,可以替他们阅读那份文档,并指出他们最应该关注的 5 到 10 个事项。比如不明显的过敏反应就是其中一个例子——我们实际上发现了一位患者的某种过敏,这种过敏从文档中并不容易被注意到,而且这种过敏实际上会与他们即将被开具的药物产生冲突。所以这个 AI 工具提取出了这种关联,而这种关联即使是人类也很难发现。要让这个工具越来越好,用现成模型能走的路是有限的,实际上这个医疗系统内部的人员必须自己做标注。所以我们谈到了为模型构建者做标注,但我们现在开始看到标注工作正在向企业和政府迁移,因为仅靠现成模型加上 RAG,再加上基于已有数据的微调,能走到的程度是有限的。人们经常忽略关于这些系统的一点是,我们假设——因为你经常听到这样的数字,比如”哦,某家银行每年就有 200 PB 的数据”或者什么令人难以置信的数字——我们忽略的是:那些是正确的数据吗?其中哪些数据对模型有用?大部分并不有用。有一些有用,但我们在谈论知识工作、谈论做出判断时,很大一部分是人类判断——基于对”这位医生在这种情况下会怎么做”或”这位银行家在这种情况下会怎么做”的综合分析,以及他们在整个企业的背景下会如何决策。而这一点可能因银行而异、因医疗系统而异,因为文化、目标、激励机制等都不同。所以我们现在到了这样一个阶段:我们看到,将判断力——人类的判断力,真正的专业领域深度专业知识——数字化,正在成为一个瓶颈,而我们在帮助客户突破这个瓶颈。
人类标注者的未来
Lenny Rachitsky: 这真的很有意思。这个谱系从低技能通用劳动力,到专家,再到现在某一家公司里需要做这项标注工作的特定专家。
Jason Droege: 没错。理解这一点——现在有两大叙事。一种是 AGI 叙事,一切都是 AGI,一切都将成为 AGI;另一种是怀疑论者,“嘿,这都是胡扯,这是一个泡沫”等等。当然,我的观点是大多数事情就像很多情况一样——真相在两者之间,两端的极端部分可能各有道理,但现实是,在智能体(agent)系统中——智能体(agent)与智能体(agent)对话的系统——要让机器关键用例达到完成目标所需的准确度,是非常困难的。其中一个主要问题是,即使只是理解一份文档,同一个文档,在 A 公司和B 公司,完全相同的文字会有不同的含义和重要性。你怎么让一个系统知道这一点?所以这一切都需要构建起来,才能做出好的决策。
Lenny Rachitsky: 这个问题衔接得很好,人们看你们这样的公司以及这个领域的其他参与者时,心里总有一个疑问:我们还需要人类做这件事多久?AI 什么时候才能聪明到自己来做?我知道你的利益所在——你会说我们永远不会缺人,因为这和你们的增长一致——但我们该怎么思考这个问题?为什么 10 年后我们还需要人?这些专家告诉 AI 它不知道的东西,我们还需要多久?
Jason Droege: 首先,数据标注的历史就是一部不断重新开始的历史。自动驾驶现在需要的数据标注已经不像过去那么多了。我的意思是,Scale AI 是一家相信数据将永远重要的公司——除非有一天你完全不需要把外部的人类数据放进模型里。我认为我们已经达到了一种几乎难以想象的进步水平,因为你实际上是在说,没有任何新的人类技能和人类知识重要到需要放进这些模型中。这感觉相当遥远。所以对于我们这样的业务,我们一直在思考如何构建一种运营体系,能够不断发现新的需求,然后与我们称为”专家贡献者”的贡献者网络合作,挖掘那些数据,挖掘那些信息。有时候是新人,有时候是在我们现有的人才库中发现某些人拥有我们之前不知道的专业技能,也许一年前对模型还没有用,但现在有用了。
所以这是一个不断将越来越多数据注入模型的过程。是的,我们在经济上有动力相信人类将始终参与其中,但这不仅仅是一种商业信念,也是我个人的信念。这些系统需要为我们服务,如果这些系统为我们服务,那么我们需要对这些系统做出的任何决策保持参与或监督。至于劳动力这个更广泛的话题——我觉得它经常和”白领末日”之类的说法一起出现——我在这方面可能更偏向务实派,也许是因为我的性格,也许是因为我在一线看到那些据说将在未来一两年内经历这场变革的客户实际上正在发生什么。我只是觉得这件事可能会发生。这个领域发展极其迅速,但我认为不会那么快。
肯定不会在明年发生。认为两年内就会发生的想法我觉得非常牵强,但这里没有什么是不可能的。长远来看,如果你回顾那些——我不知道——悲观主义档案馆之类的,那些发布”收音机被发明了,然后一切都会被淘汰”这类内容的账号。变化是会有的,但人类非常擅长适应变化。所以我认为,在所有那些悲观论调中,我们低估了人类的适应能力。我们作为一家公司具有高度的适应力,而技术的历史也表明人类是具有适应力的。
评测(evals)的作用
Lenny Rachitsky: 我很喜欢这个观点。我也是个乐观主义者,所以总在寻找乐观的理由。在继续那个话题之前,我想问一个非常具体的问题——评测(evals)这个词最近经常出现,尤其是在你们这个领域的公司中。我还在了解这一切到底是什么,特别是在你们的市场中。你或者专家们提供的东西中,评测(evals)和其他类型的数据各占多大比例?
Jason Droege: 很大一部分是评测(evals)。在企业客户和政府客户中,几乎全部都是评测(evals),因为必须有人建立”什么是好的”这个基准。这是理解评测(evals)最简单的方式——“好的”是什么样的,你是否有一套全面的评测(evals),让系统知道什么是好的。就这么简单。
Lenny Rachitsky: 那么在你之前分享的医疗案例中,基本上就是这位医生坐在那里看所有这些报告,创建评测(evals),比如”这份报告、这份病历中应该发现这些内容”。可以这样理解吗?
Jason Droege: 对,这是非常重要的一部分,就是”好的”是什么样的。
Lenny Rachitsky: 好的,明白了。
Jason Droege: 我必须把事情简化到最简单的表述。
Lenny Rachitsky: 有意思的是你用了”好的”而不是”正确的”。这是你特意喜欢用的词吗——“好的”,而不是”这是正确答案”?
Jason Droege: 我不是刻意用这个词的,但这些是概率系统,所以取决于……是的,我可以在这里展开谈谈 AI 擅长解决哪些类型的问题。如果你有一个准确率只有 10% 或 20% 的人类流程,或者说只有 10% 或 20% 的满意度,那 AI 就非常厉害。因为如果你能做到 50%、60%、70%、80% 的准确率,你就稳赚不赔了,一片大好,皆大欢喜。当然,系统还需要知道,对于剩余的部分,如何确保有人类参与剩下的决策。但从净增价值的角度来看,人类在这个场景中是非常满意的。
如果你有一个准确率已经达到 98% 的人类流程和工 作流,而你期望 AI 系统帮你补上剩下的 2%——目前还没完全做到。所以当我说”好的”是什么样的,很多流程和人们要求这些系统做的事情、以及要求我们构建的系统,都是在替人做判断。所以就像我们会问一个人”嘿,你觉得在这种场景下我们应该怎么做?“一样,你要找的是在当前信息条件下最好的建议或行动方案。
AI 模型的进步与基础设施
Lenny Rachitsky: 对你来说这很明显,对你这个市场中的人来说也很明显。我觉得很多人想到 AI 训练时,想的就是给它一堆数据,看看吧,从人类历史和所有文字记录中学习你能学到的一切。但真正令人惊讶的是,本质上就是一群人坐在那里教 AI 它不知道的东西,填补空白。AI 现在就是这样变得更聪明的。它已经没有更多真实数据可以消化了。就是——“这是我不知道的”,或者”这是专家发现你弄错的地方,我来教你”。而这种方式可以规模化,并且让模型持续进步,这真的很令人震撼。
Jason Droege: 是的,没错,我同意。就像任何一次重大的技术革命一样,头条新闻讲的是一个故事,而实际上在地面层面,铺设宽带意味着你需要挖开美国每一条公路来铺设线路。就是这么回事。总得有人去挖路,或者有人去铺设海底光缆。所有这些行业中总有一些运营层面上的凿石般的苦活在进行。想想这些模型有多神奇——如果你在科技行业待得够久,即使到今天它们能持续正确使用标点符号都让我觉得不可思议。在现在这个市场阶段说这话听起来可能有点蠢,但如果你回到三年前从技术角度想这个问题,很多我们现在觉得微不足道的事情其实非常复杂。真正的答案是,这是算力、模型改进和数据三者的结合,而且三者同时在变得更好。
未来两三年的展望
Lenny Rachitsky: 我们顺着这个话题继续。你在 Scale AI 已经很长时间了,担任 CEO——你说了——13 个月。我觉得你能看到更多关于事情走向的信号,因为你和实验室在他们还没发布的东西上合作。你看到的比大多数人都多,我知道你能分享的有限,但关于 AI 模型未来两三年会发展到什么程度,你有没有觉得人们还没有真正理解或把握的东西?
Jason Droege: 你看,现在各种讨论太多了。我觉得这取决于你刷多少 X 或者看多少新闻。所以我就说说我们的视角。当前的大趋势是从模型”知道事情”走向模型”做事情”。我们确实在知识边界上不断推进——我们发布的评测以及其他机构发布的评测都表明,这些模型的知识储备已经相当扎实了。接下来问题就变成了:它能为我做什么?一旦进入这个世界,我们之前谈到的那些环境就开始发挥作用了。智能体(agent)如何操作一个 Salesforce 实例?如何操作一个医疗系统?甚至如何操作你手机上的天气应用?它如何替你做决策?
我们才刚刚进入这个阶段的起点。这个过程推进的速度会非常值得关注。我认为这也是为什么各种预测差异如此之大的原因——因为我们还处于起步阶段。人们对其进步速度的预估各不相同。如果你取最激进的路径,也就是说,在这些方面训练智能体实际上会相当容易,那么接下来就是经济体系中的变革管理问题了——顺便说一句,变革管理绝不可以低估。世界上至今还有人没有电子邮件地址。所以采纳曲线最终变成一个人和政策问题,而非技术问题。从技术角度来看我们还没到那一步,但我确实认为在未来两三年内,如果我接下这个话茬非要做个预测的话——技术会发展到迫使变革管理者和政策制定者去思考:“我们该怎么应对这个东西,它已经离我们很近了。“这大概还有两三年的时间。
AI 是否被过度炒作
Lenny Rachitsky: 最近有很多关于 AI 未能兑现承诺的讨论,尤其是在企业层面。MIT 刚刚发布了一项研究,显示有大量人们兴奋启动的试点项目最终都没能落地,企业并没有真正采用这些工具。有数据显示工程师使用这些工具后生产力并没有实际提升,有时反而会拖慢速度。你和大量公司合作实施各种 AI 项目。你在一线看到了什么?你看到的是什么样的收益?你觉得 AI 是被高估了,还是被低估了?
Jason Droege: 市面上确实有很多炒作。我们的工作实际上是打造能用的产品,为客户交付价值,搞清楚到底哪里真正能落地。要实现一个复杂的工作流——我举的医疗场景是一个例子,我们还做其他复杂的工作流,比如保险公司的理赔管理。这是一个涉及财务决策的流程,但也是可以自动化的。但基本上情况是这样的:概念验证(POC)能做到 60% 到 70%,然后人脑就会觉得,剩下的没什么大不了的。但这就像数据中心的可用性一样——每多一个九,在可靠性、备份等方面就是数量级的投入。一个九基本上就是我们在 UCLA 时宿舍里的那台 Web 服务器,而五个九则是一个极其疯狂的高标准,但从数字上看似乎只是很小的进步。
这里出现了类似的动态。概念验证之所以失败,原因之一——首先存在一个分母效应,因为做这件事太容易了——“嘿,我启动了一个项目,我又启动了一个项目,我又启动了一个项目。“所以人们很容易去尝试。所以我不一定认为那个 95% 的数字——我觉得它在某种程度上有点标题党。它讲述的方向是对的,但确实有些夸张。因为如果你看那些真正在内部推进的项目,他们找了像我们这样真正有实力的合作伙伴,或者自己做——如果他们有和模型合作过的工程师,而且他们投入了时间,我说的可是几个月,不是你在那些视频里看到的几分钟——去真正获得法律审批、政策审批、监管审批、变革管理,达到所有人都认可的准确度。如果你真的这么做,这些事情需要 6 到 12 个月才能足够健壮,才能真正自动化一个重要流程。
所以我觉得炒作的这一点是对的——当你真正做成了,影响力确实是”哇,我自己根本不可能做到这个”,比如以我作为世界上最受过教育的医生之一为例。但到达那个阶段的时间比人们兜售的要长得多。
Lenny Rachitsky: 这话说得太好了。不仅仅是这些东西容易尝试,而是每个人都在做,所以大家都有 FOMO 心态——“我也得试试这些,我得试试所有这些原型工具,Cursor 什么的。“然后就一头扎进去,结果实际并不管用。
Jason Droege: 易学难精。这是我的总结。
回归产品与客户
Lenny Rachitsky: 好的,我们聊聊 AI 之外的话题。关于 AI 的讨论可以无休止地继续下去,但你还有很多经验要分享。你帮助打造了 Uber Eats,之前也有过几家创业公司,我们也聊了一点 Scour。我和这些年跟你共事过的人聊了很多,得到了很多关于你极其擅长的事情的有趣洞察。我就一个一个来。第一个是你对贴近客户、与客户交流的执念。我很喜欢这个话题,因为每个人都觉得自己在这方面做得很好,觉得完全理解这为什么重要。他们都说”我在做这件事,别担心……没在做的是别人,不是我。“聊聊你觉得人们在这方面可能忽略了什么——做得好的时候是什么样的,以及为什么这件事这么重要。
Jason Droege: 我的意思是,我大概也属于你刚才描述的那类人,而这也许正是开始做任何新事情所需要的某种自信。但我不认为这是一个干净利落的过程。我的做法是,在任何事情的初期,我对听到的每一个信息都会持续质疑。我不会字面理解客户说的话。从产品管理的角度,关于这个话题已经有很多讨论了——比如,不要照他们说的做,要做他们真正需要的,去看真正的问题和底层的东西。我觉得我对这个问题的看法,可能对这个讨论有所补充的地方在于:我会去看客户的底层激励机制。客户的底层激励并不总是财务层面的,有时候是自我意识,有时候是职业发展。
举个例子,如果你向企业销售软件,背后会有一个高管发起人,这个人需要相信你能为他做好这件事。你怎么让他跟你一起跳上这个大项目?这就不只是产品本身的旅程了——他们需要从我们这里听到什么?我们需要向他们提供什么?我们需要做什么才能真正 unlock 实施产品的机会?所以我认为这里有一个激励对齐的基准。我很信奉那句老话,虽然听起来是陈词滥调——给我看激励,我就告诉你结果。我觉得这绝对正确。甚至当客户告诉你一些事情的时候,我举个例子。我已经离开这行有一段时间了,所以可以公开说——Uber Eats。
我们推出 Uber Eats 的时候,我从贴近客户的角度来看待这个业务。实际上我们当时请不到餐厅导览,我对这个行业一无所知。在 Uber,我的工作是弄清楚我们应该进入哪些其他业务。所以我们考察了大量的业务方向,Uber Eats 即外卖配送是我们认为最有趣的一个——后来证明我们是对的,算我们走运。
Lenny Rachitsky: 确实非常对。
逆向推算餐厅经济学
Jason Droege: 然后我们找不到餐厅导览来帮我们理解他们的单店经济模型。他们会说,“哦,大概是这个比例或者那个比例吧,你为什么想知道?“然后我们去找另一家餐厅导览,他们会解释一下,但也会有点疑虑——这些 Uber 的人为什么要来问我火腿多少钱?于是我们就从这些地方点了一堆食物,又找了一个餐厅供应商拿到了基础目录,然后我们就去匹配——火腿有多重?芝士有多重?面包有多重?上面有几片生菜?我们试图自己独立构建一套认知:食材成本是多少,人工成本是多少?然后我们把我们得出的”地面真相”与餐厅导览告诉我们的、以及网站方面告诉我们的关于餐厅经济学的信息进行三角交叉验证。
如果这些东西都重叠在一起,我们就说,好,我们有了关于这里该怎么做的洞察。这跟 Uber Eats 有什么关系?我们发现的是,一家餐厅大约把每顿饭的 20% 到 30% 用于食材,大约 20% 或 30% 用于人工,大约 10% 用于房租,等等其他开销,顺着链条往下。但关键的是——增量需求的价值是多少?
于是我们进去说,“我们要收取账单的 30%。“他们就说,“天哪,这又来一个团购吗?这太高了。“我们给他们解释了经济模型,他们说,“好吧,我们试试,但这太高了。“他们说得对,真正的数字、真正的清算价格不是 25%,但我们也没差太远。所以当你去找产品市场契合(product-market fit)或者贴近客户的时候,它是一个组合——最有价值的东西是什么?对于餐厅导览来说,给我增量需求。因为如果你能让一个餐厅位置在相同人工的基础上把需求增加到三倍,只是按比例增加食材,那你就拥有了一个 70%、80% 增量毛利率的产品。
餐厅导览很讨厌我们这么说,因为现实中不会完全是这样。但正是因为我们有了这个洞察,我们才有信心这样推向市场——我们需要收你这么多,这样配送费才能是那个水平。然后如果配送费是那个水平、我们收你这么多,我们觉得消费者就会接受,而这就是你获得增量需求所需要的,然后我们就能付给司机这么多。于是你把整个拼图拼在一起——在交易平台的情况下,你不会百分之百满足任何一个参与者百分之百的需求。你满足的是让他们参与市场的一个清算价格。这就是一个例子。
Lenny Rachitsky: 我特别喜欢这个例子——你几乎是在帮他们弄清楚他们自己都还没有完全意识到的东西。你站在他们的角度替他们思考他们的目标,拆解经济模型,然后给出解决方案,而不是问”嘿,你们需要我们做什么?“
买家视角的紧迫性
Jason Droege: 对。如果你走进一家餐厅,他们会告诉你很多东西。他们会说排班是个问题,会说房租是个问题,会说食材价格太高了,占了 20% 到 30%,如果能砍掉 3% 就太好了。你可能就拿着这个去做,“我要建一个帮你省 10% 食材成本的生意。”
但这并没有真正走进他们的脑子——日常最重要的事情是什么。这在年度层面可能很重要,但日常层面呢?他们在看数字,看有没有人来,昨天赚钱了吗,明天能赚钱吗?所以紧迫性——我觉得人们在打造新产品时最容易忽略的,就是买家视角的紧迫性。你可以做出一个提供很多价值的东西,但如果它不是客户在忙碌日常中最挂念的事,那你就是一条漫长的路通向一个小镇。
Lenny Rachitsky: 这正好触及了我从很多人那里听到的一个主题——独立思考,以及你多么重视这一点。这感觉是一个很好的例子。关于这种思维方式为什么如此关键,还有其他方面可以分享的吗?
独立洞察与 alpha
Jason Droege: 我认为创始人的工作——我把这个词用得宽松一点,因为在 Uber 我们享受了 Uber 的所有资源,所以我并不算是真正的创始人,我只是在内部启动了那个业务。但创始确实有一些要素——你要在市场中寻找 alpha。我们 97 年创办第一家公司的时候,创业这件事还没那么酷。在硅谷可能还行,但在洛杉矶绝对不酷。现在创业变得超级酷了,所以所有人都在尝试一切。那你怎么在这个市场上获得 alpha?如果你的研究很大程度上受周围世界言论的影响,你就不可能拥有独立的洞察。你必须走出去,做自己的事。
Jason Droege: 正因如此,从创业的角度来看,我对创办公司应该采取什么方法有非常强烈的看法,这可能非常个人化。但核心确实在于——我有什么洞察?因为我凭什么这么幸运拥有这个洞察?在一个拥有一百万思考着、聪明着、尝试着一切的创业者的世界里,为什么我处于这样一个位置——很可能拥有别人没有的洞察?然后,为什么我是去做这件事的人?
答案可能是,我身处一个狭窄而偏远的领域。另一个答案可能是,我骨子里就是一个逆向思维型人格,所以我总是在寻找那些为真但人们不相信其为真的事情,这有时候确实管用。但接下来的第二部分同样极其重要——我为什么要在这个问题上投入 5 到 10 年?人们在这方面经常犯错。他们跑去跟客户聊,然后说:“他们有痛点,我去解决它。” 这并不是一个很好的创业方式。你真的必须有一种不断自我质疑的强烈驱动力。
关于独立思考的另一件事是,你不能爱上自己的想法。我并不自诩为世界上最伟大的思考者——这话是你们听别人说的——但这其中的一部分,本质上就是放下你是谁、你过去是谁、你所有既有的想法,为了你所承担的使命——即为客户达成某件事。
创业标准要多高
Lenny Rachitsky: 这太好了,我很高兴你谈到了这里。这正好触及了我经常听到的关于你的另一个主题——你对新业务的标准设得有多高。我觉得这个建议对创始人很有用,正如你所说,对公司内部启动新业务线的人也同样适用。你刚才已经谈到了一些,但关于启动新事物时这个标准到底需要多高才能让事情大概率成功,还有什么可以补充的吗?
Jason Droege: 这样说吧,如果你想给自己最好的机会——事情并不总是这样运作的——但如果你像我一样,职业生涯超过 25 年,想给自己最好的机会,我认为公司最终能成功的方式基本上有两条。第一条,坦率地说,可能是最重要的——就是创始人自身在很长一段时间内就是一股不可阻挡的力量。因为你必须能够转型,你必须有那个能量去转型。你可能要年复一年地承受艰难,这大概是最重要的事情。
但第二重要的是,你可以很容易地自学——什么是好的商业模式,什么是坏的商业模式,什么是好市场,什么是坏市场。即使你是那股不可阻挡的力量,如果你要带着全部精力冲进一个糟糕的市场,你至少应该知道。也许无知是福,你一股脑投入其中,最终时间会解决一切。但那不是我的行事方式。我的方式是:交易平台是好生意。SaaS,至少在历史上是这样的——我们后面再看这会不会改变——但 SaaS 历史上一直是很好的生意,经常性收入生意,高粘性生意,有网络效应的生意。
如果你看看顶级风险投资(VC)投资什么,是的,他们确实有大量的组合构建,但在他们相信能够价值数百亿美元的商业模式类型上,是有共性的。这些模式有网络效应,有锁定效应,在规模大时比规模小时更有价值。所以如果你对新业务加一层过滤——这也是我在 Uber 做的事——就是如果你有一个筛选机制来审视新业务,淘汰坏想法其实花不了多长时间。然后在剩下的里面,你可以挑选——“我对这个非常有热情”,即使它可能在纸上看起来问题比另一个更多。但你必须对它有热情。不过我觉得人们最基本的问题是,根本不了解什么样的生意才有机会做到千亿美元级别。
Uber Eats 是如何被选中的
Lenny Rachitsky: 所以你启动了 Uber Eats,你判断这是值得押注的方向。作为局外人来看,这感觉很理所当然——外卖当然会大获成功,这想法多好。但我知道你在这个过程中考察了大量想法。你能谈谈你探索了什么、为什么最终选择了 Uber Eats 吗?
Jason Droege: 在搞清楚这些事情方面,我绝对不是房间里最聪明的人。所以我在想法上会尽可能长时间地保持非常非常宽的口径,直到我觉得——好,一切都在汇聚了。我认为有很多理由让你必须对那些一开始看起来不怎么样的想法保持开放的口径,你只需要不断深挖,看看你判断它们不好是对了还是错了。所以作为一个总体的哲学原则,我先从这里说起。我们看了很多,做了些疯狂的事。有一天我在旧金山走路,沿着 Market Street 望去,看到一个 CVS、一个 7-Eleven、又一个 CVS、一个 Walgreens、又一个 7-Eleven,我就想——“这些东西里面到底有多少 SKU 是人们真正想要的?你能不能把这些东西放进一辆货车里,按一下按钮,货车开过来,你就能拿到想要的便利店商品?它们本来就是便利商品,这怎么会有问题?”
我们在华盛顿特区上线了这个项目。我们投放了 10 辆货车,装了 250 个 SKU。怎么说呢,说”门可罗雀”都算是轻描淡写了。我们根本接不到订单。我们意识到,我们其实完全没有研究过便利店到底是什么。真相是——如果你没有香烟,没有啤酒,没有 Slurpee 冰沙饮料,没有这些东西,你就不会把人吸引进来,也就卖不了其他所有的东西。所以我们对零售一无所知,完全外行。这是一个想法。我们还看过杂货配送,但说实话,所有的拣货打包之类的单位经济模型(unit economics)让我非常害怕。我认为 Instacart 在把单位经济模型做到一个好的水平方面做得非常出色,那可能是你能 tackles 的最难的运营难题。
我们还做了通用配送、点对点配送——就是现在 Uber 的那个产品,我忘了叫什么,好像叫 Uber Direct——就是你有东西需要在城市里从 A 点送到 B 点。那从一开始就失败了,因为真相是,消费者并不真的有这个需求,企业倒是有一些这样的需求,而在 2014 年我们做这件事的时候,几乎没人有这个需求。但我们把这些东西尝试了 15 个版本,最终才说——“好吧,外卖这个方向在所有信号上都在爆发,而且我们能跑通单位经济模型。人们看起来想要它。这是一个超级酷的问题,因为我们可以用所有这些工具赋能独立餐厅,让它们能与大连锁竞争。我们可以把地段因素从竞争中剔除——你可以拥有一个非黄金地段的位置,但如果你有顶级品质的食物,你就能够竞争。” 所以我们觉得,“哦,这是一个非常有意思的问题,而且我们真的能帮助地方经济。”
Lenny Rachitsky: 如果我没记错的话,这在 COVID 期间基本上拯救了 Uber。Lyft 没有这样的业务。这个业务现在有多大?关于它对 Uber 最终有多重要,你能分享些什么吗?
Jason Droege: 当然。我们在 2015 年 12 月于多伦多上线,两小时内销售额就达到了两万美元。我们以惊人的速度意识到这是正确的方向,而且单位经济模型是好的。四年半之后——我在 Uber 大约待了六年,但我们花了大约一年半才摸索到这个方向——四年半后,规模达到了大约 200 亿美元。从 0 到 200 亿,用了四年半,相当不错。Uber 非常擅长规模化扩张,但市场竞争激烈。其他玩家也做得不错。我们打败了很多人,也有人打败了我们。到现在,我估计这个数字已经逼近 800 亿了,那是我离开后又过了四年半的成绩。COVID 把它从 200 亿推到了——我恰好就在 COVID 之前离开,纯属巧合——一年之内从 200 亿到 500 亿。我的意思是,网约车是这样走的(手势向下),而外卖直接飞到了冥王星。
Lenny Rachitsky: 运气真好。干得漂亮。
Jason Droege: 运气是游戏的一部分。这是另一件需要意识到的重要事情。运气是游戏的一部分,所以不要因为别人的运气而心生嫉妒。这个行业很难。我们做的所有这些事情都非常非常难。运气就是游戏的一部分。
麦当劳的故事
Lenny Rachitsky: 说到这个,也许也许不算。你的一位同事 Stephen Chau——我在他的新公司有投资——他和你一起在 Uber Eats 工作了很长时间。他让我问你麦当劳的故事。我猜那对你们来说是一个很大的里程碑,一个非常重要的时刻。你们为什么决定把麦当劳放到 Uber Eats 上?据说你们拿下那个合作的故事也很有意思。
Jason Droege: 这很有意思,这也恰好说明有时候无知反而会让你意外地走到正确答案。我们上线了 Uber Eats,而 Uber 拥有全球业务版图,我们是除中国之外唯一拥有全球版图的外卖网络。Uber 内部所有业务都需要全球上线,这是公司文化中非常重要的一部分,诸如此类。这工作量巨大,你可能会把自己铺得太薄,引发其他问题。但在这一点上,它反而成了优势。我的愿景是,好,让我们帮助小商户与这些连锁品牌竞争。它们拥有系统化的餐饮体系,而食物才是一座城市之所以精彩的原因。没有人会谈论自己在巴黎去过的连锁餐厅,人们谈论的都是自己发现的本地小店。我们想成为其中的一部分。这就是我们想成为的样子。
然后,麦当劳主动找到了我们,他们说:“嘿,我们很想和你们做外卖。“我说:“不要。“他们愣了一下:“等等,我们每天有八千万消费者,你不想一起做吗?“我说:“这不太符合我们现在的调性。“于是我把他们推脱了四五个月,直到我的团队受不了了:“你疯了吧。这些人会在背后投入营销资源。他们真的想做这件事,他们想全力投入。”
正因为我们之前的那种态度——这些事情很难说有直接的因果关系——我们最终和他们达成了独家合作关系,获得了海量的客户……当时连锁品牌其实并不怎么出现在外卖平台上,因为所有人都非常担心单位经济模型,因为它们对客单价极其敏感。
我的态度是,嗯,想办法搞定,这也是非常 Uber 文化的一套。好,客单价就 17 美元,那我们的工作就是让它跑得通——缩小配送半径,把经济模型算清楚,也许在某些地方把食物价格标高一点。总有办法能搞定。所以我们做了,三个月后业务又开始以完全不同的量级曲棍球棒式飙升。我的团队就说:“兄弟,你在这一点上太固执了。“但我认为这最终反而是净收益,因为我们和他们谈出了一个非常好的合作条件。
Lenny Rachitsky: 所以你之前把他们晾在一边,反而帮你拿到了更好的合作条件,我是这么理解的。太厉害了。
Jason Droege: 对,我想他指的就是这个故事。然后,上线的过程简直疯狂,因为我们基本上在六个月内就和他们在全球范围内推出了合作,而当时这个业务还不到两岁。你要激活一家我不确定有没有 80 年历史的老公司——人家期望一切流程都已经到位,而我们这边是两个纽约的办公室行政在负责对接。简直一团混乱。
Lenny Rachitsky: 我到现在还是很伤心 In-N-Out 还是没上任何一个外卖平台。
Jason Droege: 对,我也是。
Lenny Rachitsky: 我记得有人搞过破解。大家找到了各种曲线救国的办法,然后他们就说:“不行不行。好吧,你是 Postmates,我们知道,我们不会给你提供食物的。”
Jason Droege: 是的,超爱 In-N-Out。
毛利率思维
Lenny Rachitsky: 你提到了毛利率和利润率的概念,说你对此非常痴迷。我想花点时间聊聊这个。我听说你在投入任何项目之前都痴迷于理解毛利率。大多数创始人在这方面完全不知道自己在做什么。关于人们应该关注什么、在思考一个业务的可行性时可能会忽略什么,你有什么经验之谈?
Jason Droege: 对,你看,这只是众多筛选器中的一个。当然存在毛利率低但依然很出色的生意。Costco、Walmart 等等。Amazon 一直在谈论这个——有些公司是提价的,有些公司是降价的。但我想说,总体而言,高毛利率结合健康的流失曲线,是企业非常健康的信号。你想想看,如果我卖给你一样东西,而我无法加价太多,那我在手中这件商品之外到底创造了多少价值?如果我没有创造太多价值,那我到底在做什么生意?我做生意的目的是创造价值。当然事情没那么简单。这只是一个粗略的试金石——当有人来找你,尤其是做一个新业务的时候,我们经常遇到这种情况。我在 Uber 处理过,在各处都处理过。
有人提出一个想法,说:“我们可以进入这个业务,我觉得我们可以定这样的价格,能达到 40% 的毛利率。“然后我的下一个问题是:从 60% 的毛利率开始,为什么不行?他们就说:“哦,嗯,客户会……”然后立刻就短路到了真正的问题所在。哦,客户有替代方案。好,替代方案是谁?哦,是某个离岸外包公司。那他们的毛利率是多少?哦,不知道。你去查查。结果是 20%,而且他们已经存在很长时间了,拥有成熟的大规模运营。你就会意识到,好,你的毛利率会比你想象的更快从 40 掉到 20,你会陷入很困难的境地,除非你能做出差异化。
所以我把毛利率当作一个非常粗略的工具——不是完美的工具——来思考:我创造的价值够不够?我够不够有差异化?它不完美,但它是一个非常快速的短路筛选器,甚至可以用来判断一个人向你推介想法时,他有没有想清楚这个动态关系。因为如果对方的回答是毛利率现在非常低,但这是我瞄准的动态趋势,然后你就会说:“哦,好吧。“有时候对方会说”我们靠规模来弥补”,然后毛利率会有一段时间变成负数,你就会说:“等等,这行不通。”
Lenny Rachitsky: 我喜欢这个思路的地方在于,它提供了一个视角来判断——我的想法够不够好,能不能维持高毛利率?这个领域里的人为什么一直没能做到更高的利润率,是否有其原因?
Jason Droege: 对,没错。而且就像我说的,它的目的是帮你排除掉那些不可行的想法——你面对的是大公司,每个人都有想法。所以这是一种快速筛选的方式:你有没有想清楚两三年后需要搭建什么样的体系?你现在可能有 70% 的毛利率,但接下来的问题是——为什么别人做不了这个?如果你的回答是:“嗯,他们现在能做,但如果我们跑得足够快,两年后他们就做不了了。“好,那我们可能真有戏。如果他们现在能做,两年后也能做,你的毛利率就会被压缩。
Lenny Rachitsky: 顺着这个思路,我刚听了——好像是 a16z 的播客——Alex Rampell 讲了一个 Costco 的故事:就像你说的,他们的策略实际上是保持非常非常低的利润率,因为他们所有的收入都来自会员费。他们大概有五千万会员,每人每月付一百美元,这就是他们整个生意的基础。所以他们并不打算、也不希望从产品上赚钱。
Jason Droege: 对,没错。他们玩的是一个稍微不同的游戏。我不是 Costco 的专家,跟这家公司有过一些接触,但他们把价格当作获取规模的方式。他们的逻辑基本上是:如果我们压低价格——Walmart 也是一样——我们就能获得巨大的体量,把所有竞争对手的生存空间全部抽干。那问题就变成了:好,如果你现在毛利率很低,两三年后,一旦你在某个市场开了一家店,为什么你的利润率不会被进一步侵蚀?答案是:因为我们已经吸纳了所有的需求。你想去做 8% 对比 10% 的毛利率——我大概认为这就是他们的毛利率水平——那会是一门非常难做的生意。如果客户已经养成了习惯,每周的购物行程都围绕着你,你已经和供应商建立了关系,你的店长们已经知道怎么管理库存——那不是一个轻松的事情。所以他们率先做到规模,然后祝你好运去跟他们竞争。
“不输是赢的前提”
Lenny Rachitsky: 好,还有几个问题。其中一个是我听到你经常说、也非常认同的一个概念——“不输是赢的前提”。
Jason Droege: 是的,是的。
Lenny Rachitsky: 聊聊这个。
Jason Droege: 科技行业的文化中,投资组合是由投资者构建的,而且坦率地说,很多叙事也是由投资者主导的。创始人当然也参与其中,但”放手一搏”这个理念已经成了共识。放手一搏,管它呢。但我不确定——如果这是我的人生,我只有一个机会去尝试,我可能不会只是放手一搏。我可能想先想一想。而且我认为最优秀的企业家——我没有数据来支撑这个观点——但这些是我的朋友圈子,我身边的观察——最优秀的企业家和最优秀的生意人,会审视自己决策的风险特征,并始终努力做出非对称正向的决策。
很多时候我觉得我们会忽略一个决策的风险。这里面还有更多可以展开的,因为我实际上认为做出高风险决策然后碰巧成功了,这也是一种奇怪的文化现象——因为那样你怎么培养人去做这件事?做出高风险决策并且足够多地猜对,是一件非常困难的事,因为它会带来很大的波动性。但这又回到了我之前说的关于创始人最重要的品质——就是那种坚持到底的能力。生存本身就是这场游戏的一部分,大多数人都在时机对上之前、在获得对客户的正确洞察之前、在对市场推出正确的产品之前就放弃了。科技行业里,事情可以变化得很快。你可以在极短的时间内从无人问津变成英雄,但你是在一条非常漫长的旅途上,而你必须活到那个条件被满足的时候。
所以问题就变成了:当你处于一个炒作周期中——我认为我们现在就处于这样一个周期——每个人都想放手一搏,然后更大力度地搏,再更大力度地搏,再更大力度地搏,但你没有意识到——伙计们,我们所有的客户五年后都还在。他们只是想让我们解决他们的问题。我们必须活着才能帮他们解决问题。所以生存是这一切的前提。那我们就不要把自己置于可能危及企业的处境中。这不是说不要冒险,而是要思考你怎么计算风险。
Lenny Rachitsky: 我很喜欢的一点是,这个教训以及很多类似的教训,都来自于失败、事情没做成、东西坏掉了——而这是最好的结果。
Jason Droege: 如果你曾经站到过一项高风险高回报决策的错误一端,那种痛苦是难以言喻的——你就彻底完了,很多时候根本没有退路。
Lenny Rachitsky: 有没有这方面的故事或例子浮现在脑海中?
从卖高尔夫球杆中学到的教训
Jason Droege: 这就是我为什么会如此……我觉得你可以在前端多花一点时间思考,来省掉后面很多痛苦。我有过一个生意——不值得细说——但在 2001 年互联网泡沫破裂之后,我想:“我要自己出资做一个生意。我要做一个赚钱的生意。我想证明自己能做到这件事。“我们之前创办了 Scour,那里面有我们之前谈到的所有那些东西。所以我当时想——我是一个高尔夫球手,而且坦率地说,科技行业那时候没什么事可做。
所以我开始在互联网上卖高尔夫球杆,而且真的赚到了钱。这门生意可能比其他任何一门生意都让我学到了更多——我在 eBay 上起步,那时我 22 岁,我并没有真正意识到我的利润率会下降,因为任何人都能做这个,但我是最早做的一批人之一,所以赚了很多钱。然后我把这个生意做大了,但我完全没有意识到自己的傲慢。我想:“哦,只要我能把全美国的二手高尔夫球杆都买下来,我就能成为价格的做市商。“——不是吗,人们不就是这么做的吗?
Lenny Rachitsky: 我喜欢这种野心,太棒了。
Jason Droege: 但如果认真想想那个想法的实际可行性,简直荒唐。我就是没有花时间去思考,然后就陷入了这门生意。生意是赚钱的,做到了几百万美元的收入什么的,还给我发了一段时间的分红,但整个过程都很痛苦。
搭建正确的团队比寻找最优人才更重要
Lenny Rachitsky: 我喜欢你经历的这种跨度——你卖过高尔夫球杆,又在帮助实现 AGI——可以这么说。你的职业生涯中还有一整个部分我们还没聊到:你做过泰瑟枪、执法记录仪、无人机,还有所有这些东西。还有在所有人之前做 P2P 文件共享。最后一个话题,基于这些经验我想花一点时间聊聊——就是招聘和搭建团队,我知道你在这方面有非常强的观点。最近在这个播客上我听到很多的一个理念是:搭建正确的团队比找到最优的顶尖人才更重要。聊聊这个,为什么它如此有趣和重要。
Jason Droege: 最近我对这个问题形成了一个更细致的看法:在某些岗位上,你绝对需要在当前市场中有相关经验的人。研究人员就是如此,因为市场变化太快,你没有时间去培养人,所以你确实需要去找那些——要么与你想获取的客户有正确关系的人,要么在其他方面可能不太符合标准但在这方面非常出色的人。他们可能不符合我理解你所指的那种经典标准——解决问题能力强、能随公司成长、成长曲线高等。我会说这大约占公司 5% 的岗位,但在上市速度至关重要的时候,这些岗位非常关键。
对于面试,我只考察三件事,而且我需要跨各种专业领域去面试,这其实很难。我不可能样样精通。所以我把它简化成三件事:第一,你是不是一个有好奇心的解决问题的人,而且能把这一点清晰地口头表达出来?第二,你能不能与他人协作?你是否有足够的谦逊去和不同的人合作?第三,你是不是一个好的领导者?如果你做到这三件事,我认为你成功的概率相当高,至少在我管理的组织中是这样的,因为世界在不断变化,所以你确实需要适应性强的人。之前的经验不一定是逐一对应的、完全适用的。
然后关于你提到的团队协作这一点,这在 Uber Eats 时期就有过实际案例。当我搭建 Uber Eats 的管理团队时——我不确定那个团队的人有没有跟你提过——但每当我招人的时候,我都在尝试组合出一个几乎像一个有机体一样的优势互补体,然后把冲突降到最低。那个管理团队,除了运营方面的部分人员外,基本上从第一天我们什么都没有,一直到做到 200 亿美元规模,都是同一拨人。我就是相信,团队成员彼此了解各自的长处和短处、能够互相补位,比那种经典建议——“那个人没见过这么大规模的业务”——更重要。你会说:“是,但他能学会吗?“我自己也是学来的。所以你确实需要稍微相信人,这是我的工作,不一定是他们的工作。说到底,这些都是人的系统,不是可以简单套用规则的东西。
Lenny Rachitsky: 我特别喜欢这条建议,因为现在有太多关于在 AI 将取代我们所有工作的世界里哪些技能还重要的讨论,而感觉你这三个维度其实就是同一件事——他们善于解决问题吗?他们是好的领导者吗?他们能和别人很好地协作吗?
Jason Droege: 是的,我认为人性的核心不会改变,这些东西是人类长期以来之所以成功的核心所在。
AI 快问快答环节
Lenny Rachitsky: 说到这个,我要带大家进入这个播客的一个固定环节,我叫它 AI 角,我会问每位嘉宾同一个问题:你在日常工作生活中有没有找到什么使用 AI 的方式,让你更高效、完成更多事情、做得更好?
Jason Droege: 说实话,当我来到 Scale 的时候,我的背景一直在消费领域,也做过一些面向政府的应用层面的事情,而这个领域变化太快了。AI 是我的——我把它当导师用。每当新概念出现,公司里有很多人可以给我讲解数据的技术细节和产品的技术特性,但他们时间有限。而且说实话,新概念一直在不断涌现,我需要跟上节奏。
所以这听起来可能有点疯狂,但我工作中很大一部分并不是处理与 AI 相关的工程问题——我是在管理一个组织——但我很享受理解技术本身。这是我工作中最有趣、最有成就感的部分之一,就是从这些 AI 研究人员那里学习,但他们不总有时间来教我,所以我就把 AI 当导师。我打开语音模式,在上班路上和它对话。所以这大概是我用 AI 做的最有价值的事,也是和今天这个话题最相关的。
Lenny Rachitsky: 我完全做同样的事,尤其是在准备这档播客的时候。“这个到底是什么?“当你说到这个的时候,我就想到几年前我采访 Perplexity 的创始人,问他们在 Perplexity 是怎么工作的,创始人们说他们有一条规矩:在向团队任何人提问之前,必须先问 AI。我当时就想:“这也太疯狂了吧。“而现在,这已经是理所当然的事了。但当时我确实觉得:“这是一种我闻所未闻的工作方式。“这也说明了他们当时有多么领先。
Jason Droege: 对,我想第二个用途是我会把内部文档拿过来,问它:这份文档里最重要的内容是什么?结果让我很惊讶。然后我会自己去读一遍确认,但它在提取要点方面的表现让我震惊。组织里有大量的信息是这样的——我不确定你想让我说什么,我也不确定我需要知道什么,但每个人都有各自的关注点,所以就存在一个匹配问题,进而产生了一个巨大的信息广播问题——在所有你可能想接收的信息中,哪些对你来说才是真正重要的?所以我也经常在这方面使用 AI。
Lenny Rachitsky: 太棒了,这个建议非常好。我用它来看法律文件,比如——他们知道什么、他们想对我做什么、是帮我还是对我不利?Jason,在我们进入非常令人期待的快问快答环节之前,你还有什么想分享的,或者在某个观点上想再强调一下的吗?
Jason Droege: 当然。我想说的是,我来做这个访谈很重要的原因——除了我一直想上这个节目、做了很长时间的听众之外——是 Scale 正在进行大量令人惊叹的工作。团队们非常努力,我们在为客户交付大量价值。而外界的公开叙事并没有真正反映出这里的人所做的工作,以及我们的客户借助我们为他们做的事情所取得的成果。我觉得这些人的付出应该得到应有的尊重和回报,我们希望大家了解这一点。
Lenny Rachitsky: 感谢你说这些。接下来,我们进入非常令人期待的快问快答环节。五个问题,准备好了吗?
Jason Droege: 好,来吧。
Lenny Rachitsky: 你最常推荐给别人的两三本书是什么?
Jason Droege: 有些答案可能会让你觉得有意思。《自私的基因》是我最喜欢的书之一。
Lenny Rachitsky: 那本书我太喜欢了。我记得好像没有其他嘉宾提过,但它也是对我影响最大的书之一。抱歉,你继续。
Jason Droege: 嗯。《自私的基因》。《少有人走的路》,我读过不止一遍,这是经典的人类心理学著作。然后在商业方面,我觉得是《从优秀到卓越》。它不是你在度假时最兴奋去读的那种书,但它说的基本都对。我觉得我们应该听取那些分析过这些商业问题的人的建议,因为很多东西并没有真正改变,但我们总表现得好像一切都变了。
《思考,快与慢》与认知偏差
Lenny Rachitsky: 那本书最疯狂的地方在于,你去看他们讨论的所有公司——我已经有一阵子没读了——但整本书讲的就是那些基业长青的公司,至少我记得是这样,也可能是另一本书,我不确定。但不管怎样,他们提到的所有那些公司,我不知道现在是否还存在。一家企业要长久存活真的太难了。
Jason Droege: 我还想推荐《思考慢与快》,就是那本……对。
Lenny Rachitsky: 《思考,快与慢》。
Jason Droege: 《思考,快与慢》,不好意思说反了。距离我读那本书已经大概十年了,但它的核心观点就是,人类的认知偏差非常重要,值得深入理解。
Lenny Rachitsky: 那本书以及 Kahneman 本人让我最震惊的是,有人问他,了解了人类有这么多认知偏差之后,你的生活受到了什么影响?他的回答是,“没什么影响。我还是有同样的偏差。知道它们并不能真正帮我避免它们。”
Jason Droege: 你看,我发现我会自我审视。现在每当我对自己某件事特别笃定的时候,我会停下来问自己:我倾向于做的那些事情里,有哪些是我需要警惕以捕捉自己失误的?因为我觉得我们最容易在这些时候做出糟糕的冲动决策,我想这也是那本书大部分内容在讲的。当然了,那是一本很长的书。
Lenny Rachitsky: 真的太长了,天哪。感觉这正是 AI 未来能帮到我们的地方。就像,“嘿,Jason,你确定这不是一个框架效应之类的偏差吗?”
Jason Droege: 是的。
最近喜欢的影视作品
Lenny Rachitsky: 好,下一个问题。你最近有没有特别喜欢的一部电影或电视剧?
Jason Droege: 我看的电影大多是跟孩子们一起看的,所以我希望我能给出什么深刻有内涵的推荐。
Lenny Rachitsky: 没关系,儿童内容也是完全可以的——
Jason Droege: 我觉得那部 F1 的电影很不错。我的意思是,它是一部经典的动作片。我不觉得它对 AI 或商业有什么启发,但偶尔从科技圈的疯狂节奏中抽离出来看看也挺好的。
最近发现的产品
Lenny Rachitsky: 有没有你最近发现的、特别喜欢的产品?可以是 App,可以是衣服,可以是厨房小工具,任何这类东西都行。
Jason Droege: VO3。不算完全新的东西了,但我高中的时候想当编剧。我其实在湾区长大,周围的人都是工程师,但我想做编剧。后来我翻出了一部以前写的剧本的第一页,写得并不好,但我拿了第一页,拍了张照片喂给 VO3,然后说”把这个场景做出来”,它竟然做对了。
Lenny Rachitsky: 哇。
Jason Droege: 我非常震惊。简直不敢相信你只需要拍一张剧本的照片就行了。所以现在我就在想,怎么用这些工具来做家庭视频?现在一些图像工具可以让静态照片活起来,我觉得很有意思。它们可能还需要再迭代一步,但我认为这些工具对人们来说会产生改变生活的情感冲击,因为一张来自祖父母或亲戚的照片里哪怕只有一点点动态,对你情感上的影响都会非常大。
Lenny Rachitsky: 我很喜欢这一点,当你玩这些工具的时候,你可能会想到,哦,这些就是帮助训练这个东西的人,这些就是帮它解决那些问题的人。
Jason Droege: 我之前刚好在跟一个做 VO3 的人聊天,我跟他说了剧本那个用法,他说,“哦,剧本啊,对,剧本里数据的格式其实非常好。“因为剧本一开始就是”场景:昏暗的室内”,然后这个角色用沙哑的声音说了什么,所以剧本本身就给了你所有的指令。
Lenny Rachitsky: 天哪,这直接解锁了一个全新的业务线。还有两个问题。第一个是,你有没有最喜欢的人生格言,在工作和生活中经常想到的、觉得很有用的?
人生格言
Jason Droege: 有的。“终点从来不是终点。“这是我最喜欢的内心信条,它和之前说的”生存是繁荣的前提”是一脉相承的。你必须先活下来才能繁荣——但你的大脑会告诉你已经完了。在这些创业旅程中,我觉得这句话最适用。这是任何人能走上的最艰难的旅程。如果你在这条路上走了五年,你在精神上已经比 99.9% 的人都更坚韧。人们不理解那种日复一日的自我怀疑和不断出问题的中国水刑式折磨。
更具体地说,你健身的时候也会有这种感觉,“哦,我太累了,我需要停下来。“但事实是你可以继续,世界也会继续转。所以在那些最艰难的时刻,或者面对一个看似无法逾越的艰难决策、你的身体在对”这不可能”产生本能反应的时候,提醒自己:我明天还会醒来。这不是终点。某个地方还有另一个终点。我发现这样能帮我解锁自己,让我觉得,好吧,可能没有完美的方案,但也许有一个不完美的方案,可它终究是个方案,那就继续走下去。
Uber Eats 与告别
Lenny Rachitsky: 最后一个问题。你参与创建了 Uber Eats,我猜你现在仍然是 Uber Eats 的重度用户。你在 Uber Eats 上有没有最爱的餐厅,也许大家应该知道的,或者你下单最多的?
Jason Droege: 说出来你可能吃惊,我点麦当劳点得惊人地多。尽管有我最初的故事,但在我们家那就是家庭犒赏。我觉得那大概是我们点得最多的东西了。
Lenny Rachitsky: 天哪,我真为你的健康担心,不过我也很喜欢麦当劳,我已经好久没吃了。也许我应该再试试——
Jason Droege: 实际上更日常的情况下我们会点混合沙拉或者轻食沙拉之类的,但我觉得更值得注意的是,尽管我最初对与全球连锁品牌合作有抵触心理,偶尔吃一次确实是不错的犒赏。只是不应该天天吃。
Lenny Rachitsky: Jason,这次对话太精彩了。非常感谢你抽出时间。能在你接管 Scale 之后的第一场访谈就是我,我深感荣幸。如果大家想联系你,了解更多信息,或者也许想加入 Scale,去哪里能找到你?你想指引大家去哪里,以及听众怎样才能帮到你?
Jason Droege: 当然。我在 X 上的账号是 @jdroege,J-D-R-O-E-G-E。这大概是关注我、了解动态最简单的方式,你也可以给我发私信。我觉得这就是保持联系的方式,抱歉,你另一个问题是什么来着?抱歉。
Lenny Rachitsky: 比如你们是否在招人,如果有的话大家应该去哪里看看,还有——
Jason Droege: 当然。直接去 scale.com,进入我们的招聘页面,我们有 250 个开放职位。回到之前说的,我们在做业务而且在增长,正在大量招人。我们的数据业务在增长,应用和服务业务增长非常迅猛,所以这段旅程上我们需要很多人来帮忙。
Lenny Rachitsky: 我读到你们刚签了跟政府的超级大合同。
Jason Droege: 两份一亿美元的合同。
Lenny Rachitsky: 一亿美元的合同。
Jason Droege: 一亿美元,没错。而且我们签的不只是一份,一个月内签了两份。所以没错,我们的联邦业务做得很好,企业业务做得很好,国际政府业务也做得很好。市场需求非常旺盛。
Lenny Rachitsky: 有些销售应该拿到了非常可观的提成。干得漂亮。Jason,非常感谢你来参加节目。
Jason Droege: 谢谢,很荣幸能作为嘉宾来到这里。非常高兴能和你交流,尤其是在这段旅程的早期——至少是我领导 Scale 这段旅程的早期。
Lenny Rachitsky: 非常感谢。谢谢你的到来,谢谢收听。大家再见。
非常感谢大家的收听。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本节目。也请考虑给我们打分或留下评价,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本节目的信息。下期再见。
术语表
| 原文 | 中文 |
|---|---|
| a16z | 保留原文(Andreessen Horowitz 风险投资公司简称) |
| agent | 智能体(agent) |
| AI Corner | AI 角 |
| Alex Rampell | 保留原文(a16z 合伙人) |
| Alex Wang | 保留原文(Scale AI 前任 CEO,现领导 Meta 超级智能团队) |
| Allen Penn | 保留原文 |
| Brendan | 保留原文(Workhorse 相关人物) |
| churn curves | 流失曲线 |
| Costco | 保留原文(美国连锁仓储式超市) |
| evals | 评测(evals) |
| frontier labs | 前沿实验室 |
| Good to Great | 《从优秀到卓越》(Jim Collins 著作) |
| hockey sticking | 曲棍球棒式增长(指突然加速的指数级增长) |
| In-N-Out | 保留原文(美国知名汉堡连锁品牌) |
| Instacart | 保留原文(美国杂货配送平台) |
| Jason Droege | 保留原文(Scale AI 新任 CEO) |
| Kahneman | 保留原文(Daniel Kahneman,诺贝尔经济学奖得主、行为经济学家) |
| Lenny Rachitsky | 保留原文(播客主持人) |
| Lyft | 保留原文(美国网约车平台) |
| McDonald’s | 麦当劳 |
| Meta | Meta(保留原文) |
| Mike Ovitz | 保留原文(投资人) |
| Perplexity | 保留原文(AI 搜索引擎公司) |
| Postmates | 保留原文(美国外卖配送平台) |
| RAG | RAG(检索增强生成) |
| RL environments | RL 环境 |
| Ron Burkle | 保留原文(投资人) |
| Salesforce | 保留原文(产品名) |
| Sand Hill Road | 沙山路(Sand Hill Road) |
| sandbox | 沙盒 |
| Scale AI | Scale AI(保留原文) |
| Scour | 保留原文(Travis Kalanick 早期创业公司) |
| SKU | SKU(库存量单位) |
| Slurpee | Slurpee 冰沙饮料(7-Eleven 品牌产品) |
| Stephen Chau | 保留原文 |
| The Information | 保留原文(科技媒体) |
| The Road Less Traveled | 《少有人走的路》(M. Scott Peck 著作) |
| The Selfish Gene | 《自私的基因》(Richard Dawkins 著作) |
| Thinking, Fast and Slow | 《思考,快与慢》(Daniel Kahneman 著作) |
| Travis Kalanick | 保留原文(Uber 及 Scour 联合创始人) |
| Uber Eats | Uber Eats(保留原文) |
| unit economics | 单位经济模型 |
| VC | 风险投资(VC) |
| VO3 | 保留原文(AI 视频生成工具) |
| Walmart | 保留原文(美国连锁零售巨头) |
| Workhorse | 保留原文(公司名) |
| X | X(社交媒体平台) |
此文档由 AI 分片翻译(translate_long_document)