Welcome to Decoder. This is Alex Heath. For my final episode as your Thursday guest host, I recently sat down with Bret Taylor, the CEO of AI startup Sierra and the chairman of OpenAI, for a live event in San Francisco, California, hosted by Alix Partners.
Very few people have seen the tech industry up close like Bret has. He was an early engineer at Google before starting FriendFeed, a social network he sold to Facebook in 2009, where he then served as chief technology officer. He later founded Quip, which he sold to Salesforce.
After eventually becoming co-CEO of Salesforce, he left to start Sierra, which is rethinking how businesses use AI for customer support. Along the way, he led Twitter’s board during Elon Musk’s takeover and became chairman of the OpenAI board after the firing and rehiring of CEO Sam Altman.
As you’ll hear in our conversation, Bret is all in on AI. Just this week, Sierra raised a new round of funding, valuing it at $10 billion. In this episode, we get into Sierra’s origins and what it’s doing with AI agents. I also asked Bret about OpenAI and the overall trajectory of the AI industry. We covered a lot of ground, and I hope you find Bret’s perspective as fascinating as I do.
Okay, here’s my conversation with Bret Taylor:
This interview has been lightly edited for length and clarity.
So I hope people here are familiar with Decoder. I’ve been guest-hosting over the summer. Nilay Patel, our editor-in-chief, has been out on parental leave, and I’m very happy to be here talking with Bret. Thanks for being on the show.
Thank you for having me.
I would like to start by going back to early 2023. You’re leaving Salesforce, you were the co-CEO. Talk about that process of deciding to make a new company and what you looked at. Why did you land on Sierra at that time?
I happened to announce I was leaving Salesforce within a few days of ChatGPT coming out. I don’t know if you believe in cosmic forces. But like for every single human being, particularly a geek like me, who first uses a product like that, it’s all I could think about.
So I was honestly not 100 percent sure what I wanted to do when I left Salesforce. I was trying to leave and then figure it out, which is a good thing to do in life. And I immediately just became obsessed with the technology. I was using it personally. [LinkedIn cofounder] Reid Hoffman is a friend of mine, and he was showing me early versions of GPT-4, and I just couldn’t believe the level of empathy and how it just truly sounded human. I had been following AI for years, but honestly, if you had told me in October [2023] or the month before to define what a large language model was, I would’ve given you a blank stare.
I ended up realizing that this technology, which I had not been following as closely as I wish I had been, was really going to change the world. I knew I wanted to work in it. I didn’t know what I wanted to do, but that was okay. It reminded me a little bit of when I first discovered the internet, and I think everyone knew it was going to change everything. At least I felt that way, and I was excited to work in that space, and that’s all I knew.
I ended up having lunch with Clay Bavor, whom I’d known for 20 years and was not planning to start a company with, but I found out through the course of the lunch that he was equally obsessed. He was working for [Google CEO] Sundar Pichai at the time, and by the end of lunch, we had had a couple of more courses than we had originally planned.
We had decided to start a company, and we had no idea what we were going to do, but I think it was really based on the premise that when you have a seismic shift in technology, a lot of business opportunities present themselves, because it kind of shuffles the deck of what consumers want, what companies need, and what software vendors have the resources to support that need.
If you look at the advent of the internet, it gave birth to some of the largest names in the stock market today, like Amazon and Google. It disrupted companies like Microsoft, which got through it quite strong. It disrupted companies like Siebel Systems, which emerged a little bit less strong. So you end up where the incumbent insurgent dynamic changes quite a bit, and huge markets open up. In the case of the internet, it was search and e-commerce.
I think with large language models, and we will probably talk about that more as I have given too long-winded of an answer here, it really stands to reason that a lot of different markets, from software engineering to customer service, are going to be completely transformed and upended. What an interesting time to start a company.
So, we left and gave ourselves a few months of just recovering from our jobs and then talked to a lot of customers and decided to build Sierra. At Sierra, we’re building AI agents for customer experiences. Everyone from ADT home security to Ramp in New York to SiriusXM are using agents to answer the phone when you call them up, or in their digital properties, or just to have a conversation — they’re doing everything from helping you upgrade or downgrade your SiriusXM plan to calling you when your ADT alarm goes off, which I think is pretty exciting.
Talk to me about Sierra and how you work practically with a new customer. Walk me through that process, because this is all a very new field. I mean, customer support is not new, but the way you’re doing it is new. So what is unique about how you work with a customer versus how you would’ve done it at another company before?
I’ll start with our business model, because I think it will help to answer your question. One of the things that we do differently at Sierra from traditional software companies is we charge only for outcomes. So for most of our customers, that means when the AI agent autonomously resolves the case that the customer called about or chatted in about, there’s a fee for that. If the AI agent has to transfer to a real person, it’s free.
We really like this as a business model, and I think it will become the standard business model for agents because the word “agent” comes from the word “agency,” and the principle of it implies some degree of autonomy. I think most of the most sophisticated agents will actually start and complete a task, whether it’s generating a new lead for your sales team or solving a customer service inquiry or doing a legal analysis for an antitrust review, whatever it might be.
If an AI agent isn’t only helping a person become more productive but is actually accomplishing a task, why not just pay for a job well done? If you look at most of your companies, if you have a job where the outcome is measurable, like sales, you tend to pay a commission, right? There’s not just a salary. So I think agents sort of being paid on commission, if you will, is not only a great incentive alignment between a vendor and a partner and a company but also just feels right from first principles. That’s why I think it will be just like the advent of cloud-based software. When Mark [Benioff] and Parker [Harris] started Salesforce, and it was a subscription-based service rather than a perpetual license, it changed the landscape of software; the same will happen with agents.
Now going back to how we work with customers, it raises the question of what the relationship is between a software vendor and a company if you get paid only when it works. There’s a certain degree of arm’s-length relationship that most software vendors have with their customers. If you’ve ever seen someone who’s done a big ERP implementation, I don’t know much about ERP systems, but apparently, they’re really hard to execute because for everyone I’ve ever met who’s done one, it’s taken two years longer than expected and cost a lot more money than expected.
If you go and talk to the 10,000 people involved in one of those projects, the systems integrator points to the software vendor. The software vendor points to the systems integrator. No one’s really pointing at the company, because the company is the one paying the bills. So everyone’s like, “Oh, you’re great. No, I’m sure everything’s fine.” And it’s like success has a thousand fathers, but failure is an orphan. Part of the issue is the only party in that relationship that cares about the outcome is the company.
So everyone’s blaming everyone else, and it requires a good CIO or CTO to navigate that, but you can see all the perverse incentives involved. Maybe the partner’s getting paid by the hour; that’s not a great incentive. The software vendor’s already made the sale, so good luck to you in getting it successfully deployed.
I think going toward outcome-based pricing demands a different relationship between a software company and the companies it works with. I think it’s trendy right now in AI, in part for this reason. No software company wants to be a professional services firm. So you can’t turn that knob all the way up to 11 and build a company that I think you want to build. But there is a different level of accountability.
So in our relationship with our customers, we’ve really focused on a couple of different things. One is product usability. I think to make your outcome, you need to make it as easy as possible to achieve that outcome. We’re somewhat unparalleled in the market in having a product for technology teams as well as a product for operations teams. You can build agents without any technical knowledge at all. Again, we’re trying to empower as many customer experience professionals as possible. And then on the partnership side, we have a lot of support with what we call agent development. So if you need help getting your agent out the door, we show up in a bus to help you do it. That’s unique.
I’m not sure how everything will play out, but I’m really bought into this vision. When I talk to our customers, I love the idea that they know exactly the value that we’re providing for them because they pay us only when the agent works. I just love the simplicity of that relationship, and I’m really bought into it.
So you have hundreds of customers, and 50 percent have revenue over $1 billion, and 20 percent have revenue over $10 billion a year. Is that right?
That’s correct.
Why focus on customers like that instead of taking a huge Shopify-like approach to this? Why are you going toward the big companies?
Big companies have big problems. I love first principles thinking, and if you are a large consumer brand and you have 100 million consumers around the globe, before large language models, you could not have had a conversation with all of them. If you just do the math, there’s a phrase in call centers called cost per contact, and it essentially measures how much all-in labor and technology costs to answer the phone or answer the chat. It really depends on how complex the conversation is, and the qualifications of the person answering the phone. It depends on whether it’s onshore or offshore.
But say it costs somewhere between $10 and $20 to answer the phone. For most consumer brands, their average revenue per user is less than that phone call. So you literally can’t afford to have a conversation. It’s why if you’ve ever tried to call any consumer brand, you’ll find you can’t.
There’s entire websites devoted to finding the phone numbers for many consumer brands. It’s not because they don’t care about you. It’s just not economical. If everyone who wanted to call them called them, they would go out of business, which is probably not good for you either. Now with large language models, that’s totally different. You bring down the cost of a phone call by not one but two orders of magnitude, and all of a sudden the economics of having a conversation change dramatically.
So the reason we’ve pursued larger enterprise brands is that’s the type of step-change function in customer experience that is relevant to a company that has tens of millions or hundreds of millions of customers. Those are the larger companies in the world. What’s really exciting is, I think for a lot of people, when they think about AI agents for customer experience, they think contact-center automation, and that’s a huge part of it.
But if you think about it through the lens of what I just said, you can now have an order or two orders of magnitude more conversations with your customers than you could before, for the same cost. And that’s really remarkable. And if you think about all the companies that are competing for whether it’s, let’s say, a mobile phone company, you’re competing for a fixed pie of customers trying to decide which company to align themselves with. And if you can improve customer attention by 100 basis points, that’s a lot of value. If you can reduce your attrition and churn by 500 basis points, that changes the lifetime value equation of your company.
So I think people are thinking about it. I think it is really the first-order effect of reducing the cost of a phone call, which is great; you can save that money and return it to shareholders. But I think the more sophisticated companies are asking, “Can I actually gain market share?” And that’s really, really exciting, and that’s what we’re trying to do for some of the largest brands in the world.
Do you have agents right now that are doing things for customers without human involvement? I’m talking like beyond a chatbot, but actually doing things that have economics tied to it, or that would be something you would think a human would be involved in but is actually not. Is there an example of this right now?
I’ll give a few. We have retailers for whom you can submit a photo of your damaged goods and immediately adjudicate a warranty claim, and it’ll connect to the inventory system and ship you a new product. You can refinance your home with an AI agent powered on our platform end to end.
Without a human in the loop?
Without a human in the loop. These agents are remarkable at what they’re doing, and you can take action with an agent built on the Sierra platform — 100 percent of our customers are doing it. To some degree, there’s this technique in AI called retrieval augmented generation, which is a fancy way of saying it’s answering questions. It turns out that that’s kind of a commodity at this point. Slapping ChatGPT together with a knowledge base is not that hard. Most engineers nowadays could do that in a weekend, which by the way is mind-blowing. It was science fiction three years ago; now it’s a weekend project. Welcome to technology. It’s mind-blowing.
Actually, being able to put sophisticated guardrails around a regulated process… We work in the health-insurance payer industry, we work in the provider space, we work with banks, we work with property and casualty insurance companies. If you’re talking about sophisticated, regulated conversations like claims processing, that’s not retrieval augmented generation, right? That’s a very complex conversation with regulatory oversight. How do you put AI-based guardrails around it? How do you put deterministic-based guardrails around it? How do you solve the mundane problems of transcription accuracy in 40-plus languages?
It turns out transcription accuracy doesn’t really matter if it misses the word “and” or “or,” but it really matters if it’s your account number. So how do you get the hard parts? We do roadside assistance, and it turns out if you’ve ever chatted with an AI agent and a car horn honks, it’ll often stop talking because it thinks it’s being interrupted, because it can’t distinguish the difference between a car horn and you talking.
Our platform is really designed to solve those problems: effective guardrails, multilingual conversations over chat and voice, deterministic guardrails, AI-based guardrails, which are called supervisor models — and are really, really effective and interesting. And simple stuff like knowing, “Hey, that’s the television in the background, no one’s talking right now,” or “That’s a car horn, someone’s not interrupting me.” It turns out that I’m sure in three or four years, that’ll be easy. Right now, it’s really hard, which is why we have a lot of demand for our product.
I’m glad you brought up voice. I’d be curious to hear how voice is entering this mix beyond chat, and do you think voice will actually be a bigger piece of the pie for agents than text?
I do. Voice is already a bigger part of our platform than text, which is kind of remarkable, because we launched it in November of last year. I think it stems from a couple of reasons. One is, first, I’ll just go to the human parts of it. I mean, if you watch movies about computers in the future, or science fiction authors’ vision of the future, you’re usually talking to a computer. I think it is the most ergonomic interface. We’re all born with it. We all know how to talk. As a consequence, I think it’s quite low friction, it’s quite accessible. We talk a lot about the digital divide, and I think if most of the ways you interact with digital technology is just speaking, what a great way to make it accessible to everyone, especially if it’s multilingual and patient.
If you look at the telecommunications industry, the health-insurance industry, and things like that, a lot of customer service still goes over the phone. It’s not just as patients or consumers but providers to payers. A lot of this is still running over the phone. And what AI has done is it’s taken one of the oldest analog channels, which is the publicly switched telephone network, and made it digital for the first time. It used to be that almost every company I talked to had a digital self-service team, which is a fancy way of saying, “Hey, can you please use our website rather than calling us, because gosh, it’d be better for you, and it’s better for us. It’s cheaper. It’s faster.” And there’s entire teams devoted to that. Now it’s like, “Maybe call us. It’s all good. It turns out the same agent on our website is picking up the phone,” which is kind of crazy.
You always talk about TCP/IP, which is like TCP running over the internet protocol. There’s some name for this. We’ve basically put the internet on the phone; we’ve just made the phone a channel for digital technology for the first time. And so as a consequence, if you look at… There’s a proverb in entrepreneurialism that says, “You want to make a painkiller, not a vitamin,” because people buy painkillers and people think about buying vitamins.
This is truly a painkiller. You’ve just taken the most expensive, the most tedious, channel — and everyone hates it too, by the way, even if you talk to the best customer service agent of all time on the phone, it’s usually after you’ve been waiting on hold for 10 minutes. Because the economics of making a call center where customers don’t have to wait on hold are just untenable.
So it’s just one of those things where consumers, companies, there’s no one defending the current landscape of phone calls at all. Everyone hates it on all sides, and now you have this technology that just solves the problem. So that’s why I think it’s going to have a big impact. But looking forward, it’s really unclear. I’m kind of in the center of a lot of this AI stuff, and I couldn’t tell you where the world is going, but I think it is really exciting. If you look at the way WhatsApp is used in Brazil and India, you wonder with conversational agents whether that style of digital interaction will be as pervasive in other markets.
I was blown away when I went to Brazil, I don’t know, four years ago and saw someone do a mortgage over WhatsApp. I was like, “Tell me what you’re doing.” And it’s like uploading the PDF. All of a sudden, if you think about every company in the world’s customer experience having a conversational agent, then maybe every company in the world will have a WhatsApp presence to do that, or maybe smart speakers will make a comeback.
I think about driving into work and CarPlay. I love the product in a lot of ways, but you can’t really do anything with it. Imagine triaging your email, having a conversation with a personal agent while you’re driving into work, and all of a sudden your commute just got super productive. It’d be like talking to a personal assistant with a PhD in everything. I mean, that’s pretty cool. So I think that’s exciting. We talk about phones, because I do think it is the area that is just economically impactful. Right now, we’re making computers conversational, and I think it is a user-interface paradigm as much as a technical change. And I’ve never felt sold.
The other day I was talking on the phone like this. For people online, I’m touching the phone to the side of my face, which until this moment I thought was normal. My kid was like, “You’re touching your phone to your face?” It would be like someone licking their phone or something. All kids just talk on the phone differently. I never thought of that as abnormal until that moment. And then I thought, “I’m fucking old.”
You realize that I just think that kids who grew up with these technologies who never saw a rotary dial, they just have a different style of interaction with these new technologies. Younger kids today are going to grow up in a world where of course computers can understand what I say when I talk to them with nuance and sarcasm, and of course I have a personal AI agent that can go do my research for me for my next trip.
I think we are not even contemplating the second and third order effects that led to my child thinking that touching a phone to the side of my face was weird, which just still boggles my mind. But I think we’re at the start of a really significant trend, and I’m hopeful in a lot of ways because I, like many others, read things like The Anxious Generation, and I catch myself being mildly addicted to staring at the glowing screen in my pocket. You wonder if you fast-forward four or five years, will software melt away into the background? Will a lot of things that are tedious, like waiting on hold and not being able to find a phone number, will this technology make all that go away?
“Yeah, call us anytime.” Oh, and it knows everything about me and whether I want to talk to it over chat because I’m on the BART train and I don’t want people to hear me, or I want to talk on the phone because I am holding things in my hands. All of that will just be available.
So I’m excited for it because I think like with all technology trends, we’re on the bottom rung of Maslow’s hierarchy of needs, and it’s very hard to see self-realization, or whatever the top is, but I think we’re going to get there relatively quickly. Our hope at Sierra is that we can help every company in the world navigate that. Step one is to create an amazing customer service experience for your customers that makes them feel respected and valued, and is truly personalized. Step two is to set up your company for whatever the future holds. What does conversational commerce mean? What does it mean when people are doing their consumer research on OpenAI rather than search engines?
I’m sure many of you have done that, when, for example, you get a lab result. I just upload it into ChatGPT immediately before I talk to my doctor, and I don’t know how he feels about that. But I promise you 100 percent of his other patients are doing that too. The whole world’s changing. So a lot of what we think about at Sierra is how do we set up every company in the world to be successful in that new world?
Technically, though, are you developing your own models? What is the actual tech secret sauce here that you have? Is it models, or something else?
We do a lot of fine-tuning. We don’t pretrain any models. I think most applied AI companies shouldn’t. It’s a very fast-appreciating asset, and probably would not produce a meaningful return for your shareholders, but it is quite complex. For any given message to one of the agents on our platform, that’s probably 20-plus inference calls just to generate one response. Just to give you a sense of the complexity, there are lots of different models under the hood. There’s not one provider or even one parameter count, which is a measure of the complexity of these models.
I think that’s where the world is going for an applied AI company like Sierra, because it’s almost like saying, “What’s the right way to store my data?” And for the technologists in the room, there’s a trillion different databases and data storage systems from Snowflake and Databricks to traditional transactional databases. We’ve gotten to the point now where a modern technologist would know, “Hey, for this use case, this is the correct choice.”
That’s where I think we’re going in the applied AI space, not artificial general intelligence but in the applied AI space, where these models are truly pieces of infrastructure, and sometimes you want something that is really fast, and sometimes you want something that’s really cheap, and sometimes you want something that’s really high-quality. And with this price-performance or latency choice, there’s really an option everywhere in that matrix for whatever you want for your business.
I think it will end up like the database market. It will be the practitioners of building these agents and other things. They’re not going to be the researchers who know how to pretrain a model. My intuition, for what it’s worth, is that even fine-tuning will wane over time just as the context of windows and the quality of rules adherence improves in these models. But what it will mean to build an application on these models will be like saying, “Hey, do you know how to use a database?” Not “do you know how to write a database?” Those are two very different skill sets today, and I think that’s kind of where the applied AI market is going.
I think we saw with the release of GPT-5 that the models are still getting better, but the step changes are not as dramatic as they used to be. Maybe that will change as the space moves faster?
I don’t totally agree with you on this one, but finish your question because I’m rudely interrupting.
Nor should you agree; you’re on the board of OpenAI. But I guess what I’m saying is, do you agree with the thesis that the models themselves are becoming commodified? I mean, you talked about it as infrastructure, but I guess what I’m getting at is, what are the second-order effects, if that is true? If the models are really just becoming anything plug and play, yes, they have certain attributes that are better, but they’re not dramatically step-function changing like they used to.
Well, the reason I was disagreeing wasn’t about being an OpenAI homer, which I am by the way. So happy to play that role. Actually, it’s more just saying, I think it really depends on the task. If anyone was using GPT-4.0 or 4.1 for coding and then swapped in GPT-5 for coding afterward, you saw a dramatic improvement in performance.
So through the lens of that task, it was very much a step change in performance. And so for people who are using this for coding agents, I think through the lens of that use case, what you said was definitely not true. There was absolutely a step change in performance. I planned one of our vacations on ChatGPT earlier this year, and I think I was using 4.0 to do it. And my guess, if I’d used GPT-5 for that same trip planning, it would’ve been like, yeah, okay, it’s whatever, slightly better. I had a great vacation, so maybe I just didn’t have high enough standards. Maybe it would’ve gotten a lot better.
I think that we’re getting to the point that for a lot of tasks, we’ve reached sufficient intelligence. So when new models come out, if you’re measuring it relative to planning my vacation, you’d be like, “Gosh, I don’t see a huge change in the quality of this model.” If you’re trying to discover a new therapy and you’re doing drug discovery, or you’re trying to autonomously write a complex piece of software, or you’re trying to do a complex asynchronous agentic task, your perspective on how big of a step change there was may change. So my intuition, but it’s just one person’s intuition, is that perception of how big of a step change these models bring will increasingly be a function of how complex of a problem you’re trying to solve with them.
If you think about what it means to build artificial general intelligence, we need some more improvements, right? There was a really interesting thread on X from an OpenAI researcher [Sebastien Bubeck] who gave it a math paper, and it actually had a relatively novel approach for a type of math I don’t understand. So that’s the limit of what I can say about that, but it was really interesting. It was really creative. It really had that sort of almost alpha go moment of like, “Wow, that’s interesting. It’s sort of novel new mathematics.” Certainly, if we want to get to the point of developing new AI research, finding new drug therapies, proving some of the unproven math problems in the world, we have some work to do. We haven’t gotten to that point.
But my guess for what motivated your comment, probably for a big bunch of tasks, the models have sort of gotten to the point of sufficiency. So going back to your question, which is, what does it mean? I think OpenAI is a mission-driven company. Our mission is to ensure that artificial general intelligence benefits humanity, and we want to work toward beneficial AGI, and we’re not there yet. We need to continue to do that research and development. There are parts of it that are already superintelligent, but there are a lot that aren’t. That’s really what we’re working on.
Does it mean that for different tasks that Sierra solves, or that you do in your personal life, we need those really powerful models? Maybe not, and I think that will just result in an ecosystem of models and what they’re used for. But what’s exciting, just around here in San Francisco, is we’re not done yet. We want to create AGI, and that’s really exciting. I think despite the perception of these models slowing down, I don’t really subscribe to it. You can see in some of the true research breakthroughs, the Math Olympiad results — I mean, these are really meaningful new changes that weren’t possible with previous models, and I think they’re dang exciting.
I’m glad you brought up AGI. I am increasingly of the opinion that no one knows what AGI means, but I think as the chairman of OpenAI, it actually really matters what you think AGI means. I would love to know what AGI means, and I would love to know what you think it means, and if that has changed at all for you, especially in the last year or so. Do you have a sense of “this is AGI,” and when we achieve this, we have hit it?
First of all, I’ll answer the last question, which is, has it changed? Yes, it has changed for me. I think we are already at what I would’ve defined AGI as three years ago. Actually, by the way, I think we’re already at what almost anyone in the world would have defined AGI as three years ago. There’s this thing called the Turing test, which I think, actually, I don’t know what the original one was in the paper, but the way it was taught to me in computer science was having a conversation with an AI and having it be basically indistinguishable from a human conversation. We’ve been past that for years. This was a big thing in AI for a long time. We passed that and yeah, just forget the Turing test. That was a dumb idea made by the smartest computer scientist of all time.
So we just keep on moving the goalposts, because we have exceeded our own collective expectations about what this technology can do so many times that what we had intuitively thought of as AGI, we’ve lapped it four or five times. The way I think about it now — and it may change again — is that in the domain of digital technology and ideas, are these models exceeding human intelligence or at human intelligence in almost all domains? I say that in the digital domain of ideas because I think it’s one thing to invent new types of math, which I think a lot of people would put in the domain of AGI and superintelligence. But interacting with the physical world is a whole different thing. I think that’s a separate problem that’s unrelated to intelligence per se. Just being concrete about it. You can invent a new therapy, but a clinical trial is a completely independent process. So I think the intelligence part of it is really what I was trying to find a measure of.
The other part of the G in AGI is generalization. So one of the things that I don’t believe, but I’ve talked to a lot of researchers — and that’s what’s interesting about AI, some of the smartest people don’t agree on all these things — if you make something that is really, really good at math, the question is how good will it be at a lot of other things? You’ll talk to some researchers who think, “Well, math is sort of the basis of reasoning, and it will be great at a lot of things.” You talk to other people who wonder, “I don’t know, will it generalize to different parts of biology and other things like that?”
So I am more in the camp of thinking that as long as the model isn’t trained for something like the Math Olympiad and is a byproduct of the model, it will generalize. But I think we’ll have to see. I think we’ll see how it generalizes to other forms of science and things like that. But more and more, just looking at the rate of progress and seeing some of the results coming out of OpenAI and the other research labs, I’m more optimistic, and I’m looking forward to the first true scientific breakthrough to come from an AGI. I think it will happen in the next couple of years. I don’t know. But it feels as if it will. It’s a better question for the researchers at OpenAI than me. But certainly, if you start to see some of these early results, it certainly feels possible.
Why are people like your old boss Mark Zuckerberg now talking about superintelligence? What is the difference there? I mean, this is a thing when you’re out here in Silicon Valley and San Francisco, now people are saying superintelligence. It’s like, well, is it because everyone is kind of like, “Well, we did it,” shrug, “We passed the Turing test?”
[Laughs] It’s a rebrand, yeah.
Yeah, rebrand. What is the difference? I don’t really understand, to be honest.
Superintelligence, I think, literally just means that it is more intelligent than humans. So I guess if there’s a subtle distinction, it’s if you made something that was generally intelligent and functioned as well as you and me, is that now lackluster? No offense, Alex, by the way. I think it would be great if we made it. You’re sufficiently intelligent for me.
So I think it’s a higher bar that is truly exceptional. There’s a few reasons from a research and safety standpoint. It’s useful to talk about superintelligence because it’s a reminder that if the models exceed your own capacity to reason, how do you monitor them? How do you make them safe? You really have to use technology to monitor the technology if it exceeds your own capacity to do so. There’s lots of precedent in non-AI things. You have lots of things in an airplane or a car that are monitoring for things you can’t understand or are operating too fast, but that is a really important area of research.
So I think it’s useful to talk about. There’s the public relations part of it that I don’t really have an opinion on or care to think about, but it’s useful when you think about safety. There’s a real question of how do you know that it’s aligned if you can’t understand it? How important is it that a human being understand it versus a supervisor AI that we made to understand it? There’s a lot of both technical and philosophical questions that I think are really important to answer as we develop.
I was at a recent dinner with Sam, Sam Altman. This dinner got a lot of headlines because Sam said that he thinks we’re in an AI bubble. His exact quote was, “Someone is going to lose a phenomenal amount of money. We don’t know who. And a lot of people are going to make a phenomenal amount of money.”
It’s like the old marketing quote.
Yeah?
Only 50 percent of my marketing is useful.
Which one?
I just don’t know which 50, yeah.
Right. Do you agree with that? And if so, why?
Oh, absolutely, yes. I’ve given this analogy before, so I apologize, Alex, if you’ve heard it, but I think there’s a lot of parallels to the internet bubble. If you look at the internet bubble, a lot of people think about the flops, like Pets.com and Webvan. Through the lens of the past 30 years, though, we’ve now gotten most of the largest companies in the world, including Amazon and Google, two of the largest companies in the world. But then you look at how much of Microsoft’s market cap is from Cloud and others, and you start to look and you say, “Actually, if you look at the GDP of the world, how much has actually been created or influenced by the existence of the internet?” One could argue that all the people in 1999 were kind of right. It was as impactful on pretty much every measure.
Even things like Webvan, there’s now, as the internet became more distributed, really healthy businesses like Instacart and DoorDash and others that were built now that the smartphone and the scale of the internet has matured. So even some of the specific ideas were actually not that bad, but maybe a little early.
But if you look at the internet, if you were an Amazon shareholder from its IPO to now, you’re looking pretty good. If you’re a Webvan shareholder, you might feel a little differently. So both exist at the same time, and I think right now you have modern large language models and modern AI that are absolutely going to have a huge impact on the economy, if you just look at software engineering and customer service by themselves.
I mean, we haven’t seen a world in which we’ve reached a sufficient number of software engineers, and we probably will with coding agents, just because we’ve taken something scarce and we’re making it more plentiful. What is the market for developing software? I don’t know. I mean, I don’t even know how to measure that because every company in the world is now a software company to some degree.
So as a consequence, I think just for me, it almost has to be that there’s going to be huge winners in this. Because of the amount of economic opportunity, you just end up with a ton of investors, and some companies will fail and some will succeed. If you look at the people who built out fiber in the early days of the internet, a lot of them went bankrupt, but that fiber ended up getting used, just by the next person or the private equity firm or whatever entity bought it.
I think it is both true that AI will transform the economy, and I think it will, like the internet, create huge amounts of economic value in the future. I think we’re also in a bubble, and a lot of people will lose a lot of money. I think both are absolutely true at the same time, and there’s a lot of historical precedent for both of those things being true at the same time.
Does it worry you at all that the bubble could be in the sector of AI you’re in, in the enterprise? There was that MIT report that everyone’s been talking about where a lot of spend [on AI] is not seeing results. I know you have a different pricing model that’s more geared toward success. But I don’t know, it seems like the bubble could be all the enterprises that have rushed in and spent a ton of money on stuff that’s not working. What happens when that reverses?
I’ll decouple whether I worry about it from that study, which I disagree with. Because I do worry about it, but I don’t worry about that study, so I’ll decouple the two. So I’ll end with the study, because it’s more optimistic than me worrying about my existential issues around my business. But I’ll start with that.
Yeah, I mean it’s weird. So there’s this story that goes on around me about rewriting Google Maps, and it’s mostly true and a little embellished like many great stories are. And it’s interesting to me because people like to tell the story because they’re like, “Oh, wow, one person wrote a lot of software over a weekend.” And now if you’ve used Codex or Cloud Code, you’re like, “Yeah, I can just have an AI agent do that over a weekend.”
So it’s like the thing that’s so exciting, that was actually part of my own personal identity is now an AI agent. Maybe not quite yet. I wrote some pretty good code. But probably in a couple of years, yeah, an AI agent could totally do that. So it’s going to go from, “Wow, that was impressive,” to “Wow people did that?” over the next couple of years.
There’s the business thing, which is what is the software market of the future? I think it’s a really good question, because if you pull the thread… And we reach plateaus, like self-driving cars, we’re really excited. It took a long time. So even smart people can be wrong on these things or too overoptimistic, but with agents doing software engineering, we’re taking the scarcest resource and one of the highest-paying jobs and we’re literally making AI agents that do that. So what will that do? I have a lot of people ask, “Should I study computer science in school?” I have a bunch of opinions. I think the answer’s yes, but honestly no one really knows.
Are we going to reach a world where generating software — and generating’s not the hardest part of software as most software people know — will largely become a commodity? Maybe. A lot of people think that. What does that do to the software market? My hypothesis is actually it doesn’t change it a lot. I don’t think when you buy an ERP system — going back to my ERP; I don’t know why ERP systems are on my mind this morning — you’re not buying the bits and bytes that did it. You’re buying the fact that a lot of companies have their ledger on it and that you can close your books every quarter on it and it’s reliable, and there’s a patch to the servers so that you know that your cloud-based ERP is not going to have a security vulnerability, and the system has these compliance certifications and all these other things that aren’t particularly exciting, but they’re kind of the boring but important part of enterprise software.
If you could write your own ERP system as a major CPG [consumer packaged goods] company, is that a good idea? I’m not totally convinced it is. I always like to say software’s like a lawn, you have to tend to it. And so if you build it, you bought it, right? You have to own it and maintain it and deal with all of it. There’s a new accounting standard that comes out, and all of a sudden you have to do that yourself. So I think it will change the way we write software. Do I think it will completely upend the landscape of the existence of an enterprise software market? I don’t totally believe that. Might be wrong.
It’s really new. We’re just in a really new world because we’re taking something scarce and making it plentiful. I have thought about this movie a lot recently, and I recommend people watch it, which is Hidden Figures. It’s a great movie about putting people on the moon, but it focuses particularly on the women who did the math calculations to do it, and they were called computers. I’m a computer. I didn’t know that until that movie, and I watched it with my kids, that that was a job title.
It’s interesting. One of the women in there, they’re putting in an IBM computer, which is the size of a living room, and she in a sort of a savvy way learns how to use punch cards to program it, basically for job security. We’re all kind of going through these moments right now. Like I’m a computer, a calculator, basically. And that story of me with Google Maps is like a story of a computer or calculator, right? But I think the second and the third order effects are a little fuzzy. I believe the enterprise software market will change from software to agents, but I believe companies want to buy solutions to their problems and not build software. So I believe the market will continue to exist.
On that study, I don’t know the basis for the data. I think it was problematic because it conflated people building their own software with AI and buying solutions off the shelf. I think those are two very different types of AI adoption. We have basically a 100 percent success rate with our customers doing a proof of concept and going live with our platform. And I think it’s because we’re not selling AI, we’re selling customer experience, and you just turn it on and it works.
There’s an amazing company called Harvey somewhere here in San Francisco. I actually don’t know where it’s headquartered, but it’s a really great company. I’ve talked to so many law firms who’ve deployed Harvey for a lot of their legal analyses and they’re all successful, but it’s not an AI project. Like Harvey uses AI. That’s the only reason the company can exist, but it’s doing it because you want a better antitrust review process, and that’s what they’re buying from it.
I think we’re just in the early days of AI, where there isn’t yet a wonderful vendor for every problem that you have in your business. So as a consequence, you either have to wait, or you have to build it yourself. And so I don’t know what the right answer is. And when you’re trying to build these things yourself, it’s just a glacial amount of complexity. And what you end up having is a lot of these, I call it AI tourism, like a lot of people doing performative AI projects, but to actually go to that last mile is quite difficult. I think the solution will be applied AI companies. I think if you want to build a better customer experience, buy Sierra. If you want a legal AI agent, buy Harvey.
I think we need to go through every use case, from supply chain optimization to accounting to maybe an auditor for your accounting department. All of those will be AI agents, but I think there’s a company in every single one of those domains. And I think that is the correct way to package and purchase AI software. I think that study is a measure of all the people who are trying to… which is just a lack of maturity in the market. There’s just not a solution to every space yet. But there’s a few VCs in the room, and hopefully thanks to all of you, in a couple of years there will be. And I think that will be the new what was software as a service, those new agent companies will be that next frontier of business solutions for enterprises.
All right, Bret, we have to leave it there. Thank you so much.
Thanks for having me.
Questions or comments about this episode? Hit us up at [email protected]. We really do read every email!