#224 Bret Taylor – A Vision for AI’s Next Frontier

0:00:03 Technology companies aren’t entitled to their future success.
0:00:06 AI, I think, will change the landscape of software.
0:00:10 And I think it will help some companies and it will really hurt others.
0:00:15 And so when I think about what it means to build a company that’s enduring,
0:00:19 that is a really, really tall task in my mind right now.
0:00:24 Because it means not only making something that’s financially enduring over the next 10 years,
0:00:31 but setting up a culture where a company can actually evolve to meet the changing demands
0:00:36 of society and technology when it’s changing at a pace that is like unprecedented in history.
0:00:40 So I think it’s one of the most fun business challenges of all time.
0:00:45 I just get so much energy because it’s incredibly hard and it’s harder now than it’s ever been
0:00:49 to do something that lasts beyond you. But that, I think, is the ultimate measure of a company.
0:01:02 Welcome to the Knowledge Project Podcast. I’m your host, Shane Parish.
0:01:06 In a world where knowledge is power, this podcast is your toolkit for mastering the best
0:01:08 what other people have already figured out.
0:01:13 If you want to take your learning to the next level,
0:01:17 consider joining our membership program at fs.blog/membership.
0:01:24 As a member, you’ll get my personal reflections at the end of every episode, early access to episodes,
0:01:29 no ads including this, exclusive content, hand-edited transcripts, and so much more.
0:01:32 Check out the link in the show notes for more.
0:01:36 Six months after Brett Taylor realized AI was about to change everything,
0:01:41 he walked away from his co-CEO job at Salesforce to start from scratch.
0:01:46 That’s how massive this shift really is. The mastermind behind Google Maps and the
0:01:51 former chief technology officer at Facebook, Brett reveals the brutal truths about leadership,
0:01:57 AI, and what it really takes to build something that endures long after you’ve reached the top.
0:02:03 Brett’s led some of the most influential companies in tech and seen exactly what makes businesses scale,
0:02:08 what kills them from within, and why most founders don’t survive their own success.
0:02:13 In this conversation, you’ll discover why so many companies are already on life support without
0:02:19 realizing it, how first principles thinking separates the next wave of winners from everyone else,
0:02:25 and the hidden reason most acquisitions fail. We’ll explore why AI is bigger than anyone suspects,
0:02:31 plus the mindset shift that turns great engineers into exceptional CEOs. Whether you’re a founder,
0:02:37 an operator, or simply someone who wants to think sharper, this episode will change how you see your
0:02:42 business, technology, and the future. It’s time to listen and learn.
0:02:53 What was your first real aha moment with AI where you realized, holy shit, this is going to be huge?
0:02:59 I had two separate aha moments. One that I don’t think I really appreciated how huge it would be,
0:03:07 but it kind of reset my expectation, which was the launch of Dolly in the summer of 22. Is that right?
0:03:15 I might be off by Europe. I think summer of 22. And the avocado chair that they generated. And I had
0:03:22 been, well, my background is in computer science and pretty technically deep. I hadn’t been paying
0:03:31 attention to large language models. I just didn’t follow the progress after the Transformers paper. And
0:03:38 I saw that and my reaction was, I had no idea computers could do that. And that particular launch,
0:03:43 you know, seeing a generated image of an avocado chair, I don’t think I extrapolated to what,
0:03:51 you know, where we are now. But it, for me, shook me and realized I need to pay more attention to this
0:03:57 space and open AI specifically than I had been. I think I had, that was the moment where I realized like,
0:04:01 I clearly have been not paying attention to something significant. And then it was,
0:04:08 you know, six months later, coincidentally, like the month after I left Salesforce, the chat GBD came out
0:04:14 and, uh, before it became a phenomenon, although it did so quickly, but I was already, you know,
0:04:20 plugged into it. And, and, uh, and I was, uh, from then on, you know, I could not stop thinking about it.
0:04:26 Um, but that avocado chair, I don’t know why I think it was the, there was a bit of an emotional
0:04:33 moment where you saw a computer doing something that wasn’t just rule-based, but, uh, creative and,
0:04:39 um, the idea of a computer doing something and creating something from scratch, uh, was, uh,
0:04:43 well, it doesn’t seem so novel. Uh, you know, a few years later just blew my mind at the time.
0:04:49 One of the unique things about you is that you’ve started companies. You’ve been acquired by Facebook
0:04:56 and Salesforce inside those companies. You rose up to be the CTO at Facebook, the co-CEO at Salesforce.
0:05:03 Talk to me about founders working for founders and founders working within a company.
0:05:10 Yeah. It’s, uh, it’s a very, um, challenging transition for a lot of founders to make. I think
0:05:17 there’s lots of examples of acquisitions that have been really transformative from a business standpoint.
0:05:23 Uh, I think YouTube, Instagram being two of the more prominent that have clearly changed the shape of,
0:05:30 of the acquiring company. But even in those cases, you know, the founders didn’t stay around that long.
0:05:33 Uh, and those guys, that’s maybe a little unfair, you know, stick around for a little bit.
0:05:39 I think the interesting thing about being a founder is it’s not just building a business,
0:05:43 but it’s very much your identity. Um, and I think it’s very hard for people aren’t founders to
0:05:48 experience it. If you take everything very personally, you know, from the product to the
0:05:55 customers, to the press, to, uh, your competitors, uh, the, the both in inner and outer measures of,
0:06:02 of, of success. And I think when you go to being acquired, there’s a business aspect to it. And,
0:06:06 you know, can you operate within a larger company, but that’s intertwined with a sense of identity.
0:06:12 You go from being a, the founder of a company and the CEO of a company or CTO of a company,
0:06:16 whatever your, your title happens to be as one of the co-founders to be in a part of a larger
0:06:22 organization. And to fully embrace that, you actually need to change your identity. Um, you need to go from
0:06:29 being, you know, the head of Instagram or my case, the head of quip to being an employee of Salesforce or
0:06:36 going from being the CEO of friend feed to being an employee of Facebook. And what I’ve observed is
0:06:42 it’s that identity shift is a prerequisite for most of the other things. It’s not simply your ability to
0:06:48 handle the politics and bureaucracy of a bigger company or to navigate a new structure. I actually
0:06:54 think most founders don’t make that leap where they actually identify, uh, with that new thing.
0:06:57 It’s even harder for some of the employees too, because most of the time in an acquisition,
0:07:02 an employee of an acquired company didn’t choose that path. And in fact, they chose to work for
0:07:07 a different company and they, you know, the, the acquisition determined a different outcome.
0:07:11 And that’s why integrating acquisitions is so nuanced. And I would say that, uh,
0:07:18 having the experience of having been acquired, uh, you know, before and having acquired some
0:07:23 companies before when I got to Salesforce, I really tried to be self-aware about that and really tried
0:07:29 to, you know, be a part of Salesforce, you know, and tried to shift my identity and, and not be a
0:07:35 a single issue voter around quip, you know, I’d really tried to embrace it. Um, and, uh, and I
0:07:38 think it’s really hard for some founders to do it. Some founders don’t want to, honestly, you know,
0:07:43 they, uh, maybe cash the check and, and, you know, that’s the, it’s more of a transactional relationship.
0:07:50 I, um, I, I really actually am so grateful for the experience of having been at Facebook and Salesforce.
0:07:53 I learned so much, but it really took a lot of effort on my part to just, um,
0:07:59 transform my perception of myself and who I am to get that value out of the company that acquired us.
0:08:04 How did you, how did it change how you did acquisitions at Salesforce? You guys did a
0:08:09 lot of acquisitions while you were there and you’re acquiring founders and sort of startups. And I think
0:08:15 Slack was while you were there too. How did that change how you went about integrating that company
0:08:20 into the Salesforce culture? I’ll talk abstractly about talking about some specific acquisitions too, but
0:08:27 first, I think I tried to approach it with more empathy, um, and more realism. You know, uh,
0:08:35 one of the nuanced parts about acquisitions is there’s the period of, um, doing the acquisition.
0:08:39 There’s sort of the period, uh, after you’ve decided to do it, of doing due diligence. And then there’s
0:08:44 a period when it’s done and you’re integrating the company and sort of the period after one of the
0:08:51 things that I have observed is that, uh, companies doing acquisitions, often the part of deciding to do
0:08:58 it is a bit of a mutual sales process. Um, uh, you’re trying to find a fair value for the company
0:09:04 and, and there’s some back and forth there, but at the end of the day, there’s usually some objective
0:09:09 measure of that, um, influenced by a lot of factors, but, but there’s some fair value of that.
0:09:15 But what you’re trying to do is what are, uh, and corporate speak could be synergies, but like,
0:09:19 why do this? Why is one plus one greater than two? You know, that’s, that’s why you do an acquisition
0:09:26 just from first principles. It’s often a exercise in storytelling. You know, uh, you, you know,
0:09:31 bring this product together with our product and customers will, you know, find the whole greater
0:09:38 than the sum of its parts. This team applied to our sales channel, or if you’re a Google acquisition,
0:09:44 you know, imagine the traffic we can drive to, to this product experience. Uh, you know, in the case of
0:09:49 something like an Instagram, imagine our ad sales team attached to your, you know, amazing product
0:09:55 and how quickly we can help you realize that value, whatever it might be. I find that people,
0:10:01 because there’s sort of a craft of storytelling to, uh, for both sides to come to the same conclusion
0:10:08 that they should, uh, do this acquisition sometimes, uh, either simplifies or sugarcoats,
0:10:14 like some of the realities of it. Um, you know, little things like, you know, how much control
0:10:21 will the founding team of the acquired company have over those decisions? Um, uh, will it be operated
0:10:27 as a standalone business unit or will your team be sort of broken up into functional groups within the
0:10:33 larger company? And it’s sort of those little, they’re not little, but those I’ll say boring,
0:10:38 but important things that often people don’t talk enough about. And you don’t need to figure
0:10:42 out every part of an acquisition to make it successful, but often you can end up running
0:10:47 into like true third rails that you didn’t find because you were having the storytelling discussions
0:10:51 rather than getting down to brass tacks about how things will work and what’s important.
0:10:55 The other thing that I think is really important is being really clear what success looks like.
0:11:02 Um, and you know, I think, uh, sometimes it’s a business outcome, sometimes it’s a product goal,
0:11:11 but I found that, um, if you went to most of the like larger acquisitions, uh, in the valley and you,
0:11:18 two weeks after it was closed, interviewed the management team of the acquiring company and the
0:11:21 acquired company and you asked them like, what does success look like two years from now?
0:11:28 My guess is like 80% of the time you get different answers. Um, and I think, uh, it goes back to this
0:11:31 sort of storytelling thing where you’re talking about the benefits of the acquisition. We’re talking
0:11:35 about like, what does success look like? So I really tried to approach it. I tried to, um,
0:11:40 pull forward some harder conversations when, when we’re, you know, when I’m doing acquisitions or
0:11:45 even when I’m being acquired since it’s happened to be not twice so that, you know, when you’re approaching
0:11:50 it, you’d not only get the, Hey, why is one plus one equal greater than two? Everything’s gonna be
0:11:57 awesome. You know, but no, for real, like what, you know, what does success look like here? And then,
0:12:02 you know, as a founder, your job of an acquiring acquired company is to tell your team that and align
0:12:07 your team to that. And I think founders don’t take on enough accountability towards making these
0:12:11 acquisitions successful as I think they should. And, um, and it goes back to again, a certain,
0:12:17 uh, naivete, you know, it’s like, you’re, you’re, you’re not your company anymore.
0:12:22 You’re a part of something larger. And I think, you know, successful ones work when everyone embraces,
0:12:27 um, embraces that. What point in the acquisition process is that conversation? Is that after we’ve
0:12:33 signed our, our binding, you know, sort of commitment or is it, we should have that conversation
0:12:39 before. So I know what I’m walking into. My personal take is it’s not something you have,
0:12:44 you have to get to the point where the two parties want to merge, you know, and that’s a,
0:12:49 obviously a financial decision, particularly if it’s like a public company, there’s a board and
0:12:56 shareholders. Most acquisitions in the Valley are a larger firm acquiring a private firm. That’s not
0:13:00 all of them, but I would say that’s the vast majority. And in those cases, there’s often a
0:13:04 qualitative threshold where someone’s like, yeah, let’s do this. We’ve kind of have the high level
0:13:09 terms, sometimes a term sheet, you know, formally, I think it’s right after that. Um,
0:13:16 so where people have really committed to the, the key things, how much value, why are we doing this,
0:13:23 the big stuff. And there’s usually, uh, you know, many, lots of lawyers being paid lots of money to
0:13:28 turn those term sheets into, you know, uh, uh, more complete set of documents, usually more complete
0:13:33 due diligence, stuff like that. That’s a, there’s an awkward waiting period there. And
0:13:38 that’s a time I think where like the strategic decision makers in those moments can get together
0:13:44 and say, let’s talk through what this really means. And, um, the nice part about having them for all
0:13:48 parties is you’ve kind of made the commitment to each other. So it’s, you’ve, I think you have more
0:13:54 social permission to have real conversations at that point. Um, but you also haven’t consummated the
0:14:01 relationship, you know? And so, uh, there’s a, the power imbalance isn’t totally there and, and you can
0:14:06 really talk through it. And it also, I think engenders trust just because by having harder
0:14:11 conversations in those moments, you’re learning how to have real conversations and learning how each
0:14:16 other works. So that’s my, my personal opinion when to have it. So you mentioned the board, you’ve been
0:14:22 on the board of Shopify, you’re on the board of open AI, you’re a founder. What’s the role of a board
0:14:26 and how is it different when you’re on the board of a founder led company?
0:14:34 I, um, really like being involved, um, in a board. Um, and I’ve been involved in multiple boards
0:14:40 because I think I am an operator through and through. I probably self-identify as an engineer first
0:14:48 more than anything else. And I love to build learning and how to be an advisor, um, is a very different
0:14:54 vantage point that I think, uh, you see how other companies operate and you also learn how to
0:15:00 have an impact and add value without doing it yourself. Um, and it’s a very, and I’ve really,
0:15:05 I think become a better leader, you know, having learned to do that. I have really only joined,
0:15:12 uh, boards that were led by founders because typically I think they, you can speak to them,
0:15:17 but I think they sought me out because I’m a founder and I like working with founder led companies. Um,
0:15:26 I, I think the, uh, founders, I’m sure there’s lots of studies on this, but I think founders drive better
0:15:34 outcomes for, uh, companies. Um, there’s a, I think founders tend to have permission to make bolder,
0:15:39 more disruptive decisions about their business than a professional manager. There’s exceptions
0:15:44 like Satya, I think is, you know, uh, one of the greatest and not the greatest CEO of, you know,
0:15:50 our generation and, uh, as a professional manager, but you know, you look at, uh, everyone from, uh,
0:15:57 Toby Lukey to Mark Benioff to Mark Zuckerberg to, uh, Sam at open AI. And I think when you have founded
0:16:04 a company, it’s all your stakeholders, employees in particular, uh, give you the benefit of the
0:16:09 doubt. You know, you created this thing. And if you say, Hey, we need to, um, do a major shift in
0:16:16 our strategy, even hard things like, uh, layoffs, founders tend to get a lot of latitude and are
0:16:21 judged, I think differently. And, and I think rightfully so in some ways, because of the interconnection
0:16:26 of their identity to the thing that they’ve created. And so I actually really believe in founder led
0:16:32 companies. Um, one of the real interesting challenges is going from a founder led company to
0:16:35 not, and you know, Amazon has gone through that transition. Microsoft has gone through that
0:16:42 transition, um, for that reason. Uh, but I love working with founders. Um, and I, I love working
0:16:49 with people like Toby and Sam because they’re so different than me yet. Um, uh, and I can see how
0:16:54 they operate their businesses and I am inspired by it. I learned from it and obviously working for, for
0:16:58 market Salesforce, you, you have like, wow, that’s really interesting. Like most like an
0:17:03 anthropologist. Like, why did you do that? You know, I want to learn more. And so I love working
0:17:07 with founders that inspire me because I just learned so much from them. It’s such an interesting front
0:17:11 row seat into what’s happening. Do you think founders go astray when they start listening to
0:17:16 too many outside voices? And this goes back to the, I’m sure you’re aware of the Brian Chesky,
0:17:20 founder mode, the founder mode. Do you think, talk to me about that.
0:17:27 I have such a nuanced point of view on this because it is decidedly not simple. Uh, so
0:17:34 broadly speaking, I really like the spirit of founder mode, which is just having
0:17:42 deep founder led accountability for every decision at your company. Um, I think that that’s how great
0:17:49 companies operate. Uh, and when you, you know, proverbially make decisions by committee or you’re
0:17:55 more focused on process than outcomes, um, that produces all the experiences we hate as employees,
0:17:59 as customers, you know, that’s the proverbial DMV, right? You know, it’s like process over outcomes.
0:18:06 Um, and then similarly, uh, you look at the disruption in all industries right now because
0:18:12 of AI, you know, the companies that will recognize where things are clearly going to change. Like
0:18:18 everyone can see it. It’s like a slow motion car wreck. Everyone knows how it ends. You need that
0:18:24 kind of decisive breakthrough boundaries, layers of management, um, to actually make change as fast
0:18:30 as required in business right now. The issue I have not with Brian’s statements, Brian’s amazing,
0:18:37 um, is how people can sort of interpret that and sort of execute it as a caricature of what I think
0:18:44 it means. Uh, you know, there was a, I remember after Steve jobs passed away and you know, um, I don’t
0:18:50 know, I’ve met Steve a couple of times. I haven’t never worked with him in any meaningful way, you know,
0:18:54 know, but he was sort of, uh, if you believe the story is like kind of, uh, pretty hard on his
0:18:59 employees and, and very exacting. And I think a lot of founders were like mimicking that, you know,
0:19:04 done to wearing a black turtleneck and yelling at their employees. I’m like, not sure that was the
0:19:10 cause, you know, uh, I think Steve jobs taste and judgment through, you know, executed through that,
0:19:15 you know, packaging was the cause of their success and somehow. And then similarly, I think founder
0:19:21 mode can be weaponized as an excuse for just like overt micromanagement. And that probably won’t
0:19:27 lead to great outcomes either. And most great companies are filled with extremely great individual
0:19:35 contributors who make good decisions and work really hard. And, uh, uh, companies that are like solely
0:19:40 executing through the judgment of individual probably aren’t going to be able to scale to be truly great
0:19:46 companies. So I have a very nuanced point because I actually believe in founders. I believe in actually
0:19:52 that accountability that comes from the top. I believe in cultures where, you know, founders have
0:19:58 license to go in and all the way to a small decision and fix it, the infamous question mark emails from
0:20:02 Jeff Bezos, you know, that type of thing. That’s, that’s the right way to run a company, but that doesn’t
0:20:08 mean that you don’t have a culture where individuals are accountable and empowered. And, uh, you don’t
0:20:12 want, uh, you know, people trying to decide, make business decisions because of what will please our
0:20:17 dual leader, you know, which is like the caricature of this. And so, you know, after that came out,
0:20:20 I could sort of see it all happening, which is like, some people will take that and be like, you know what,
0:20:24 you’re right. I need to go down and be in the details. And some people will do it and probably make
0:20:28 everyone who works for them miserable and probably both will happen as a consequence. So.
0:20:34 That’s totally. Thank you for the detail and nuance that I love that too. Do you think engineers make
0:20:39 good leaders? I do think engineers make good leaders, but one thing I’ve seen is that
0:20:50 I think that I really believe that great CEOs and great founders, um, start usually with one specialty,
0:20:57 but become, uh, more broadly specialists in our parts of their business. Um, you know,
0:21:04 I think the, uh, businesses are multifaceted and rarely is a business’s success due to one
0:21:08 thing, uh, like engineering or product, which is where a lot of founders come from.
0:21:14 Often your go to market model is important, uh, for consumer companies, how you engage with the
0:21:22 world and public policy becomes extremely important. And I think as you see, um, uh, founders, you know,
0:21:27 grow from doing one thing to growing, to be in a real meaningful company like Airbnb or Meta or
0:21:32 something, you can see those founders really transform from being one thing to many things.
0:21:38 Um, so I do think engineers make great leaders. I think the first principles thinking the system
0:21:46 design thinking, um, really benefits things like organization design strategy. Um, and, but I also
0:21:53 think that, you know, uh, when we were speaking earlier about identity, I think one of the main
0:21:58 transitions founders need to make, especially engineers, uh, is you’re not like the product
0:22:06 manager for the company or the CEO. And at any given day, do you spend time recruiting an executive
0:22:12 because you have a need? Do you spend time, uh, on sales because that will have the biggest impact?
0:22:19 Um, do you spend time on public policy or regulation? Because if you don’t, uh, it will happen to you and,
0:22:25 and could really impact your business in a negative way. And I think engineers who are
0:22:31 unwilling to elevate their identity from what they were to what it needs to be in the moment
0:22:36 often leads to sort of plateaus, uh, in companies growth. So a hundred percent, I think engineers
0:22:43 make great, um, leaders and it’s not a coincidence. I think that most of the Silicon Valley great Silicon
0:22:49 Valley CEOs came from engineering backgrounds. Um, but I also don’t think that’s sufficient either
0:22:54 as your company scales. And I think that making that transition as all the great ones have is incredibly
0:23:00 important. To what extent are all business problems, engineering problems? That’s a deeper philosophical
0:23:06 question that I think I have the capacity to answer. Um, what is engineering? What I like about
0:23:14 approaching problems, uh, as an engineer is, uh, first principles thinking and understanding,
0:23:21 uh, the root causes of issues rather than simply addressing the symptoms of the problem. And I do think
0:23:27 that coming from a background in engineering, that is, um, everything from process, like how engineers do a
0:23:32 root cause analysis of a outage on a server is a really great way to analyze why you lost a sales
0:23:39 deal. You know, like I love the systematic approach of engineering. One thing that I think going back to
0:23:45 good ideas that can become caricatures of themselves, like one thing I’ve seen though with engineers who
0:23:53 go into other disciplines is, um, sometimes you can overanalyze decisions in some domains. Let’s just take
0:24:00 modern communications, which is driven in social media and, and very fast paced. Um, having a
0:24:07 systematic first principles discussion about every, you know, tweet you do is probably not a great comms
0:24:15 strategy. Um, and so, uh, and then similarly, um, you know, there are some aspects of say enterprise
0:24:22 software sales that, you know, aren’t rational, but they’re human, you know, like forming personal
0:24:27 relationships, you know, and, and the importance of those to building trust with a partner. It’s not
0:24:33 all just, you know, product and technology. And so I would say, I think a lot of things, uh, coming
0:24:40 with an engineer mindset can really benefit, but I do think that, uh, taking that to its like logical
0:24:47 extreme can lead to analysis paralysis, can lead to, uh, over intellectualizing some things that are
0:24:52 fundamentally human problems. And so, yeah, I think a lot can benefit from engineering, but I wouldn’t say
0:24:57 everything’s an engineering problem in my experience. You brought up first principles a couple times. You’re
0:25:03 running your third startup now, Sierra. It’s going really well. How do you use first principles in terms
0:25:12 of how to use that at work? Yeah, it’s, it’s particularly important right now because the market of AI is
0:25:22 changing so rapidly. So if you rewind two years, you know, most people hadn’t used chat GPT yet. Uh,
0:25:31 most companies hadn’t heard the phrase large language models or generative AI yet. And in two years, you have chat GPT
0:25:38 becoming one of the most popular consumer services in history faster than his than any service in history.
0:25:46 And you have across so many domains in the enterprise, uh, really rapid transformation.
0:25:52 The law is being transfer transformed. Marketing is being transformed. Customer service, which is where my
0:26:00 company Sierra works is being transformed. Software engineering is being transformed. And the amount of
0:26:06 change in such a short period of time is, uh, I think unprecedented. Uh, and, uh, perhaps I lack the
0:26:12 historical context, but it feels faster than anything I’ve experienced in my career. And so as a consequence,
0:26:19 I think, uh, if you’re responding to the facts in front of you and not thinking from first principles about
0:26:25 why we’re at this point and where it will probably be 12 months from now, the likelihood that you’ll make the
0:26:33 right strategic decision is almost zero. Uh, so, uh, as an example, uh, it’s really interesting to me that
0:26:39 with modern large language models, one of the careers that is being most transformed is software engineering.
0:26:47 Uh, and, uh, you know, one of the things I think a lot about is how many software engineers will we have
0:26:54 our company three years from now? What will the role of a software engineer be as we go from being authors of code to
0:26:57 operators of code generating machines? Um,
0:27:01 What does that mean for the type of people we should recruit?
0:27:06 And if I look at the actual craft of software engineering that we’re doing right now, um,
0:27:11 I think it’s literally a fact that it’ll be completely different two years from now.
0:27:16 Yet I think a lot of people building companies hire for the problem in front of them rather than doing
0:27:22 that. But two years is not that long. Those people that you hire now will just be getting really
0:27:28 productive a couple of years from now. So we try to think about most of our long-term business from
0:27:34 first principles, everything from, I’ll say a couple examples in our business. Our pricing model is really
0:27:40 unique and comes from first principles thinking rather than having our customers pay a license for the
0:27:45 privilege of using our platform. We only charge our customers for the outcomes. Uh, meaning if the
0:27:50 AI agent they’ve built for their customers solves the problem, there’s like a usually a pre-negotiated
0:27:55 rate for that. And that comes from the principle that in the age of AI, software isn’t just helping
0:28:02 you be more productive, but actually completing a task. Uh, what is the right and logical business
0:28:06 model for something that completes a task? Well, charging for a job well done rather than charging for
0:28:12 the privileges using the software. Um, similarly, um, you know, we, with a lot of our customers,
0:28:18 you know, we help deliver them a fully working AI agent. We don’t hand them a bunch of, uh, software
0:28:25 and say, good luck, you know, configure it yourself. And the logic there is, you know, uh, in a world where,
0:28:32 uh, making software is easier than it ever is before. And you’re delivering outcomes for your customer.
0:28:37 Um, the delivery model of software probably should change as well. And we’ve really tried to
0:28:41 reimagine what like the software company of the future should look like and trying to,
0:28:46 you know, model that in everything that we do. That’s brilliant. How do you think software
0:28:50 engineering will change? Is it you’re going to have fewer people or the people are going to be
0:28:57 organized differently or how do you see that? How geeky can I get as geeky as you want?
0:29:04 I actually wrote a blog post, uh, right before Christmas about this. I think this is an area
0:29:09 that deserves a lot more research. Uh, I’ll describe where I think we are today and smart people may
0:29:18 disagree, but a lot of the modern large language models, both the traditional large language models
0:29:24 and sort of the new reasoning models are trained on a lot of source code. And it’s a, an important input
0:29:29 to all of the knowledge that they’re trained on. Um, as a consequence, even the early models were
0:29:37 very good at generating code. Um, so, you know, uh, every single engineer at, at CIRA uses a cursor,
0:29:43 which is a great product that basically integrates with the IDE visual studio code to help you
0:29:50 generate code more quickly. Um, it feels like a local maximum, uh, in a really obvious way to me,
0:29:57 which is you have a bunch of code written by people, um, written in programming languages that
0:30:04 were designed to make it easy for people to tell a computer what to do. Probably this funniest example,
0:30:10 this is Python. Um, it almost looks like natural language, but it’s notoriously not robust. Um,
0:30:16 you know, most Python bugs are found by running the program because there’s not static type checking.
0:30:25 Um, similarly, there’s most bugs, uh, while you could run a fancy static analysis, like most bugs show up
0:30:32 simply at runtime because, uh, it’s just not designed. Um, it’s designed to be ergonomic to write. Um,
0:30:39 yet we’re using AI to generate that. Uh, we, and so we’ve sort of designed most of our computer
0:30:47 programming systems to make it easy for the author of code to type it quickly. Um, and we’re in a world
0:30:53 where actually generating code is going to like the marginal cost of doing that is going to zero,
0:30:58 but we’re still generating code and programming languages that were designed for human authors.
0:31:07 And similarly, um, if you’ve ever like looked at someone else’s code, um, which a lot of people
0:31:13 do professionally, it’s called the code review. It’s actually quite hard to do a code review. Um,
0:31:18 you know, you’re end up interpreting, you’re trying to basically put the system in your head and simulate
0:31:25 it as you’re reading the code to find errors in it. So the irony now that I’ve taken things that are
0:31:30 code programming languages that were designed for authors and now having humans do the job of
0:31:36 essentially code reviewing code written by an AI. And, and yet all of the AI is being in the code
0:31:42 generation part of it. I’m like, I’m not sure it’s, it’s great, but we’re generating a lot of
0:31:47 code with similar flaws to that we’ve been generating before from security holes to just functional bugs
0:31:54 and in greater volumes. And I think we’re, uh, what I would like to see is if you start with the
0:32:01 premise that generating code is free or, or, or going towards free, what would be the programming
0:32:06 systems that we would design? So for example, uh, you know, Rust is an example of a programming
0:32:13 language that was designed for safety, not for programming convenience. Uh, you know, my understanding is
0:32:18 that the, you know, Mozilla project, you know, there were so many security holes in
0:32:24 Firefox. They said, let’s make a programming language that’s very fast. Uh, you know, uh,
0:32:28 but everything can be checked statically, including memory safety. Well, it’s a really interesting
0:32:34 direction where you weren’t operating, like optimizing for authorship convenience or optimizing for correctness.
0:32:40 Are there programming language designs that are designed so a human looking at it can very quickly
0:32:46 evaluate, does this do what I intended it to do? There’s an area of computer science I studied in
0:32:50 college called formal verification, which at the time was turning a lot of computer programs into
0:32:55 math proofs and finding inconsistencies. And it sort of worked well, not as well as you’d hope.
0:33:02 But, you know, in a world where AI is generating a lot of code, you know, should we be investing in
0:33:08 more informal verification so that the operator of that code generator machine can more easily verify
0:33:14 that it does in fact to do what they intended is to do? And could a combination of a programming
0:33:19 language that is more structurally correct and structurally safe and exposes more primitives for
0:33:26 verification plus a tool to verify? Could you make an operator of a code generated machine 20 times
0:33:30 more productive, but more importantly, make the robustness of their output 20 times greater?
0:33:36 And then similarly, you know, there’s themes, things go in and out of fashion, but like test driven
0:33:40 development, you know, where you write your unit test first or your integration test first and then
0:33:44 write code until it fulfills the test. Most programmers I know who are really good,
0:33:49 not despise it, but it’s just like a, it sounds better than it, than it is in practice. But
0:33:54 again, writing code is free, you know, so writing tests is free, you know, how can you create a
0:34:00 programming system where the combination of great programming language design, formal verification,
0:34:05 robust tests, because you didn’t have to do the tedious part of writing them all. Could you make
0:34:12 something that made it possible to write increasingly complex systems that were increasingly robust?
0:34:15 And then similarly, like the elephant in the room for me is the anchor tenant of most of
0:34:19 these code generating systems are an IDE right now, you know, and
0:34:26 that obviously doesn’t seem as important in this world. And even with coding agents,
0:34:30 which is sort of where the world is going, it doesn’t change the fact that like, you know,
0:34:33 who’s accountable for the quality of it, who’s fixing it. And I think
0:34:39 there is a world where we can make reasonable software by just automating what we as software
0:34:46 engineers do every day. But I have a strong suspicion that if we designed these systems with the role of a
0:34:53 software engineer in mind, being an operator of a machine rather than the author of the code,
0:34:58 we could make the process much more robust and much more productive. And it feels like a research
0:35:03 problem to me, it doesn’t feel. And I think a lot of people, and for good reason, including me,
0:35:08 are just excited about the efficiency of software development going up. And I want to see the new
0:35:11 thing, though. I’m constructively dissatisfied with where we are.
0:35:16 It’s so interesting that if software AI is good enough to write the code, should we get enough to
0:35:17 check the code?
0:35:23 That’s a great, great question. But actually, I’ll, you know, it’s still funny to me that we’d be
0:35:28 generating Python, you know, just because for anyone who’s listening right now has ever operated a web
0:35:34 service running Python, it’s CPU and intensive, really inefficient. You know, should we be taking most of
0:35:41 the unsafe C code that we’ve written and converting it to a safer system like Rust? You know, if authoring
0:35:46 these things and checking it are relatively free, shouldn’t all of our programs be incredibly
0:35:52 efficient? Should they all be formally verified? Should they all be analyzed by a great agent? I do
0:35:58 think it can be turtles all the way down. You can use AI to solve most problems in AI. The thing that I’m
0:36:04 trying to figure out is like, what is the system that a human operator is using to orchestrate all those
0:36:10 tasks. And, you know, I go back to the history of software development, and most of the really
0:36:15 interesting metaphors in software development came from breakthroughs in computing. So, you know, the C
0:36:20 programming language came from Unix. And when these time sharing systems were really, it went from sort
0:36:27 of punch cards to something that were a lot more agile. Small talk came out of the development of the
0:36:34 graphical user interface at Xerox PARC. And, you know, there was a sort of a confluence of message
0:36:39 passing as a metaphor and the graphical user interface. And then there was a lot of really
0:36:46 interesting principles that came out of networking, you know, and sort of distributed systems, distributed
0:36:52 locking, sequencing. I think we should recognize that we’re in this brand new era as significant
0:36:57 as the GUI. You know, it’s like a completely new era of software development. And if you were just to say,
0:37:03 I’m going to design a programming system for this new world from first principles, what would it be?
0:37:06 And I think when we develop it, I think it will be really exciting because rather than
0:37:13 automating and turning up the speed of just generating code and with the same processes we
0:37:21 have today, I think we’ll feel native to this system and give a lot more control to the people who are
0:37:24 orchestrating the system in a way that I think will really benefit software overall.
0:37:29 Let’s dive into AI a little bit. How would you define AGI to the layman?
0:37:39 I think a reasonable definition of AGI might be that any task that a person can do at a computer,
0:37:49 that system can do on par or better. I’m not sure it’s a precise definition, but I’ll tell you where that
0:37:54 comes from and it’s flaws, but there’s not a perfect definition of AGI in my opinion,
0:37:59 or there’s not a precise definition of AGI. I’m sure there’s good answers.
0:38:06 One of the things about the G and AGI is about generalization. So can you have a system that is
0:38:14 intelligent in domains that it wasn’t explicitly trained to be intelligent on? And so I think that’s
0:38:22 one of the most important things is like given a net new domain, can this system become more competent
0:38:31 and more intelligent than a person sort of trained in that domain? And I think that’s sort of the,
0:38:35 you know, at or better than a person is certainly a good standard there. And that’s sort of the
0:38:40 definition of super intelligence. The reason I mentioned at a computer is I do think that
0:38:51 it is a bar that means like if there’s a digital interface to that system, it affords the ability
0:38:59 for AI to interact with it, which is why that’s a bar that’s reasonable to hit. I say that because
0:39:08 one of the interesting questions around AGI is how quickly it does generalize. And there are domains in
0:39:19 the world that the progress in that domain isn’t necessarily limited by intelligence, but by other
0:39:24 social artifacts. So as an example, and I’m not an expert in this area, but if you think about
0:39:33 the pharmaceutical industry, my understanding is, you know, the one of the main bottlenecks
0:39:42 is clinical trials. So no matter how intelligent a system would be in discovering new therapies,
0:39:50 it may not materially change that. And so you may have something that’s discovering new insights in math,
0:39:55 and that would be delightful and amazing. But the existence of that
0:40:01 system that’s super intelligent in one domain may not translate to all domains equally.
0:40:06 I just heard at least a snippet of a talk by Tyler Cohen, the economist. And
0:40:12 it was really interesting to hear his framing on this about which parts of the economy
0:40:18 could sort of absorb intelligence more quickly than others. And so I choose that definition of AGI,
0:40:24 recognizing that there’s not a perfect definition, because it captures the ability of
0:40:31 this intelligence to generalize, while also recognizing that the domains of society might
0:40:36 not apply with equal velocity, even once we reach that point of a system being able to have that level
0:40:37 of intelligence.
0:40:44 When I think about what artificial intelligence is limited by, or the bottlenecks, if you will,
0:40:50 I keep coming back to a couple of things. There’s regulation, there’s compute, there’s energy,
0:40:54 there’s data, and there’s LLMs. Am I missing anything?
0:40:57 Uh, so you’re saying the ingredients to AGI?
0:41:03 Yeah, like, there’s limitations on each aspect of those things. And there seem to be the main
0:41:09 contributors to the, what’s limiting us from even accelerating at this point. Is that, how do you
0:41:10 think about that?
0:41:14 Yeah, what you said is roughly how I think about it. I’ll put it into my own words, though.
0:41:25 I think the three primary inputs are data, compute, and algorithms. And data is probably obvious,
0:41:30 but you know, one of the things after the Transformers model was introduced is it afforded
0:41:37 an architecture with just much greater parallelism, which meant models could be much bigger and train
0:41:44 more quickly on much more data, um, which just led to a lot of the breakthroughs with, that’s the LLM,
0:41:50 just they’re large. And, uh, the scaling laws, you know, a couple years ago, you know, indicated like the
0:41:58 larger you make the model, um, the more intelligent it would be, and at a degree of efficiency that was, uh, tolerable.
0:42:06 Uh, and there, we are, you know, there’s lots of stuff written about this, but you know, there’s
0:42:11 in terms of just like textual content to train on, you know, the availability of new content is
0:42:16 certainly waning. And some people would say, I think there’s like a data wall. Uh, I’m not an expert in
0:42:21 that domain, but it’s been talked about a lot and you can read a lot about it. There’s a lot of interesting
0:42:27 opportunities though to generate data too. Um, so, uh, there’s a lot of people working on simulation.
0:42:31 If you think about a domain, like self-driving cars, simulation is a really interesting way to
0:42:36 generate. Uh, is that synthetic data? Is that what? Yeah, I would say that synthetic data,
0:42:42 though synthetic data has a, uh, uh, simulation and synthetic data are a little different. So
0:42:49 you can generate synthetic data, like generate a novel. Um, simulation, I would put at least in my
0:42:56 head and I’m sure that like academics might critique what I’m saying, but, uh, I’ve used simulation is
0:43:00 based on a set of principles like the laws of physics. So if you were to build a real world
0:43:06 simulation for training, a self-driving car, um, you’re not just generating arbitrary data,
0:43:10 like the roads don’t turn into loop-de-loops, you know, because that’s not possible with physics.
0:43:18 So by constraining a simulation with a set of, uh, real world constraints, the data has more efficacy,
0:43:24 you know, and so, uh, and there’s, uh, sort of a, it constrains the different permutations
0:43:29 of data you can generate from it. So it’s, I think a little bit higher quality, but then along those
0:43:37 lines, you know, uh, a lot of people wonder if you generate, uh, synthetic data, um, how much value can
0:43:43 that add to a training process? Um, you know, is it sort of, uh, regurgitating information it already
0:43:49 had? What’s really interesting about, you know, reasoning and reasoning models is I think, uh,
0:43:54 I feel really optimistic these models are generating net new ideas. And so it really affords the opportunity
0:44:00 to break through, uh, some of these, the data wall as well. So data is one thing. And I think both
0:44:06 synthetic data and simulation are really interesting opportunities to, to grow there. Then you have compute.
0:44:15 And, uh, this is, um, something that, you know, it’s why, uh, there’s so many data center investments.
0:44:21 It’s why Nvidia as a company has, has grown so much. Um, the, probably the more interesting kind
0:44:27 of breakthroughs there are these reasoning models where, uh, there’s not quite such a formal separation
0:44:33 between the training process and the inference process where you can spend more compute at the
0:44:38 time of inference to generate more intelligence, um, which has really been a breakthrough in a variety
0:44:42 of ways, I think is really interesting, but it shows you how you can run up against walls and,
0:44:47 and find new opportunities to use it. And then finally algorithms. And the biggest breakthrough
0:44:52 is obviously the transformers model attention is all you need that paper from Google that sort of led to
0:44:56 where we are now. But there’s been a number of really important papers since then, from
0:45:02 the idea of chain of thought reasoning into, um, what, uh, at open AI, what we did with the O1 model,
0:45:08 which is to, um, do some reinforcement learning on those chains of thought to, uh, really reach new
0:45:14 levels of intelligence. Um, and so I do think that I mentioned some anecdotes about some breakthroughs
0:45:20 there because my view is that each one of them has their own problems. You know, compute,
0:45:27 it’s very capital intensive. Um, and a lot of these models, the half-life of their value is pretty
0:45:32 short because new ones come out so frequently. And so, you know, you’re, you wonder like, you know,
0:45:37 can we afford, uh, what’s the business case for investing this capex? And then you have a
0:45:44 breakthrough like, uh, you know, O1 and you’re like, gosh, you know, with a distilled model and
0:45:48 moving more to inference time, it changes the economics of it. You have data. You say, gosh,
0:45:53 we’re running out of textual data to train on. Well, now we can generate reasoning. We can do
0:45:56 simulations. Oh, that’s an interesting breakthrough. And then on the algorithm side, as I mentioned,
0:46:02 just the idea of these reasoning models is really novel itself. And each of these at any given point,
0:46:06 if you talk to an expert and in one of them, and I’m an expert in none of them,
0:46:12 they will tell you the sort of current plateau that they can see on the horizon. And there usually is
0:46:16 one, I mean, you’ll talk to different people about how long the scaling laws for something
0:46:19 will continue and you’ll get slightly different opinions, but no one thinks it’s looking to last
0:46:26 forever. Um, and at each one of those, because you have so many smart people working on them,
0:46:32 you often have people discovering a breakthrough, um, in each of them. And so as a consequence, I,
0:46:39 I really do feel optimistic about the progress towards HEI because one of those plateaus might
0:46:44 extend a while if we just don’t have the key idea that we need to break through the idea that we will
0:46:50 be stuck on all three of those domains feels very unlikely to me. And in fact, what we’ve seen
0:46:55 because of the potential economic benefits of AGI is we’re in fact seeing breakthroughs in all three of
0:47:02 them. And, um, as a consequence, um, you know, you’re just seeing just the blistering pace of progress
0:47:09 that we’ve seen over the past couple of years. At what point does AI start making AI better than
0:47:15 we can make it or making it better while we’re sleeping or we can’t be too far from that?
0:47:19 Well, it might reflect back to our software engineering discussion, but you know,
0:47:25 the broadly, this is the area of AGI around self-improvement, which is meaningful from a
0:47:31 improvement standpoint, but also obviously from a safety standpoint as well. So I, um,
0:47:36 I don’t know when that will happen, but I do think, you know, by some definition,
0:47:42 you could argue that it’s happening already in the sense that every engineer in Silicon Valley is already
0:47:49 using coding agents and, um, platforms like Cursor to help them code. So it’s contributing already.
0:47:54 Um, and I imagine as, uh, coding assistants go to coding agents in the future,
0:47:58 most engineers in Silicon Valley will show up in the morning and,
0:48:04 but this is sort of the difference between, uh, you know, the assisted driving and Tesla versus like
0:48:10 self-driving, right? Like at what point do we leap from, I’m a co-pilot in this to,
0:48:12 I’m, I don’t have to do anything.
0:48:17 I mean, it’s a question that’s, there’s so much nuance to the answer. I’m not sure to answer
0:48:21 because I’m not sure you’d want to necessarily, like, I think for some software applications,
0:48:25 that’s important. But when we brought up, you know, we were talking about the active software
0:48:32 development, people have to be accountable, um, for the software that they produce. Um, and that
0:48:38 means if you’re doing something simple, like a software as a service application, that it’s secure,
0:48:45 that it’s reliable, that it, the functionality works as intended for something as meaningful as,
0:48:51 uh, you know, uh, an agent that is, um, somewhat autonomous, does it have the appropriate guard
0:48:57 rails? Um, does it actually do what the operators intended? Is there appropriate safety measures?
0:49:02 So I’m not sure there’s really any system where you’d want to turn a switch and, and go get your
0:49:08 coffee. But I do think to the point on, you know, uh, these broader safety things is I think that when
0:49:15 you think about, uh, more advanced models, we need to be developing not only more and more advanced,
0:49:22 um, safety measures and safety harnesses, but also, um, using AI to supervise AI and things like that.
0:49:28 So it’s a part, uh, probably my colleague on the board, Zico Coulter is probably a better person to talk
0:49:32 through some of the technical things, but there’s a lot of prerequisites to get to that point. And I’m not
0:49:37 sure it’s simply like the availability of the technology. Um, just because it is, uh, that at
0:49:41 the end of the day, we are accountable for the safety of the systems we produce, not just opening
0:49:47 I like every, every engineer. Um, and, and, and that’s a principle that should not change.
0:49:53 What does that mean? Like when we say safety and AI, that seems so vague in general that everybody
0:49:58 interprets it quite differently. Like, how do you think about that? And how do you think about that
0:50:05 in the world where, uh, let’s say we regulate safety in the United States and another country
0:50:11 doesn’t regulate safety? How does that affect the dynamic of it? I’ll answer broadly and then
0:50:18 go into the regulatory question. So I really like opening eyes mission, uh, which is to ensure that
0:50:24 AGI benefits all of humanity. That isn’t only about safety. Um, and, and I, and I believe intentionally.
0:50:29 So though the, the, obviously the mission was created prior to my arrival because it’s both
0:50:34 about safety, kind of Hippocratic oath first to no harm. And I don’t think one could credibly achieve
0:50:39 that mission if we created something unsafe. So I would say that’s the most important part of the
0:50:45 mission, but there’s also a lot of other aspects of benefiting humanity. Um, is it universally accessible?
0:50:52 Is there a digital divide where some people have access to AGI and some don’t? Um, uh,
0:50:58 similarly, you could argue that does it, are we maximizing the benefits and minimizing the downsides?
0:51:04 Uh, clearly, uh, AI will disrupt some job, but it also could democratize access to healthcare,
0:51:11 education, expertise. Um, so as I think about the mission, it starts with safety, but I actually like
0:51:16 thinking about it more broadly because I think at the end of the day, benefiting humanity is the mission
0:51:21 and, um, uh, safety is a prerequisite, but it’s almost like going to my analogy of the Hippocratic
0:51:28 oath. A doctor’s job is, you know, uh, to cure you first do no harm, but then to cure you and a doctor
0:51:32 that did no harm, but didn’t cure you wouldn’t be great either. So I really like to think about the,
0:51:40 uh, holistically and, um, again, uh, Zika or Sam might have a more complete answer here, but broadly,
0:51:48 I think about does the system that represents AGI, um, align with the intentions of the people
0:51:53 created it and the intentions of the people operating it, um, so that it, it does what we want and it’s
0:52:00 a tool, um, uh, that benefits humanity that, um, a tool that we’re actively using, um, to affect the
0:52:05 outcomes that we’re looking for. And that’s kind of the way I think about safety. Um, and, uh,
0:52:10 it can be meaningful things like misalignment or, or more subtle things like unintended consequences.
0:52:16 Um, and I think that latter part is probably the areas, uh, really interesting from, uh, um,
0:52:23 intellectual and ethical standpoint as well. If I look at, um, uh, what was the bridge in Canada
0:52:28 that fell down where it motivated the ring that a lot of engineers. Oh yeah. I forget the name of it,
0:52:34 but just like the, whether it’s the Tacoma Narrows Bridge in Washington or, uh, three mile island or
0:52:43 these intersections where, um, uh, we’ve engineered these, um, you know, what, what at the time people
0:52:49 hope would be positively impact humanity, but something went horribly wrong. Um, sometimes it’s
0:52:54 engineering, sometimes it’s bureaucracy, sometimes it’s a lot of things. And so I don’t think when I think
0:53:00 about safety, I don’t just look at the technical measures of it, but how does this technology
0:53:05 manifest in society? How do we make decisions around it? And you could take, put another way,
0:53:10 technology is rarely innately good or bad. It’s sort of what we do with it. Um, and I think those
0:53:15 social constructs and, uh, matter a lot as well. Um, so I think it’s a little early to tell because we
0:53:21 don’t have this kind of super intelligence right now. Um, and I think it won’t just be a technology
0:53:28 company defining how it manifests in society. And you could imagine, uh, taking a very well aligned
0:53:35 AI system and a human operator directing it towards something, um, that would, uh, objectively hurt
0:53:40 society. And, and there’s a question of like, who gets to decide who’s accountable? And it’s a
0:53:47 perennial question. I mean, it’s whether you’re deciding, uh, you know, uh, uh, should you use
0:53:51 your smartphone in school? You know, who, who should decide that? And I, there’s parents
0:53:55 who will tell you, Hey, it’s my decision. It’s my kid. And then there’s principals who will tell
0:54:00 you it’s not benefiting the school. And I’m not sure that’s going to be my place or our place,
0:54:04 but there’ll be a number of those conversations that are much deeper than that question that I
0:54:12 think we’ll need to answer. Um, as it relates to regulation, uh, there’s two, uh, not conflicting
0:54:16 forces, but two forces that exist somewhat independently, but relate to each other. One
0:54:22 is the pace of progress in AI and ensuring that, you know, uh, the, the folks working on frontier
0:54:29 models are ensuring those models do benefit humanity. And, uh, and then there’s the, uh,
0:54:36 sort of geopolitical landscape, which is, you know, do you want, uh, AGI to be created by the freedom,
0:54:43 uh, sort of, uh, the West, um, by democracies, um, or do you want it to be created by more
0:54:50 totalitarian governments? And so I think the inherent tension for regulators will be, um,
0:54:57 a sense of obligation to ensure that, you know, uh, the technology organizations creating AGI
0:55:04 are in fact focusing enough on, um, that I’m fitting humanity, all the other, uh, uh, stakeholders
0:55:11 there that whose, uh, interests that they’re accountable for and ensuring that the West remains
0:55:16 competitive. Um, and, and, uh, I think that’s a really nuanced thing. And I think, uh, you know,
0:55:23 my, my view is it’s very important that the West is, uh, leads in AI and I’m very proud of the fact that,
0:55:28 um, you know, open AI is based here in the United States and we’re investing a lot in the United
0:55:32 States. And I think that’s very important. And I also, you know, having sort of seen the inside of,
0:55:37 I think we’re really focused on benefiting humanity. So I, I tend to think that, you know,
0:55:41 it needs to be a multi-stakeholder dialogue, but I think there’s a really big risk that some
0:55:47 regulations could have the unintended consequence of, of, uh, slowing down this larger conversation.
0:55:51 But I don’t say that to be dismissive of it either. It’s actually just a impossibly hard,
0:55:55 uh, problem. And I think you’re seeing it play out as you said, and in really different ways in
0:56:00 Canada, United States, Europe, China, elsewhere. I want to come back to compute and the dollars
0:56:07 involved. So, I mean, on one hand you have, um, if I just, I could start an AI company today by,
0:56:12 you know, going, putting my credit card down and using AWS and leveraging their infrastructure,
0:56:17 which they’ve, they’ve built, they’ve spent the hundreds of billions of dollars and I get to use
0:56:24 it on a time-based model. On the other hand, you have people like open AI, uh, and Microsoft investing
0:56:32 tons of money into it that may be more proprietary or, um, how do you think about the different models
0:56:38 competing? And then the one that really throws me for a bit of a loop is Facebook. So Facebook has
0:56:47 spent meta, I’m like aging myself here. So meta comes along and, you know, possibly for the good
0:56:53 of humanity, but like I tend to think Zuck is like incredibly smart. So I don’t think, I don’t think
0:56:59 he’s spending, you know, a hundred billion dollars to develop a free model and give it away to society.
0:57:05 How do you think about that in terms of return on capital and return on investment?
0:57:11 It’s a really complicated business to be in just given the capex required to build
0:57:15 a frontier model. But let me just start with a couple definitions of terms that I think are useful.
0:57:22 Um, I think most large language models I would call foundation models. And I like the word foundation
0:57:28 because I think it will be foundational to most intelligent systems going forward. And
0:57:37 most people building modern models, uh, particularly if they involve language image or, or audio shouldn’t
0:57:42 start from building a model from scratch. They should pick a foundation model, either use it off the shelf
0:57:47 or fine tune it. Um, and so it’s truly foundational in many ways. Uh, in the same way,
0:57:52 most people don’t build their own servers anymore. They lease them from one of the cloud infrastructure
0:57:58 providers. I think foundation models will be something trained by companies that have a lot
0:58:06 of capex and leased by a broad range of customers who have a broad range of use cases. Um, and I think
0:58:12 that leads in the same way that data center builders having a lot of data centers enabled you to have
0:58:17 the capital scale to build more data centers. I think the same will largely be true of, uh, you know, building
0:58:23 the huge clusters to, to do training and things like that. Foundation models, I think are somewhat distinct
0:58:30 from frontier models and frontier models. I think it’s a term credited to, to Reed Hoffman, but, uh, I may be mistaken
0:58:35 on that. That’s where I heard it from. And these are the models that are usually like the one or two that are
0:58:43 clearly the leading edge. Oh, three, as an example from open AI. And these frontier models are being built by labs
0:58:52 who are trying to build AGI that benefits humanity. And I think if you’re deciding whether you’re building a
0:59:00 foundation model and, uh, what your business models around it, it’s very different business than I’m going to go pursue AGI.
0:59:07 Uh, because if you’re pursuing AGI, really, there’s only one answer, which is to build and train and move to
0:59:13 the next front, you know, horizon, because if you can truly build something that is AGI, the economic value
0:59:20 is so great. Uh, I think there’s a really clear business case there. If you’re pre-training a foundation
0:59:28 model, that’s the fourth best, uh, that’s going to cost you a lot of money. And the return on that
0:59:36 investment is, is probably fairly questionable because why use your fourth best large language model versus a frontier
0:59:43 model or an open source one, uh, for meta. And as a consequence of that, I think we have probably have too many people
0:59:49 building models right now. There’s already been some consolidation actually of companies being folded into Amazon and
0:59:56 Microsoft and others. But I do think it will play out a bit like the cloud infrastructure business where a very small number of
1:00:04 companies with very large CapEx budgets, uh, are responsible for both building and operating these data centers.
1:00:10 And then developers and, and in consumers will use things like chat GPT as a consumer or as a developer,
1:00:16 you’ll license, uh, and rent, you know, one of these models, uh, in the cloud. Um,
1:00:21 how it will play out is a really great question. You know, I think the, uh, I heard one investor
1:00:28 talk about these as like the fastest depreciating assets of all time. Um, uh, on the other hand,
1:00:34 you know, I, uh, if you look at the revenue scale of something like an open AI and, and what I’ve read
1:00:41 about places like Anthropic, let alone Microsoft and Amazon, it’s pretty incredible as well. And so you
1:00:46 can’t really, if you’re one of those firms, you can’t afford to sit on the sidelines as, as the
1:00:51 world transforms. But I, I would have a hard time personally like funding, uh, a startup that says
1:00:56 I’m going to do pre-training. You know, it’s, it’s, uh, it’s, I don’t really know like what’s your,
1:01:00 what’s your, um, differentiation in this marketplace. And I think a lot of those companies,
1:01:05 you’re already seeing them consolidate because they have the cost structure of a pharmaceutical
1:01:09 company, but not the business model. But this is just it though, right? Like open AI has a revenue
1:01:16 model around a revenue model. Microsoft has a revenue model around their AI investments. They
1:01:24 just updated the price of teams with copilot. You know, uh, Amazon has a revenue model around
1:01:29 AI in a sense to getting other people to pay for it through AWS. And then they’re getting the advantages
1:01:35 of it, uh, at Amazon too, from a consumer point of view and all the millions of projects. Bezos
1:01:40 was doing an interview last week. He said, there’s every project at Amazon basically has an AI component
1:01:46 to it now. Facebook on the other hand has spent all of this money already. And with, you know, an endless
1:01:54 amount, presumably insight or like not insight and endless amount to go, but they don’t have a revenue
1:01:59 model specifically around AI where it would have been cheaper obviously for them to use a different
1:02:05 model, but that would have required presumably giving data away or like, you know, I’m just trying to work
1:02:12 through it from Zuck’s point of view. If you know, I actually will take Mark at his word and you know,
1:02:18 that post he wrote about open source, I think was very well written and encourage people to read it.
1:02:23 I think that’s a strategy. And you know, if you look at Facebook, um, you know, you’ve got me saying
1:02:28 Facebook too. So that was, that was what it was called. You know, the company has always really
1:02:34 embraced open source. And if I look at really popular things from react to, uh, you know, now
1:02:40 the llama models, it’s always been a big part of their strategy to court developers around sort of their
1:02:46 ecosystem. And Mark articulated some of the strategy there and I’m sure there’s elements of commoditizing
1:02:52 your compliment. But I also think that, you know, if you can attract developers towards models, there’s
1:02:58 a strength. Um, I, you know, I’m not really on the inside there, so I don’t really have a perspective
1:03:04 on it other than I actually think it’s really great that there’s different players with different
1:03:10 incentives, all investing so much. And I think it is really furthering the cause of like bringing these
1:03:18 amazing tools to society. And, um, but a lot changes. I mean, if you look at the price of GPT
1:03:24 4.0 mini, you know, it is, uh, so much higher quality than like the highest quality model two years
1:03:30 ago and much cheaper. Um, I, I haven’t done the math on it, but it’s probably cheaper to use that than
1:03:37 to host self host any of the open source models. So even, even the existence of the open source models,
1:03:43 it’s not free. I mean, inference costs money. And so there’s a lot of complexity here. And, and
1:03:48 actually I have the email even being relatively close to stuff. Like I have no idea where things
1:03:53 are going, but you know, it’s, um, you could talk to a smart engineer and they’ll tell you,
1:03:59 oh yeah, if you built your own servers, you’ll spend less than renting them from say Amazon Web
1:04:04 Services or Azure. That’s sort of true in absolute terms, but misses the fact, like, do you want someone
1:04:09 on your team building servers? Oh, and in fact, if you change the way your service works and you
1:04:14 need a different SKU, like you all of a sudden are doing training and you need Nvidia, uh, you know,
1:04:21 H 100s. Now all of a sudden you’re built servers like this, you know, asset that’s worthless. So
1:04:26 I think with a lot of these models, you know, the presence of open source is incredibly important and,
1:04:32 and, uh, and, uh, and I really appreciate it. I also think like the economics of AR pretty complex
1:04:39 because the hardware is very unique. The cost to serve is much higher. Um, techniques like
1:04:45 distillation have really changed the economics of models, whether or not it’s open source or, or,
1:04:52 uh, hosted and, and, and leased. Um, so it’s, I think broadly speaking for developers,
1:04:57 it’s kind of an amazing time right now because you have a, like a menu of options that’s incredibly
1:05:02 wide. And I actually think of it as, you know, just like in cloud computing, you’ll end up with
1:05:06 a price performance quality trade off. And for any given engineering talents, they’ll have a different
1:05:13 answer and that’s appropriate. And some people use, uh, open source Kafka. Some people work with
1:05:19 confluent. Um, great. You know, like that’s just the way these things work, you know, and, um,
1:05:22 so you don’t think AGI is going to be like a winner take all. You think there’s going to be multiple
1:05:30 options that have by definition, whatever the definition is of AGI. Well, first I think open AI,
1:05:36 I believe will play a huge part in it because, uh, there’s both the technology, which I think open AI
1:05:44 continues to lead on. Um, but also chat GPT, which has become synonymous with AI for most consumers.
1:05:50 But more than that, um, it is the way most people access AI, um, today. And so one of the interesting
1:05:57 things like what is AGI, we talked about, you know, opinions on what the definition might be. But the
1:06:02 other question is like, how do you use that? Like, what do you, what is, uh, what is the packaging? Um,
1:06:09 and some of, uh, intelligence will be simply the outcomes of it, like a discovery of a new drug,
1:06:15 which would be, you know, remarkable and hopefully we can cure some illnesses. Uh, but others will be
1:06:20 just how you as an individual access it. And, you know, I, most of the people I know, like if they’re
1:06:26 signing an apartment lease, we’ll put it into chat GPT, you get a legal opinion. Uh, if you get, you know,
1:06:34 lab results from your doctor, you can get a second opinion on, on chat GPT, um, Clay and I use, uh,
1:06:42 the O1 pro mode for like criticizing our strategy at Sierra all the time. And so for me, what’s so
1:06:48 remarkable about chat GPT, which was this, you know, quirkly named research preview that has come to be
1:06:56 synonymous with AI as I do think that it will be the delivery mechanism for AGI when it’s produced and, uh,
1:07:02 not just because of the many researchers at open AI, but because of the amazing like utility it’s become
1:07:06 from individuals. And I think that’s really neat because I don’t know if it would have been obvious
1:07:12 if we were having this conversation three years ago, um, you know, and you were talking about
1:07:17 artificial general intelligence. I’m not sure either of us would have envisioned something so simple as
1:07:24 a form factor, uh, to absorb it that you just talk to it. Um, so I think it’s great. And especially as I
1:07:30 think about the mission of open AI, which is to ensure that AGI benefits humanity, what a simple,
1:07:37 accessible form factor, there’s free tiers of it. Like what a kick ass way to benefit humanity. So I
1:07:42 really think that will be central to what we come as society to define, uh, as a AGI.
1:07:50 You mentioned using it at Sierra to critique your, your business strategy. What do you know about prompting
1:07:54 that other people miss? I mean, you must have the best prompts.
1:07:56 People think that, you know, cause I’m affiliated with it.
1:07:59 You’re not going like, here’s my strategy. What do you think? What are you putting in there?
1:08:07 Um, I often with the, uh, reasoning models, which are slower, we’ll use a faster model
1:08:17 first GPT four O to refine my prompts. Um, so, uh, over the holidays, um, partly because I was thinking
1:08:21 about the future of software engineering, I’ve, I’ve written a lot of compilers in my time. I’m like
1:08:27 written enough that I, you know, it’s like, uh, uh, uh, it’s, it’s easy for me. So I decided to see
1:08:36 if I could have a one pro mode, um, generate end to end, a compiler front end, parsing the grammar,
1:08:42 um, checking for semantic correctness, generating an intermediate representation, and then using,
1:08:49 uh, uh, LLVM, which is sort of a compiler collection, um, that’s very popular to actually do,
1:08:56 you know, run, run it all. And I would spend a lot of time iterating on four O to sort of like
1:09:02 refine and, and make more complete and specific what I was looking for. And then I would put it
1:09:07 into a one pro mode, go get my coffee and, you know, come back and get it. I’m not sure if that’s
1:09:11 a viable technique, but it’s really interesting because I do think in the spirit of AI being the
1:09:20 solution to more problems than AI, um, having a, uh, lower latency, simpler model help refine,
1:09:24 essentially. I like to think of it as like, you’re like a product manager and you’re asking,
1:09:29 you know, an engineer what to do is your, is your product requirements document complete and specific
1:09:35 enough. And, uh, waiting for it is sometimes slower than, and so I, I like doing it in stages like
1:09:38 that. So that’s my, my chip at some point. There’s probably someone from open AI listening. He’s gonna
1:09:43 like roll their eyes, but that’s just, uh, that’s, uh, who can I talk to at open AI? That’s like the prompt.
1:09:50 Yeah. I’m like so curious about this because I’ve, I’ve actually taken recently to, uh, getting
1:09:56 open AI or chat GBT, I guess, if you want to call it that I’ve been getting chat GBT to write the prompt
1:10:03 for me. So I’ll prompt it with, I’m prompting an AI. Yeah. Here are the key things similar to my
1:10:08 technique. I want to accomplish what would an excellent prompt look like. And then I’ll copy
1:10:15 paste that prompt that it gives me back into the system. Uh, but I’m like, I wonder what I’m missing
1:10:18 here. Right. Like it’s a good technique. I mean, there’s lots of engine techniques like that. Like
1:10:25 self-reflection is a technique where you have a model observe and critique, you know, a decision
1:10:30 like a chain of thought. Uh, so in general, you know, that mechanism of self-reflection is I think
1:10:39 a really effective technique. You know, at Sierra, we help companies build customer facing AI agents. So, um,
1:10:44 if you’re setting up a Sonos speaker, you’ll now chat with an AI. If you’re a SiriusXM subscriber,
1:10:50 you can chat with Harmony, who’s their AI to manage your account. Um, we use all these tricks,
1:10:56 you know, self-reflection to detect things like hallucination or decision-making, generating
1:11:02 chains of thought for more complex tasks to ensure that it’s, you know, you’re putting as much, uh,
1:11:10 compute and cognitive load into important tricks. So, uh, you know, we’re the, there’s a whole
1:11:15 industry around sort of figuring out how do you exact the, like robustness and, and, um,
1:11:18 precision out of these models. So it’s really fun, but changing rapidly.
1:11:25 Hypothetical question. You, you’ve been hired to lead or advise a country, uh, that wants to become
1:11:33 an AI superpower. What sort of, um, steps would you take? What sort of policies would you think
1:11:37 would help create that? How would you bring investment from all over the world into that country?
1:11:41 And researchers, right? Like, so now all of a sudden you’re competing. It’s not the United States.
1:11:46 Like, how do you, how do you sort of set up a country like from first principles all the way
1:11:48 back to like, what does that look like? What are the key variables?
1:11:54 Well, I mean, especially, uh, this is definitely outside of my domain of expertise, but I would say
1:12:04 one of the key ingredients to modern AI is compute, um, which is a noun that wasn’t a noun until recently,
1:12:11 but now compute is a noun. And, uh, you know, I do think that’s one area where policymakers can,
1:12:19 um, uh, because it involves a lot of things that, uh, touch, uh, federal and local governments like power,
1:12:28 land. Um, and then similarly attracting the capital, which is immense to finance, uh, to the real estate,
1:12:35 to purchase the, uh, you know, uh, compute itself, um, and then to sort of operate the data center. And again,
1:12:41 there’s really immense power requirements for these data centers as well. Um, and then, you know,
1:12:47 it’s attracting sort of the right researchers and research labs to, you know, leverage that. But in
1:12:52 general, where there is compute, the research labs will find you, you know, and so I think that’s it.
1:12:56 And then there’s a lot of national security implications too, just because, you know,
1:13:01 these models are very sensitive, at least the frontier models are. And so, um, you know,
1:13:08 how you, your place in the geopolitical landscape is quite important. Like will research labs and,
1:13:13 uh, will the U S government be comfortable with training happening there and, and export restrictions
1:13:20 and things like that. But I think a lot of it comes down to infrastructure, uh, as it relates to policy,
1:13:28 is my intuition. Uh, you know, I think right now so much of AI is constrained on, on infrastructure that,
1:13:35 that is the input to a lot of, uh, of this stuff. Um, uh, and then there’s a lot around,
1:13:39 you know, attracting talent and all that. But as I said, you know, you look at the research labs,
1:13:44 it’s not that many people, actually, it’s a lot, but the compute is a limited resource right now.
1:13:49 That’s a really good way to think about it. I think about this from the lens of Canada,
1:13:56 right? Which is like, we don’t have enough going on in, in AI. We, we tend to lose most of our great
1:14:02 people to the states, uh, who then go to set up infrastructure here for whatever reason and don’t
1:14:09 bring it back to Canada. And I, I wonder how Canada can compete better. So this is like sort of the lens
1:14:15 I like look at these questions through, how do you see that the next generation of education?
1:14:20 Like if you were setting up a school today from scratch and again, hypothetical, not your domain
1:14:27 of expertise, but like using your lens on AI, how do you think about this? So like what skills will kids
1:14:32 need in the future and what skills do we probably don’t need to teach them anymore that we have been
1:14:39 teaching them? Well, I’ll start with, uh, the benefits that I think are probably obvious, but I’m incredibly
1:14:48 excited about. I think education can become much more personalized. Oh, totally. Have you seen synthesis
1:14:53 tutor by the way? No, I have not. Oh, so they developed this, uh, synthesis, this AI company developed this
1:15:00 tutor, which actually teaches kids. And it’s so good that El Salvador, the country just recently adopted
1:15:06 it and replaced their teachers and, uh, like it’ll teach you, but it teaches you specific to what you’re missing.
1:15:11 So it’s not like every lesson’s the same. It’s like, well, you’re not understanding this foundational concept.
1:15:17 So it’s like K through five or six right now. That’s amazing. And you know, I actually, and the results are like off the charts.
1:15:22 Well, it doesn’t surprise me. And I, I don’t actually view it as like necessarily replacing a teacher,
1:15:26 but my view is if you have a teacher with 28 kids in his or her class,
1:15:33 the likelihood that they all learn the same way or learn at the same pace is very unlikely.
1:15:39 And, you know, I can really think of a, say an English teacher or history teacher orchestrating
1:15:45 their learning journeys through a topic, say AP European history in the United States,
1:15:50 there’s a curriculum, they need to learn it. Um, how someone will remember something or understand
1:15:57 the significance of Martin Luther, you know, is very different. And, um, you can, you know,
1:16:03 generate a audio podcast for someone who might be an audio auditory learner. Um,
1:16:08 you can create cue cards for someone who needs that kind of repetition, repetition. Um,
1:16:14 you can visualize, uh, key moments in history, um, for people who just maybe want to more viscerally
1:16:19 appreciate why this was a meaningful event rather than this dry piece of history. And all of that,
1:16:23 as you said, can be personalized to the way you learn and how you learn. And I think it’s just
1:16:30 incredibly powerful. And so one of the things I think is neat about AI is it’s, uh, democratizing
1:16:35 access to a lot of things that used to be fairly exclusive. A lot of wealthy people, if their child
1:16:40 was having trouble in school, would pay for a tutor, a math tutor, science tutor. Um, and, you know,
1:16:46 you know, if you look at, uh, kids who are trying to get into, you know, uh, big name colleges, you
1:16:51 know, if you have the means, you’ll have someone prep you for the SATs or help you with your college
1:16:58 essays, all of that should be democratized if we’re doing our jobs well. And it means that we’re not
1:17:05 limiting people’s opportunity from by their means. And I think that’s a, just the, the most, uh,
1:17:09 American thing ever, uh, Canadian as well. Um, it’s the most incredible thing.
1:17:14 It’s the most incredible thing, humanity. And, and so I, I just think education will change for
1:17:20 the positive in so many ways. Um, because, uh, I, I actually, with my kids walking around when they
1:17:24 ask, uh, you know, if you have little kids, they ask why, why, why? And, you know, there’s some point
1:17:28 a parent just starts making up the answer or being dismissive. And like, we have chat TVT out and it’s
1:17:34 like the best when you’re traveling and put on advanced voice mode and be like, ask a hundred percent.
1:17:39 And I’m listening to, you know, it’s like you’re, uh, you live through your children’s curiosity.
1:17:44 And, um, you know, my daughter went to high school and came home with Shakespeare for the first time.
1:17:49 And I was, she asked me a question. I was like, I, I felt this is like total inadequacy. I was like,
1:17:54 I was very bad at this the first time. And then we put it into chat GPT and it was the most thoughtful
1:17:59 answer. And she could ask follow-up questions. And I actually was, you know, with her because I was
1:18:02 like, Oh, I forgot about that. You know, didn’t even think about that. So I, I just think it’s
1:18:09 incredible. And I would like to, uh, in public school systems, uh, I think it’s really, I think
1:18:16 it would be a really great, uh, when public school systems formally adopt these things so that they lean
1:18:25 into, uh, tools like chat GPT, uh, as mechanisms to L like raise, uh, the performance level of their
1:18:30 classroom. And, and hopefully you’ll see it in things like test scores and other things because,
1:18:35 uh, kids can get the extra time, even if the school system can’t afford it for everyone.
1:18:40 Uh, and then most importantly, kids care getting explanations according to their style of learning,
1:18:44 which I think will be, um, quite, uh, important as well. As it relates to skills,
1:18:50 it’s really hard to predict right now. And I, I would say that I do think learning how to learn and
1:18:54 learning how to think will continue to be important. So I think most of, you know,
1:19:00 primary and secondary education shouldn’t and is not vocational necessarily. Um, some of it is,
1:19:06 uh, you know, I took auto shop and all of that and I’m glad I did, but I couldn’t fix my electric
1:19:10 car today with that knowledge, you know, things change and I don’t think it needs to be purely,
1:19:16 you know, um, non-vocational, but you know, the basics of learning how to think, uh, learning,
1:19:25 um, uh, uh, writing, reading, math, physics, uh, chemistry, biology, not because you need to memorize
1:19:32 it, but understand the mechanisms that, uh, uh, create the world that we live in is, is quite
1:19:42 important. Um, I do think that the, there’s a risk of people sort of, uh, becoming ossified in the tools
1:19:49 that they use. Um, so, you know, uh, let’s go back to our discussion of software engineering for a
1:19:54 second, but I’ll give other examples. You know, if you define your role as a software engineer is how
1:20:01 quickly you type into your IDE, the next few years might leave you behind, you know, because that, um,
1:20:06 that is no longer a differentiated, you know, part of the software engineering experience or will not
1:20:13 be, but your judgment as a software engineer will continue to be, uh, incredibly important in your
1:20:21 agency and making a decision about what to build, how to build it, um, how to architect it, uh, maybe
1:20:27 using AI models as a creative foil. And so I think that, uh, just in the same way, if you’re an accountant,
1:20:32 you know, using Excel doesn’t make you less of an accountant, uh, and, um, and just because you
1:20:38 didn’t, you know, handcraft that math equation, it doesn’t make the results any less valuable to your
1:20:44 clients. Um, and so I think we’re going to go through this transformation where I think the, um,
1:20:50 the tools that we use to create value in the world will change dramatically. And I think some people who
1:20:57 define their jobs by their ability to use the last generation’s tools really, really effectively,
1:21:06 um, will, will be disrupted. But I, I think we, if we can empower people and, and to reskill, um, and also
1:21:10 broaden the aperture by which they define the value they’re providing to the world, I think a lot of
1:21:16 people can make the transition. The thing that is sort of uncomfortable, not really in education,
1:21:23 or it’s just earlier in, in most people’s lives. It’s just, I think the pace of change, uh, exceeds
1:21:29 that of most technology transitions. And I think it’s unreasonable, um, to expect most people to
1:21:34 change the way they work that quickly. And so I think the, the next five years, I think will be,
1:21:40 you know, for some jobs really disruptive and tumultuous. But if you take the longer view and you
1:21:46 fast forward 25 or 50 years, I’m incredibly optimistic. I think it’s the, the change will
1:21:53 require, um, from society, from companies and from individuals, like an open-mindedness about
1:21:58 reskilling and, and re-imagining their job to the lens of this like dramatically different new technology.
1:22:04 At what point do we get to, I mean, we’re probably on the cusp of it now and it’s happening in pockets,
1:22:10 but what point do we start solving problems that humans haven’t been able to solve or eliminating
1:22:15 paths that we’re on maybe with medical research that it’s like, no, that, that the, this whole
1:22:22 thing you’ve spent $30 billion on, you know, based on this 1972 study that was fabricated. But that one
1:22:27 study had all these derivative studies and like, I’m telling you it’s false, you know, because I can look
1:22:32 at it through an objective lens and get rid of these 30 billion while you’re smiling.
1:22:40 So, Oh no, I just, I hope soon. I mean, I hope, I mean, I, uh, there was a, a lot of, there’s a,
1:22:44 one of the models I can’t remember which one introduced a very long context window. And there’s a lot of
1:22:50 people on X over the weekend putting in their, uh, thesis, you know, like grads, grad school thesis in
1:22:55 there. And, and it was actually critiquing them with like surprising levels of fidelity. Uh,
1:23:02 and, uh, I think we’re sort of there perhaps with the right, um, right tools, but certainly over the
1:23:06 the next few years, I, you know, we talked about how, what does it mean to generalize AI?
1:23:15 Certainly in the areas of, um, science that are, you know, largely represented through text and
1:23:20 digital technology, like math being probably the most, uh, most applicable, there’s not really
1:23:24 anything keeping AI from getting really good at math. There’s not really an interface to the real
1:23:30 world. You don’t need to do a clinical trial to verify something’s correct. So I feel a ton of optimism
1:23:36 there. Um, it’ll be really interesting in like, you know, areas of like theoretical physics. Um,
1:23:40 uh, you’ll tend to, you’ll continue to have the divide between the applied and the theoretical
1:23:45 people. But I think there could be like really interesting new ideas there and perhaps some,
1:23:50 uh, finding logical inconsistencies with some of the, you know, uh, fashionable theories,
1:23:56 which has happened many times over the past few decades. Um, I think, I think we’ll get there soon.
1:24:01 And I actually, I, um, what’s really neat about is most of the scientists, I know people
1:24:05 who are actually like doing science, like they’re the most excited about these technologies and I,
1:24:09 they’re using them already. And I think that’s really neat. And I think we’re hopefully going to be,
1:24:15 I really hope we see more breakthroughs in science. One of the things I am not an expert in, but I’ve
1:24:25 read a lot, like, uh, a lot as a amateur about is just the slowdown in scientific breakthroughs over
1:24:29 the past, you know, a few decades and, and some theories that it’s because of the degree of
1:24:34 specialization that we demand of grad students and things like that. And I hope with, you know,
1:24:43 in general with AI, um, democratizing access to expertise, um, I, I have a completely personal
1:24:50 theory that it will benefit deep generalists in a lot of ways too, because your ability to understand
1:24:57 a fair amount and a lot of domains and leveraging AI, um, knowing where to prompt the AI to,
1:25:04 to go explore and, and, um, bringing together those domains, it will start to shift sort of the
1:25:10 intellectual power from people who are extremely deep to people who actually can, uh, orchestrate,
1:25:15 uh, intelligence between lots of different domains for breakthroughs. I think that’ll be really good
1:25:20 for society because most scientific breakthroughs aren’t, they tend to be, you know, cross-pollinating
1:25:24 very important ideas from a lot of different domains, which I think will be really exciting.
1:25:26 How important is the context window?
1:25:33 I think it could be quite important. Um, especially it certainly simplifies working with an AI. If
1:25:39 you can just give it everything and instruct it to do something. Um, and so, and, and assuming it
1:25:46 works, you know, you can extend a context window and it can, um, uh, the, the tension can be spread
1:25:52 fairly thin and, and, and the robustness of the answer can be questionable. So, but assuming,
1:25:58 let’s just for argument’s sake, you know, perfect robustness, um, I think it can really simplify the
1:26:05 interface, uh, to AI. The, not all uses, I also think that we’re talking about open source models and
1:26:13 APIs. Um, I also think that, you know, most, what I’m excited about in the software industry is not
1:26:19 not necessarily a large language model with a prompt and a response being the product of AI,
1:26:25 but actually end in closed loose systems that use large language models as pieces of infrastructure.
1:26:30 And I actually think that a lot of the value in software will be that. And for many of those
1:26:35 applications, the context window size can matter, but often because you have contextual awareness of
1:26:40 the process that you’re executing, um, yeah, context window is a little bit less important.
1:26:45 So I think it matters a lot to intelligence. Um, you know, there’s a, I can’t remember someone,
1:26:49 one of the, some researcher said, you know, you put all of human knowledge in the context window,
1:26:54 and you ask it to invent the next thing. You know, it’s a, uh, obviously a reductive, uh, thought,
1:26:59 but, but interesting. Um, uh, but I actually, I’m equally excited about sort of the industrial
1:27:02 applications of large language models, sort of like my company, Sierra. And if you’re,
1:27:08 you’re, um, returning a pair of shoes at a retailer and it’s a process that’s fairly complicated and,
1:27:15 uh, you know, is it within the return window? Uh, you know, uh, do you want to return it in store?
1:27:19 Do you want to send it? Do you want to print to your QR code? Blah, blah, blah, blah. Um,
1:27:23 the orchestration of that is as significant as the models themselves. And I actually think as we,
1:27:28 um, just like, uh, uh, computers, you know, there’s going to be a lot of things where computers are a
1:27:32 part of the experience, but it’s not like manifesting itself as a computer. So I,
1:27:36 I’m actually equally excited about those. And I think context window is slightly less important
1:27:41 than those applications. Do you think that the output from AI should be copyrightable
1:27:46 or patentable or let, let me just take an example. If I go to the U S patent office,
1:27:53 I download a patent for, let’s say the AeroPress and I upload it to, uh, Oh one pro. And I say,
1:27:58 I can’t upload it yet. Cause you don’t let me do the PDS, but I upload it to four. And, uh,
1:28:05 so I say, Hey, what’s the next logical leap that I could patent office. It would give me back diagrams
1:28:10 and an output. And presumably if I look at that and I’m like, yeah, that’s legit. I want to file
1:28:14 that patent. Can I, I don’t know to answer that question. I’m not an expert in sort of intellectual
1:28:21 property, but I, uh, uh, I think there will be an interesting question of, was that your idea
1:28:27 because you used a tool to do it? I think the answer is probably yes, that you, you use the tool
1:28:37 to do it. But I also think that the, um, uh, uh, in general, like the sort of marginal cost of intelligence
1:28:43 will go down a lot. So a lot of the, you know, I think in general, like we’ll, we’ll be in this
1:28:50 renaissance of, of new ideas and intelligence being produced. And so, uh, I think that’s broadly a good
1:28:55 thing. And I think, you know, the marginal value of that insight that you had might be lower than it
1:29:00 was, you know, and years ago, what I was hoping you would say is that, you know, that’s going to become
1:29:05 less and less important because I feel like all the patent trolls and all of this stuff that slows down
1:29:10 innovation in some ways, uh, obviously like there’s legitimate patents that people infringe on and
1:29:15 there should be legal recourse. But if I could just go and patent like a hundred things a day,
1:29:19 because it seems like that should not be allowed. This is what I’m saying though.
1:29:24 Well, in general, I think that, you know, companies, you know, I think patents make sense
1:29:29 if it’s protecting something that’s an active use that you, you know, invented and you’re, you’re trying to
1:29:35 uh, you know, like the standard, you know, uh, legal rationale for patents, just generating a
1:29:40 bunch of ideas and patenting. It seems destructive to the value of it. So here’s the idea I had last
1:29:45 night to counter this because I was like, I don’t want somebody doing this. Uh, and I was thinking
1:29:50 like, what if prior art eliminates patents? Yeah. So I was like, what if I just set off like an
1:29:55 instance and just publish it on a website? Nobody has to read that website. Here’s a billion ideas.
1:30:00 Exactly. But it’s like basically patenting like anything, not patenting,
1:30:05 but it’s creating prior art for everything. So like, you can’t compete on that anymore.
1:30:10 I don’t know. I was like thinking about that. I thought it was fun. Um, tell me about the
1:30:16 Google maps story. This is like now legend and I want to hear it from you. Uh, this is my weekend
1:30:22 coding. Is that what you want to hear about? Yeah. Um, yeah. So, uh, I’ll start with just
1:30:28 like the story of Google maps, the abbreviated version. Uh, we had launched a product at Google
1:30:33 called Google local, which was sort of a yellow pages, uh, search engine. Uh, you probably,
1:30:37 probably most listeners don’t even know what yellow pages are, but it was a thing back then.
1:30:44 And, um, we had licensed maps from map quest, which was the dominant sort of mapping provider at the
1:30:49 time. And it was sort of an eyesore on the experience and also always felt like it could
1:30:54 be a more meaningful part of the kind of local search and navigation experience on Google. So
1:30:58 Larry Page in particular was really pushing us to really invest more in maps.
1:31:05 Um, we found this, uh, small company with a, like four people in it, if I’m remembering correctly,
1:31:10 started by Lars and Jens Rasmussen called where to technologies where, um, they had made a windows
1:31:16 application called expedition. Um, that was just a beautiful, uh, mapping product. Um,
1:31:20 it was running on windows long after it was sort of out of fashion to make windows apps, but they,
1:31:25 they were sort of where the technology they’re comfortable with, but they’re really, um,
1:31:32 their, their maps modeled the A to Z maps and in the UK were just beautiful. And they just had a lot
1:31:39 of passion for mapping. So we did a little aqua hire of them and took together the Google local team and
1:31:44 Lars and Jens’s team and, and said, okay, like, let’s take the good ideas from this windows app and
1:31:49 the good ideas from Google local. And like, let’s bring them together to make something completely
1:31:54 new. And that, and that’s what became Google maps. But there was a couple of idiosyncrasies
1:32:00 in the integration because it was a, um, Windows app. It really helped and hurt us in a number of
1:32:05 ways. Like one of the ways it helped us is the reason why Google maps, we were able to drag the map
1:32:11 and it like, uh, was so much more interactive than any web application that preceded it was
1:32:17 the standard that we needed to hit from interactivity was set by a native windows app, not set by the
1:32:25 legacy, uh, you know, websites that we had used at the time. And I think that by having the goalposts
1:32:30 so far down the field, because they had just started with this windows app, which was sort of a quirk of
1:32:36 Lars and Jens, just like technical choices. We made much bolder technical bets than we would
1:32:41 have. Otherwise, I think we would have ended up much less interactive had we, uh, not started with
1:32:47 that quirky technical sort of a decision. But the other thing was this windows app. There’s a lot of
1:32:52 like, it’s hard to describe the like early two thousands. We wouldn’t live it, but like XML was
1:32:58 like really in fashion. So like most things and windows and other places like XML and XSLT,
1:33:03 which was a way of transforming XML into different XML was the basis of everything. It was like all
1:33:09 of enterprise software was like XML this, XML that. So similarly, when we were taking some of these ideas
1:33:14 and putting them in a web browser, we kind of like went into autopilot and used like a ton of XML
1:33:23 and it made everything just like really, really tedious. And so Google maps launched with some
1:33:27 really great ideas like the draggable maps. And, and we did a bunch of stuff for the local search
1:33:32 technology. So you could, you know, overlay restaurant listings. It was really great. It was a really
1:33:38 successful launch. Uh, we were like the hot shots within Google afterwards. And, uh, but it really
1:33:41 started to show its craft. And we got to this point where we decided we wanted to support the Safari
1:33:45 web browser, which was relatively new at the time. This is before, you know, mobile phones.
1:33:52 And, uh, there was much less XML support in Safari than there was an internet explorer in Firefox. And
1:34:00 so one of the engineers implemented like a full XSLT transform engine in JavaScript to get it to work.
1:34:07 And it was just like shit on top of shit on top of shit. And so what was a really elegant fat,
1:34:12 like fast web application had sort of quickly become something, you know, there’s a lot of
1:34:17 dial up modems at the time and other things. So like you’d show up to maps and it just was slow.
1:34:21 And like, it just bothered me as like someone who takes a lot of pride in their craft. And so
1:34:29 I got really, uh, energized and like over a weekend and a lot of coffee, like rewrote it.
1:34:34 Um, but it rewrote the whole thing there rewrote. Yeah. More or less the whole thing. And it took
1:34:38 probably another week of like, you know, working through the bugs, but yeah, I sent it out to the,
1:34:43 you know, the team after that weekend. And it was, it was nice. The reason I was able to do it,
1:34:50 yeah, I’m like a decent programmer, but you know, you’d also like lived with every bad decision up to
1:34:56 that point too. So I knew exactly the output I was going to like, I had simulated in my head,
1:35:01 like if I could do it over again, this is the way I do it. So by the time I like put my hands on the
1:35:07 keyboard on like, you know, Friday night, I, it wasn’t like I was designing a product. Like I knew
1:35:11 I had been in every detail of that product since the beginning and including me, the bad decisions
1:35:16 too. They’re not all the bad decisions. And so it was just very clear. I knew what I wanted to
1:35:20 accomplish. And for any, you know, any engineers worked on a big system, you have the whole system
1:35:27 mapped out in your head. So I knew, knew everything. And I, and I, and, and I also knew that, you know,
1:35:33 there’s a lot of pride of authorship with engineering and code. So I sort of knew I really wanted to
1:35:38 finish it over the weekend so that people could use it and see how fast it was and kind of overcome
1:35:44 anyone who was like, you know, you know, protective of the code they had written a few months ago.
1:35:50 And so I really wanted the prototype to go out. And so I did it. And then I didn’t,
1:35:53 it’s funny, I never talked about it again, but I think Paul Buhite, who’s was a co-creator,
1:36:00 Gmail, and, and I worked and started friend feed with me. He was on an interview and mentioned this
1:36:03 story. So now all of a sudden, it’s like, everyone’s talking about it. And I was like,
1:36:07 well, thank you, Paul. It’s a little embarrassed that people know about it, but it was, it was,
1:36:10 it’s a true story. And, and, and XML is just the worst.
1:36:17 Did you get a lot of flack from the people who had built the system you effectively replaced? Like,
1:36:21 you were part of that team, but everybody else had so much invested in it, even though it was like,
1:36:23 shit on top of shit, on top of shit.
1:36:28 you know, um, I wrote a lot of it too. So yeah, I’m sure there was some around it, but actually,
1:36:33 I think good teams want to do great work. And so, uh, I think there was a lot of people constructively
1:36:41 dissatisfied with the state of things too. And, um, uh, you know, I think, uh, you know, the engineer
1:36:46 had written that XSLT transform, I think was like, you know, a little bit, it’s a lot of work. So you have
1:36:52 to throw out a lot of work, which feels bad, but particularly, you know, um, Lars and Jens and I,
1:36:56 like, we want to make great products. And so I don’t think there was a, you know, at the end of the day,
1:37:01 everyone’s like, wow, that’s great. You know, we went from a bundle size of 200 K to a bundle size
1:37:06 of 20 K and it was a lot faster and better. So, you know, broadly speaking, I think good engineering
1:37:12 cultures. You don’t want a culture of, um, you know, ready, fire, aim, but I also think you just
1:37:20 need to be really outcomes oriented. And I think people, if they become, they’d start to treat
1:37:26 their code is too precious. It can really, uh, impede forward progress. Um, and yeah, I’ll just take,
1:37:33 like, I, my understanding is like a lot of the early self-driving car software was a lot of hand-coded
1:37:38 heuristics and rules. And, you know, a lot of smart people think that eventually it’ll probably
1:37:43 be a more monolithic model that, uh, encodes many of the same rules. You have to throw out a lot of
1:37:47 code in that transition, but it doesn’t mean it’s not the right thing to do. And so I think in general,
1:37:51 um, yeah, there might’ve been some feathers ruffled, but at the end of the day, everyone’s like,
1:37:55 that’s faster and better. Like, let’s, let’s do it, you know, which is, I think the right decision.
1:37:59 That’s awesome. I’m going to give you another hypothetical. I want you to share your
1:38:05 inner monologue with me as you think through it. So if I, uh, told you, you have to put 100% of
1:38:11 your net worth into a public company today and you couldn’t, you couldn’t touch it for at least 20
1:38:15 years, what company would you invest in? And like, walk me through your thinking.
1:38:20 I literally don’t know how to answer that question. Um, how would you think about it without giving me
1:38:24 an answer? Like, what? Yeah, that’s a good question. I, first of all, I’ll give you how
1:38:29 I think about it, but I’m so, uh, having not been a public company CEO for a couple of years, I’m
1:38:34 blissfully don’t pay attention as much, um, to the public markets. And in particular right now,
1:38:40 it’s obviously valuations have gone up a lot. So there’s a, but because it’s a long-term question,
1:38:45 maybe that doesn’t matter. I think what I’d be thinking about right now is, um,
1:38:53 over the next 20 years, like what are the parts of the economy that will most benefit from this
1:38:57 current wave of AI? That’s not the only way to invest over a 20 year period, but certainly it’s
1:39:03 a domain that I understand. And in particular, you know, I mentioned that, uh, talk, I heard a snippet of
1:39:09 from Tyler Cohen, which is like, it will probably AI will probably benefit different parts of the economy.
1:39:13 Um, disproportionately, there will be some parts of the economy that can essentially,
1:39:20 um, where intelligence is a limiting factor to its growth and where you can absorb
1:39:26 almost arbitrary levels of intelligence and generate almost arbitrary levels of growth.
1:39:30 Obviously there’s limits to all of this just because you change one part of the economy,
1:39:36 it impacts other parts of the economy. And that was what, uh, uh, Tyler’s point was, uh, in his talk.
1:39:39 But I would probably think about that because I think that
1:39:44 over a 20 year period, there are certain parts of society that won’t be able to change extremely
1:39:49 rapidly. Um, but there will be some parts that probably will, and it’ll probably be
1:39:56 domains where intelligence is, is the scarce resource, uh, right now. And then I would probably
1:40:00 try to find companies that will disproportionately benefit from it. And I assume this is why like
1:40:06 Nvidia stock is so high right now, because if you want to sort of get downstream, you know,
1:40:11 Nvidia will probably benefit from all of the investments in AI. Um, I’m not sure I would
1:40:15 do that over a 20 year period, just assuming that the infrastructure will shift. So I don’t have
1:40:19 an intelligent answer, but that’s the way I would think about it if I were, if we’re doing that exercise.
1:40:26 I love that. Where do you think, like, what’s your intuition say about what areas of the economy are
1:40:33 limited by intelligence and not just economy? I mean, perhaps politicians, uh, might be limited by,
1:40:40 by this and, and aid and benefit from, in which case countries could benefit enormously from AI and unlock
1:40:45 growth and potential in their economy. But I think maybe just to scope the question, like what areas of
1:40:51 the economy do you think are limited by intelligence or workers, like smart workers, in which case, like,
1:40:59 that’s another limit of intelligence. Yeah. I mean, uh, two that are, I think probably going to benefit a
1:41:07 lot are technology and finance. Um, you know, where you’re, you know, if you can make better financial
1:41:12 decisions than competitors, you’ll generate outsized returns. And that’s why over the past, you know,
1:41:18 30 years, you know, of machine learning, um, you know, uh, hedge funds and financial service
1:41:24 institutions, everything from fraud prevention to true investment strategies, it’s already been an area
1:41:30 of domain, uh, domain of investment, um, software, similar, as we talked about, I think that, uh,
1:41:36 at some point we will be, um, we will no longer be supply constrained in software, but we’re not anywhere
1:41:42 close to it right now. And you’re taking something that has always been the scarce resource, which is
1:41:47 software engineers and you’re making it not scarce. And I think as a consequence,
1:41:53 you just think of like, how much can that industry grow? We don’t know. Um, but we’ve been so constrained
1:41:59 on software engineering as a resource, uh, who knows over the next 20 years, but we’ll find out, uh,
1:42:04 where, where the limits are. But to me, intellectually, there’s just a ton of growth there. Um,
1:42:09 and then broadly, I think areas of like processing information are areas that will
1:42:16 really benefit, um, quite a bit here. And so that, and I think the, the thing that I would think about
1:42:19 over a 20 year period is like second and third order effects, which is why I don’t have an intelligent
1:42:22 answer. And if you’re asking me to put all my money in something, I would think about it for a
1:42:28 while, um, probably use a one pro a little bit to help me. Um, but, uh, you know, because you can end
1:42:33 up, uh, generated a bunch of growth in the short term, but then, you know, if everyone does it,
1:42:39 it commoditizes the whole industry, you know, type of thing. So, you know, there used to be, you know,
1:42:44 before the introduction of the freezer ice was like a really expensive thing and now it’s free,
1:42:49 you know? And so I think it is really important to actually think through those. If you’re talking
1:42:52 in a timeframe of like 20 years. And that’s why having not thought about this question ahead of
1:42:57 time, I, um, you could be quite simplistic elsewhere, but I would say software and finance
1:43:02 are areas that I, I think stand to reason should benefit quite a bit. I love that response. How do
1:43:12 you balance, uh, having a young family with also running a startup again? I work a lot. Um, I don’t,
1:43:20 uh, I really care and love care about and love working. Um, so one thing is that I, um,
1:43:27 well, there’s always trade-offs in life. Um, if I didn’t love working and I wouldn’t do it as much
1:43:34 as I do, but I, I just love, uh, love to create things and love to have an impact. And so I like
1:43:39 jump out of bed in the morning and, um, work out, go to work and then spend time with my family
1:43:45 broadly, probably, you know, being honest first. I’m not perfect. I think for a second,
1:43:48 I don’t have a ton of hobbies. You know, I basically work and spend time with my family.
1:43:54 Um, the first time we talked, you saw a couple of guitars in my background. Uh, I haven’t picked
1:44:00 one of those up in a while. Um, uh, I, I mean, I literally pick it up occasionally, but I, you know,
1:44:05 do not devote any time into it. And I don’t regret that either. Like I am so passionate about what
1:44:10 we’re building at Sierra. I’m so passionate about opening. I am so love my family so much. I don’t
1:44:15 really have any regrets about it, but I basically just like life is all about where do you spend your
1:44:20 time and mine is at work and with family. And so that’s how I do it. I don’t know if I’m particularly
1:44:26 balanced, but I don’t strive to be either. I really take a lot of pride and I love, I love to work.
1:44:32 Having sold the companies you started twice, how does that influence what you think of Sierra? Like,
1:44:37 are you thinking like, Oh, I’m building this in order to sell it? Or do you think differently? Like,
1:44:41 this is my life’s work. I’m building this with, that’s not going to happen.
1:44:47 I absolutely, uh, intend Sierra to be an enduring company and an independent company,
1:44:54 but to be honest, every entrepreneur, every company starts that way. And so, um, you know,
1:45:00 uh, I’m really grateful for both Facebook and Salesforce for having acquired my previous companies
1:45:04 and hopefully I had an impact about those companies, but you don’t start off. Well,
1:45:11 at least I never started off saying, Hey, I want to make a company to sell it. Um, uh, and, uh,
1:45:15 but I actually think with Sierra, we have just a ton of traction in the marketplace. Uh,
1:45:19 I really do think Sierra is the leader in helping consumer brands build customer facing
1:45:25 AI agents. And I’m really proud of that. So I really see a path to that. And I joke with Clay,
1:45:30 I want to be, you know, an old man sitting on his porch, you know, complaining how
1:45:33 the next generation of leaders at Sierra don’t listen to us anymore. You know,
1:45:37 I want this to be something that not only is enduring, but outlives me. Um, and I think
1:45:43 just actually, I don’t think we’ve ever talked about this, but it was really interesting, um,
1:45:47 moment for me when Google went from its one building in mountain view to its first corporate
1:45:53 campus. It, uh, we moved into the Silicon graphics campus, which was right over near shoreline
1:46:00 Boulevard and in mountain view. And, uh, SGI had been a really successful company enough to build a
1:46:04 campus. And when we, it was actually quite awkward. We moved into like half the campus,
1:46:08 they were still in half and they’re like, we’re this up and coming company. They’re declining.
1:46:14 And then when Facebook, when we moved out of the second building, we were in Palo Alto,
1:46:19 it was slightly larger building. I think we leased it from HP, but when we finally got a campus,
1:46:24 it was from Sun Microsystems who had gone through an Oracle acquisition and had been sort of on the
1:46:32 decline. And it was interesting to me because both SGI and Sun, um, had been started and grown to
1:46:36 prominence in my lifetime. Uh, obviously I was maybe like a little younger, obviously,
1:46:41 but in my lifetime enough to build a whole corporate campus and then declined fast enough
1:46:48 to sell that corporate campus to a new software company. And for me, I, it was just so interesting
1:46:54 to have done that twice to move into like a, you know, a used campus, you know, for the previous,
1:47:00 uh, uh, uh, owners. It was a very stark reminder that technology companies aren’t entitled to their
1:47:07 future success. And I think we’ll see this actually now with AI, AI, I think will change the landscape
1:47:14 of software to be, um, tools of productivity that to agents that actually accomplish tasks. And I think it
1:47:20 will help some companies who, for whom that’s a, uh, uh, amplifies their existing value proposition
1:47:25 and it will really hurt others where it will essentially the seat based kind of model of
1:47:32 legacy software will wane, um, very quickly and then really harm them. And so when I think about the,
1:47:40 what it means to build a company that’s enduring, um, that is a really, really tall, um, task in my mind
1:47:44 right now, because it means not only making something that’s financially enduring over the
1:47:52 next 10 years, but setting up a culture where a company can actually evolve to meet the changing
1:47:58 demands of, uh, society and technology at a, when it’s changing at a pace that is like unprecedented
1:48:03 in history. So I think it’s one of the most fun business challenges of all time. And I think it has
1:48:10 as much to do with culture as it has to do with technology because every line of code in Sierra
1:48:15 today will be completely different, you know, probably five years from now, let alone 30 years
1:48:19 from now. Um, and, uh, I think that’s really exciting. So when I think about it, I just get
1:48:26 so much energy because, um, it’s incredibly hard and it’s harder now than it’s ever been, um, to do
1:48:30 something that lasts beyond you. Um, but that I think is the ultimate measure of a company.
1:48:34 You mentioned AI agents. How would you define that? What’s an agent?
1:48:39 I’ll define it more broadly and then I’ll tell you how we think about it at Sierra, which is a more
1:48:44 more narrow view of it. The word agent comes from agency. And I think it means affording
1:48:53 a software, the opportunity to reason and make decisions autonomously. Um, and I think that’s
1:48:59 really all it means to me. And I think there’s lots of different, uh, applications of it. The three
1:49:04 categories that I think are meaningful and I’ll end with the Sierra one just so I can talk about it a
1:49:10 a little more, but one is personal agents. So I do think that most people will have
1:49:18 probably one, but maybe a couple AI agents that they use on a daily basis that are, uh, essentially
1:49:25 amplifying themselves as an individual. Um, you can do the rote things like help you triage your email to
1:49:32 helping you schedule a vacation. You know, you’re flying back to, um, Edmonton and help you arrange
1:49:38 your travel. Um, two more complex things like, you know, I’m going to go ask my boss for promotion,
1:49:44 like help me role play. And, um, you know, uh, I’m setting up my resume for this job. Help me do that
1:49:49 too. I’m applying for a new job. Help me find companies I haven’t thought of that I should be applying to.
1:49:55 Uh, and I think these agents will be really powerful. I think it might be a really hard
1:50:00 product to build because when you think about all the different services and people you interact with
1:50:06 every day, it’s kind of everything. So it’s not, it has to generalize a lot to be useful to you. And
1:50:12 because of the personal privacy and things like that, it has to work really well for you to trust it.
1:50:15 So I think it’s going to take a while to go. I think it’ll be a lot of demos. I think it’ll take
1:50:22 a while to be robust. The second category of agent is I would say, um, really filling a persona, uh,
1:50:32 within a company. So a coding agent, a paralegal agent, um, a analyst agent. Um, I think these
1:50:37 already exist. I mentioned cursor. There’s a company called Harvey that makes a legal agent. I’m sure
1:50:44 there’s a bunch in the analyst space. Um, these do a job and they’re more narrow. Um, but they, uh,
1:50:49 they’re really commercially valuable because most companies hire people or consultants that do those
1:50:55 things already, like analyze the contracts of the supply chain, right? That’s a kind of a rote
1:51:01 kind of law, but it’s really important and AI can do it really well. So I think that’s why, uh,
1:51:06 this is the area of the economy that I think is really exciting. And, and as, uh,
1:51:11 I’m really excited about all the startups in this space because you’re essentially, um, taking what
1:51:16 used to be a combination of people and software and really making something that solves a problem.
1:51:24 Uh, and by narrowing the domain of, of autonomy, you can have more robust guard rails and even with
1:51:29 current models actually achieve something that’s effective enough to be commercially viable today.
1:51:35 Um, and, uh, and by the way, it changes the total addressable market of these models too. Like,
1:51:39 I don’t know what the total addressable market of legal software was three years ago, but it
1:51:43 couldn’t have been that big. I couldn’t tell you like a legal software company. I probably should,
1:51:48 I just can’t think of one, but if you think about the money we spend on lawyers, that’s a lot. And so
1:51:56 you end up where you’re broadening the, the addressable market quite a lot. The domain we’re
1:52:03 in, um, I think is somewhat special, which is, um, a company’s branded customer facing agent. And
1:52:08 the reason why I think it’s, one could argue we’re sort of, uh, helping with customer service,
1:52:14 which is a, a, a persona, a role, but I do think it’s broader than that. Because if you think about,
1:52:20 um, a website, you know, like your insurance company’s website, try to list all the things
1:52:24 you can do on it. You can look up the stock quote, you can look up the management team,
1:52:31 you can compare their insurance company to all their competitors. You can file a claim. You can,
1:52:39 you know, uh, buy, you can bundle your home and auto. You can, um, uh, um, add a member of your
1:52:44 family to your premium. There’s a million things you can do on it. Essentially over the past 30 years,
1:52:50 websites, a company’s website singular has come to be the universe of everything that you can do with
1:52:54 that company. I like to think it was like the digital instantiation of the company.
1:53:01 And that’s what we’re helping our customers do at Sears, help them build a conversational AI that does
1:53:05 all of that. So, you know, most of our customers start with customer service and it’s a great
1:53:10 application because no one likes to wait on hold and, and having something that has perfect access
1:53:16 to information is multilingual and empathetic is just amazing. But you know, when you put a conversational
1:53:22 AI as your digital, um, front door, people will say anything they want to it. And, um,
1:53:28 we’re now doing product discovery, consider purchases, going back to the insurance example.
1:53:34 Hey, you know, I’ve got a 15 year old daughter. I really am concerned about the cost of her premium
1:53:39 until she grows up. Tell me, um, which plan I should be on. Tell me why you’ll be better than
1:53:43 your competitors. That’s a really complex interaction, right? That’s not something that,
1:53:47 can you make a webpage that does that? No, that’s, but that’s a great conversation.
1:53:54 And so we really aspire that when you encounter a branded agent in the wild, we want Sierra to be
1:54:00 the platform that powers it. And it’s super important because there was a case, at least in Canada,
1:54:06 where an AI agent for Air Canada hallucinated a bereavement policy, right? But they were found liable
1:54:12 to hold themselves to what the agent said. Yeah. I mean, it turns out, and it was an AI
1:54:17 agent. There was no human involved in the whole thing. Well, look, it’s one thing if chat GPT
1:54:22 hallucinate something about your brand. It’s another if your AI agent hallucinate something about your
1:54:29 brand. So the bar just gets higher. So the robustness of these agents, the guardrails, everything is more
1:54:35 important when it’s yours and it has your brand on it. And so it’s harder, but I also, I’m just so
1:54:41 excited for it because this is a little overly intellectual, but I really like the framing.
1:54:49 If you think about a modern website or mobile app, it’s essentially you’ve created a directory of
1:54:56 functionality from which you can choose. But the main person with agency in that is the creator of
1:55:02 the website. Like what are the universe of options that you can do? When you have an AI agent represent
1:55:08 your brand, the agency goes to the customer. They can express their problem any way they want in a
1:55:12 multifaceted way. And so it means that like your customer experience goes from the
1:55:18 enumerated set of functionality you’ve decided to put on your website to whatever your customers ask.
1:55:23 And then, you know, you can decide how to fulfill those requests or whether you want to.
1:55:29 But I think it will really change the dynamic to be really empowering to consumers. As you said,
1:55:35 I mean, the reason that that Air Canada case is the reason we exist. You know, companies,
1:55:42 if they try to build this themselves, there’s a lot of ways you can shoot yourself in the foot.
1:55:48 But in particular, too, your customer experience should not be wedded to one model, let alone even
1:55:54 this current generation of models. So with Sierra, you can define your customer experience once in a
1:56:00 way that’s abstracted from all of the technology. And it can be a chat. It can be, you can call you on
1:56:05 the phone. It can be all of those things. And as new models and new technology comes out,
1:56:10 our platform just gets better. But you’re not like re-implementing your customer experience. And I
1:56:14 think that’s really important because, you know, we were talking about what’s happened over the past
1:56:19 two years. Can you imagine if you’re a consumer brand like ADT Home Security and thinking about,
1:56:23 like, how can you maintain your AI agent in the face of all of that, right? It’s just not even,
1:56:28 it’s not tenable. I mean, it’s not what you do as ADT. So they’ve worked with us to build their AI agent.
1:56:35 Like, how do you fend off complacency? Like, a lot of these companies, and maybe not in tech
1:56:44 specifically, but they get big, they get dominant, and then they take their foot off the gas. And that
1:56:50 opens the door to competitors. And there’s like a natural entropy almost to bureaucracy in some of
1:56:57 these companies that, and the bureaucracy sows the seeds of failure and competition. How do you,
1:56:58 how do you fend that off constantly?
1:57:05 It is a really challenging thing to do at a company. One of the, there’s two things that
1:57:15 I’ve observed that I think manifest as corporate complacency. One is bureaucracy. And I think the
1:57:22 root of bureaucracy is often when something goes wrong, companies introduce a process to fix it.
1:57:32 And over the sequence of 30 years, the layered sum of all of those processes that were all created for
1:57:42 good reason, with good intentions, end up being a bureaucratic sort of machine where the reasons for
1:57:48 many of the rules and processes are rarely even remembered by the organization. But it creates this
1:57:55 sort of natural inertia. Sometimes that inertia can be good. You know, it’s like, you know, if you end up
1:58:01 with you, there’s definitely been stories of executives coming in and ready, fire, aim, new strategies that
1:58:08 backfire massively. But often, it can mean in the face of a technology shift or a new competitor,
1:58:13 you just can’t move fast enough to address it. The second thing that I think is more subtle is as a
1:58:20 a company grows in size. Often its internal narrative can be stronger than the truth from customers.
1:58:28 I remember one time when this sort of peak of the smartphone wars, and I ended up visiting a friend
1:58:37 on Microsoft’s campus. And I got off the plane and, you know, Seattle Tacoma Airport, drove into Redmond,
1:58:43 went on to the campus. And all of a sudden, everyone I saw was using Windows phones.
1:58:50 I assume it must have been a requirement or formal or social, like you were definitely uncool if you’re
1:58:56 using anything else. And from my perspective at the time, like the war had already been lost.
1:58:57 Yeah.
1:59:04 Like, it was definitely a two-horse race between Apple and Google on iOS and Android.
1:59:10 And I remember sitting in the lobby waiting for my friend to get me from the security check-in.
1:59:15 And I made a comment, like it wasn’t a confrontation, but I made a comment to someone who’s at Microsoft.
1:59:22 I was like, you know, something along the lines of, are you required to use Windows phones?
1:59:27 How these are? And I just sort of like curious. And then I got a really bold answer, which is like,
1:59:31 yeah, we’re going to win. Like, we’re taking over the smartphone market. And I was like,
1:59:36 I didn’t say anything because it was like a little socially awkward. I was like, no, you’re not. Like,
1:59:38 you lost like four years ago.
1:59:44 But there’s something that’s happening that’s preventing you from getting reality.
1:59:48 Well, that’s the thing is, if you think about it, if anyone, if you’ve ever worked for like a large
1:59:55 company, you know, when you work at a small company, you care about your customers and your
2:00:02 competitors and you feel every bump in the road. When you’re a, you know, junior vice president of
2:00:10 whatever, and you’re, you know, eight levels below your, you know, CEO, and you have a set of
2:00:15 objectives and results, uh, your, you might be focused as I want to go from junior vice president
2:00:20 to senior vice president. That’s what success looks like for me. And you end up with this sort
2:00:28 of myopic focus on this internal world in the same way your kids will focus on, you know, the social
2:00:33 dynamics of their high school, not the world outside of it. And it’s probably rational by the way, because
2:00:38 like, you know, probably their social life is more determined by those, you know, 1,000 kids in their
2:00:42 high school than it is like all the things outside. But this is, that’s the life of a
2:00:48 person inside of these big places. And so you end up where, uh, you know, if you have a very senior
2:00:53 head of product, who’s like, are this competitor says they’re faster, but this next version we’re
2:00:58 so much better. And then everyone says, all of a sudden that’s like the windows phone is going to
2:01:04 win. That’s what everyone says. And, and you truly believe it because everyone you meet says the
2:01:10 same thing and you end up reflecting, you know, uh, customer anecdotes through that lens and you end up
2:01:16 with this sort of reality distortion field manifested from the sum of, of this sort of myopic
2:01:22 storytelling that, um, exists within, within companies. The, what’s interesting about that
2:01:27 is like, you know, the ability for a culture to believe in something is actually a great strength of
2:01:33 a culture, but it can lead to this as well. And so the combination of bureaucracy and inaccurate
2:01:39 storytelling, um, I think is the reason why companies sort of die. Uh, and, and it’s really
2:01:45 remarkable to look at, you know, the blackberries of the world or the TiVos or the, you know, there,
2:01:52 you can really, um, you know, as the plane is crashing, like tell the story that you’re not. And,
2:01:58 um, and, and, and then similar, as I said, like culturally, you can still have like the person
2:02:03 in the back of that crashing plane being like, when am I going to get promoted to SVP? And then you’re
2:02:08 like, you know, and, and that’s, I mean, this is like, I mean, I’ve seen it a hundred times. And so
2:02:14 I think it really comes down to leadership, you know, and I think that one of the things that most
2:02:19 great companies have is they are obsessed with their customers. Um, and I think the, the free market
2:02:24 doesn’t lie. And so I think the, one of the most important things I think for any like enduring
2:02:29 culture, particularly in an industry that changes as rapidly as software is how close are your
2:02:35 employees to customers and how much can customer like the direct voice of your customers be a part of,
2:02:42 uh, your decision-making. Um, and that is something that I think you need to constantly work out because
2:02:52 that, you know, person employee number 30,462, you know, how does he or she actually, actually
2:02:55 directly hear from customers is, it’s not actually a simple question to answer.
2:02:58 Is it direct? Is it filtered? How many filters are there?
2:03:04 That’s exactly right. And then, um, I think the other part on leadership is, you know,
2:03:09 we talked about bureaucracy is process is there to serve the needs of the business.
2:03:18 And, uh, uh, often, um, mid-level managers, uh, don’t get credit for removing process.
2:03:24 They often are held accountable for things going wrong. Um, and I think it really takes top-down
2:03:31 leadership to, uh, you know, remove bureaucracy. Um, and, uh, it is not always comfortable, you know,
2:03:40 when companies remove spans of control or, uh, all the people impacted will, it’s like antibodies.
2:03:44 Uh, and for good reason, I mean, it makes sense their lives are negatively impacted or whatever it is,
2:03:51 but it almost has to come from the top because you need to give, uh, air cover, uh, almost certainly
2:03:56 something will go wrong by the way. I mean, like processes usually exist for a reason. Um,
2:04:01 but when they accumulate, um, without end, you end up with bureaucracy. So those are the two things
2:04:06 that I always, uh, and you could smell it when you go into a really bureaucratic company, the,
2:04:13 the inaccurate storytelling, the process over outcomes. And it’s just, uh, it sort of sucks the
2:04:19 energy out of you when you feel it. That’s a great answer. We always end these interviews with the exact
2:04:24 same question, which is what is success for you? Success for me, we talked about how I spend my
2:04:31 time with my family at work is, you know, having a happy, healthy family and being able to work with
2:04:35 my co-founder Clay for the rest of my life, making Sierra into an enduring company. That would be success for me.
2:04:47 Thanks for listening and learning with us. The Farnham street blog is where you can learn more
2:04:54 about my new book clear thinking, turning ordinary moments into extraordinary results. It’s a transformative
2:05:00 guide that hands you the tools to master your fate, sharpen your decision-making and set yourself up for
2:05:14 unparalleled success. Learn more at fs.blog slash clear until next time.