Bret Taylor: A Vision for AI’s Next Frontier

27:36Rather than having our customers pay a license for the privilege of using our platform, we only charge our customers for the outcomes. Meaning if the AI agent they've built for their customers solves the problem, there's usually a pre-negotiated rate for that. And that comes from the principle that in the age of AI, software isn't just helping you be more productive, but actually completing a task. What is the right and logical business model for something that completes a task? Well, charging for a job well done rather than charging for the privileges of using the software. Similarly, with a lot of our customers, we help deliver them a fully working AI agent.

28:16We don't hand them a bunch of software and say, good luck, configure it yourself. And the logic there is, you know, in a world where making software is easier than it ever is before and you're delivering outcomes for your customer, the delivery model of software probably should change as well. And we've really tried to reimagine what like the software company of the future should look like and trying to, you know, model that in everything that we do. That's brilliant. How do you think software engineering will change? Is it you're going to have fewer people or the people are going to be organized differently?

28:52Or how do you see that? How geeky can I get? As geeky as you want to. I actually wrote a blog post right before Christmas about this. I think this is an area that deserves a lot more research. I'll describe where I think we are today. And smart people may disagree. But a lot of the modern large language models, both the traditional large language models and sort of the new reasoning models are trained on a lot of source code. And it's an important input to all of the knowledge that they're trained on. As a consequence, even the early models were very good at generating code. So, you know, every single engineer at CIRA uses Cursor, which is a great product that basically integrates with the IDE Visual Studio Code to help you generate code more quickly.

29:46it feels like a local maximum in a really obvious way to me, which is you have a bunch of code written by people, written in programming languages that were designed to make it easy for people to tell a computer what to do. Probably the funniest example of this is Python. It almost looks like natural language. But it's notoriously not robust. Most Python bugs are found by running the program because there's not static type checking. Similarly, there's most bugs, while you could run a fancy static analysis, like most bugs show up simply at runtime because it's just not designed. It's designed to be ergonomic to write.

30:33Yet we're using AI to generate that. So we've sort of designed most of our computer programming systems to make it easy for the author of code to type it quickly. And we're in a world where actually generating code is going to, like the marginal cost of doing that is going to zero. But we're still generating code in programming languages that were designed for human authors. And similarly, if you've ever looked at someone else's code, which a lot of people do professionally, it's called the code review. it's actually quite hard to do a code review. You end up interpreting, you're trying to basically put the system in your head and simulate it as you're reading the code to find errors in it.

31:22So the irony now of taking things that are code programming languages that were designed for authors and now having humans do the job of essentially code reviewing code written by an AI and yet all of the AI is being in the code generation part of it. I'm like, I'm not sure it's great, but we're generating a lot of code with similar flaws to that we've been generating before from security holes to just functional bugs and in greater volumes. And I think what I would like to see is if you start with the premise that generating code is free or going towards free, what would be the programming systems that we would design?

32:03So, for example, Rust is an example of a programming language that was designed for safety, not for programming convenience. My understanding is that the Mozilla project, there were so many security holes in Firefox. They said, let's make a programming language that's very fast, but everything can be checked statically, including memory safety. Well, it's a really interesting direction where you weren't optimizing for authorship convenience or optimizing for correctness. Are there programming language designs that are designed so a human looking at it can very quickly evaluate, does this do what I intended it to do?

32:44There's an area of computer science I studied in college called formal verification, which at the time was turning a lot of computer programs into math proofs and finding inconsistencies. And it sort of worked well, not as well as you'd hope. But, you know, in a world where AI is generating a lot of code, you know, should we be investing in more informal verification so that the operator of that code generator machine can more easily verify that it does in fact do what they intended it is to do? And could a combination of a programming language that is more structurally correct and structurally safe and exposes more primitives for verification plus a tool to verify, could you make an operator of a code-generated machine 20 times more productive, but more importantly, make the robustness of their output 20 times greater?

33:30And then similarly, you know, there's themes, things go in and out of fashion, but like test driven development, you know, where you write your unit test first or your integration test first and then write code until it fulfills the test. Most programmers I know who are really good, not despise it, but it's just like it sounds better than it is in practice. But again, writing code is free, you know, so writing tests is free. How can you create a programming system where the combination of great programming language design, formal verification, robust tests because you didn't have to do the tedious part of writing them all?

34:04Could you make something that made it possible to write increasingly complex systems that were increasingly robust? And then similarly, like the elephant in the room for me is the anchor tenant of most of these code generating systems are an IDE right now. And that obviously doesn't seem as important in this world. And even with coding agents, which is sort of where the world is going, it doesn't change the fact that who's accountable for the quality of it, who's fixing it. And I think there is a world where we can make reasonable software by just automating what we as software engineers do every day.

34:41But I have a strong suspicion that if we designed these systems with the role of a software engineer in mind being an operator of a machine rather than the author of the code, we could make the process much more robust and much more productive. And it feels like a research problem to me. It doesn't feel, and I think a lot of people, and for good reason, including me, are just excited about the efficiency of software development going up. And I want to see the new thing. I'm constructively dissatisfied with where we are. It's so interesting that if software AI is good enough to write the code, should we get enough to check the code?

35:18That's a great, great question. But actually, it's still funny to me that we'd be generating Python, Just because for anyone who's listening right now has ever operated a web service running Python, it's CPU and intensive, really inefficient. Should we be taking most of the unsafe C code that we've written and converting it to a safer system like Rust? If authoring these things and checking it are relatively free, shouldn't all of our programs be incredibly efficient? Should they all be formally verified? Should they all be analyzed by a great agent? I do think it can be turtles all the way down.

35:54You can use AI to solve most problems in AI. The thing that I'm trying to figure out is, like, what is the system that a human operator is using to orchestrate all those tasks? And, you know, I go back to the history of software development, and most of the really interesting metaphors in software development came from breakthroughs in computing. So, you know, the C programming language came from Unix, and when these time-sharing systems were really, It went from sort of punch cards to something that were a lot more agile. Smalltalk came out of the development of the graphical user interface at Xerox PARC.

36:29And, you know, there was a sort of a confluence of message passing as a metaphor and the graphical user interface. And then there was a lot of really interesting principles that came out of networking, you know, and sort of distributed systems, distributed locking, sequencing. I think we should recognize that we're in this brand new era as significant as the GUI. You know, it's like a completely new era of software development. And if you were just to say, I'm going to design a programming system for this new world from first principles, what would it be? And I think when we develop it, I think it will be really exciting because rather than automating and turning up the speed of just generating code and with the same processes we have today, I think will feel native to this system and give a lot more control to the people who are orchestrating the system in a way that I think will really benefit software overall.

37:24Let's dive into AI a little bit. How would you define AGI to the layman? I think a reasonable definition of AGI might be that any task that a person can do at a computer, that system can do on par or better. I'm not sure it's a precise definition, but I'll tell you where that comes from and its flaws. But there's not a perfect definition of AGI, in my opinion, or there's not a precise definition of AGI. I'm sure there's good answers. One of the things about the G and AGI is about generalization. So can you have a system that is intelligent in domains that it wasn't explicitly trained to be intelligent on.

38:12And so I think that's one of the most important things is like given a net new domain, can this system become more competent and more intelligent than a person sort of trained in that domain? And I think that's sort of the, you know, at or better than a person is certainly a good standard there. And that's sort of the definition of superintelligence. The reason I mentioned at a computer is I do think that it is a bar that means like if there's a digital interface to that system, it affords the ability for AI to interact with it, which is why that's a bar that's reasonable to hit. I say that because one of the interesting questions around AGI is how quickly it does generalize.

39:04And there are domains in the world that the progress in that domain isn't necessarily limited by intelligence, but by other social artifacts. So as an example, and I'm not an expert in this area, but if you think about the pharmaceutical industry, my understanding is, you know, the one of the main bottlenecks is clinical trials. So no matter how intelligent a system would be in discovering new therapies, it may not materially change that. And so you may have something that's discovering new insights in math, and that would be delightful and amazing. But the existence of that system that's super intelligent in one domain may not translate to all domains equally.

40:01I just heard at least a snippet of a talk by Tyler Cohen, the economist, and it was really interesting to hear his framing on this about which parts of the economy could sort of absorb intelligence more quickly than others. And so I choose that definition of AGI, recognizing that there's not a perfect definition because it captures the ability of this intelligence to generalize, while also recognizing that the domains of society might not apply with equal velocity, even once we reach that point of a system being able to have that level of intelligence.

40:57When I think about what artificial intelligence is limited by, or the bottlenecks, if you will, I keep coming back to a couple things. There's regulation, there's compute, there's energy, there's data, and there's LLMs. Am I missing anything? So you're saying the ingredients to AGI? Yeah, there's limitations on each aspect of those things, and those seem to be the main contributors to what's limiting us from even accelerating at this point. How do you think about that? Yeah, what you said is roughly how I think about it. I'll put it into my own words, though. So I think the three primary inputs are data, compute, and algorithms.

41:42And data is probably obvious, but, you know, one of the things after the Transformers model was introduced is it afforded an architecture with just much greater parallelism, which meant models could be much bigger and train more quickly on much more data, which just led to a lot of the breakthroughs with, that's the L, just they're large yeah and uh the scaling laws you know a couple years ago you know indicated like the larger you make the model um the more intelligent would be and at a degree of efficiency that was uh tolerable uh and there we are you know there's lots of stuff written about this but you know there's in terms of just like textual content to train on you know the availability of new content is certainly waning.

42:32And some people would say, I think there's like a data wall. I'm not an expert in that domain, but it's been talked about a lot and you can read a lot about it. There's a lot of interesting opportunities though to generate data too. So there's a lot of people working on simulation. If you think about a domain like self-driving cars, simulation is a really interesting way to generate. Is that synthetic data? Is that what? Yeah, I would say that's synthetic data. though synthetic data has a uh simulation and synthetic data are a little different so you can generate synthetic data like generate a novel um simulation i would put at least in my head and i'm sure that like academics might critique what i'm saying but uh i've used simulation is based on a set of principles like the laws of physics so if you're a build a real world simulation for training a self-driving car um you're not just generating arbitrary data like the roads don't turn into loop-de-loops, you know, because that's not possible with physics.

43:31So by constraining a simulation with a set of real-world constraints, the data has more efficacy, you know, and so, and there's sort of a, it constrains the different permutations of data you can generate from it. So it's, I think, a little bit higher quality. But then along those lines, you know, a lot of people wonder if you generate synthetic data, how much value can that add to a training process? You know, is it sort of regurgitating information already had? What's really interesting about, you know, reasoning and reasoning models is I think, I feel really optimistic these models are generating net new ideas.

44:12And so it really affords the opportunity to break through some of these, the data wall as well. So data is one thing, and I think both synthetic data and simulation are really interesting opportunities to grow there. Then you have compute. And this is something that, you know, that's why there's so many data center investments. It's why NVIDIA as a company has grown so much. Probably the more interesting kind of breakthroughs there are these reasoning models where there's not quite such a formal separation between the training process and the inference process where you can spin more compute at the time of inference to generate more intelligence, which has really been a breakthrough in a variety of ways.

44:59I think it's really interesting, but it shows you how you can run up against walls and find new opportunities to use it. And then finally, algorithms. And the biggest breakthrough is obviously the Transformers model attention is all you need that paper from google that sort of led to where we are now but there's been a number of really important papers since then from the idea of chain of thought reasoning to um what uh at open ai what we did with the o1 model which is to do some reinforcement learning those chains of thought to uh really reach new levels of intelligence um and so i do think that i mentioned some anecdotes about some breakthroughs there because my view is that each one of them has their own problems.

45:40Compute, it's very capital intensive. And a lot of these models, the half-life of their value is pretty short because new ones come out so frequently. And so you wonder, can we afford, what's the business case for investing this capex? And then you have a breakthrough like 01 and you're like, gosh, with a distilled model and moving more inference time, it changes the economics of it. You have data. You say, gosh, we're running out of textual data to train on. Well, now we can generate reasoning. We can do simulations. Oh, that's an interesting breakthrough. And then on the algorithm side, as I mentioned, just the idea of these reasoning models is really novel itself.

46:20And each of these at any given point, if you talk to an expert in one of them, and I'm an expert in none of them, they will tell you the sort of current plateau that they can see on the horizon. And there usually is one. I mean, you'll talk to different people about how long the scaling laws for something will continue and you'll get slightly different opinions but no one thinks it's going to last forever um and at each one of those because you have so many smart people working on them you often have people discovering a breakthrough um in each of them and so as a consequence i i really do feel optimistic about the progress towards hei because one of those plateaus might extend a while if we just don't have the key idea that we need to break through, the idea that we will be stuck on all three of those domains feels very unlikely to me.

47:09And in fact, what we've seen because of the potential economic benefits of AGI is we're in fact seeing breakthroughs in all three of them. And as a consequence, you know, you're just seeing just the blistering pace of progress that we've seen over the past couple of years. At what point does AI start making AI better than we can make it or making it better while we're sleeping. We can't be too far from that. Well, it might reflect back to our software engineering discussion, but broadly this is the area of AGI around self-improvement, which is meaningful from an improvement standpoint, but also obviously from a safety standpoint as well.

47:49So I don't know when that will happen, but I do think by some definition you could argue that it's happening already in the sense that every engineer in Silicon Valley is already using coding agents and platforms like Cursor to help them code. So it's contributing already. And I imagine as coding assistants go to coding agents in the future, most engineers in Silicon Valley will show up in the morning. But this is sort of the difference between the assisted driving in Tesla versus self-driving. At what point do we leap from I'm a co-pilot in this to I don't have to do anything? I mean, it's a question that there's so much nuance to the answer.

48:37I'm not sure to answer because I'm not sure you'd want to necessarily. Like I think for some software applications, that's important. But when we brought up, you know, we were talking about the active software development, people have to be accountable for the software that they produce. And that means if you're doing something simple like a software as a service application that it's secure, that it's reliable, that the functionality works as intended for something as meaningful as, you know, an agent that is somewhat autonomous. Does it have the appropriate guardrails? Does it actually do what the operators intended?

49:16Is there appropriate safety measures? So I'm not sure there's really any system where you'd want to turn a switch and go get your coffee. But I do think to the point on these broader safety things is I think that when you think about more advanced models, we need to be developing not only more and more advanced safety measures and safety harnesses, but also using AI to supervise AI and things like that. So it's a part, probably my colleague on the board, Zico Coulter, is probably a better person to talk through some of the technical things. But there's a lot of prerequisites to get to that point.

49:51And I'm not sure it's simply like the availability of the technology. Just because it is that at the end of the day, we are accountable for the safety of the systems we produce. Not just OpenAI, like every engineer. And that's a principle that should not change. What does that mean? Like when we say safety and AI, that seems so vague in general that everybody interprets it quite differently. Like, how do you think about that? And how do you think about that in the world where, let's say we regulate safety in the United States and another country doesn't regulate safety? How does that affect the dynamic of it?

50:30I'll answer broadly and then go into the regulatory question. So I really like OpenAI's mission, which is to ensure that AGI benefits all of humanity. That isn't only about safety, and I believe intentionally so, though. Obviously, the mission was created prior to my arrival because it's both about safety, kind of Hippocratic, first to no harm. And I don't think one could credibly achieve that mission if we created something unsafe. So I would say that's the most important part of the mission. But there's also a lot of other aspects of benefiting humanity. Is it universally accessible? Is there a digital divide where some people have access to AGI and some don't?

51:12Similarly, you could argue that are we maximizing the benefits and minimizing the downsides? Clearly, AI will disrupt some job, but it also could democratize access to healthcare, education, expertise. piece. So as I think about the mission, it starts with safety, but I actually like thinking about it more broadly because I think at the end of the day, benefiting humanity is the mission and safety is a prerequisite, but it's almost like going to my analogy of the Hippocratic Oath. A doctor's job is to cure you. First do no harm, but then to cure you. And a doctor that did no harm but didn't cure you wouldn't be great either.

51:50So I really like to think about the holistically. and again Zika or Sam might have a more complete answer here but broadly I think about does the system that represents AGI align with the intentions of the people created it and the intentions of the people operating it so that it does what we want and it's a tool that benefits humanity a tool that we're actively using to affect the outcomes that we're looking for and that's kind of the way I think about safety. And it can be meaningful things like misalignment or more subtle things like unintended consequences. And I think that latter part is probably the area that's really interesting from an intellectual and ethical standpoint as well.

52:39If I look at, what was the bridge in Canada that fell down where it motivated the ring that a lot of engineers read? Oh, yeah, I forget the name of it. Just like whether it's the Tacoma Narrows Bridge in Washington or Three Mile Island or these intersections where we've engineered these, you know, what at the time people hope would be positively impact humanity, but something went horribly wrong. Sometimes it's engineering. Sometimes it's bureaucracy. Sometimes it's a lot of things. And so I don't think when I think about safety, I don't just look at the technical measures of it. But how does this technology manifest in society?

53:21How do we make decisions around it? And you could take, put another way, technology is rarely innately good or bad. It's sort of what we do with it. And I think those social constructs matter a lot as well. So I think it's a little early to tell because we don't have this kind of super intelligence right now. And I think it won't just be a technology company defining how it manifests in society. And you could imagine taking a very well-aligned AI system and a human operator directing it towards something that would objectively hurt society. And there's a question of, like, who gets to decide who's accountable?

53:59And it's a perennial question. I mean, it's whether you're deciding, you know, should you use your smartphone in school? You know, who should decide that? And there's parents who will tell you, hey, it's my decision. It's my kid. And then there's principals who will tell you it's not benefiting the school. And I'm not sure that's going to be my place or our place. But there will be a number of those conversations that are much deeper than that question that I think we'll need to answer. As it relates to regulation, there's two not conflicting forces, but two forces that exist somewhat independently but relate to each other.

54:36One is the pace of progress in AI and ensuring that the folks working on frontier models are ensuring those models do benefit humanity. And then there's the sort of geopolitical landscape, which is, do you want AGI to be created by the freedom, sort of the West, by democracies? Or do you want it to be created by more totalitarian governments? And so I think the inherent tension for regulators will be a sense of obligation to ensure that, you know, the technology organizations creating AGI are in fact focusing enough on better and fitting humanity, all the other stakeholders whose interests that they're accountable for, and ensuring that the West remains competitive.

55:33and I think that's a really nuanced thing and I think my view is it's very important that the West leads in AI and I'm very proud of the fact that OpenAI is based here in the United States and we're investing a lot in the United States and I think that's very important and I also, having sort of seen the inside, I think we're really focused on benefiting humanity. So I tend to think that it needs to be a multi-stakeholder dialogue but I think there's a really big risk that some regulations could have the unintended consequence of slowing down this larger conversation. But I don't say that to be dismissive of it either.

56:09It's actually just an impossibly hard problem. And I think you're seeing it play out, as you said, in really different ways in Canada, United States, Europe, China, elsewhere. I want to come back to compute and the dollars involved. So, I mean, on one hand, you have, if I could start an AI company today by putting my credit are down using AWS and leveraging their infrastructure, which they've built. They've spent the hundreds of billions of dollars, and I get to use it on a time-based model. On the other hand, you have people like OpenAI and Microsoft investing tons of money into it that may be more proprietary.

56:48How do you think about the different models competing? And then the one that really throws me for a bit of a loop is Facebook. So Facebook has spent... Meta. You know the mother made a new. Oh, God. So Meta, I'm like aging myself here. So Meta comes along and, you know, possibly for the good of humanity, but like I tend to think Zuck is like incredibly smart. So I don't think he's spending, you know,$100 billion to develop a free model and give it away to society. How do you think about that in terms of return on capital and return on investment? It's a really complicated business to be in just given the capex required to build a frontier model.

57:32But let me just start with a couple definitions of terms I think are useful. I think most large language models I would call foundation models. And I like the word foundation because I think it will be foundational to most intelligent systems going forward. And most people building modern models, particularly if they involve language, image, or audio, shouldn't start from building a model from scratch. They should pick a foundation model, either use it off the shelf or fine-tune it. And so it's truly foundational in many ways. In the same way, most people don't build their own servers anymore.

58:10They lease them from one of the cloud infrastructure providers. I think foundation models will be something trained by companies that have a lot of capex and leased by a broad range of customers who have a broad range of use cases. and I think that leads in the same way that data center builders having a lot of data centers enabled you to have the capital scale to build more data centers. I think the same will largely be true of building the huge clusters to do training and things like that. Foundation models I think are somewhat distinct from frontier models and frontier models I think it's a term credited to Reid Hoffman but I may be mistaken on that that's where I heard it from And these are the models that are usually like the one or two that are clearly the leading edge, O3, as an example from OpenAI.

59:00And these frontier models are being built by labs who are trying to build AGI that benefits humanity. And I think if you're deciding whether you're building a foundation model and what your business model is around it, it's a very different business, then I'm going to go pursue AGI. because if you're pursuing AGI, really there's only one answer, which is to build and train and move to the next horizon because if you can truly build something that is AGI, the economic value is so great. I think there's a really clear business case there. If you're pre-training a foundation model that's the fourth best, that's going to cost you a lot of money.

59:46And the return on that investment is probably fairly questionable because why use your fourth best large language model versus a frontier model or an open source one for meta? And as a consequence of that, I think we probably have too many people building models right now. There's already been some consolidation, actually, of companies being folded into Amazon and Microsoft and others. But I do think it will play out a bit like the cloud infrastructure business where a very small number of companies with very large CapEx budgets are responsible for both building and operating these data centers.

1:00:24And then developers and consumers will use things like ChatGPT as a consumer or as a developer. You'll license and rent one of these models in the cloud. How it will play out is a really great question. I heard one investor talk about these as the fastest depreciating assets of all time. On the other hand, if you look at the revenue scale of something like an open AI and what I've read about places like Anthropic, let alone Microsoft and Amazon, it's pretty incredible as well. And so you can't really, if you're one of those firms, you can't afford to sit on the sidelines as the world transforms.

1:01:07But I would have a hard time personally funding a startup that says, I'm going to do pre-training. I don't really know what's your differentiation in this marketplace. And I think a lot of those companies, you're already seeing them consolidate because they have the cost structure of a pharmaceutical company, but not the business model. This is just it, though, right? Like, OpenAI has a revenue model around a revenue model. Microsoft has a revenue model around their AI investments. They just updated the price of Teams with Copilot. You know, Amazon has a revenue model around AI. In a sense, they're getting other people to pay for it through AWS, and then they're getting the advantages of it at Amazon, too, from a consumer point of view and all the millions of projects.

1:01:55Bezos was doing an interview last week. He said every project at Amazon basically has an AI component to it now. Facebook, on the other hand, has spent all of this money already and with an endless amount, presumably in sight, or not in sight, an endless amount to go. But they don't have a revenue model specifically around AI where it would have been cheaper, obviously, for them to use a different model, but that would have required presumably giving data away. or like, I'm just trying to work through it from Zuck's point of view. Yeah, I actually will take Mark at his word and that post he wrote about open source, I think was very well written and encouraged people to read it.

1:02:38I think that's a strategy. And if you look at Facebook, now you've got me saying Facebook too. So that was what it was called. I messed you up. You know, the company has always really embraced open source. And if I look at really popular things from React to now the Llama models, it's always been a big part of their strategy to court developers around sort of their ecosystem. And Mark articulated some of the strategy there, and I'm sure there's elements of commoditizing your complement, but I also think that if you can attract developers towards models, there's a strength. I'm not really on the inside there, so I don't really have a perspective on it other than I actually think it's really great that there's different players with different incentives all investing so much and i think it is really furthering the cause of like bringing these amazing tools to society and um but a lot changes i mean if you look at the price of gpt4o mini yeah you know it is uh so much higher quality than like the highest quality model two years ago and much cheaper yeah um i i haven't done the math on it, but it's probably cheaper to use that than to host, self-host any of the open source models.

1:03:54So even, even the existence of the open source models, it's not free. I mean, inference costs money. And so there's a lot of complexity here. And actually I have the email even being relatively close to stuff. Like I have no idea where things are going, but you know, it's, um, you could talk to a smart engineer and they'll tell you, oh yeah, if you built your own servers, you'll spend less than renting them from say, Amazon Web Services or Azure. That's sort of true in absolute terms, but misses the fact, like, do you want someone on your team building servers? Oh, and in fact, if you change the way your service works and you need a different SKU, like you all of a sudden are doing training and you need NVIDIA, you know, H100s, now all of a sudden you're built servers like this, you know, asset that's worthless.

1:04:41So I think with a lot of these models, you know, The presence of open source is incredibly important, and I really appreciate it. I also think the economics of AR are pretty complex because the hardware is very unique. The cost to serve is much higher. Techniques like distillation have really changed the economics of models, whether or not it's open source or hosted and leased. so it's I think broadly speaking for developers it's kind of an amazing time right now because you have a like a menu of options that's incredibly wide and I actually think of it as you know just like in cloud computing you'll end up with a price performance quality trade-off and for any given engineering talents they'll have a different answer and that's appropriate and some people use open source Kafka some people work with Confluent great you know like that's just the way these things work you know and so you don't think agi is going to be like a winner take all you think there's going to be multiple options that have by definition whatever the definition is of agi well first i think open ai i believe will play a huge part in it because uh there's both the technology which i think open ai continues to lead on um but also chat gpt which has become synonymous with AI for most consumers.

1:06:04But more than that, it is the way most people access AI today. And so one of the interesting things like what is AGI, we talked about, you know, opinions on what the definition might be. But the other question is like, how do you use that? Like, what is the packaging? And some of intelligence will be simply the outcomes of it, like discovery of a new drug, which would be, you know, remarkable and hopefully we can cure some illnesses. But others will be just how you as an individual access it. And, you know, I, most of the people I know, like if they're signing an apartment lease, we'll put it in chat GPT, you get a legal opinion.

1:06:45If you get, you know, lab results from your doctor, you can get a second opinion on chat GPT. Clay and I use the O1 pro mode for like criticizing our strategy at CIRA all the time. And so for me, what's so remarkable about ChatGPT, which was this quirkily named research preview that has come to be synonymous with AI, is I do think that it will be the delivery mechanism for AGI when it's produced. And not just because of the many researchers at OpenAI, but because of the amazing utility it's become from individuals. And I think that's really neat because, I don't know if it would have been obvious if we were having this conversation three years ago and you were talking about artificial general intelligence.

1:07:34I'm not sure either of us would have envisioned something so simple as a form factor to absorb it that you just talk to it. So I think it's great and especially as I think about the mission of OpenAI which is to ensure that AGI benefits humanity. What a simple accessible form factor. There's free tiers of it. What a kick-ass way to benefit humanity. So I really think that will be central to what we come as society to define as an AGI. You mentioned using it at Sierra to critique your business strategy. What do you know about prompting that other people miss? I mean, you must have the best prompts.

1:08:14People think that, you know, because I'm affiliated with it. You're not going like, here's my strategy. What do you think? What are you putting in there? I often, with the reasoning models, which are slower, will use a faster model first, GPT-4.0, to refine my prompts. So over the holidays, partly because I was thinking about the future of software engineering, I've written a lot of compilers in my time. I've written enough that it's easy for me. So I decided to see if I could have 01 pro mode generate end-to-end a compiler, front-end parsing the grammar, checking for semantic correctness, generating an intermediate representation, and then using LLVM, which is sort of a compiler collection that's very popular to actually do, you know, run it all.

1:09:12and I would spend a lot of time iterating on 4.0 to sort of like refine and make more complete and specific what I was looking for and then I would put it into one promo go get my coffee and you know come back and get it I'm not sure if that's a viable technique but it's really interesting because I do think in the spirit of AI being the solution to more problems than AI having a lower latency, simpler model, help refine. Essentially, I like to think of it as like, you're like a product manager and you're asking an engineer what to do. Is your product requirements document complete and specific enough?

1:09:50And waiting for it is sometimes slower than, and so I like doing it in stages like that. So that's my trip. At some point, there's probably someone from OpenAI listening who's going to like roll their eyes, but that's just, that's I've heard. Who can I talk to at OpenAI that's like the prompt? Yeah. I'm so curious about this because I've actually taken recently to getting OpenAI or ChatGPT, I guess, if you want to call it that. I've been getting ChatGPT to write the prompt for me. So I'll prompt it with, I'm prompting an AI. Yeah. Here are the key things. So similar to my technique. I want to accomplish.

1:10:25What would an excellent prompt look like? And then I'll copy paste that prompt that it gives me back into the system. but I'm like, I wonder what I'm missing here, right? It's a good technique. I mean, there's lots of techniques like that. Like self-reflection is a technique where you have a model, observe and critique, you know, a decision like a chain of thought. So in general, you know, that mechanism of self-reflection is I think a really effective technique. You know, at Sierra, we help companies build customer-facing AI agents. So if you're setting up a Sonos speaker, you'll now chat with an AI.

1:11:01If you're a SiriusXM subscriber, you can chat with Harmony, who's their AI, to manage your account. We use all these tricks, you know, self-reflection to detect things like hallucination or decision-making, generating chains of thought for more complex tasks to ensure that it's, you know, you're putting as much compute and cognitive load into important tricks. So, you know, there's a whole industry around sort of figuring out how do you exact the robustness and precision out of these models. So it's really fun, but changing rapidly. Hypothetical question. You've been hired to lead or advise a country that wants to become an AI superpower.

1:11:48What sort of steps would you take? What sort of policies would you think would help create that? How would you bring investment from all over the world into that country? And researchers, right? So now all of a sudden you're competing. It's not the United States. How do you sort of set up a country from first principles all the way back to what does that look like? What are the key variables? Well, I mean, especially this is definitely outside of my domain of expertise, but I would say one of the key ingredients to modern AI is compute, which is a noun that wasn't a noun until recently, but now compute is a noun.

1:12:26And I do think that's one area where policymakers can, because it involves a lot of things that touch federal and local governments like power, land, and then similarly attracting the capital, which is immense to finance the real estate, to purchase the, you know, compute itself, and then to sort of operate the data center. And again, there's really immense power requirements for these data centers as well. And then, you know, it's attracting sort of the right researchers and research labs to, you know, leverage that. But in general, where there is compute, the research labs will find you, you know, and I think that's it.

1:13:12And then there's a lot of national security implications too, just because, you know, These models are very sensitive, at least the frontier models are. And so, you know, how you your place in the geopolitical landscape is quite important. Like will research labs and will the U.S. government be comfortable with training happening there and export restrictions and things like that? But I think a lot of it comes down to infrastructure as it relates to policy is my intuition. you know i think right now so much of ai is constrained on on infrastructure that that is the input to a lot of of this stuff um uh and then there's a lot around you know attracting talent and all that but as i said you know you look at the research labs it's not that many people actually it's a lot but the compute is a limited resource right now that's a really good way to I think about this from the lens of Canada, right?

1:14:09Which is like, we don't have enough going on in AI. We tend to lose most of our great people to the States who then go to set up infrastructure here for whatever reason and don't bring it back to Canada. And I wonder how Canada can compete better. So this is like sort of the lens I like look at these questions through. How do you see the next generation of education? Like if you were setting up a school today from scratch, and again, hypothetical, not your domain of expertise, but like using your lens on AI, how do you think about this? So like what skills will kids need in the future? And what skills do we probably don't need to teach them anymore that we have been teaching them?

1:14:54Well, I'll start with the benefits that I think are probably obvious, but I'm incredibly excited about. I think education can become much more personalized. Oh, totally. Have you seen Synthesis Tutor, by the way? No, I have not. Oh, so they developed this Synthesis, this AI company, developed this tutor, which actually teaches kids. And it's so good that El Salvador, the country, just recently adopted and replaced their teachers. That's amazing. And it'll teach you, but it teaches you specific to what you're missing. So it's not like every lesson's the same. It's like, well, you're not understanding this foundational concept.

1:15:32So it's like K through five or six right now. That's amazing. And the results are like off the charts. Well, it doesn't surprise me. And I don't actually view it as like necessarily replacing a teacher. But my view is if you have a teacher with 28 kids in his or her class, the likelihood that they all learn the same way or learn at the same pace is very unlikely. and you know i can really think of a say an english teacher a history teacher orchestrating their learning journeys through a topic say ap european history in the united states there's a curriculum they need to learn it um how someone will remember something or understand the significance of martin luther you know is very different and um you can you know generate a audio podcast for someone who might be an audio, auditory learner.

1:16:23You can create cue cards for someone who needs that kind of repetition, repetition. You can visualize key moments in history for people who just maybe want to more viscerally appreciate why this was a meaningful event rather than this dry piece of history. And all of that, as you said, can be personalized to the way you learn and how you learn. And I think it's just incredibly powerful. And so one of the things I think is neat about AI is it's, democratizing access to a lot of things that used to be fairly exclusive. A lot of wealthy people, their child was having trouble in school would pay for a tutor, a math tutor, science tutor.

1:17:00And, you know, if you look at kids who are trying to get into, you know, big name colleges, you know, if you have the means, you'll have someone prep you for the SATs or help you with your college essays. All of that should be democratized if we're doing our jobs well. And it means that we're not limiting people's opportunity from by their means and i think that's just the the most uh american thing ever uh canadian as well it's the most incredible thing it's the most incredible thing humanity and and so i i just think education will change for the positive in so many ways because uh i actually with my kids walking around when they ask uh you know if you have little kids ask why why why and you know there's some point a parent just starts making up the answer being dismissive and like we have chat gpt out it's like the best when you're traveling and put on advanced voice mode and be like ask away a hundred percent and i'm listening to you know it's like you're you live through your children's curiosity and um my daughter went to high school and came home with shakespeare for the first time and i was she asked me a question i was like i i felt this is like total inadequacy i was like i was very bad at this the first time and then we put it into chat GPT and it was the most thoughtful answer and she could ask follow-up questions.

1:18:16And I actually was, you know, with her because I was like, oh, I forgot about that. You know, didn't even think about that. So I just think it's incredible. And I would like to, in public school systems, I think it's really, I think it would be a really great when public school systems formally adopt these things so that they lean into tools like chat GPT as mechanisms to like like raise the performance level of their classroom. And hopefully you'll see it in things like test scores and other things because kids can get the extra time, even if the school system can't afford it for everyone. And then most importantly, kids are getting explanations according to their style of learning, which I think will be quite important as well.

1:19:04As it relates to skills, it's really hard to predict right now. And I would say that I do think learning how to learn and learning how to think will continue to be important. So I think most of primary and secondary education is not vocational necessarily. Some of it is. I took auto shop and all of that, and I'm glad I did, but I couldn't fix my electric car today with that knowledge. Things change, and I don't think it needs to be purely non-vocational. But the basics of learning how to think, learning, writing, reading, math, physics, chemistry, biology, not because you need to memorize it, but understand the mechanisms that create the world that we live in is quite important.

1:19:54I do think that there's a risk of people sort of becoming ossified in the tools that they use. So, you know, let's go back to our discussion of software engineering for a second, but I'll give other examples. You know, if you define your role as a software engineer is how quickly you type into your IDE, the next few years might leave you behind, you know, because that is no longer a differentiated part of the software engineering experience or will not be. But your judgment as a software engineer will continue to be incredibly important. your agency and making a decision about what to build, how to build it, how to architect it, maybe using AI models as a creative foil.

1:20:43And so I think that just in the same way, if you're an accountant, using Excel doesn't make you less of an accountant. And just because you didn't handcraft that math equation, it doesn't make the results any less valuable to your clients. and so I think we're going to go through this transformation where I think the tools that we use to create value in the world will change dramatically and I think some people who define their jobs by their ability to use the last generation's tools really really effectively will be disrupted but I think if we can empower people and to reskill and also broaden the aperture by which they define the value they're providing to the world, I think a lot of people can make the transition.

1:21:32The thing that is sort of uncomfortable, not really in education, or it's just earlier in most people's lives, is just I think the pace of change exceeds that of most technology transitions. And I think it's unreasonable to expect most people to change their way to work that quickly. And so I think the next five years, I think will be, you know, for some jobs really disruptive and tumultuous. But if you take the longer view and you fast forward 25 or 50 years, I'm incredibly optimistic. I think it's the change will require from society, from companies and from individuals like an open mindedness about reskilling and reimagining their job to the lens of this like dramatically different new technology.

1:22:19At what point do we get to, I mean, we're probably on the cusp of it now and it's happening in pockets, but what point do we start solving problems that humans haven't been able to solve or eliminating paths that we're on maybe with medical research that it's like, no, that, that, this whole thing you've spent$30 billion on, you know, based on this 1972 study that was fabricated, but that one study had all these derivative studies and like, I'm telling you it's false, you know, because I can look at it through an objective lens and get rid of these 30 billion while you're smiling 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 one of the models i can't remember which one introduced a very long context window and there's a lot of people on x over the weekend putting in their uh thesis you know like grads grad school thesis in there and and it was actually critiquing them with like surprising levels of fidelity.

1:23:16And I think we're sort of there, perhaps, with the right tools. But certainly, over the next few years, we talked about what does it mean to generalize AI. Certainly, in the areas of science that are largely represented through text and digital technology, like math being probably the most applicable, there's not really anything keeping AI from getting really good at math. There's not really an interface to the real world. You don't need to do a clinical trial to verify something's correct. So I feel a ton of optimism there. It'll be really interesting in areas of theoretical physics. You'll continue to have the divide between the applied and the theoretical people.

1:24:01But I think there could be really interesting new ideas there and perhaps some finding logical inconsistencies with some of the fashionable theories, which has happened many times over the past few decades. I think we'll get there soon. And actually, what's really neat about it is most of the scientists I know, people who are actually doing science, they're the most excited about these technologies. And they're using them already. And I think that's really neat. And I think we're hopefully going to be, I really hope we see more breakthroughs in science. One of the things I am not an expert in, but I've read a lot as an amateur about is just the slowdown in scientific breakthroughs over the past few decades and some theories that it's because of the degree of specialization that we demand of grad students and things like that.

1:24:53I hope in general with AI democratizing access to expertise, I have a completely personal theory that it will benefit deep generalists in a lot of ways too because your ability to understand a fair amount in a lot of domains and leveraging AI, knowing where to prompt the AI to go explore and bringing together those domains, it will start to shift sort of the intellectual power from people who are extremely deep to people who actually can orchestrate intelligence between lots of different domains for breakthroughs. I think that'll be really good for society because most scientific breakthroughs aren't.

1:25:37They tend to be cross-pollinating very important ideas from a lot of different domains, which I think will be really exciting. How important is the context window? I think it could be quite important, especially it certainly simplifies working with an AI. You can just give it everything and instruct it to do something. And assuming it works, you can extend a context window and the attention can be spread fairly thin and the robustness of the answer can be questionable. But assuming, for argument's sake, perfect robustness, I think it can really simplify the interface to AI. Not all uses. I also think that we're talking about open source models and APIs.

1:26:28I also think that most, What I'm excited about in the software industry is not necessarily a large language model with a prompt and a response being the product of AI, but actually end-to-end closed-leaf systems that use large language models as pieces of infrastructure. And I actually think that a lot of the value in software will be that. And for many of those applications, the context window size can matter. But often, because you have contextual awareness of the process that you're executing, context window is a little bit less important. So I think it matters a lot to intelligence. You know, there's a, I can't remember some, one of the, some researcher said, you know, you put all of human knowledge in the context window and you ask it to invent the next thing.

1:27:11And it's obviously a reductive thought, but interesting. but I actually I'm equally excited about sort of the industrial applications of large language models sort of like my company Sierra and if you're returning a pair of shoes at a retailer and it's a process that's fairly complicated and you know is it within the return window you know do you want to return it in store do you want to send it do you want to print a QR code blah blah blah the orchestration of that is as significant as the models themselves and I actually think as we, just like computers, you know, there's going to be a lot of things where computers are a part of the experience, but it's not like manifesting itself as a computer.

1:27:52So I'm actually equally excited about those. And I think context window is slightly less important in those applications. Do you think that the output from AI should be copyrightable or patentable? Or let me just take an example. If I go to the US patent office, I download a patent for, let's say the AeroPress and I upload it to 01 Pro and I say, I can't upload it yet because you don't let me do the PDFs, but I upload it to 4 and so I say, hey what's the next logical leap that I could patent off this? It would give me back diagrams and an output and presumably if I look at that and I'm like, yeah, that's legit, I want to file that patent, can I?

1:28:31I don't know to answer that question, I'm not an expert in sort of intellectual property but I I think there will be an interesting question of was that your idea because you used a tool to do it I think the answer is probably yes that you you used the tool to do it but I also think that the

1:28:53in general like the sort of marginal cost of intelligence will go down a lot so a lot of the you know i think in general like we'll we'll be in this renaissance of of new ideas and intelligence being produced and so uh i think that's broadly a good thing and i think you know the marginal value of that insight that you had might be lower than it was you know and years ago what i was hoping you would say is that you know that's going to become less and less important because i feel like all the patent trolls and all of this stuff that slows down innovation in some ways. Obviously, like there's legitimate patents that people infringe on and there should be legal recourse.

1:29:32But if I could just go and patent like 100 things a day, because it seems like that should not be allowed. This is what I'm saying, though. In general, I think that, you know, companies, you know, I think patents make sense if it's protecting something that's an active use that you, you know, invented and you're you're trying to, you know, like the standard, you know, So legal rationale for patents, just generating a bunch of ideas and patenting, it seems destructive to the value of it. So here's the idea I had last night to counter this because I was like, I don't want somebody doing this. And I was thinking like, what if prior art eliminates patents?

1:30:08So I was like, what if I just set up like an instance and just publish it on a website? Nobody has to read that website. Here's a billion ideas. Exactly, but it's basically patenting anything. PriorArt.org. But it's creating prior art for everything. So you can't compete on that anymore. I don't know. I was thinking about that. I thought it was fun. Tell me about the Google Maps story. This is now legend, and I want to hear it from you. This is my weekend coding. Is that what you want to hear about? Yeah. Yeah, so I'll start with just the story of Google Maps, the abbreviated version. We had launched a product at Google called Google Local, which was sort of a Yellow Pages search engine.

1:30:53You probably, most listeners don't even know what Yellow Pages are, but it was a thing back then. We had licensed maps from MapQuest, which was the dominant mapping provider at the time. It was sort of an eyesore on the experience and also always felt like it could be a more meaningful part of the local search and navigation experience on Google. So Larry Page in particular was really pushing us to really invest more in maps. We found this small company with like four people in it, if I'm remembering correctly, started by Lars and Jens Rasmussen called Where2 Technologies, where they had made a Windows application called Expedition that was just a beautiful mapping product.

1:31:36It was running on Windows long after it was sort of out of fashion to make Windows apps, but that was sort of where the technology they're comfortable with. But their maps modeled the A to Z maps in the UK and were just beautiful. And they just had a lot of passion for mapping. So we did a little aqua hair of them and took together the Google Local team and Lars and Jens' team and said, OK, let's take the good ideas from this Windows app and the good ideas from Google Local and let's bring them together to make something completely new. And that's what became Google Maps. but there's a couple of idiosyncrasies in the integration because it was a windows app it really helped and hurt us in a number of ways like one of the ways it helped us is the reason why google maps we were able to drag the map and it like uh was so much more interactive than any web application that preceded it was the standard that we needed to hit from interactivity was set by a native windows app not set by the legacy uh you know websites that we had used at the time.

1:32:42And I think that by having the goalposts so far down the field, because they had just started with this Windows app, which was sort of a quirk of Lars and Jens, just like technical choices, we made much bolder technical bets than we would have otherwise. I think we would have ended up much less interactive had we not started with that quirky technical sort of decision. But the other thing was this Windows app, there's a lot of like, it's hard to describe the early 2000s, people didn't live it, but XML was really in fashion. So most things in Windows and other places was like XML and XSLT, which was a way of transforming XML into different XML, was the basis of everything.

1:33:23It was like all of enterprise software was like XML this, XML that. So similarly, when we were taking some of these ideas and putting them in a web browser, we kind of went into autopilot and used a ton of XML. and it made everything just like really, really tedious. And so Google Maps launched with some really great ideas like the draggable maps. And we did a bunch of stuff for the local search technology so you could overlay restaurant listings. It was really great. It was a really successful launch. We were like the hotshots within Google afterwards. But it really started to show its craft.

1:33:59And we got to this point where we decided we wanted to support the Safari web browser, which was relatively new at the time. This was before mobile phones. And there was much less XML support in Safari than there was in Internet Explorer and Firefox. And so one of the engineers implemented a full XSLT transform engine in JavaScript to get it to work. And it was just like shit on top of shit on top of shit. And so what was a really elegant, fast web application had sort of quickly become something. And, you know, there's a lot of dial-up modems at the time. and other things. So like you'd show up to maps and it just was slow.

1:34:37And like, it just bothered me as like someone who takes a lot of pride in their craft. And so I got really, uh, energized and like over a weekend and a lot of coffee, like rewrote it. Um, but it was the whole thing there rewrote. Yeah. More or less the whole thing. And it took probably another week of like, you know, working through the bugs, but yeah, I sent it out to the, you know, the team after that weekend. And it was, it was nice. The reason I was able to do it, yeah, I'm like a decent programmer, but you know, you'd also like lived with every bad decision up to that point too. So I knew exactly the output I was going to, like I had simulated in my head, like if I could do it over again, this is the way I do it.

1:35:19So by the time I like put my hands on the keyboard on like, you know, Friday night, I, it wasn't like I was designing a product. Like I knew I had been in every detail of that product since the beginning, including many of the bad decisions, too. They're not all the bad decisions. And so it was just very clear. I knew what I wanted to accomplish. And for any engineers who worked on a big system, you have the whole system mapped out in your head. So I knew everything. And I also knew that there's a lot of pride of authorship with engineering and code. So I sort of knew I really wanted to finish it over the weekend so that people could use it and see how fast it was and kind of overcome anyone who was like, you know, you know, protective of the code they had written a few months ago.

1:36:04And so I really wanted the prototype to go out. So I did it. And then I didn't, it's funny, I never talked about it again, but I think Paul who was the co-creator of Gmail and, and I worked and started friend feed with me. He was on an interview and mentioned this story. So now all of a sudden it's like, everyone's talking about it. And I was like, well, thank you, Paul. It's a little embarrassed that people know about it. But it's a true story. And XML is just the worst. Did you get a lot of flack from the people who had built the system you effectively replaced? Like you were part of that team, but everybody else had so much invested in it, even though it was like shit on top of shit on top of shit.

1:36:44You know, I wrote a lot of it, too. So, yeah, I'm sure there was some around it. But actually, I think good teams want to do great work. And so I think there was a lot of people constructively dissatisfied with the state of things, too. And, you know, I think, you know, the engineer had written that XSLT transform, I think, was like, you know, a little bit. It's a lot of work. So you have to throw out a lot of work, which feels bad. But particularly, you know, Lars and Jens and I, like, we want to make great products. And so I don't think there was a, you know, at the end of the day, everyone was like, wow, that's great.

1:37:18We went from a bundle size of 200K to a bundle size of 20K, and it was a lot faster and better. So broadly speaking, I think good engineering culture is you don't want a culture of ready, fire, aim. But I also think you just need to be really outcomes-oriented. And I think people, if they become – they start to treat their code as too precious, it can really impede forward progress. and I'll just take like my understanding is like a lot of the early self-driving car software was a lot of hand-coded heuristics and rules and you know a lot of smart people think that eventually it'll probably be a more monolithic model that encodes many of the same rules.

1:38:02You have to throw out a lot of code in that transition but it doesn't mean it's not the right thing to do and so I think in general yeah there might have been some feathers ruffled but at the end of the day everyone's like that's faster and better like let's do it. which is, I think, the right decision. That's awesome. I'm going to give you another hypothetical. I want you to share your inner monologue with me as you think through it. So if I told you you have to put 100 % of your net worth into a public company today and you couldn't touch it for at least 20 years, what company would you invest in?

1:38:34And walk me through your thinking. I literally don't know how to answer that question. How would you think about it without giving me an answer? Yeah, that's a good question. First of all, I'll give you how I think about it, but having not been a public company CEO for a couple years, I blissfully don't pay attention as much to the public markets. And in particular right now, obviously valuations have gone up a lot. But because it's a long-term question, maybe that doesn't matter. I think what I'd be thinking about right now is over the next 20 years, like what are the parts of the economy that will most benefit from this current wave of AI?

1:39:14That's not the only way to invest over a 20-year period, but certainly it's a domain that I understand. And in particular, you know, I mentioned that talk I heard a snippet of from Tyler Cohen, which is like it will probably AI will probably benefit different parts of the economy disproportionately. There will be some parts of the economy that can essentially where intelligence is a limiting factor to its growth and where you can absorb almost arbitrary levels of intelligence and generate almost arbitrary levels of growth. Obviously, there's limits to all of this just because you change one part of the economy, it impacts other parts of the economy.

1:39:52And that was what Tyler's point was in his talk. But I would probably think about that because I think that over a 20-year period, there are certain parts of society that won't be able to change extremely rapidly. But there will be some parts that probably will and will probably be domains where intelligence is the scarce resource right now. And then I would probably try to find companies that will disproportionately benefit from it. And I assume this is why like NVIDIA stock is so high right now, because if you want to sort of get downstream, you know, NVIDIA will probably benefit from all of the investments in AI.

1:40:31I'm not sure I would do that over a 20 year period, just assuming that the infrastructure will shift. So I don't have an intelligent answer, but that's the way I would think about it if I were doing that exercise. I love that. where do you think, like what's your intuition say about what areas of the economy are limited by intelligence and not just economy? I mean, perhaps politicians might be limited by this and aid and benefit from, in which case countries could benefit enormously from AI and unlock growth and potential in their economy. But I think maybe just to scope the question, what areas of the economy do you think are limited by intelligence or workers, like smart workers, in which case, like, that's another limit of intelligence.

1:41:14Yeah, I mean, two that are, I think, probably going to benefit a lot are technology and finance, you know, where you're, you know, if you can make better financial decisions than competitors, you'll generate outsized returns. And that's why over the past, you know, 30 years, you know, of machine learning, hedge funds and financial service institutions, everything from fraud prevention to true investment strategies. It's already been an area of domain of investment. Software, similar as we talked about, I think that at some point we will no longer be supply constrained in software, but we're not anywhere close to it right now.

1:41:58And you're taking something that has always been the scarce resource, which is software engineers and you're making it not scarce. And I think as a consequence, if you just think of like, how much can that industry grow? We don't know. But we've been so constrained on software engineering as a resource, who knows over the next 20 years, but we'll find out where the limits are. But to me, intellectually, there's just a ton of growth there. And then broadly, I think areas of like processing information are areas that will really benefit quite a bit here. And so that, and I think the thing that I would think about over a 20 year period is like second and third order effects, which is why I don't have an intelligent answer.

1:42:39And if you're asking me to put all my money in something, I would think about it for a while. Probably use a one pro a little bit to help me. But, you know, because you can end up generating a bunch of growth in the short term, but then, you know, if everyone does it, it commoditizes the whole industry, you know, type of thing. So, you know, there used to be, you know, before the introduction of the freezer, ice was like a really expensive thing and now it's free, you know. And so I think it is really important to actually think through those if you're talking a timeframe of like 20 years. And that's why having not thought about this question ahead of time, you could be quite simplistic elsewhere.

1:43:16But I would say software and finance are areas that I think stand to reason should benefit quite a bit. I love that response. How do you balance, uh, having a young family with also running a startup again? I work a lot. Um, I don't, uh, I really care and love, care about and love working. Um, so one thing is that I, um, well, there's always trade-offs in life. If I didn't love working, I wouldn't do it as much as I do. But I just love to create things and love to have an impact. And so I like jump out of bed in the morning and work out, go to work, and then spend time with my family. Broadly, probably being honest.

1:44:03First, I'm not perfect but second, I don't have a ton of hobbies. I basically work and spend time with my family. the first time we talked you saw a couple guitars in my background yeah I haven't picked one of those up in a while I mean I literally pick it up occasionally but I you know do not devote any time into it and I don't regret that either like I am so passionate about what we're building at Sierra I'm so passionate about opening I am so love my family so much I don't really have any regrets about it but I basically just like life is all about where do you spend your time and and mine is at work and with family.

1:44:38And so that's how I do it. I don't know if I'm particularly balanced, but I don't strive to be either. I really take a lot of pride, and I love to work. Having sold the companies you started twice, does that influence what you think of Sierra? Like, are you thinking like, oh, I'm building this in order to sell it, or do you think differently? Like, this is my life's work. I'm building this. That's not going to happen. I absolutely intend Sierra to be an enduring company and an independent company. But to be honest, every entrepreneur, every company starts that way. And so, you know, I'm really grateful for both Facebook and Salesforce for having acquired my previous companies.

1:45:20And hopefully I had an impact about those companies. But you don't start off, well, at least I never started off saying, hey, I want to make a company to sell it. but I actually think with Cira we have just a ton of traction in the marketplace I really do think Cira is the leader in helping consumer brands build customer facing AI agents and I'm really proud of that so I really see a path to that and I joke with Clay I want to be an old man sitting on his porch complaining how the next generation of leaders at Cira don't listen to us anymore I want this to be something that not only is enduring but outlives me And I think just actually, I don't think we've ever talked about this, but it was a really interesting moment for me when Google went from its one building in Mountain View to its first corporate campus.

1:46:09We moved into the Silicon Graphics campus, which was right over near Shoreline Boulevard in Mountain View. And SGI had been a really successful company enough to build a campus. And it was actually quite awkward. we moved into like half the campus. They were still in half and they're like, we're this up and coming company. They're declining. And then when Facebook, when we moved out of the second building, we were in Palo Alto, a slightly larger building. I think we leased it from HP. But when we finally got a campus, it was from Sun Microsystems who had gone through an Oracle acquisition and had been sort of on the decline.

1:46:45And it was interesting to me because both SGI and Sun had been started and grown to prominence in my lifetime. Obviously, I was maybe like a little younger, obviously, but in my lifetime, enough to build a whole corporate campus and then decline fast enough to sell that corporate campus to a new software company. And for me, it was just so interesting to have done that twice to move into like a, you know, a used campus for the previous owners. It was a very stark reminder that technology companies aren't entitled to their future success. And I think we'll see this actually now with AI. AI, I think, will change the landscape of software to be tools of productivity to agents that actually accomplish tasks.

1:47:33And I think it will help some companies for whom that amplifies their existing value proposition. and it will really hurt others where it will essentially the seat-based kind of model of legacy software will wane very quickly and then really harm them. And so when I think about the, what it means to build a company that's enduring, that is a really, really tall task in my mind right now, because it means not only making something that's financially enduring over the next 10 years, but setting up a culture where a company can actually evolve to meet the changing demands of society and technology when it's changing at a pace that is like unprecedented in history.

1:48:18So I think it's one of the most fun business challenges of all time. And I think it has as much to do with culture as it has to do with technology because every line of code in Sierra today will be completely different, you know, probably five years from now, let alone 30 years from now. And I think that's really exciting. So when I think about it, I just get so much energy because it's incredibly hard and it's harder now than it's ever been to do something that lasts beyond you. But that I think is the ultimate measure of a company. You mentioned AI agents. How would you define that? What's an agent?

1:48:56I'll define it more broadly and then I'll tell you how we think about it at CIRA, which is a more narrow view of it. The word agent comes from agency and I think it means affording a software the opportunity to reason and make decisions autonomously. And I think that's really all it means to me and I think there's lots of different applications of it. The three categories that I think are meaningful and I'll end with the Sierra one just so I can talk about it a little more but one is personal agents. So I do think that most people will have probably one, but maybe a couple AI agents that they use on a daily basis that are essentially amplifying themselves as an individual.

1:49:41You can do the rote things like help you triage your email to helping you schedule a vacation. You know, you're flying back to Edmonton and help you arrange your travel. two more complex things like, you know, I'm going to go ask my boss for promotion, like help me role play. And, you know, I'm setting up my resume for this job, help me do that too. I'm applying for a new job, help me find companies I haven't thought of that I should be applying to. And I think these agents will be really powerful. I think it might be a really hard product to build because when you think about all the different services and people you interact with every day, it's kind of everything.

1:50:22So it's not, it has to generalize a lot to be useful to you. And because of the personal privacy and things like that, it has to work really well for you to trust it. 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 a while to be robust. The second category of agent is I would say really filling a persona within a company. So a coding agent, a paralegal agent, a analyst agent. I think these already exist. I mentioned Cursor. There's a company called Harvey that makes a legal agent. I'm sure there's a bunch in the analyst space. These do a job and they're more narrow, but they're really commercially valuable because most companies hire people or consultants that do those things already, like analyze the contracts of your supply chain, right?

1:51:13That's kind of a rote kind of law, but it's really important. And AI can do it really well. So I think that's why this is the area of the economy that I think is really exciting. And I'm really excited about all the startups in this space because you're essentially taking what used to be a combination of people and software and really making something that solves a problem. and by narrowing the domain of autonomy, you can have more robust guardrails and even with current models actually achieve something that's effective enough to be commercially viable today. And by the way, it changes the total addressable market of these models too.

1:51:55I don't know what the total addressable market of legal software was three years ago, but it couldn't have been that big. I couldn't tell you a legal software company. I probably should. I just can't think of one. But if you think about the money we spend on lawyers, that's a lot. And so you end up where you're broadening the addressable market quite a lot. The domain we're in, I think, is somewhat special, which is a company's branded customer-facing agent. And the reason why I think it's – one could argue we're sort of helping with customer service, which is a persona, a role. But I do think it's broader than that because if you think about a website like your insurance company's website, try to list all the things you can do on it.

1:52:41You can look up the stock quote. You can look up the management team. You can compare their insurance company to all their competitors. You can file a claim. You can bundle your home and auto. You can add a member of your family to your premium. There's a million things you can do on it. Essentially, over the past 30 years, websites, a company's website, singular, has come to be the universe of everything that you can do with that company. I like to think it was like the digital instantiation of the company. And that's what we're helping our customers do at Cira, is help them build a conversational AI that does all of that.

1:53:22So, you know, most of our customers start with customer service. and it's a great application because no one likes to wait on hold and having something that has perfect access to information is multilingual and empathetic is just amazing. But when you put a conversational AI as your digital front door, people will say anything they want to it. And we're now doing product discovery, consider purchases, going back to the insurance example. Hey, I've got a 15-year-old daughter. I really am concerned about the cost of her premium until she grows up. Tell me which plan I should be on. Tell me why you'll be better than your competitors.

1:54:00That's a really complex interaction, right? That's not something that, can you make a web page that does that? No, but that's a great conversation. And so we really aspire that when you encounter a branded agent in the wild, we want Sierra to be the platform that powers it. And it's super important because there was a case, at least in Canada, where an AI agent for Air Canada. Hallucinated a bereavement policy. Right. But they were found liable to hold themselves to what the agent said. Yeah. I mean, it turns out... And it was an AI agent. There was no human involved in the whole thing. Well, look, it's one thing if ChatGPT hallucinates something about your brand.

1:54:39It's another if your AI agent hallucinates something about your brand. So the bar just gets higher. So the robustness of these agents, the guardrails, everything is more important when it's yours and it has your brand on it. And so it's harder, but I also, I'm just so excited for it because this is a little overly intellectual, but I really like the framing. If you think about a modern website or mobile app, it's essentially you've created a directory of functionality from which you can choose. But the main person with agency in that is the creator of the website. Like what are the universe of options that you can do?

1:55:20When you have an AI to represent your brand, the agency goes to the customer. They can express their problem any way they want in a multifaceted way. And so it means that like your customer experience goes from the enumerated set of functionality you've decided upon your website to whatever your customers ask. And then, you know, you can decide how to fulfill those requests or whether you want to yeah um but i think it would really change the dynamic to be really empowering to consumers as you said i mean the reason that that air canada case um is the reason we exist you know companies if they try to build this themselves um there's a lot of ways you can shoot yourself in the foot but in particular um to your customer experience should be wedded should not be wedded to one model let alone even this current generation of models so with sierra you can define your customer experience once in a way that's abstracted from all of the technology.

1:56:17And it can be a chat. It can be, you can call you on the phone. It can be all of those things. And as new models and new technology comes out, our platform just gets better. But you're not like re-implementing your customer experience. And I think that's really important because, you know, we were talking about what's happened over the past two years. Can you imagine if you're a consumer brand like ADT Home Security and thinking about how can you maintain your AI agent in the face of all that? It's just not even, 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.

1:56:48How do you fend off complacency? Like a lot of these companies, and maybe not in tech specifically, but they get big, they get dominant, and then they take their foot off the gas. And that opens the door to competitors. and there's like a natural entropy almost to bureaucracy in some of these companies that and the bureaucracy sows the seeds of failure and competition how do you how do you fend that off constantly it is a really challenging thing to do at a company um one of the there's two things that i've observed that i think manifest as corporate complacency one is bureaucracy um and i think the root of bureaucracy is often when something goes wrong, companies introduce a process to fix it.

1:57:43And over those like a sequence of 30 years, the layered sum of all of those processes that were all created for good reason with good intentions end up being a bureaucratic sort of machine where the reasons for many of the rules and processes are rarely even remembered by the organization, but it creates this sort of natural inertia. Sometimes that inertia can be good. You know, it's like, you know, if you end up with, there's definitely been stories of executives coming in and ready, fire, aim, new strategies that backfire massively. But often it can mean in the face of a technology shift or a new competitor, you just can't move fast enough to address it.

1:58:30The second thing that I think is more subtle is as a company grows in size, often its internal narrative can be stronger than the truth from customers. I remember one time when the sort of peak of the smartphone wars and I ended up visiting a friend on Microsoft's campus. And I got off the plane in Seattle Tacoma Airport, drove into Redmond, went onto the campus and all of a sudden everyone I saw was using Windows phones. I assume it must have been a requirement or formal or social like you were definitely uncool if you're using anything else. And from my perspective at the time like the war had already been lost.

1:59:18Like it was definitely a two-horse race between Apple and Google on iOS and Android and I remember sitting in the lobby waiting for my friend to get me from the security check-in and I made a comment like it wasn't a confrontation but I made a comment to someone who's at Microsoft I was like you know something along the lines are you required to use Windows phones how these other and I just sort of like curious and then I got a really bold answer which is like yeah we're gonna win like we're taking over the smartphone market and I was like I didn't say anything because it It was like a little socially awkward of like, no, you're not.

1:59:56Like you lost like four years ago. But there's a process. There's something that's happening that's preventing you from getting reality. Well, that's the thing is if you think about it, if anyone – if you've ever worked for like a large company, you know, you're – when you work at a small company, you care about your customers and your competitors and you feel every bump in the road. when you're a junior vice president of whatever and you're eight levels below your CEO and you have a set of objectives and results, you might be focused as I want to go from junior vice president to senior vice president.

2:00:37That's what success looks like for me. And you end up with this sort of myopic focus on this internal world in the same way your kids will focus on the social dynamics of their high school not the world outside of it and it's probably rational by the way because like you know probably their social life is more determined by those you know 1 000 kids in their high school than it is like all the things outside but this is that's the life of a person inside of these big places and so you end up where uh you know if you have a very senior head of product who's like are this competitor says they're faster but this next version we're so much better.

2:01:14And then everyone says, all of a sudden, that's like, the Windows phone is going to win. That's what everyone says. And, and you truly believe it, because everyone you meet says the same thing. And you end up reflecting, you know, customer anecdotes through that lens, and you end up with this sort of reality distortion field manifested from the sum of this sort of myopic storytelling that exists within within companies. The what's interesting about that is like, you know, the ability for a culture to believe in something is actually a great strength of a culture, but it can lead to this as well.

2:01:49And so the combination of bureaucracy and inaccurate storytelling, I think is the reason why companies sort of die. And, and it's really remarkable to look at, you know, the BlackBerrys of the world or the TiVos or the, you know, there, you can really, um you know as the plane is crashing like tell the story that you're not and and and and and and and Similar as I said, like culturally, you can still have like the person in the back of that crashing plane being like, when am I going to get promoted to SVP? And you're like, you know, and that's, I mean, this is like, I've seen it a hundred times.

2:02:28And so I think it really comes down to leadership, you know, and I think that one of the things that most great companies have is they are obsessed with their customers. And I think the free market doesn't lie. And so I think one of the most important things I think for any like enduring culture, particularly in an industry that changes as rapidly as software, is how close are your employees to customers and how much can customer – like the direct voice of your customers be a part of your decision making. And that is something that I think you need to constantly work out because that person, employee number 30 ,462, how does he or she actually directly hear from customers?

2:03:13It's not actually a simple question to answer. Is it direct? Is it filtered? How many filters are there? That's exactly right. And then I think the other part on leadership is, you know, we talk about bureaucracy is process is there to serve the needs of the business. And often mid-level managers don't get credit for removing process. They often are held accountable for things going wrong. And I think it really takes top-down leadership to, you know, remove bureaucracy. and it is not always comfortable when companies remove spans of control or all the people impacted it's like antibodies and for good reason, it makes sense, their lives are negatively impacted or whatever it is, but it almost has to come from the top because you need to give air cover almost certainly something will go wrong by the way processes usually exist for a reason but when they accumulate without end you end up with bureaucracy so those are the two things that i always and you could smell it when you go into a really bureaucratic company the the inaccurate storytelling the process over outcomes and it's just uh it sort of sucks the energy out of you when you feel it that's a great answer we always end these interviews with the exact same question which is what is success for you?

2:04:42Success for me, we talked about how I spend my time with my family at work is, you know, having a happy, healthy family and being able to work with my co-founder, Clay, for the rest of my life, making Sierra into an enduring company. That would be success for me.

2:05:03Thanks for listening and learning with us. The Farnham Street blog is where you can learn more about my new book, Clear Thinking, turning ordinary moments into extraordinary results. It's a transformative guide that hands you the tools to master your fate, sharpen your decision-making, and set yourself up for unparalleled success. Learn more at fs.blog.com. Until next time.