
Bret Taylor: Facebook, Salesforce, OpenAI, and Sierra
What this covers
The conversation is a roughly two-hour exchange between host Shane Parrish and Brett Taylor, whose résumé runs through Google Maps, Facebook, Salesforce, and the chairmanship of OpenAI, and who now runs the agent company Sierra. It moves between management lessons and AI forecasting. The management half covers why companies decline: Taylor points to SGI and Sun Microsystems, whose campuses Google later leased, and to Microsoft employees using Windows Phones out of social pressure rather than preference. His account of corporate complacency) centers on distorted internal narratives and on how many layers filter customer voice. He also argues that removing bureaucracy has to come from the top, since middle managers get blamed for failures but not credited for cutting process.
The technical half starts with software engineering. Taylor expects engineers to become operators of code-generating machines rather than authors. He calls today's code generation a local maximum because languages like Python were built for human authors, and he suggests that nearly free code makes formal verification and test-driven development economical. Humans, he insists, should remain accountable for correctness. From there he turns to AGI): the limits of data, compute, and algorithms, and how reasoning models shift compute to inference and make distillation possible. He is skeptical of the economics of building a fourth- or fifth-place foundation model.
Contested territory includes safety framed as alignment with inseparable governance questions, whether AI output should be copyrightable, and his 20-year bet on sectors where intelligence is the bottleneck, such as software and finance. He expects AI to help most in digital domains like mathematics, less in pharmaceuticals, where clinical trials are the constraint. He also discusses AI tutoring and education. The closing chapters define an AI agent as software that reasons and decides autonomously, and describe how Sierra guards branded agents against hallucination through self-reflection and verification.
Brett Taylor argues that AI agents—software systems capable of autonomous reasoning and decision-making—will become the primary interface through which companies deliver customer value, replacing traditional websites and apps, and that building enduring companies in this era requires obsessive customer focus and cultural resilience against bureaucracy.
- AI agents shift agency from creators (who enumerate options) to customers (who express problems naturally)
- Companies fail not from technological disruption alone but from internal narrative distortion and accumulated bureaucratic inertia that prevent them from seeing market reality
- The future of software is delivering outcomes (task completion) rather than tools, which fundamentally changes business models, skill requirements, and organizational structure
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Technology companies are not entitled to their future success—illustrated by the rise and fall of companies like SGI and Sun Microsystems, which grew large enough to build corporate campuses but declined fast enough that Google later leased those same campuses from them.
“Technology companies aren't entitled to their future success. AI, I think, will change the landscape of software.”
Removing bureaucratic process requires top-down leadership because mid-level managers are typically held accountable for things going wrong but rarely get credit for removing process, so they have no incentive to eliminate it even when it's obsolete.
“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 uh often u mid-level managers uh don't get credit for removing process. they often are held accountable for things going wrong. Um and I think it really takes top down leadership to uh you know remove bureaucracy.”
Keeping employees close to customers is essential to maintaining accuracy of internal narrative and preventing complacency, but it's not a simple problem—it requires deliberate structural and cultural choices about how many layers of filtering exist between front-line employees and customer voice.
“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 uh your decision-m um and that is something that I think you need to constantly work out because that you know person employee number 30,462 you know how does he or she actually actually directly hear from customers is it's not actually a simple question to answer. Is it direct? Is it filtered? How many filters are there?”
Education will be transformed by AI in at least two positive ways: (1) personalization—AI tutors can adapt to how each student learns, providing audio for auditory learners, flashcards for visual learners, etc.; and (2) democratization—tutoring and test prep, which used to be available only to wealthy families, can now be provided at scale to all students.
“I think education can become much more personalized... if you have a teacher with 28 kids in his or her class... I can really think of a say an English teacher, history teacher, orchestrating their learning journeys... 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... you can generate a audio podcast for someone who might be an audio auditorial learner. Um you can create qards for someone who needs that kind of reputition repetition.”
OpenAI's mission—'ensure that AGI benefits all of humanity'—is fundamentally about safety first (analogous to the Hippocratic Oath's 'first, do no harm') but extends beyond safety to include access (avoiding digital divide), job democratization, and economic benefit, making it a holistic benefit framework rather than a safety-only framework.
“I really like OpenAI's mission uh which is to ensure that AGI benefits all of humanity that isn't only about safety um and and I and I believe intentionally so though obviously the mission was created prior to my arrival because it's both about safety kind of hypocratic first do no harm and I don't think one could credibly achieve that mission if we created something unsafe.”
AI safety should be understood as alignment—the AI system doing what we intend it to do—rather than a narrowly technical problem, because even a perfectly aligned AI system can be directed by human operators toward harmful outcomes, making the social and governance dimensions of AI safety inseparable from the technical ones.
“broadly I think about does the system that represents AGI um align with the intentions of the people created it and the intentions of the people operating it. Um so that it does what we want and it's a tool um uh that benefits humanity”
Synthetic data and simulation are important mechanisms to address data scarcity, with simulation being higher quality than arbitrary synthetic data because it encodes real-world constraints (like physics), limiting the permutations of invalid data that could be generated.
“I would say that synthetic data the synthetic data has 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... I view simulation as based on a set of principles like the laws of physics. So if you 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.”
Even if AI could operate fully autonomously on software development, humans should remain accountable for the safety and correctness of systems—not because we lack capability but because accountability and human responsibility are principles that should not change as technology advances.
“I'm not sure there's really any system where you'd want to turn a switch and and go get your coffee... we are accountable for the safety of the systems we produce not just open like every every engineer um and and and that's a principle that should not change.”
Modern large language models are trained on massive amounts of source code, so they are already very good at code generation, but this represents a 'local maximum' because we are still using programming languages designed for human authorship (like Python) even though the marginal cost of code generation is approaching zero, requiring a rethinking of programming language design itself.
“every single engineer at at Sierra uses, uh, cursor, which is a great product that basically integrates with the IDE, Visual Studio Code, to help you generate code more quickly. Um it feels like a local maximum uh in a really obvious way to me which is you have a bunch of code written by people um written in programming languages that were designed to make it easy for people to tell a computer what to do... we're still generating code in programming languages that were designed for human authors... we're still generating code with similar flaws to that we've been generating before from security holes to just functional bugs and in greater volumes.”
First principles thinking is essential for navigating rapidly changing domains like AI because if you respond only to current facts without understanding root causes and future trajectories, your strategic decisions will almost certainly be wrong.
“if you're responding to the facts in front of you and not thinking from first principles about why we're at this point and where it will probably be 12 months from now, the likelihood that you will make the right strategic decision is almost zero.”
AI agents at Sierra use techniques like self-reflection (detecting hallucinations), chain-of-thought reasoning, and decision-making verification to ensure robustness and precision in customer-facing interactions, because hallucination by a branded agent has much higher stakes than hallucination by ChatGPT.
“At Sierra we help companies build customerf facing AI agents. So um if you're setting up a Sonos speaker you'll now chat with an AI. If you're a SiriusXM subscriber you can chat with Harmony who's their AI to manage your account. Um, 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 uh compute and cognitive load into important tricks.”
Skills that will continue to be important in the AI era are 'learning how to learn' and foundational knowledge (reading, writing, math, science, physics, chemistry, biology) not because students need to memorize specific facts but because understanding mechanisms creates the ability to reason from first principles in novel situations.
“I do think learning how to learn and learning how to think will continue to be important... the basics of learning how to think uh learning um uh writing reading math physics uh chemistry biology not because you need to memorize it but understand the mechanisms that uh uh create the world that we live in is is quite important.”
Formal verification and test-driven development, which are currently underutilized because writing tests and proofs is expensive and tedious for humans, become economically rational to scale when code generation is free, enabling dramatically more robust and complex software systems.
“Most programmers I know who are really good not despise it but it's just like a it sounds better than it than it is in practice. But again writing code is free you know so writing tests is free you know 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. could you make something that made it possible to write increasingly complex systems that were increasingly robust?”
Reasoning models like OpenAI's O1 change the economics of AI by moving compute from training time to inference time, allowing model distillation to smaller models, which reduces the capital requirements for building and deploying frontier AI systems.
“the probably the more interesting kind of breakthroughs there are these reasoning models where uh 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. Um which has really been a breakthrough in a variety of ways I think is really interesting but it shows you how you can run up against walls and and find new opportunities to use it.”
AI will help solve previously unsolved problems especially in domains where work is represented digitally and doesn't require real-world experimentation—mathematics, theoretical physics, logical analysis of written work—more than in domains like pharmaceuticals where intelligence isn't the bottleneck (clinical trials are).
“certainly in the areas of um science that are you know largely represented through text and digital technology like math being probably the most 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.”
An AI agent is software that affords the opportunity to reason and make decisions autonomously, which is the definition Taylor uses to describe systems that move beyond rule-based logic to independent judgment and action.
“The word agent comes from agency and I think it means affording a software the opportunity to reason and make decisions autonomously.”
For a 20-year investment horizon, the most promising sectors are those where intelligence is currently the limiting factor to growth—such as software and finance—because these sectors can absorb and monetize arbitrary levels of AI capability.
“what are the parts of the economy that will most benefit from this current wave of AI... There will be some parts of the economy that can essentially um 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.”
Building a foundation model (4th or 5th place) is economically questionable because it costs pharmaceutical-company-scale capex but lacks pharmaceutical-company-scale returns, resulting in a misaligned cost structure and leading to consolidation of model-building into a small number of well-capitalized companies.
“Foundation models I think are somewhat distinct from frontier models. And frontier models I think it's a term credit to to Reed Hoffman but uh I may be mistaken on that but 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 03 as an example from open AI. And 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 uh what your business model is around it, it's very different business than I'm going to go pursue AGI. uh because if you're pursuing AGI really there's only one answer which is to build and train and move to the next front you know horizon because if you can truly build something that is AGI the economic value is so great uh I think there's a really clear business case there if you're pre-training a foundation model that's the fourth best uh that's going to cost you a lot of money and the return on that investment is is probably fairly questionable because why use your fourth best large language model versus a frontier model or an open- source one uh for meta and as a consequence of that I think we have probably have too many people building models right now there's already been some consolidation actually of companies being folded into Amazon and and Microsoft and others”
Conversational AI shifts agency from the company (enumerating features on a website) to the customer (expressing needs in natural language), fundamentally changing the user experience from constrained to open, which benefits consumers but requires companies to be prepared to handle arbitrary customer requests.
“If you think about a modern website or mobile app, um it's essentially you've created a um directory of functionality from which you can choose. But the the main uh person with agency in that is the the creator of the website. Like what are the universe of options that you can do? When you have an AI agent representing your brand, the agency goes to the the cons 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 to put on your website to whatever your customers ask.”
Microsoft's failure to compete in smartphones (losing the smartphone wars to Apple iOS and Google Android) was enabled by an internal reality distortion where employees on the Microsoft campus were using Windows Phones not out of preference but because of social/career pressure, creating a feedback loop where company insiders' perception of the product diverged from market reality.
“I remember one time uh when the sort of peak of the smartphone wars and I end up visiting a friend on Microsoft's campus and uh I got off the plane in, you know, Seattle Tacoma airport, drove into Redmond, went onto the campus and all of a sudden everyone I saw was using Windows phones. Um, 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, you know, um, like it was definitely a two-horse race between uh, Apple and and Google on on iOS and Android.”
The primary inputs limiting AGI progress are data (textual training data is 'waning' with a potential 'data wall'), compute (capital-intensive, with hardware costs that may not be recovered), and algorithms (with potential plateaus in scaling laws even after breakthroughs like reasoning models).
“I think the three primary inputs are data, compute, and algorithms... in terms of just like textual content to train on you know the availability of new content is certainly waning and and some people would say I think there's like a data wall uh I'm not an expert in that domain but it's been talked about a lot”
The price-to-performance ratio of large language models has improved dramatically—GPT-4 mini is probably cheaper to use than self-hosting open-source models—which means the economics of using proprietary vs. open-source models depend on complex factors beyond just upfront cost.
“I mean if you look at the price of GPT4 mini you know it is uh so much higher quality than like the highest quality model 2 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.”
The role of a software engineer will shift from being an author of code to an operator of code-generating machines, which means companies need to rethink hiring, skill development, and architectural approaches to account for this transition over the next few years.
“I think it's really interesting to me that with modern large language models, one of the careers that is being most transformed is software engineering. uh and you know one of the things I think a lot about is how many software engineers will we have at our company 3 years from now what will the role of a software engineer be as we go from being authors of code to operators of code generating machines um what does that mean for the type of people we should recruit”
The question of whether AI-generated output should be copyrightable or patentable is complex, and Taylor's intuition is that the answer is likely yes (because the person using the AI tool made the creative choice to ask), but the marginal value of AI-generated ideas will be lower than human-generated ideas because the marginal cost of intelligence will approach zero.
“I think the answer is probably yes that you you used the tool to do it. But I also think that the um uh uh in 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 years ago.”
Corporate complacency stems from two mechanisms: (1) accumulated bureaucracy where processes created for good reasons over decades lose their original justification and create inertia; and (2) internal narrative that becomes stronger than customer truth, where employees' perception of the company's position diverges from reality.
“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 uh when something goes wrong, companies introduce a process to fix it. And over the like uh sequence of 30 years, the layered sum of all of those processes... end up being a um bureaucratic sort of machine where um the reasons for many of the rules and processes are are rarely even remembered by the organization.”
There is a tension between enabling rapid AI progress (competitive advantage for Western democracies against totalitarian regimes) and ensuring companies are sufficiently focused on beneficial outcomes and safety, and overregulation risks slowing progress and ceding leadership to less-constrained regimes, but this is not a reason to dismiss regulation entirely—it requires nuanced, multistakeholder dialogue.
“there's two uh not conflicting forces but two forces that exist somewhat independently but relate to each other. one is the pace of progress in AI and ensuring that you know uh the the folks working on frontier models are ensuring those models do benefit humanity and uh and then there's the uh sort of geopolitical landscape which is you know do you want uh AGI to be created by the freedom uh sort of uh the west um by democracies um or do you want it to be created by more totalitarian governments and So I think the inherent tension for regulators will be um a sense of obligation to ensure that you know uh the technology organizations creating AGI are in fact focusing enough on um better and fitting humanity all the other uh stakeholders that that whose uh interests that they're accountable for and ensuring that the west remains competitive um and and uh I think that's a really nuanced thing and I uh you know my my view is it's very important that the west is uh leads in AI... and I also you know having sort of seen the inside I think we're really focused on benefiting humanities I I tend to think that you know it needs to be a multistakeholder dialogue but I think there's a really big risk that some regulations could have the unintended consequence of of uh slowing down this larger conversation. But I don't say that to be dismissive of it either. It's actually just a impossibly hard uh problem”
Prior art generation—publishing generated ideas on a website without anyone needing to read them, just to establish prior art and prevent patent trolling—could be a mechanism to limit patenting of abundant AI-generated ideas.
“here's the idea I had last night to counter this because I was like, I don't want somebody doing this. Uh, and I was thinking like, what if prior art eliminates patents? Yeah. So 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 like basically patenting like anything, but it's creating prior art for everything.”
Being acquired by larger companies (Facebook, Salesforce) as a founder is fundamentally difficult because it requires a deep shift in identity—from seeing yourself as the founder/CEO of your company to seeing yourself as part of a larger organization, which is a prerequisite for success in the acquisition.
“it's not just building a business, but it's very much your identity... when you go to being acquired, there's a business aspect to it... that's intertwined with a sense of identity... to fully embrace that you actually need to change your identity... that identity shift is a prerequisite for most of the other things.”
Building a company that endures is one of the hardest business challenges because it requires not just sustained financial success but cultural resilience that allows the company to evolve its product and strategy as technology and society change, particularly in eras like AI where change is unprecedented in pace.
“when I think about the what it means to build a company that's enduring um that is a really really tall um 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 uh society and technology at a when it's changing at a pace that is like unprecedented in history.”
Sierra's pricing model is based on first principles: rather than licensing software (charging for privilege of access), they charge only when the AI agent solves a customer problem—a model that reflects the insight that in the age of AI, software completes tasks rather than just enabling productivity.
“Our pricing model is really unique and comes from first principles thinking rather than having our customers pay a license for the privilege of using our platform. We only charge our customers for the outcomes. Uh meaning if the AI agent they've built for their customers solves the problem. There's like a 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. uh 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 privilege of using the software.”
Sierra's platform design allows companies to define customer experience abstractly (independent of specific models or technologies) so that as new models and capabilities emerge, the customer experience improves without requiring reimplementation.
“your customer experience should be wed should not be wetted 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 and 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”
Engineer-founders make excellent leaders but must transition their identity from being deep specialists in engineering/product to becoming broad generalists across all business functions (go-to-market, sales, policy, recruiting) as their company scales, and those who fail to make this transition often plateau their company's growth.
“I do think engineers make great leaders, but one thing I've seen is that I think that I really believe that great CEOs and great founders um start usually with one specialty but become uh more broadly specialists in all parts of their business.”
To build an AI superpower country, policy-makers should prioritize compute infrastructure (data centers, power, land) as the load-bearing variable, since where compute exists, research labs and talent will follow.
“one of the key ingredients to modern AI is compute. Um, which is a a noun that wasn't a noun until recently, but now compute is a noun. And uh you know I do think that's one area where policy makers can um because it involves a lot of things that uh touch uh federal and local governments like power, land um and then similarly attracting the capital which is immense to finance uh to the real estate to purchase the uh you know uh compute itself”
The value in future AI-powered software may be less in single large language models with prompt-response interactions and more in end-to-end closed systems where LLMs are embedded as infrastructure components, orchestrating complex processes—a shift from model-centric to application-centric thinking.
“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 toend closed loose systems that use large language models as pieces of infrastructure.”
Founders tend to produce better outcomes for companies than professional managers because they have greater permission and latitude to make bold, disruptive decisions, especially difficult ones like layoffs, because stakeholders attribute these decisions to the founder's deep commitment to the company they created rather than to political self-interest.
“I think founders drive better outcomes for uh companies. Um there's a I think founders tend to have permission to make bolder more disruptive decisions about their business than a professional manager... you know you have founded a company it's all your stakeholders employees in particular uh give you the benefit of the doubt you know you created this thing and if you say hey we need who um do a major shift in our strategy even hard things like uh layoffs founders tend to get a lot of latitude and are judged I think differently”
The slowdown in scientific breakthroughs over past decades may be partly due to excessive specialization required of grad students and researchers, creating a system optimized for depth over breadth, which AI could help reverse by democratizing expertise access.
“I've read a lot like a lot as a amateur about is just the slowdown in scientific breakthroughs over the past you know few decades and and some theories that it's because of the degree of specialization that we demand of grad students and things like that”
AI will disproportionately benefit deep generalists who can orchestrate intelligence across multiple domains, rather than continuing the trend toward deep specialization, because generalists can use AI to fill knowledge gaps while their broad understanding allows them to see novel connections across domains.
“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 um knowing where to prompt the AI to to go explore and and um 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 uh intelligence between lots lots of different domains”
Acquisitions fail because the acquiring and acquired companies often have fundamentally different definitions of success two years after closing, and this misalignment is often obscured by storytelling about synergies during the deal process rather than concrete discussion of what specific outcomes are expected.
“if you went to most of the like larger acquisitions uh, in the valley and you two weeks after it was closed, interviewed the management team of the acquiring company and the acquired company, and you asked them like, 'What does success look like two years from now?' My guess is like 80% of the time you get different answers.”
OpenAI (via ChatGPT) will likely be the primary delivery mechanism for AGI when it is achieved, not just because of the researchers building it, but because ChatGPT has become the canonical interface by which most people access AI, making it the 'packaging' that will be synonymous with AGI.
“I do think that it will be the delivery mechanism for AGI when it's produced and uh not just because of the many researchers at OpenAI but because of the amazing like 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 um you know and you were talking about artificial general intelligence. I'm not sure either of us would have envisioned something so simple as a form factor uh to absorb it that you just talk to it.”
Context window size (the amount of information an AI model can consider at once) matters significantly for raw intelligence and generalization, but is less important for industrial applications (like customer service agents) where the system already has contextual awareness of the process it's executing.
“I think context window could be quite important um especially it certainly simplifies working with an AI if you can just give it everything and instruct it to do something... but I also think that you know 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 toend closed loose 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.”
The engineering mindset (first principles thinking, root cause analysis, systematic approach) is broadly applicable and valuable across business problems, but can lead to analysis paralysis and overintellectualization when applied to fundamentally human domains like sales relationships and social media strategy.
“one thing I think I've seen though with engineers who go into other disciplines is um sometimes you can overanalyze decisions in some domains... having a systematic first principles discussion about every you know tweet you do is probably not a great comm strategy. Um and so uh and then similarly um you know there are some aspects of say enterprise software sales that you know aren't rational but they're human you know like forming personal relationships... taking that to its like logical extreme can lead to analysis paralysis”
Frontier models (the one or two clearly leading-edge models like OpenAI's O3) differ fundamentally from foundation models (broader LLMs like GPT-4 mini that can be licensed and fine-tuned) in their business model: frontier model builders pursuing AGI must continuously invest in the next generation, while foundation model builders should think more like cloud infrastructure companies licensing to diverse customers.
“Foundation models I think are somewhat distinct from frontier models. And frontier models I think it's a term credit to to Reed Hoffman but uh I may be mistaken on that but 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 03 as an example from open AI... if you're deciding whether you're building a foundation model and uh what your business model is around it, it's very different business than I'm going to go pursue AGI. uh because if you're pursuing AGI really there's only one answer which is to build and train and move to the next front”
The concept of 'founder mode' is valuable in spirit (deep founder accountability for every decision) but is frequently weaponized as an excuse for micromanagement, which likely does not produce great outcomes and is distinct from the kind of decisive judgment and taste that created actual value (e.g., Steve Jobs' example).
“I really like the spirit of founder mode which is just having deep founder accountability for every decision at your company... The issue I have not with Brian's statements. Brian's amazing. Um is how people can sort of interpret that and sort of execute it as a caricature... I think a lot of founders were like mimicking that you know down to wearing a black turtleneck and yelling at you their employees I'm like not sure that was the cause... I think Steve Jobs taste and judgment through you know executed through that you know packaging was the cause of their success”
The ultimate measure of a company is whether it lasts beyond its founder, achieving longevity through institutional design and culture rather than depending on the founder's individual capability.
“it's incredibly hard to do something that lasts beyond you, but that I think is the ultimate measure of a company.”
The best time to have difficult strategic conversations about an acquisition is after both parties have committed to the high-level terms (term sheet stage) but before the deal closes, when there is sufficient commitment to enable candor but the power imbalance isn't yet fully established.
“I think it's right after that. Um, so where people have really committed to the the key things, how much value, why are we doing this, the big stuff... That's a there's an awkward waiting period there. And that's a time I think where like the strategic decision makers in those moments can get together and say let's talk through what this really means... you've kind of made the commitment to each other. So it's you've I think you have more social permission to have real conversations at that point. Um, but you also haven't consummated the relationship.”
Reasoning models represent a breakthrough in generating net new ideas and synthetic data quality because they're not just regurgitating training data, which could help overcome the 'data wall' facing traditional scaling approaches.
“What's really interesting about you know reasoning and reasoning models is I think uh I feel really optimistic these models are generating net new ideas and so it really affords the opportunity to break through some of these the data wall as well.”
Building interactive web applications was a local maximum in the early 2000s because Google Maps started with a Windows application (Expedition) that set a high bar for interactivity that exceeded typical web applications of the time, forcing the web version to be much more interactive than it would have been otherwise.
“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. And 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 Yen's just like technical choices, we made much bolder technical bets than we would have otherwise.”
A useful prompting technique is to first use a faster, cheaper model (GPT-4o) to refine and complete your prompt, then pass the refined prompt to a slower, more capable reasoning model (O1 Pro), treating the cheaper model as a 'product manager' that helps clarify specifications before the expensive engineer begins work.
“I often with the uh reasoning models which are slower will use a faster model first GBT40 to refine my prompts. Um so uh over the holidays um partly because I was thinking about the future of software engineering... I would spend a lot of time iterating on 40 to sort of like refine and and make more complete and specific what I was looking for and then I would put it into one pro mode, go get my coffee and you know come back and get it.”
The launch of DALL-E in summer 2022 and the avocado chair image it generated was a pivotal moment that made Taylor realize he needed to pay much closer attention to generative AI and the pace of progress in large language models, which he had previously been ignoring despite his computer science background.
“one that I don't think I really appreciated how huge it would be but it kind of reset my expectation which was the launch of Dolly uh in the summer of 22 is that right I might be off by year but I think summer of 22 and uh the avocado chair that they generated... I had no idea computers could do that”
Meta's strategy of building and open-sourcing large language models (Llama) is consistent with their historical approach of using open source to court developers and attract ecosystem lock-in, and while this may involve 'commoditizing complements,' it also genuinely furthers the development and democratization of AI tools.
“I actually will take Mark at his word and you know that post he wrote about open source I think was very well written and encourage people to read it I think that's his strategy and you know if you look at Facebook um the now you've got me saying Facebook too so that was what it was called I was there you know the company's always really embraced open source and if I look at really popular things from react to uh you know now the llama models it's always been a big part of their strategy to court developers around sort of their ecosystem”
The Google Maps rewrite—where Taylor rewrote the entire Google Maps frontend over a weekend, reducing bundle size from 200k to 20k and dramatically improving speed—succeeded because he had lived with the original codebase and understood all its design decisions, allowing him to redesign from first principles with clarity about the intended outcome.
“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 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'd do it.”
Good engineering cultures are outcomes-oriented and pragmatic about code replacement; they don't fetishize code authorship if there's a demonstrably better path forward, and they recognize that throwing out code for a better architecture is often the right decision (as with self-driving car transitions from handcoded heuristics to monolithic models).
“at the end of the day everyone was like wow that's great you know we went from a bundle size of 200k to a bundle size of 20k and it was a lot faster and better so you know broadly speaking I think good engineering cultures you don't want a culture of um you know ready fire aim but I also think you just need to be really outcomesoriented and I think people if they become they start to treat their code as too precious it can really uh impede forward progress. Um and yeah I'll just take like I my understanding is like a lot of the early self-driving car software was a lot of handcoded heruristics and rules and you know a lot of smart people think that eventually it'll probably be a more monolithic model that uh encodes many of the same rules. you have to throw out a lot of code in that transition, but it doesn't mean it's not the right thing to do.”
Personal AI agents (that assist individuals with email, scheduling, job search, preparation) are a meaningful category of AI agents but will be harder to commercialize because they must generalize across many user contexts and integrate with diverse services.
“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 uh essentially amplifying themselves as an individual. Um, you can do the wrote things like help you triage your email to helping you schedule a vacation. You know, you're flying back to um, Edmonton and help you arrange your travel. Um, to more complex things like, you know, I'm going to go ask my boss for promotion. Like, help me role play and um, you know, I'm setting up my resume for this job. Help me do that to I'm applying for a new job. Help me find companies I haven't thought of that I should be applying to. Uh, 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. So, 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.”
Self-improvement in AI (where AI systems improve themselves faster than humans can) is already happening to some degree because engineers use AI tools and agents to write code, but the difference between assisted coding and autonomous coding is significant—more like assisted driving vs. self-driving.
“you know 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 um platforms like cursor to help them code so it's contributing already um and I imagine as uh coding assistants go to coding agents in the future most engineers in Silicon Valley will show up in the morning and but This is sort of the difference between uh you know the assisted driving in Tesla versus like self-driving.”
People who define their professional identity around tools or specific techniques are at risk of disruption when tools change, whereas those who define themselves by the value they create (judgment, agency, design) can adapt across tool transitions.
“there's a risk of people sort of uh becoming oified in the tools that they use. Um so you know uh let's go back to our discussion with 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 um that is no longer a differentiated, you know, part of the software engineering experience or or will not be. But your judgment as a software engineer will continue to be uh incredibly important your agency and making a decision about what to build, how to build it, um how to architect it, uh maybe using AI models as a creative foil.”
AGI will not be a winner-take-all outcome because different use cases have different needs (price/performance/quality tradeoffs), and the existence of multiple options (open-source, smaller models, frontier models) will serve different segments, similar to how different cloud services and databases serve different needs.
“you don't think AGI is going to be like a winner take all? ... like in cloud computing you'll end up with a price performance quality trade-off and for any given engineering challenge they'll have a different answer and that's appropriate and some people use uh open-source Kafka some people work with Confluent um great you know like that's just the way these things work”
Overcoming organizational resistance to code rewrites requires living with the existing system long enough to understand all design decisions, then shipping a complete replacement prototype quickly so teams can experience the improvement directly rather than being stuck in deliberative debate.
“I and I also knew that, you know, 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. And so I really wanted the prototype to go out and so I did it”
Success for Taylor personally is defined as having a happy, healthy family, being able to work with his co-founder Clay indefinitely, and making Sierra into an enduring company that lasts beyond himself.
“success 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 Sarah into an enduring company. That would be success for me.”
The inefficiency of the original Google Maps code came from using XML and XSLT—which were in fashion in the early 2000s—despite being poorly suited for web browsers, requiring an engineer to even build a full XSLT transform engine in JavaScript just to support the Safari browser.
“when we were taking some of these ideas and putting them in a web browser, we kind of like went into autopilot and used like a ton of XML and it made everything just like really really tedious... we got to this point where we decided we wanted to support the Safari web browser which was relatively new at the time... there was much less XML support in Safari than there was in Internet Explorer and Firefox. And so one of the engineers implemented like a full XSLT transform engine in JavaScript to get it to work. And it was just like [ __ ] on top of [ __ ] on top of [ __ ]”
Work-life balance is a less relevant framing for people who love what they do; time allocation (work vs. family) is more honest than 'balance,' and the interviewer (presumably as a founder) allocates 100% of his discretionary time between work and family without guilt.
“Broadly, probably, you know, being honest. First, I'm not perfect at second, I don't have a ton of hobbies. You know, I basically work and spend time with my family. Um the first time we talked, you saw the couple guitars in my background. Uh I haven't picked one of those up in a while. Um the I 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'm 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 mine is at work and with family. And so that's how I do it. I don't know if I'm particularly balanced, but I don't strive to be either.”
OpenAI's investment in the United States is important from a geopolitical standpoint, and the company appears genuinely focused on benefiting humanity (based on Taylor's inside knowledge), which supports his view that Western dominance in AI is preferable to totalitarian regime dominance.
“I'm very proud of the fact that um you know open AI is based here in the United States and and we're investing a lot in the United States and I think that's very important and I also you know having sort of seen the inside I think we're really focused on benefiting humanities I I tend to think that you know it needs to be a multistakeholder dialogue”