The unconventional Palantir principles that catalyzed a generation of startups
Experimenting with live products
For anyone who learned product formally, you learned to experiment cheaply, meaning not with expensive engineering time and code. We’re supposed to use designs, value tests, landing pages, and all sorts of other lightweight methods to measure the value of a product before we build it. Well, at Palantir, we often used working products as the iterative units of value. When we learned about new customer needs, we didn’t spend a lot of time standing back and designing product experiments or running extensive discovery sprints. We designed and built instead. Then we shared with customers. Then we’d subtract and add and build again, and then we’d share again, and so forth. We did full-scale product iteration with real working code. To supercharge velocity, we sent our designers forward too, and generally worked in daily cycles.
Building features that magnify value over time.
We want to solve customer problems, but we also want to solve them in ways whereby the next user or customer is better off because of what the first user did. This can manifest as (1) a growing data asset, for example more valuable records or an increasingly well-trained machine learning model, or (2) network effects from user and customer interactions with one another. At Palantir, we focused on both. One key value proposition in our Gotham platform was its ability to find unknown connections across large data sets, for example between a terrorist financier and a bomb builder. Finding and disrupting a connection like this can save lives. The system already integrated key data sources, but before AI could “read” well, computers were horrible at identifying these sorts of connections in textual reports written about suspects. One of our stickiest features addressed this challenge: we called it “tagging.” If you read a document that said that John Smith was a terrorist suspect using the telephone number “123-456-7890,” you could select that text and effectively write it back to the system as a new structured entity. This information, part of the growing data asset, would then link to the other existing structured records. The next user would find evidence of John Smith, see the source referencing him as a suspect, and then easily surface connections through that telephone number, without having to first read the report. If the system previously had a billion records that made it compelling to users, it now had a billion and one, plus possibly dozens of new, unexplored connections.
Solving the hardest problems first
People are obsessed with low-hanging fruit. Solving for these was not the Palantir way. The Palantir way was to find the biggest, hairiest, and most difficult problems facing a major organization or industry and work hard on those. We literally worked on problems like finding the next terrorist and where to drill for oil. Perhaps the part of being a Palantir alumnus that makes me the most proud is that almost every company emerging from our network is focused on extremely difficult, world-changing problems. A good litmus test is as follows: if the problem you are solving doesn’t show up in the news or annual report, then it’s probably not a big enough challenge. “Impossible” challenges are opportunities in disguise. Each “unsolvable” problem you can solve becomes intellectual property that separates you from your competition and helps you establish a market of your own. When Congress publicly demanded after 9/11 that the U.S. Intelligence Community find ways to share sensitive data across agencies while protecting civil liberties, Palantir built Nexus Peering. The capability was a first-of-its-kind product that prioritized multi-directional sharing from one organization to another without divulging sensitive sources or neglecting the audit trail.
Making engineers do the sales
Why exactly is a salesperson necessary to sell a product? Is that the best way to handle the situation? Sometimes it is, but for a complex mission-driven sale, we realized that we could disrupt the old model in a positive way. Instead of forcing our government customers into opportunistic pitches from professional relationship brokers, we could put former operators and engineers in the room instead. As a result, we could speak intelligently about the mission and our product from firsthand experience. We didn’t try to hide the skeletons or answer questions “the right way.” Rather, we spoke to our clients like normal people would. Without formal sales training, we were forced to find ways to create value in sales settings and not waste time. Put in the extreme, if what you are building is incredibly compelling, and if someone truly wants to buy it, then the customer doesn’t require a person with a sales background to ask closing questions to scope the engagement and deliver the product.
How companies are bringing AI applications to life (…and another developer notes from Sequoia)
- Nearly every company in the Sequoia network is building language models into their products; visual art (Midjourney), marketing (Hubspot, Attentive, Drift, Jasper, Copy, Writer), sales (Gong), contact centers (Cresta), legal (Ironclad, Harvey), accounting (Pilot), productivity (Notion), data engineering (dbt), search (Glean, Neeva), grocery shopping (Instacart), consumer payments (Klarna), and travel planning (Airbnb). These are just a few examples and they’re only the beginning.
- Companies want to customize language models to their unique context; generalized language models are powerful, but not differentiating or sufficient for many use cases. Companies want to enable natural language interactions on their data—their developer docs, product inventory, HR or IT rules, etc. In some cases, companies want to customize their models to their users’ data as well: your personal notes, design layouts, data metrics or code base.
- The stack is becoming increasingly developer-friendly. Now that the population working with language models has meaningfully expanded to all developers. For example, LangChain helps developers build LLM applications by abstracting away commonly occurring problems: combining models into higher-level systems, chaining together multiple calls to models, connecting models to tools and data sources, building agents that can operate those tools, and helping avoid vendor lock-in by making it easier to switch language models.
- Language models need to become more trustworthy (output quality, data privacy, security) for full adoption; Before fully unleashing LLMs in their applications, many companies want better tools for handling data privacy, segregation, security, copyright, and monitoring model outputs.
- Practitioners think AI is moving too quickly to have high confidence in the end-state stack, but there was consensus that LLM APIs will remain a key pillar, followed in popularity by retrieval mechanisms and development frameworks like LangChain.
Apple Vision Pro: The most advanced personal electronics device ever
- Apple Vision is utilizing some number of its 12 cameras to capture the outside world and displaying them to the postage-stamp-sized screens in front of your eyes in a way that makes you feel like you are wearing safety goggles: you’re looking through something, that isn’t exactly like total clarity but is of sufficiently high resolution and speed that there is no reason to think it’s not real.
- Apple claims that the threshold for your brain to notice any sort of delay in what you see and what your body expects you to see (which is what causes known VR issues like motion sickness) is 12 milliseconds, and that the Vision visual pipeline displays what it sees to your eyes in 12 milliseconds or less. This is particularly remarkable given that the time for the image sensor to capture and process what it is seeing is along the lines of 7~8 milliseconds, which is to say that the Vision is taking that captured image, processing it, and displaying it in front of your eyes in around 4 milliseconds.
- This is, truly, something that only Apple could do, because this speed is function of two things: first, the Apple-designed R1 processor (Apple also designed part of the image sensor), and second, the integration with Apple’s software. None of this advanced technology could come to life without a powerful operating system called “visionOS”.
- Real-time operating systems are used in embedded systems for applications with critical functionality, like a car, for example: it’s ok to have an infotainment system that sometimes hangs or even crashes, in exchange for more flexibility and capability, but the software that actually operates the vehicle has to be reliable and unfailingly fast. This is, in broad strokes, one way to think about how visionOS works: while the user experience is a time-sharing operating system that is indeed a variation of iOS, and runs on the M2 chip, there is a subsystem that primarily operates the R1 chip that is real-time; this means that even if visionOS hangs or crashes, the outside world is still rendered under that magic 12 milliseconds.
- This is the most meaningful manifestation yet of Apple’s ability to integrate hardware and software: while previously that integration manifested itself in a better user experience in the case of a smartphone, or a seemingly impossible combination of power and efficiency in the case of Apple Silicon laptops, in this case that integration makes possible the melding of VR and AR into a single Vision.
A trivial thing can grow into a massive thing, and faster than most people realise
- It’s hard to imagine, say, our incomes doubling over the next few generations. That seems like such a massive leap, like we’d have to boil the ocean to get it done. But doubling the average income over 30 years works out to about 2.3% growth per year. It’s not crazy at all. It’s actually quite achievable. What made it seem so ambitious to begin with is that compound growth is easy to underestimate.
- If you look at the end result of a long period of compounding, it’s astounding. But all it took to get it done was little bits of incremental growth strung together for a long time. All progress is like that.
- Technological progress is easy to underestimate because it’s so counterintuitive to see how, for example, the philosophies of a guy who invented the Polaroid film would go on to inspire the iPhone. When you view it as one person coming up with a small idea, another person copying that idea and tweaking it a little, another taking that insight and manipulating it a bit, another yet taking that product and combining it with something else – incremental, tiny bits, little ideas mixing, joining, blending, mutating, and compounding together – it’s suddenly much more conceivable.
- A trivial thing can grow into a massive thing, and faster than most people realize. Small ideas mixing and compound into big ones – that’s what really drives the world. A trivial thing can grow into a massive thing, and faster than most people realize.
Notes apps are where ideas go to die (if you don’t process them regularly)
Substack edition
The unconventional Palantir principles that catalyzed a generation of startups
Experimenting with live products
For anyone who learned product formally, you learned to experiment cheaply, meaning not with expensive engineering time and code. We’re supposed to use designs, value tests, landing pages, and all sorts of other lightweight methods to measure the value of a product before we build it. Well, at Palantir, we often used working products as the iterative units of value. When we learned about new customer needs, we didn’t spend a lot of time standing back and designing product experiments or running extensive discovery sprints. We designed and built instead. Then we shared with customers. Then we’d subtract and add and build again, and then we’d share again, and so forth. We did full-scale product iteration with real working code. To supercharge velocity, we sent our designers forward too, and generally worked in daily cycles.
Building features that magnify value over time.
We want to solve customer problems, but we also want to solve them in ways whereby the next user or customer is better off because of what the first user did. This can manifest as (1) a growing data asset, for example more valuable records or an increasingly well-trained machine learning model, or (2) network effects from user and customer interactions with one another. At Palantir, we focused on both. One key value proposition in our Gotham platform was its ability to find unknown connections across large data sets, for example between a terrorist financier and a bomb builder. Finding and disrupting a connection like this can save lives. The system already integrated key data sources, but before AI could “read” well, computers were horrible at identifying these sorts of connections in textual reports written about suspects. One of our stickiest features addressed this challenge: we called it “tagging.” If you read a document that said that John Smith was a terrorist suspect using the telephone number “123-456-7890,” you could select that text and effectively write it back to the system as a new structured entity. This information, part of the growing data asset, would then link to the other existing structured records. The next user would find evidence of John Smith, see the source referencing him as a suspect, and then easily surface connections through that telephone number, without having to first read the report. If the system previously had a billion records that made it compelling to users, it now had a billion and one, plus possibly dozens of new, unexplored connections.
Solving the hardest problems first
People are obsessed with low-hanging fruit. Solving for these was not the Palantir way. The Palantir way was to find the biggest, hairiest, and most difficult problems facing a major organization or industry and work hard on those. We literally worked on problems like finding the next terrorist and where to drill for oil. Perhaps the part of being a Palantir alumnus that makes me the most proud is that almost every company emerging from our network is focused on extremely difficult, world-changing problems. A good litmus test is as follows: if the problem you are solving doesn’t show up in the news or annual report, then it’s probably not a big enough challenge. “Impossible” challenges are opportunities in disguise. Each “unsolvable” problem you can solve becomes intellectual property that separates you from your competition and helps you establish a market of your own. When Congress publicly demanded after 9/11 that the U.S. Intelligence Community find ways to share sensitive data across agencies while protecting civil liberties, Palantir built Nexus Peering. The capability was a first-of-its-kind product that prioritized multi-directional sharing from one organization to another without divulging sensitive sources or neglecting the audit trail.
Making engineers do the sales
Why exactly is a salesperson necessary to sell a product? Is that the best way to handle the situation? Sometimes it is, but for a complex mission-driven sale, we realized that we could disrupt the old model in a positive way. Instead of forcing our government customers into opportunistic pitches from professional relationship brokers, we could put former operators and engineers in the room instead. As a result, we could speak intelligently about the mission and our product from firsthand experience. We didn’t try to hide the skeletons or answer questions “the right way.” Rather, we spoke to our clients like normal people would. Without formal sales training, we were forced to find ways to create value in sales settings and not waste time. Put in the extreme, if what you are building is incredibly compelling, and if someone truly wants to buy it, then the customer doesn’t require a person with a sales background to ask closing questions to scope the engagement and deliver the product.
How companies are bringing AI applications to life
(…and another developer notes from Sequoia)
Nearly every company in the Sequoia network is building language models into their products; visual art (Midjourney), marketing (Hubspot, Attentive, Drift, Jasper, Copy, Writer), sales (Gong), contact centers (Cresta), legal (Ironclad, Harvey), accounting (Pilot), productivity (Notion), data engineering (dbt), search (Glean, Neeva), grocery shopping (Instacart), consumer payments (Klarna), and travel planning (Airbnb). These are just a few examples and they’re only the beginning.
Companies want to customize language models to their unique context; generalized language models are powerful, but not differentiating or sufficient for many use cases. Companies want to enable natural language interactions on their data—their developer docs, product inventory, HR or IT rules, etc. In some cases, companies want to customize their models to their users’ data as well: your personal notes, design layouts, data metrics or code base.
The stack is becoming increasingly developer-friendly. Now that the population working with language models has meaningfully expanded to all developers. For example, LangChain helps developers build LLM applications by abstracting away commonly occurring problems: combining models into higher-level systems, chaining together multiple calls to models, connecting models to tools and data sources, building agents that can operate those tools, and helping avoid vendor lock-in by making it easier to switch language models.
Language models need to become more trustworthy (output quality, data privacy, security) for full adoption; Before fully unleashing LLMs in their applications, many companies want better tools for handling data privacy, segregation, security, copyright, and monitoring model outputs.
Practitioners think AI is moving too quickly to have high confidence in the end-state stack, but there was consensus that LLM APIs will remain a key pillar, followed in popularity by retrieval mechanisms and development frameworks like LangChain.
Apple Vision Pro: The most advanced personal electronics device ever
Apple Vision is utilizing some number of its 12 cameras to capture the outside world and displaying them to the postage-stamp-sized screens in front of your eyes in a way that makes you feel like you are wearing safety goggles: you’re looking through something, that isn’t exactly like total clarity but is of sufficiently high resolution and speed that there is no reason to think it’s not real.
Apple claims that the threshold for your brain to notice any sort of delay in what you see and what your body expects you to see (which is what causes known VR issues like motion sickness) is 12 milliseconds, and that the Vision visual pipeline displays what it sees to your eyes in 12 milliseconds or less. This is particularly remarkable given that the time for the image sensor to capture and process what it is seeing is along the lines of 7~8 milliseconds, which is to say that the Vision is taking that captured image, processing it, and displaying it in front of your eyes in around 4 milliseconds.
This is, truly, something that only Apple could do, because this speed is function of two things: first, the Apple-designed R1 processor (Apple also designed part of the image sensor), and second, the integration with Apple’s software. None of this advanced technology could come to life without a powerful operating system called “visionOS”.
Real-time operating systems are used in embedded systems for applications with critical functionality, like a car, for example: it’s ok to have an infotainment system that sometimes hangs or even crashes, in exchange for more flexibility and capability, but the software that actually operates the vehicle has to be reliable and unfailingly fast. This is, in broad strokes, one way to think about how visionOS works: while the user experience is a time-sharing operating system that is indeed a variation of iOS, and runs on the M2 chip, there is a subsystem that primarily operates the R1 chip that is real-time; this means that even if visionOS hangs or crashes, the outside world is still rendered under that magic 12 milliseconds.
This is the most meaningful manifestation yet of Apple’s ability to integrate hardware and software: while previously that integration manifested itself in a better user experience in the case of a smartphone, or a seemingly impossible combination of power and efficiency in the case of Apple Silicon laptops, in this case that integration makes possible the melding of VR and AR into a single Vision.
A trivial thing can grow into a massive thing, and faster than most people realise
It’s hard to imagine, say, our incomes doubling over the next few generations. That seems like such a massive leap, like we’d have to boil the ocean to get it done. But doubling the average income over 30 years works out to about 2.3% growth per year. It’s not crazy at all. It’s actually quite achievable. What made it seem so ambitious to begin with is that compound growth is easy to underestimate.
If you look at the end result of a long period of compounding, it’s astounding. But all it took to get it done was little bits of incremental growth strung together for a long time. All progress is like that.
Technological progress is easy to underestimate because it’s so counterintuitive to see how, for example, the philosophies of a guy who invented the Polaroid film would go on to inspire the iPhone. When you view it as one person coming up with a small idea, another person copying that idea and tweaking it a little, another taking that insight and manipulating it a bit, another yet taking that product and combining it with something else – incremental, tiny bits, little ideas mixing, joining, blending, mutating, and compounding together – it’s suddenly much more conceivable.
A trivial thing can grow into a massive thing, and faster than most people realize. Small ideas mixing and compound into big ones – that’s what really drives the world. A trivial thing can grow into a massive thing, and faster than most people realize.