Don't build moats, build tech trees 🌲
The thrill of leveling up in a video game, applied to real life
What did Elon mean when he tweeted this?
Like many Elon tweets we’re not really sure1, so here’s our interpretation…
Rewinding a bit, our first Substack post2 explained our product pivot and revised thesis to build a tech tree for cooking skills. Since then, we’ve learned a lot and advanced significantly toward our goal.
A little history: tech trees were originally invented by Sid Meier for the game Civilization, and have been adapted across all kinds of video games, for both nations/factions with technology dependencies, and characters who learn skills—also referred to as a skill tree. Here’s the character skill tree for the long-running MMORPG RuneScape:
The underlying technical insight of Parsnip is to take the idea of a tech tree and apply it to a real life skill, in this case cooking. By modeling learning or advancement in cooking as a skill tree, several interesting things happen:
you can view your progress on the skill tree
you can see—and thus pick—what to learn next
different skills can be combined to achieve exponentially many goals (recipes)
millions of pieces of existing content can map to skills in the tech tree
Let’s dig into each of these.
Your favorite data is stats about yourself
In my academic career, I studied how to build combined systems of human and machine intelligence3. Along the way, one thing I learned is that the most appealing data to anyone is data about themselves.
Like a character in a video game, the Parsnip skill tree for cooking provides an experience that shows the skills you have and how they accumulate over time. But unlike a video game, the main character is a real person and the skills you learn are actually useful in real life. Here’s the current version of Parsnip’s cooking skill tree:
People love seeing their stats and earning badges, and so this turns cooking from a chore that seems hard and unapproachable into something satisfying, rewarding, and even fun. Other examples of this include Strava (exercise data), the Oura Ring (sleep data), or Stack Overflow (knowledge data).
As a consumer product, a user’s skill progression is also a nontransferable artifact that enhances retention. Future versions of Parsnip will also include stats about your cooking history—even better retention!
Choosing what you can learn next: personalized education
This skill tree we’re building for Parsnip can also be viewed as a cooking knowledge graph. Our goal is to allow someone to pick any recipe or food content from the Internet and see what they need to learn to cook it well.
Though the devil is in the details, the underlying process here is simple:
Pick any recipe from the internet.
Find the relevant tech tree nodes (skills) using machine learning.
Present those skills to the user in a simple way.
This combination of a tech tree & machine learning gives users the superpower of “learn to cook anything on the Internet”, a value proposition that doesn’t exist yet, anywhere.
At the same time, this also solves another major hurdle with most approaches to learning cooking: you can’t learn unknown unknowns, and you can’t search for what you don’t know. For example, did you know that if you are going to make guacamole you’ll need to buy the avocados a few days beforehand and ripen them? Or that if you overheat a nonstick pan, you’ll release noxious fumes into your home? The tech tree tells you everything you didn’t even know that you needed to know.
The exponential leverage of a knowledge graph
At this point, perhaps you buy the power of a tech tree for learning cooking skills.
“But wait”, you say. “This knowledge graph seems like it could be huge. How will you assemble all the knowledge that it needs to contain?“
And that’s where this approach truly shows its magic.
Check out the following figure. Another way to look at the tech tree is it's a “concept space” of knowledge building blocks (yellow, on the left), that supports an exponentially larger “content space” that uses those building blocks (green, on the right)4.
In other words, the yellow blocks (tech tree) are the LEGOs of cooking knowledge. Any recipe in existence is just a combination of different blocks of knowledge.
But the number of unique pieces of cooking knowledge is surprisingly small. Just like there might be red, yellow, and blue LEGO bricks of a few different shapes and sizes, there aren’t that many kinds of produce you can buy or ways you can use your stove. But you can create gazillions of different LEGO structures, or different meals, by combining these building blocks.
What does this mean for Parsnip? Well, the current app has around 80 levels and 14 dishes, which isn’t much. But we did a quick calculation and estimate that
At 500 levels, we can teach 80% of the knowledge in American cuisine
At 5,000 levels, we can capture 80% of worldwide cooking knowledge
In other words, every level added to Parsnip exponentially increases the number of recipes we can teach, after incorporating the machine learning component above.
All content creators in the food space work in this exponentially large, combinatoric recipe space. We need to create content for Parsnip too, but must only do a logarithmic amount of work in comparison. As any computer scientist would say, why do O(n) work when you can do O(log n)?
Augmenting existing content and unbundling the cooking vertical
There’s a lot of nerdy content above, so let’s get down to how we build an actual business around it. Consider this famous quote:
“There are only two ways to make money in business: bundling and unbundling.”
— Jim Barksdale, former CEO of Netscape
Currently, the way that most consumers learn to cook is on YouTube and TikTok. But it’s a middling experience:
you can’t search for what you don’t know
you can’t see your progress over time
you don’t know if that video is actually accurate or just making things up
you can’t actually see the recipe as you rewind and watch the video 40 times
The unbundling hypothesis suggests that a 10x better product can peel off users from generalized social platforms into a dedicated, verticalized cooking platform. That 10x better product could start off as a batteries-included, gamified learn-to-cook app with a built-in knowledge scaffold for any recipe on the Internet. You can read a recipe, but you might not really know how to make it until you Parsnip™ it.
The overwhelming power of this approach is that
we don’t just get to eschew creating content in the exponential space recipe space, but
we can leverage the work of everyone else who does without infringing on their content, by showing the underlying cooking concepts behind any dish.
Our end goal here is to pull more and more content onto a new vertical platform for food. We’ll explore some directions for this platform in a future post.
And by the way, if we can get this AI-boosted, personalized, and gamified learning platform to work, it’s going to have way more applications than just cooking 😉.
Did you like this post? Please share it with a friend!
An educated guess: it refers to Balaji Srinivasan’s “idea maze” and the decision tree of technological dependencies that a startup builds, rather than a single product and moat around it.
This is hopefully the only time I’ll link to my PhD dissertation from Substack! 😅
Credit goes to Sean Taylor for the idea of a “concept space” and “content space”. Check out his new machine learning for causal inference startup, Motif Analytics!
your best post yet
your best post yet