Building a startup is a machine learning problem

Source: Deep Learning on Medium


Understanding startup founders as machine learning models

If you do any machine learning or data science modelling, you’re used to solving problems with a very high number of dimensions.

Because the world is high-dimensional, that means that you’re in a uniquely good position to understand many things on a level that others don’t. And one of those things is startups.

A startup is a group of people who make a product. That product exists in what we’ll call “product space”. Product space is an insanely high-dimensional space that specifies every imaginable property that a product has:

Dimension 1: How much does the product weigh?

Dimension 2: How many wheels does the product have?

Dimension 3: How much RAM does the product have?

etc.

Every possible combination of these properties gives rise to a certain amount of demand. For the vast majority of hypothetical products, that demand is basically zero (for example, a product that weighs 20kg, has one wheel, and 10Gb of RAM is a stupid product that no one will ever want).

Peaks in product space are usually narrow. In other words, once you have a product that people already really want, tiny changes can dramatically affect its demand. To see why that is, just imagine your favourite car, and ask yourself: would it still be your favourite car if it had one wheel too few? What if the windshield was blue, instead of transparent? What if there wasn’t a driver’s seat? Probably not.

Now, you’re probably thinking: “What a ridiculous bunch of questions. Obviously I wouldn’t want a car with an opaque windshield, or no driver’s seat. But those aren’t small differences!”

And that would actually be a mistake. In the space of all possible products, a beautiful Tesla Roadster is actually pretty close to that very same Roadster, minus one wheel. The reason you feel a big difference between those two products is precisely that your demand is incredibly sensitive to the number of wheels that your Tesla has. Because the Tesla is a great product.

With terrible products, things are different: if I gave you a literal heap of garbage (which you presumably don’t want at all), you probably wouldn’t care if I added a wheel to it. Because it’s a literal heap of garbage, there isn’t much you can do to make it a worse product, and a lot more you’d have to do to make it desirable than toss a wheel on top of it.

Based on just these features, we’re already in a position to derive some tried-and-true startup wisdom, starting with why you shouldn’t dive into a space you don’t understand.

Your idea is probably terrible

If peaks in product space are narrow, and there are relatively few of them, then it follows that demand is flat (and basically zero) almost everywhere:

The vast majority of product space is filled with useless shit that no one wants and that isn’t worth building.

Now, if we’re a company, and our goal is to make something people want, then we want to find a product that has as much demand as possible. That means that demand (or negative demand) is the loss function we want to optimize.

But as we’ve just seen, demand is flat and rounds to zero almost everywhere in product space. So if we were to launch a random product tomorrow, not only would that product be almost guaranteed to flop (because demand is zero), but there wouldn’t even be any clear indications about which direction our gradient points — and therefore, which improvements we most badly need to make to get traction (because demand is flat). The way you can tell you’re in this region is usually pretty simple: your target audience sees so little value in what you’ve built that they usually can’t imagine themselves using it long enough to give you meaningful feedback.

This is a very bad zone to be in. Often, you’re so far from a peak in product space that when you put what you’ve built in front of a user, they can’t even tell which peak it’s closest to (i.e. what value you’re even trying to offer). In this territory, user feedback — if it comes at all — is sympathetic, vague, and inconsistent (“I can totally see the value that someone else would get from this…”) rather than selfish and specific (“Aw, I can’t message two people at once? Could you fix that?”), because your users — like you — can’t tell which way the gradient points.

Getting out of a rut like that is difficult, and incredibly time-consuming. Here’s what it looks like in our imaginary 2-D product space:

When you’re far from a peak, demand is flat (the gradient is zero), and users can’t agree on what they want you to build (if they even give you feedback at all).

However, things are *dramatically* different when you launch a product that has appreciably nonzero demand (even if it’s not nearly enough demand to sustain a company in the long term). If your product is just useful enough for people to want to poke at it (however fleetingly), all of the sudden, you have a gradient (communicated to your team in the form of user feedback) to follow:

When you approach a peak, feedback gets selfish and specific, and it’s relatively clear what you need to build next.

The upshot is that doubling your distance from a peak makes your startup’s job far more than twice as hard. So there’s a completely disproportionate amount of value to starting your company off with a half-decent idea of where the peaks in product space might be.

For that reason, it’s important to start a company as close to a peak as possible, which is something you can only do by: 1) being incredibly lucky (like a machine learning model with randomly initialized weights that happens to start near a good local optimum), or 2) being your own user, or spending so much time talking to your users that you know them better than they do (like a model that’s been pre-trained on your life experience in the specific region of product space that you’re targeting).

Product-market fit

As we’ve seen, there are basically two regions of product space. The first are regions where the gradient is zero, you don’t know what to build next, and your users don’t care if your company crashes and burns.

The second are areas where demand is just high enough that your target audience wants to give you feedback, and where that feedback is consistent and specific. Everyone you talk to wants the same 3 or 4 new features, and your users aren’t asking you to build a completely different kind of company every few days.

Roughly speaking, the border from the first region to the second is the mathematical definition of early product-market fit. And like any deep learning model that decreases its learning rate as it zeroes in on an interesting optimum, you’ll find yourself making more and more purposeful but less and less dramatic changes to your product as you climb the hill.

Eventually, you’ll close in on your peak. You’ll have built your Roadster. At that point, you can redirect your resources away from fine-tuning your product, and towards making more of it. Fast.

As you hang out on that sweet peak, you’ll make a lot of money. And then, gradually at first, and then all at once, you won’t. Because the peak itself will have moved.

That’s the thing about product space: it’s not static. New windshield wipers get invented. Government subsidies for electric cars are cut. Someone gets hit by a self-driving car. Suddenly, people want a new combination of product features, and the peak shifts on you. You need to keep an eye out for these shifts, because they can kill companies. And the only way to do that reliably is to keep talking to your users. They already like what you’re building, so if you catch the shift early enough, you’ll have a wonderfully clear gradient to follow. But wait too long, and you’ll be back in dangerously flat demand territory.

Why we’ll never run out of startup ideas

When you add a dimension to product space, the number of points in product space (i.e. the number of products you can build) all of the sudden increases by a multiplicative factor.

When the internet appeared, for example, a completely new axis was created in product space: to grossly oversimplify, you could now build any product you liked either “with the internet” or “without the internet”. Just like that, the number of possible products increased by 2X.

Of course, when this sort of thing happens there’s no guarantee that there will be any actual demand peaks in the newly-created region of product space. But if the axis that’s just been invented has the potential to influence a large number of sectors (as did the internet and machine learning, and as blockchain and quantum computing may do in the future), chances are that some peaks — and most likely a great many — will pop up. And that’s what investors are implicitly betting on when they drive the next tech hype cycle.

All of this is of course great news if you’re a startup founder, because it means there will always be things to do, and ideas to pursue. That is, until we build models that can search product space for us. 🤖


📝 Read this story later in Journal.

🗞 Wake up every Sunday morning to the week’s most noteworthy Tech stories, opinions, and news waiting in your inbox: Get the noteworthy newsletter >