Originally posted on jacobobryant.com

I got my bachelor's degree in 2017 and then worked for a year as a software engineer. Last January I quit and began working on a startup full-time with a friend I met while in school. I realized after a few months that we weren't a good fit as cofounders, so I bailed. I then spent some time trying to figure out what to do with my life. In June I decided to go all-in on my current startup, Lagukan.

Several days ago, I discovered a potentially critical flaw in my startup idea. I might need to take a major change of direction (though I'm thinking of moving to an idea that will still solve some of the same problems). This kind of thing has happened to me quite a lot in the past year, so I thought I'd take this opportunity to do some introspection before I go head-down coding again.

So, here are a few of the things I've been thinking about.

Spotify is actually pretty good

I've been in love with music since I was about 12. I started using Pandora while in high school. I discovered a lot of new music that way, but it was far from perfect. It was never able to fully adapt to my preferences (e.g. I love melodic hard rock and hate metal, a distinction that was utterly beyond Pandora), and it tended to play the same things over and over.

I started doing music recommendation research while in college, culminating in a prototype that played from Spotify and my MP3 collection but used my own custom algorithm. It was nice, although it left a lot to be desired. I later tried using vanilla Spotify for a while and wasn't super impressed.

Hence, I eventually decided to pursue my recommender system project as a startup (Lagukan). It seemed like a natural choice. I was encouraged by the response on Hacker News to my first prototype. Here's a snippet from one of my favorite comments:

Oh man... I haven't been able to find any recommendation algorithms that recommended songs I liked even 1% of. Spotify's is a trash fire. Last.fm's wasn't great. Pandora is okay.

Pulling off a successful music company has been historically pretty difficult, but I was alright with that since I was working on something that was deeply meaningful to me.

One of the most important issues in music recommendation (and more generally, in reinforcement learning) is balancing exploration (trying out new songs) with exploitation (playing songs you've listened to before). Lagukan has two separate algorithms for these tasks plus a parent algorithm that decides which sub-algorithm to use at any given moment. It's a delicate balance. Back when I was a Pandora user, they only really addressed exploration. Spotify seemed to have the opposite problem of doing too much exploitation.

However, I've slowly realized that I was likely just using Spotify wrong. I've always had the assumption that a music player should be totally automatic: just hit play, skip what you don't want to hear right now, then the algorithm should do everything else. That's how I've built Lagukan, and that's how I tried to use Spotify—I almost exclusively listened to Daily Mix.

This is a long-winded way of saying "I should've been using Discover Weekly."

I had read several comments on HN saying that Spotify performed much better if you listen to Discover Weekly consistently. It makes sense—that's their solution for those who want automated exploration, and Daily Mix is for exploitation. Sure, it's not all wrapped up into a single totally automatic algorithm, but the overhead is pretty low.

I've been using Spotify directly for the past few days, and it's been disturbingly effective. It's been seeded by Lagukan's exploration from the past month or so, but I'm going to try using just Spotify for several weeks and see how well it does. I'm open to the possibility of not needing Lagukan personally.

Even if Spotify turns out to work well for me, there likely are other people for whom a better recommendation algorithm would still be valuable. But if I'm not one of those people, then I don't think I'll be able to generate the insights needed to help them.

You should be comfortable with cutting your losses

It sucks to put five months into a startup built on a false hypothesis, but it's better than sticking with it and running out of savings. I'd say it's also better than never taking the leap at all, because the losses often aren't.

For one, my skills have grown far faster than they ever did while I was a student or an employee. I've been able to learn everything I wanted to in college but (ironically) never had the time for due to coursework, and I've discovered plenty of things that I'd like to dive into more. It's the most freeing educational experience I've had since I was a homeschooled teenager. I've also learned quite a bit about the industries I've been trying to enter, not to mention the emotional dynamics of not having any external validation from being an employee or a student.

Despite failing over and over in the past year, I don't see any of it as a waste. It feels more like do-it-yourself grad school (minus the degree at the end). I don't think any other path would have been as good of an investment in myself.

About three months ago, I started to get another idea for a recommender system startup that I actually think is more promising than what I've been doing. I stuck with Lagukan because I didn't want to get whipsawed, constantly switching ideas before getting traction. Even if I do ultimately decide to relegate Lagukan to side project status, maybe my work will turn out to have been optimal: I wouldn't have had this second idea if I hadn't dug into music recommendation.

What about giving up too easily? You certainly shouldn't do that, but if you, in your heart, no longer believe in the idea, I think that's a pretty good indication that it's time to switch. I've been through many discouraging experiences with Lagukan, but this is the first time I've seriously doubted the validity of my core value proposition.

It helps to articulate what you want to get out of your career

I'm a huge fan of Paul Graham. I stumbled on his essays while in high school; it was the first time I ever heard about startups (and Lisp!). I like to think I'm still as ambitious as 15-year-old me reading his essays for the first time, but actually trying to start my own business without giving up requires that I understand why it's worth my time.

First of all, I have to be a startup founder because right now there are certain things about the world that I'd like to fix, and it would kill me to spend my hours at a day job instead. But what else is important to me?

• Working on my own projects. Even if I don't do it as a startup, I have to at least have lots of time for side projects. I love exploring and I'm a very self-motivated learner. I can't give up my freedom.

• Working in Clojure. I was more language-agnostic when I graduated, but after a year each of doing Clojure and non-Clojure full-time, this is important to me.

• Playing clarinet. I'd like to join an orchestra again, though I've put it on hold while I try to build a business. At some point I need to have at least an hour a day for practicing.

• Teaching. This is another thing I love doing, and it's high-impact. At a minimum, I need time to write guides and tutorials regularly. (On the other extreme, some kind of education startup is my backup-backup-plan for Lagukan).

With all these demands, you can probably see another reason why doing a startup is so appealing to me: it'd be great to just make a lot of money all at once so I have freedom to work on whatever I want, regardless of if it's profitable or not.

But if I get through all my startup ideas and fail each time, what's next? If Stripe offered me a 30-hour-per-week position writing Clojure, I'd probably take it. Barring that, I think consulting is the most likely way to hit all the bullet points. I would be fine alternating between my own projects and clients' projects. (I've also thought seriously about grad school... but I've decided against it because I'm interested in design, not research.)

The result of all this pondering is that my decision to continue being a startup founder is based on careful reason, not emotion alone (I'm not a fan of burn-the-ships). I'm confident that if I come to a point where the rational thing is to no longer be a startup founder, then I'll do something else and still be happy. That understanding helps me to not get so depressed when bad things happen, which is important since startups run on morale. Knowing that there are other good career alternatives also helps with cutting losses.

Coming up

I've also developed my own opinions about cofounders, networking, idea validation, bootstrapping vs. VC... so let me know if you'd like to read a "part 2" of this article. Meanwhile I'll be building a prototype for my next idea. Fortunately, I should be able to release this one within a week or two. And if it works, Lagukan will likely make a good complementary product down the rode. I'll keep my fingers crossed.