SheetGenius started with a worksheet.
I was traveling with my family and homeschooling my kids when GPT‑3.5 came out. Like a lot of people, I started by asking it questions. Then I asked it to help make the next day’s schoolwork. It could help write a lesson, but it could not reliably turn one into a worksheet my kids could hold—a clean, printable PDF—without a person stitching every step together. So I built a product around that gap.
The first product taught me something I did not entirely want to learn. Each new model closed more of the gap it existed to fill. By 2026 a capable frontier agent could make a decent worksheet and PDF on its own. I moved most of my effort elsewhere. The app still has its uses, and I have not written it off, but its core trick belongs to the model providers now.
I expect this to happen to other products too. I have said the same thing about BitterClip, our strongest product today. The providers may absorb more of the underlying media work. That is a real risk. Our job is to notice what remains useful, keep that part, and stop defending the rest.
The floor is constantly moving.
I had already changed how I worked for the same reason. When GPT‑3.5 came out, I decided to stop writing code by hand even though I could still write it better than the model could. I figured I was hurting myself by continuing. The new skill was learning how to talk to these things: being precise with language and building an intuition for how a model infers. That skill compounds as the models get better.
Richard Sutton’s “The Bitter Lesson” came later. Sutton looked across decades of artificial-intelligence research and found the same pattern repeating: general methods that use more computation, especially search and learning, eventually beat systems built around a person’s hand-crafted understanding of the problem. I read it as the name for something I was already watching happen at my own expense.
I found the essay refreshing. It makes a serious academic argument without hiding it behind academic language. It is also unsentimental. The clever approach can feel better, look better, and even work better for a while. Then the more general method keeps improving and passes it.
I do not treat the essay as a law of nature. I may be stretching an argument about AI research by applying it to a small software company, and I still do not feel like I fully grasp it. It has been useful as a question: what are we hand-building because the models cannot do it yet, and what will still matter when they can?
That question changed the company. I kept building. Some products found an audience. Some did not. A few became useful enough that people paid for them. Others taught me something and then went quiet. Today SheetGenius has 58 active software products, from classroom materials and time tracking to music, data, developer tools, and games.
Plenty of software ideas were possible but not worth the cost. A tool for a few hundred people, a database about one narrow industry, or a game that does one thing well can be valuable without supporting a large team. The trouble has always been the cost of building it, putting it online, fixing it, answering questions, and keeping it alive.
Bitter is the clearest expression of what I am doing next. A customer brings the frontier agent they already use. Bitter is becoming a dedicated Linux box where that agent can work, with durable source, deployment and verification, recovery, and a history of what happened. Better models should let that environment do more.
Bitter has to face the same question as SheetGenius. Anything a stronger model can simply absorb was temporary. We should let it go. The box, the source, the live products, the customers, the domains, and the operating history are the bet, but I do not think any of that is guaranteed. Bitter has to earn its place too.
If software creation becomes abundant, other things become more valuable: a domain someone remembers, a product that still works next year, and a customer who knows who is responsible when something breaks.
I spent twelve years helping build one software company before leaving in May 2026 to work on this full time. That experience made me patient about what real companies require. They are built one customer, one outage, one renewal, and one unglamorous improvement at a time.
SheetGenius is still early. The portfolio has more breadth than traction. I need to find the products people come back to, make them excellent, and be honest about the ones they do not.
A portfolio page can look like a pile of links. What matters is what happens after each product meets reality: whether people return, pay, ask for more, or quietly leave. My job is to pay attention, move when the floor moves, and stay with the useful things long enough for them to matter.
Thank you for taking a look.