The landscape of technology has changed dramatically over the last year. I wanted to provide an update on Buckets and ask your opinion on some things.

The short summary is: sorry for the slow development of Buckets lately. We’ve been busy, but we’re still here and I’m excited for the future. And we want your thoughts on AI.

Photo of a starry sky
Photo by Benjamin Voros on Unsplash

Putting out fires

First, many people use the SimpleFIN Bridge to connect Buckets to their bank accounts. We operate the SimpleFIN Bridge and its use has grown a bit over the last year, necessitating a lot of changes and attention to keep it running smoothly. Most of my time has been sucked into putting out fires there. Thankfully, while there’s still some things to do, I foresee more time for Buckets on the horizon.

The artificial elephant in the room

Artificial Intelligence (AI) has become exponentially more powerful in the last year, mostly in the form of Large Language Models (LLMs). About 2 years ago, a friend asked me if AI could help him make a medical training app. I answered along the lines of “No, it will get you frustratingly close, but never deliver on whatever initial promise you see.” I just told him a few weeks ago, “It could do it now.” It’s surreal how quickly LLMs have improved.

How does this affect Buckets? Well, that’s a decision. And it’s a decision we’re not taking lightly. So I’ve been doing some research…

AI Experiments

Though I’ve dabbled with AI since ChatGPT burst onto the scene, I’ve largely used AI as an assistant and advisor—I ask it specific questions about pieces of code or services and incorpoate its response into my projects. But about 6 months ago, I decided to see what 100% AI would look like on 3 different projects.

1. BearCAD

Logo for BearCAD

BearCAD lets you make models for 3D printing and technical drawings of mechanical parts or structures. This app is 100% AI-made. While building our chicken coop, designed in other software, I was frustrated by the limitations of both free and paid apps, so building a CAD app seemed like a great way to test AI. I have intentionally not looked at the source code and I’m amazed at how good it is. It’s not perfect. It’s 90% great, though. I’m specifically impressed with the UI polish it has.

From building that, I’ve learned that testability is crucial. An LLM will lie to your face about what it built and if you have no way to independently test it, it will bite you.

2. Invoicing

The second AI-produced app is a custom invoicing app for my friend’s business. I looked more at the code on this one, but it’s still very much AI-built. It generates beautiful PDFs, tracks all kinds of numbers and dates, reads uploaded receipts, etc… The realization here was: I’m a capable enough programmer that I could have built every feature, but it would have taken me probably 10x as long. Being able to think, “it would be cool if the app didn’t just save receipts but read them too” and then have that feature appear is empowering.

I also learned from building that app, that while anyone can make an app nowadays, expertise and knowledge of how computers work is still valuable. After adding an account-balance tracking feature, I realized that there was a chance the app would get slower over time (because of my prior experience with databases). I asked the coding agent if that was the case. It was, and the agent fixed it.

3. Counting Carrots

Logo for Counting Carrots

Counting Carrots is the most recent experiment. And I’m impressed. This app could be seen as a kind of “Buckets for Business.” Like Buckets, you can create budgets and categorize transactions to track expenses and income. In the same spirit as Buckets, the app lives on my computer and doesn’t require an internet connection. But most importantly for me, Counting Carrots computes payroll taxes and helps filing payroll tax forms.

Building this app has reinforced an idea regarding AI that keeps tugging at me: cleaning up manmade messes. I’m a problem-solver at heart. I love fixing broken things. The challenge of overcoming an obstacle is invigorating. But I get a lot more satisfaction out of fixing real problems. Natural problems. The knowledge I gain from learning how to garden or build a chicken coop or play a sport is infinitely more rewarding than, say, figuring out how much I owe for social security. Tax law’s complexity is a manmade mess. I find less joy in overcoming the messes created and sustained by others because I can’t shake the feeling of: why make the mess in the first place?

It feels a bit like walking behind someone dropping banana peels.

I first encountered this when setting up code signing for Windows for Buckets. Without AI, I would have had to become an expert on the complexity juggernaut that is Microsoft Azure Portal. Why should I spend effort becoming an expert in that? It’s needlessly complex. I feel so good about asking AI to solve that mess for me. Likewise, I’m glad to have AI consume and compile tax law.

Lastly, the transactions-view and categorizing in Counting Carrots is frankly more user-friendly than Buckets. Seeing this emerge (I didn’t have to be very explicit in my prompting to get a great result) has really caused me to consider using AI for Buckets. It’s like I’m digging a hole with a shovel and someone offers a backhoe.

Buckets + AI?

Now I want to know your thoughts! I’m leaning toward using AI to help develop more features in Buckets, but not embedding AI features inside Buckets. How do you feel about it?

There’s people here

We’ve seen some amazing things from AI and we’re going to see even more shocking things. Regardless of what’s coming, I want you to know that we’re still here. We love Buckets and love working on Buckets. I personally love making an app that has helped so many people. AI can’t replace that.

Happy Budgeting!

— Matt