A small experiment building a home app

This is a small story about working with the circuit breakers in my house. Solving this little problem has helped illuminate the state of apps and what AI will do to them – every business app is going to get torn apart by AI, we are at the beginning a great refactoring of the SAAS market. And I see hints of a way forward for the smart home as well.

Circuit Breakers Background

We live in an older home with a detached garage; we have two breaker panels. OK, I am lying; we have six panels. The house has been added onto and remodeled many times, and now we have multiple panels in many locations — the garage, the furnace room, some closets, etc. As the house has grown and morphed, circuits have been added willy-nilly in illogical locations, and all made more complicated when we recently added a backup generator, which required rerouting critical circuits to that new panel.

Every panel is in a dimly lit location. Inside, multiple generations of old and contradictory labels are pasted on the door in tiny font with some handwritten corrections. Each panel has a mix of standard breakers, GFI breakers, and arc fault breakers (which are the bane of our existence — they are incredibly finicky and reset when we sneeze). We have local GFI outlets in the kitchen and bathrooms to protect us in those rooms.

Mostly, I can ignore this morass. Until a breaker flips, and then I have to go on a journey of discovery to figure out where the breaker is. And I have done this enough times that I have some instinct now. Other family members or guests have no idea; they certainly don’t know where all the breaker panels are.

What I Want

When a light or an outlet or an appliance won’t work, I want a simple and quick solution.

I want to ask, “Hey, the disposal is out; what breaker do I check?” And my phone (or PC or Homepod or Echo or whatever device) will tell me, “Try the panel in the garage, breaker 15.” ^[1]

Using an Existing Dedicated Mobile App

Breaker Box seems to be the leading iOS app for representing breaker panels. And it works fine; I dutifully set up all my panels. I can now search in the app and find the breakers of interest. This isn’t bad, but it is kind of representative of an old approach:

• you have to manually enter all your data in the rigid schema of the app using the app UI
• all your data is locked into the app (it offers an export option, but I have been unable to make it work)
• the app doesn’t export data to Spotlight/Siri for searching or scripting; you can only use the data in the app
• because of the above two issues, if I want family members, friends, visitors, or vendors to use the data, they all have to install the app
• the app hasn’t been updated much, and I am dependent on the vendor updating and maintaining the app

This was a reasonable way to build an app years ago, but it feels outdated. There is no reasonable way to have the entire household use it.

The AI Way

For a different and more modern approach, I started with just pictures of the panels and the labels. I named the files descriptively — “Garage Panel”, “Furnace Room Panel”, etc. I went to Google’s Notebook LM site and created a new notebook. I uploaded all the pictures into the notebook ^[2].

After ingesting the pictures, Notebook LM gives a nice overview of what it thinks the docs are all about:

Five documents, "Generator Panel Guide Sheet," "Panel A Guide Sheet," "Panel B Guide Sheet," "Panel C Guide Sheet," and "Panel D Guide Sheet," detail the electrical circuit breaker panel layouts for a house. Each document lists the circuit breaker number and corresponding electrical load description, providing a comprehensive guide to the electrical system's components. The documents likely aid in troubleshooting or maintenance. The specific loads include lighting, appliances, heating systems, and other outlets. The information is organized by panel to identify the location of each circuit.

You can type in queries like “Which breaker controls the dishwasher?” or “What are all the breakers on the garage panel?” The responses are very accurate. I did not detect any errors, even when dealing with messy handwritten labels.

Notebook LM has other odd features besides free-form querying, like generating a FAQ, a study guide, or a two-person podcast episode discussing the documents. I will let you know when I start offering this thrilling podcast.

The great things about using Notebook LM:

• I didn’t have to spend time entering, transcoding, or organizing the data.
• I didn’t have to enter it into some proprietary schema in some proprietary database — I just uploaded the pictures and got to work.
• Notebook LM creates a great understanding of what the pictures are about with no prompting from me.
• The answers to queries are very accurate.
• It is just a webpage — I can share it easily with others and other devices.
• The Notebook LM page is saved and is useable across time and devices.

The shortcomings:

• There is no voice input for queries. On the iPhone, it was easy to use the standard microphone input method to enter a voice query. However, this feature isn’t available on the Mac. I imagine Google will resolve this.
• There is no integration with Siri. I could probably write an iOS shortcut that captures voice and feeds it to the webpage. Google could also easily address this.

I also tried using ChatGPT for this solution. Notes below, it was not as good as Notebook LM. ^[3]

Reflections on App Development

We are achingly close to this scenario just working with no coding. Three years ago, the idea of being able to query across the content of a set of pictures seemed very far away. Now, the pieces are all there and almost completely lined up. And you can scale this app up — handling thousands of pictures or PDFs, getting much deeper content from them, and joining information across them. It will be very easy to build a class of apps that used to be very hard (or at least time-consuming) to build.

Microsoft, Google, or Apple is going to make this very easy. Probably each of them will. Satya has been commenting on how AI will undermine all the SAAS apps today. Many SAAS apps are just some brittle business logic built on top of a database — and AI is going to cause all those apps to be refactored and rewritten. We have the opportunity to dramatically lower the cost of app development while also tearing down the silos between apps. It will be a grand refactoring of the apps most companies use today.

Reflections on the Smart Home

Most smart home solutions today require an inordinate amount of finicky detail work. This includes naming and characterizing all the smart devices, organizing them into groups, creating rooms and scenes and automation, and so on. Almost no one has the time or interest to do this, and no one wants to be a home IT manager. I went through the pain of parsing all my data and entering it into the Breaker Box app, but no regular human is going to do this.

I can start to see the way for the smart home to really take off. Let people just upload a bunch of pictures of their home and their gear, let them voice annotate it, and stand back and let AI do all the work of making sense of it and delivering a free-form interface to humans to interact with their home — “Turn on the lights in the living room at dusk,” “Give me a list of all my furnishings for insurance purposes,” “What annual maintenance should I do for this furnace and who are the best local providers,” “How do I fix this running toilet,” etc. This is a smart home I would use.