How I used AI to create a living data visualization for our home
Putting AI to work is all well and good, but you’re more likely to push your AI learning to the limit if you pick a personal project that will inspire you to explore what’s possible.
That’s the case I make in my latest newsletter, which maps out a four-step process for learning from that kind of creative technology exploration.
One key step is to explore at least three different ways you can put a new technology to use. But I often go well beyond three, so this posts shares the nine different approaches I tried in the course of my big summer project: Designing a slat wall for our home.
A case study in learning from AI-enhanced creativity
My own creativity with AI is accelerated whenever I read an AI prompt or use case that really inspires me. So I’ve documented my summer AI project in the hope that it gives other people ideas about how they could put AI to use, plus some examples of specific prompts that help with design and data analysis.
Here is the challenge I wanted AI to solve:
You are an interior designer creating a visually appealing slat wall for an entryway.
The slats will be cut from 8, 10 and 12 foot lengths of lumber, in a mix of widths; 1.5", 3.5" and 5.5". These will be attached to the walls horizontally, so that the slats will consist of varying heights: 1.5", 3.5" and 5.5". Every slat will be exactly .75" above the slat below it.
Propose 3 different patterns for combining slats that will together complete the 92.25" height.
1. Teach GPT to do math
My first effort at coaxing a pattern from ChatGPT was a bust, because it was based on giving it the assignment as a math problem. By the time I realized the problem — “Damn it, Alex, I’m a large language model, not a calculator!” — I had used up my then-limited GPT-4 queries.
After three hours of enforced downtime, I turned on the Wolfram Alpha plugin for GPT, which extends the AI with math and data visualization capacity.
2. Ask for inspiration
More important, I decided to look to GPT for general inspiration, rather than an immediate solution. I gave it a table of my lumber sizes and quantities, and asked:
What natural or human objects, patterns or sequences reflect the proportions or quantities in the following table?
GPT came back with a few options (spider webs, leaves on a plant) but the one I really liked was “musical rhythm or melody”. I mean, how cool would it be to have a front hallway that is a piece of music? So I asked GPT to suggest some tunes:
Can you suggest a musical fragment of 25 notes (made up of no more than 3 notes in the scale, arranged in whatever way)? Ideally it would be a tune or fragment from a Star Trek theme song, from musical theater, from a progressive political anthem (like Solidarity Forever) or from 1980s rock/pop.
3. Get help with pattern recognition
Finding a matching tune also turned out to be beyond ChatGPT’s powers, but it gave me something to go on. I dug around the Internet for sheet music for the Star Trek: The Next Generation theme music, and I fed notes into GPT to see if it could find a fragment that would match my criteria.
No dice! I don’t think I can blame that one on GPT, just on the composers who saw fit to use more than three musical notes in their theme music.
4. Step away and stew
I often get the best results from AI when I actually step away and stew on what I’ve learned and observed so far.
Once I started thinking about our slat wall as symbolic language, it got me thinking about what else it might symbolize. And my experiments had shown me that just about any viable pattern was going to work out to about 25 slats. I woke up the next day to an epiphany: We’ve now been in our house for 25 calendar years!
If we treated each slat as a notional year, our baseboard could represent 1999, and the top slat could represent 2023. Then we could assign every slat a notional year and a notional weight, based on whether it was uneventful, eventful or life-changing. So Rob and I worked out the timeline and math in Google Drive, like normal people:
5. Write faster spreadsheet formulas
But I still wanted to validate that this pattern would look nice, which meant translating the S M L notes into some kind of visual, ideally by creating an Excel table where each row would represent a quarter-inch gap or a quarter-inch portion of a slat (so an L slat would occupy 22 rows, and a .75” gap would occupy 3).
Turning our sequence of S-M-L slats into Excel data seemed like it would be a much better fit for ChatGPT’s talents, so I asked it for guidance on how to create an Excel formula that would repeat different values.
Even though GPT often helps me write complex formulas, this time it wasn’t up to the job.
But one big benefit of working with AI is it lets you fail fast — and in this case, my inability to produce an effective Excel formula prompted me to try something new.
6. Ask AI to fix your data
When I got frustrated with writing formulas, I realized I could just get GPT to do the work directly. Learning to use AI means rethinking when it makes sense to do something manually or programmatically versus when it makes sense to just give the task to the AI.
In this case, I got great results from a simple instruction:
Insert repeats of each letter in this pattern based on the following multiples: G = repeat 3x, S = repeat 6 times, M= repeat 14x, L=repeat 22x
Then I just pasted in my 50-row pattern and BAM! ChatGPT spat it back as the 382-row series I needed. All it took was a little conditional formatting in Excel to turn my 382 rows into a design my carpenter and I could agree on:
That would have been the end of it: We now had our slat wall, and I’d figured out where GPT could and couldn’t help.
7. Predict the future
Our carpenter pointed out that even though our family timeline took us to the top of the entryway, our slat wall continued all the way up to the third floor — so I would need to give him a pattern that could take us up another 15 or 20 slats.
“But I don’t know what’s going to happen in 2040!” I protested. Luckily, our carpenter has been working around here for a few years now, so he isn’t fazed by my idiosyncratic design issues.
I decided that my lack of clairvoyance should not get in the way of a good data visualization: We could brainstorm a list of events that might occur in the next couple of decades, and then ask our friends to help predict what might happen when.
I asked GPT to help me come up with a list of possible future events:
What are 20 technological, social, political or environmental developments that are commonly predicted will occur between 2024 and 2040?
Then I added a bunch of more personal milestones, like seeing the kids finish university or welcoming our first grandchild.
8. Ask for software recommendations
Picking a platform for our “what will happen when?” survey sent me down a whole other tech exploration path. I used to do a lot of online survey programming but it’s been a few years, and I had a very specific survey experience in mind. I hoped ChatGPT would help me identify a platform that could meet my specs, but it only succeeded in giving me a list of survey platforms that might meet my needs, not at picking one with the exact capabilities I wanted.
So I fell back on the proven strategy of just kicking the tires myself. After testing my desired question flow on a half-dozen different survey platforms, I landed on Qualtrics:
9. Get data analysis support
From my past survey research experience, I knew that it would be a nightmare to convert the results of my matrix question into a single predictive timeline — and I thought AI could solve the problem. Since this was playtime, not serious data analysis, I was prepared to to take some risks with GPT’s data acumen.
So I turned Wolfram Alpha back on in ChatGPT (now, I’d use GPT’s “Advanced Data Analysis” option), and prompted:
You are a data journalist developing a set of event predictions based on a set of 11 survey responses from experts on this timeline. You must assign a series of 18 events to the most likely year each event will take place, running from 2024 to 2040. Because your survey responses are biased towards estimating events will happen relatively soon, events are most likely to be predicted for occurrence between 2024–2033. However you need to array these events on a timeline that runs until 2040; no event can take place in more than one year, and no year can have more than one event. Working from the attached timeline, assign each event to a year. Provide an explanation of the method used to assign events to years, working from the attached data file. Consider using the standard deviation of predicted year as an indicator of consensus around how likely it is that an event is going to happen in a specific data range.
ChatGPT struggled a bit with its constraints: I’m not going to be placing stock market bets based on these calculations or predictions. But I was able to generate a projected timeline of the future for slat-wall purposes:
And now we have a slat wall!
Curious about how this experiment helped me refine my professional use of AI? Subscribe to my newsletter here.