How I use AI

First to set the table. I’ve been building websites, doing development and contributing to open source for more than 20 years. I’ve managed hundreds of commercial sites from the server to helping update content. I’ve worked on high volume, high profile sites and built tiny tools for niche technologies. Front end, back end and all the bits in between. I also eschewed Web3 and the blockchain. I avoided “popular” web technologies because I thought they were bad for the web and the user experience. People are finally starting to come around to the school of thought on web development I’ve ascribed to and are ditching those methods.

But Large Language Models, “AI” is not something I’ve dismissed. I tracked its development and eventually began exploring using it with my own work. At first, it wasn’t very helpful. Occasionally you could coax some working code out of it but it wasn’t ready for prime time. That was 2 years ago. Starting about a year ago though, things began to change. Claude Opus was a significant shift in what the tool could accomplish. You could quickly scaffold a project and test out an idea. Ideas are cheap, committing time to building them is expensive. Cutting that time down to minutes affords the opportunity to try out more ideas and iterate faster to get a prototype of what you want to accomplish.

Over the past year, I’ve used those tools to write code. Scripts. Mostly Node and Python that parses data from a source of my determination into a machine readable format and do something with it. Census.gov data, the city clerk’s agenda, the various data sources from the city’s GeoHub. I’ve recently added the county property appraisers data. Data that I can manually confirm the output. I use AI to help write code that allows me to generate verifiable and reproducible results from the data. And then I’ve made very deliberate design and technology stack decisions on how I’m going to present that information beside the words I write. Maps, charts, tables. WordPress, markdown, stored as json, mixed into a sqlite database. HTML. Stuff I’ve been doing “by hand” for years. Different data sources are collected in a variety of methods. All decisions I’ve made for how I work and think at this point in my journey of building for the web.

But the code? Sure, I’m good with letting a machine write code to fit into the tight parameters I’ve defined. At this point, it writes much better code than I ever did. As a mostly self taught web developer (I came into the profession having been a chef, not a designer or software engineer like others who’ve been doing this as long as I have) I’ve always read code better than I could write it. I was a user of the project or library. When you’re building and managing websites, you’re not cranking out custom code all day, you’re tracing error messages and reading through code. You’re writing wrapper functions with a couple of parameters to the point you automate the code with snippets or libraries. And sometimes, you get to write a custom feature. But I was always a fixer. Someone once described the types of projects I’d worked on “the jobs no one wants to do”. One year I easily deleted 10x more lines of code from projects than I wrote. But having Claude write code? It’s a pass/fail. The code either executes or throws an error. The output is either what you expected or it’s not. Sure, it can be written in a convoluted or inefficient way, but if you put up guard rails — documentation, linters and testing — you mostly get text book code these days. There will always be ways to improve and refine. Technologies are constantly changing and so does the code.

But that’s where my use of LLMs and AI ends for Tampa Monitor. I do not use it to summarize text or documents. I read every agenda item. If it’s not self evident from the description or if have a question, I read the staff summary sheet. I’ve formatted the agenda in way that spells out the financials and the explanation from staff why the request. If that doesn’t answer my question (or generate new ones) I check for an additional staff memo. Or I read through the contract and RFP. After 3 years I’ve gotten good at finding the meat of the issue and ignoring the 90% that’s for the lawyers. All the while making notes.

If I want to do external research, I still use web search the old fashioned way, I just use Duck Duck Go (which is Bing search under the hood). No tracking and I can still generally find a broad enough set of results to base my research on. Web search quality started going down before chat AI interfaces were introduced. Do I think I could get better results than Google search using Claude Fable after writing a longer “README” file that includes context of what I’m researching? Sure. And Fable is at a level that it might not include a hidden bias to want to confirm my hypothesis. But it might. It might miss some context when evaluating a result and exclude it or go down a rabbit hole using an old source and waste my time (and money) all in an attempt to “prove” my theory.

The extent to which I use any artificial intelligence in my writing is to use Apple’s built in spelling/grammar tools in my app of choice to write. I have a couple of sticky keys on my aging laptop so it’s autocorrect while I type is nice. I keep Apple’s Dictionary/Thesaurus and calculator open. But I’ve never used any other grammar tool or writing aid. I do my best to read through and edit and read through and edit to have some flow and convey what I’m trying to emphasize. On a couple of occasions researching some complex numbers my writing has probably been closer to how I write documentation for projects than journalistic prose, especially when I start copying/pasting from the documentation and notes I’ve written. Every image I use is either stock photography or from a cited source. The icons currently in use are paid Font Awesome Pro icons funded through my freelance work.

And I’ve heard all of the arguments. Some of my web heroes are staunchly anti-generative AI for any purpose. I’ve heard about the environmental concerns and there are a multitude of things we should be doing as a society to reduce man-made climate change. Making AI the boogey man for the failures of the United States is not the solution. The current policy of the United States to embrace coal and oil over wind and solar is exponentially worse for the environment than building a data center or running an LLM to write a python script to parse a json feed of public data. If the US was going full war production level of building and installing solar, wind and geothermal; building public transportation at the local and regional level to connect inter-continental travel instead of everyone flying and driving, and real data could point to inefficiencies in data centers exist, then we can talk. I feel pretty good about my water usage and carbon footprint over the 57 years I’ve inhabited this planet.

As to the economic impacts a shift in technology this big might bring? I think at the ground level we don’t know. Will there be a shakeup in the market? Sure, it’s been overhyped and people have been making poor investment choices. Could it be 2008 real estate bubble bad? Yes if you aren’t paying attention to the warning signs. Signs far fewer people saw in 2008. Is it avoidable? Not in this gotta-get-mine now environment. And none of that takes into the speed at which open source models are advancing, mostly by the Chinese and how that will alter the course.

I was put in front of a computer in 1979 as a 5th grader with my best friend Mike and what ever manual that came with it and left to figure it out. That now I can write in natural language and have the computer do things seems like a fitting point in the journey.

I don’t know where it’s going and if it disappeared tomorrow, I’d still have all of the code and tools. I could still pipe current data from the property appraiser through a filter of the boundaries of the CRA districts to see which areas showed growth and which didn’t. It will take longer to go back and read the documentation and to modify the filter instead of using natural language parsing, but I’m still going to get the result I was after.