How I use AI

I didn’t think any of these examples were all that special, but I’ve been surprised at how surprised others are in hearing about them. By no means am I claiming to be a power user, but each of these applications has been a godsend to me. Also, most of these assume that you’re not poor and have a premium subscription to ChatGPT and the like.

Writing

The best description of AI’s role in my writing is as an ever-present editorial assistant. It’s absolutely unparalleled as a thesaurus, where I can apply an infinite number of filters on my quest of finding the perfect word to use. An example could be where I’m looking for a word to replace money, but given the particular context I want something that better conveys the filthiness aspect of wealth—AI might suggest terms like lucre, filthy lucre, or even ill-gotten gains, depending on the nuance I’m seeking. It doesn’t stop there; it can expand on these suggestions, offering synonyms or related phrases that evoke greed, corruption, or tainted origins, all tailored to the tone and style of the piece I’m crafting.

The last half of that paragraph was AI-generated, which gets me to the next best application: examples! AI is so good at generating illustrative and vivid examples of ideas I’m trying to convey. I used to agonize and waste hours trying to conjure up and hone examples so they can straddle the ideal point between simple enough to carry the point, and avoiding needlessly elaborate texture. Now I can just iterate it with AI. Most of the illustrative examples in Pronouns are Annoying were generated with the help of AI.

When I recently wrote Who Controls The Dirt? I knew I needed deliberately banal examples to illustrate what is otherwise a highly charged topic. I learned about the disagreement between Thailand and Cambodia over the Preah Vihear Temple in response to me asking ChatGPT to find me border disputes from around the world with the lowest stakes. AI also helps when you’re looking for obscure and thematically clever references to include. When I needed to skewer a parody of Portuguese nationalism, ChatGPT delivered by pointing me towards The Sleeping King and Bandeirantes.

It’s not a direct application but voice transcription has improved dramatically in recent years thanks to AI, and I’ve found it extremely useful to record myself when telling stories so I have a body of text to work with later. A bunch of my book chapters thus far have been built from me getting stoned in the evening and telling hour-long stories to my lovely and very patient wife.

I’ve tried feeding AI giant chunks of my writing to see how well it can generate something from scratch but that exercise has thus far always been a disappointment. It never gets it quite right, nor does it avoid a dry and monotone delivery. I’ve learned from my book coach that I’m firmly in the “discovery writer” camp, where the overwhelming majority of my cute and clever ideas happen during actually writing. Some of you may have noticed how my stories of the legal system follow a sort of a flipbook pattern of alternating between anecdote and meta-commentary. I fell into that randomly as I was writing Eleven Magic Words where as I was typing out the anecdote portion, I realized that some background information would be useful to contextualize what’s happening. AI seems to lack the theory of mind of identifying the utility of these asides, and the creativity of knowing how to structure them in an engaging pattern. At least for now.

Right now I’d say that at least 95% of my published writing has stemmed out from my actual fingers typing on my actual keyboard, though I’ll note that I am not ideologically opposed to a future where that number dips as AI capabilities continue to improve. I often find the writing process very painful and tedious, so this figure is actually a disappointment given my expectations and hopes from when LLMs first hit the scene. No part of me is wedded to the notion that AI-generated art, whether wholly or partially, is any less valid.

Research

What a boon! When I wrote my first post on Israel-Palestine, I was transparent about the fact that a significant portion of my research was assisted by ChatGPT. Before you strangle yourself from getting your panties in a bunch, understand that I would never do anything as foolish as accepting answers at face value. Instead, the inimitable application is using AI as a thread starter: asking it to point you to Wikipedia articles, published papers, or articles found elsewhere.

I first stumbled into this application while watching the very excellent TV show Narcos. A major plot point in season one of Narcos: Mexico involved the cartel spending enormous efforts digging for wells. I very often go Wikipedia spelunking whenever I watch historically authentic dramas, and for whatever reason, I absolutely had to find out how the Mexican legal system dealt with water property rights.

My college major was in economics, and so I was used to tracking down and devouring gigantic PDFs on the most obscure of historical minutiae, but I can’t say that was a fun or efficient process. Tracking down sources requires surviving the gauntlet of disparate academic databases with inconsistent paywalls. Tons of time is then incinerated into ashes as you read and skim through pages and pages of text that turn out to have absolutely no relevance. AI is extremely useful at cutting through that chaff and synthesizing libraries of text into precisely what you’re looking for.

A related application is asking it to generate dynamic quizzes to test your knowledge and retention. It starts with basic questions and uses your answers to calibrate difficulty until it identifies your knowledge ceiling. I tried this to see how much I retained when I “read” a book at 2.5x speed while also playing video games on another screen and I was pleasantly surprised by the results. These dynamic quizzes are also great at identifying shortcomings in your understanding that require shoring up.

One of the perennial fears any blogger has is wasting hours sketching out an idea that has already been widely elucidated to death. Whenever I think I’ve identified a novel phenomenon or dynamic, I very often ask AI to check whether what I’m describing already has a cognizable name. Identifying the name is the necessary first step to checking what contributions others may have already made to the idea. Prior to this, the only possible avenue was spending a ton of time reading tons of different sources, explicitly in search of a negative. If you happen to be fortunate enough to strike it rich on an unclaimed gold vein, AI is also very helpful in coining a name for the phenomena that not only accurately captures its essence but also finds a label that is non-duplicative.

Supposedly Coding

I’m not a coder. The last time I had any formal exposure to learning computer programming was a Javascript class in high school where missing commas constantly tripped me up. I know it’s derided as a programming language but I rely heavily on Excel spreadsheet formulas on a range of applications. Attorneys often spend exorbitant sums on expensive case management software to track deadlines, billable hours, and pending projects for clients. I’m a solo practitioner who doesn’t need most of that fluff and so my case management system is just a slightly elaborate spreadsheet with multiple formulas that keep the whole thing humming.

I’ve been using Excel formulas my whole life so I’m used to doing everything manually, but goddamn it fucking sucks having to troubleshoot anything when it (inevitably) breaks. You have to maintain a mental compartment of the different available formulas, their comparative merits (e.g. why INDEX MATCH is superior to VLOOKUP) along with their maddening quirks. It’s so much easier feeding everything to ChatGPT and asking it to do the hard work for you. It’s amusing to me how often it gets it wrong, but I just keep shoving the error messages back in its face until it gets it right. I used it to create a needlessly clever log to keep track of my weightlifting routine that I won’t bother explaining:

AMRAP is the only hint I’ll provide

I’ve now expanded into learning and writing Python scripts! I’ve been using this for work a whole bunch as a way to wrestle into submission the incoherent mess of electronic discovery I receive. We routinely receive recorded jail calls in a proprietary format where the only plausible way to review is through an included html browser list. This interface is very often broken and very poorly programmed. For those who understand, I’ve seen them use fixed rather than relative file paths, which means they only work if you run them from C:/. It’s so fucking stupid, but almost no one who uses them (read: old lawyers) has the capacity to complain.

You can actually directly access the raw audio files, but they have garbled gibberish names along the lines of 10.37.144.189-11063f610a2491bd7f0c9600b4c6502a.mp3. The fix I created with the help of AI was to create a Python script that uses the BeautifulSoup library to scrape relevant data from the HTML browser and automatically rename the garbled file names to something more legible. Now instead of relying on the shitty browser as the window into this world, I can instead listen to the jail calls directly from the folder with pertinent information about the call included in the file name itself:

I cannot tell you how much time this has saved me, and I pity my older colleagues who can’t know any better.

Personal trainer

I’ve been weightlifting throughout my adult life and though I’ve occasionally gotten a personal trainer, I’ve rarely strayed far from the basic compound lifts of squats, deadlifts, and bench press. Those have served me very well. But not one to pass up an opportunity to shore up my deficiencies, I asked ChatGPT to give me a list of exercises to serve as a physical audit to see where I might be lagging behind.

I never do any push-ups but I was able to easily blast out 30 within 30 seconds, a clear signal that I had no reason to worry about my upper body strength. But when it came to basic balance, I was absolute garbage. I barely lasted 10 seconds with a one-legged stand before I started wobbling uncontrollably. I’ve had ACL reconstruction surgery a few years ago, but relying so much on two-legged barbell squats since then has deluded me into thinking I had fully recovered. Once my strengths and deficiencies were identified, ChatGPT was also able to create a different work-out plan specifically tailored to my specifications. Barbell squats were now replaced by the thoroughly exhausting Bulgarian split squats for example.

The tailored customization cannot be understated. I built a power rack in my garage, but unfortunately the low ceiling prevents me from doing any overhead military press, so AI suggested the sitting Z-press as an alternative. It’s been so damn useful with feeding it any other equipment constraints and for it to come up with seamless alternatives. It was also really good with creating a routine that both my wife and I can do together at the same time, despite me weighing more than double what she does, and despite her pathetic noodle arms (Sorry babe it’s true 💖).

Travel Agent

Oh man, this is so good! The ability to pull disparate information and synthesizing it into a coherent response really comes into play when asking AI to create a travel itinerary. The name of the game here is the immense number of constraints that need to be navigated: you’re in a foreign place for only a narrow time window and wish to make the best of it.

We used ChatGPT when planning a trip to Europe. The initial plan was quite ambitious and involved four different countries. It was very possible, though quite hectic. We realized that our biggest prioritization should be sticking with a single city to serve as our home base to avoid having to relocate from place to place. I gave ChatGPT a list of places we ideally wanted to visit, and it narrowed it down to Florence given its central location within Northern Italy and its abundant access to high-speed train service.

Figuring out the rest of the details was trivial. It knew which destinations could easily be contained within a single day trip, and which ones were better served by overnight stays. It had awareness of how long train and airplane rides took, operating hours for various locations, and relative destination crowding depending on days of the week or time of the year. It could point you towards hidden gems off the beaten path, recommend must-see attractions based on thinly-sliced interests, and even suggest local foods filtered through dietary restrictions. It acted as both a travel guide and a logistical planner, resulting in a well-structured and reasonably predictable itinerary without the typical stress of overplanning.

What I’m describing is no different from what was already available in travel guidebooks or through travel agents, but obviously nowhere near this immediacy, breadth, or cost. The inimitable advantage with AI is again with its ability to immediately cut through the fog to quickly answer your specific question.

---

Thank you robot buddy for all the good times so far. I love you.