In fifth grade I would sometimes send emails to my math teacher. Weird, yes, and also true. Things get pretty vulnerable here on the blog.
At one point, shortly after discovering copy-and-paste, I decided to try sending the world's longest email. My strategy started with writing a sentence, "this is the world's longest email." I would copy this and paste it a few times:
"This is the world's longest email. This is the world's longest email. This is the world's longest email. This is the world's longest email. This is the world's longest email."
Then, once I had a paragraph, I would copy-and-paste that paragraph to make 10 or so paragraphs. Then I would keep snowballing: copy and paste the 10 paragraphs a few times, then copy and paste that larger chunk.
I know, I know, pretty advanced stuff.
After one of my attempts to send the world's longest email, my teacher called me over and pointed out that there were a few stray 'v's throughout the message. I had initially not confessed to using the copy-and-paste mechanism, but this was a smoking gun. A fingerprint, a signature of the tool I used to make the email. She caught me.
You may have realized by now that, unfortunately, this is another blog post about AI. Sorry! You don't have to read - the writing on this blog is mostly to help me think. I have only a slight illusion of an audience, and I am beginning to enjoy that more and more -oblivion is freedom.
I found this recent soundbite from Hank Green quite fascinating:
"I need to come to terms with the fact that the level of dopamine I’ve been getting from interacting with LLMs... with doing more and more and more and more... is not healthy for me or good for the world" (via techcrunch).
I find the quote fascinating on two levels.
First, the dopamine thing. I am growing increasingly weary of this term, because I have noticed it being used as a scapegoat. Hank is doing the opposite here, and trying to take responsibility, but it's still a bit of a material outsourcing.
Like, I doubt that Hank or his doctor actually measured the amount of dopamine in his bloodstream over the last few years, and used this as the basis to determine the effects of his LLM interactions. The same is true for my students and friends who reference "dopamine addiction" - I am not aware of any way to gain a live view of one's dopamine levels, as if it were as measurable as insulin levels or heart rate or blood pressure.
I would imagine that psychologists feel about the word "dopamine" a similar way that computer scientists feel about the word "algorithm." It used to mean something very precise, and is now used haphazardly. We can generally recognize what people are referring to when they use the word, but it's not the same referent that the experts intended when the word first leaked into the general social lexicon.
I think what Hank is getting at, and what my students and friends are getting at when they mention dopamine, is less about a chemical and more about a hollow sense of satisfaction. A fickleness or a fakeness in the experience. Something feels fraudulent and deceitful. There is a shallow reward.
So that's what is interesting to me about the dopamine thing: people seem to use the word "dopamine" to reference an experience, not a chemical.
And then the second interesting aspect is that Hank is connecting this experience to LLM interactions.
I started considering this "addictive" phenomenon of the LLM experience earlier this year, when I decided to move away from hosting my website on ghost, and enlisted LLM-based coding tools to develop parts of the new site (I'm not good at writing javascript, and I don't necessarily want to or plan to ever be good at writing javascript, and LLMs can do it pretty well).
As I experimented with Claude Code, Hermes, Codex, Copilot, and a few other coding tools, I started to get kind of sucked in. The website project went well, and my website now runs on simple html that I can edit and modify quite easily without relying on ghost. The coding agents also helped me patch some security issues, and my website now gets "extra credit" in the Mozilla HTTP observatory.
After finishing up my website migration, I cranked out a few other tools I had wanted to make for a while. I have multiple degrees in computer science, but at some point you always end up reading documentation, searching error codes, browsing stack overflow, etc., and I did not have the time to do that for each of these projects.
For example, I have a consistent need for timers in class, where I can tell students "you have 10 minutes to discuss these question," and put something useful on the board to track the time. I've seen people embed YouTube videos of timers for this purpose, and it irks me. No, it infuriates me!
There are a few decent timers online, but nothing that had all the features I wanted. I didn't like the second-by-second (or even millisecond) tracking, the sounds, the ads, etc.
With the help of an LLM coding tool, I was able to put together a delightful little made-to-order timer widget for class, with all the features I needed, and no advertisements. It may seem like a small detail, but to me, it meaningfully improves the class experience. Plus, since I themed it after the CTA, it's much more fun to use/watch compared to the other options.
(Please use it, jackbandy.com/timer, if you are ever tempted to do something embarrassing like streaming a YouTube video just to keep track of time.)
So, that's the good/fun part. I was able to make a custom tool, one that I had wanted for a long time, but which I otherwise just did not have the time to make.
The bad part was, I think, what Hank called "not healthy for me or good for the world." There is something weirdly addictive about being able to crank out software (at least for people like me and Hank and this other guy I saw on TikTok who made a video about "saying bye to my wife before I go vibe-code for 4hrs").
The truth is, I don't really know how to describe how this interaction plays out. But I have done plenty of research about addiction and compulsion, as it was my initial critique of social media apps. At this point, the DSM has eleven criteria for addiction:
I've lightly edited this list from the source above, replacing "substance" with "the thing." For some reason, I still remember in one of my undergraduate courses, someone pointed out that fixating on substance addiction did not make sene, because people also get addicted to behaviors (like putting on chapstick or plucking hairs from their face). They were right!
I am not here to finger-wag about people getting addicted to LLMs. The point is that some of the addiction criteria do seem relevant to whatever it is about LLMs that are "not healthy for me or good for the world." One could reasonably argue that Hank's case, using the tool to help with his professional responsibilities, was a case of "using the thing in risky settings."
Personally, I have experienced criteria 1 at times, firing up LLM tools in the terminal more frequently than I anticipated. I never thought I would reach a "usage limit" - they literally cut me off!
I can honestly say that the other criteria do not resonate with me at this point. But they do worry me (again, things get pretty vulnerable here on the blog)! I do not have "intense cravings," for example, but sometimes I feel bored and think, "maybe I should fire up the old LLM and crank out another widget?"
The same goes for "risky settings." This summer I am preparing materials to teach Intro to Data Science, a class I have not taught before, and I use LLM to help with some of the materials. I am using quarto for slides, to avoid Google, and sometimes I prompt an LLM tool to draft a particular format for a slide.
Although I think my use cases are reasonable and defensible, this still seems a bit risky! Ideally, using an LLM tool for some tedious formatting will "free me up" to think more carefully about content, pedagogy, etc. But if LLMs, are in some ways, to some extent, a little bit, sort of like drugs, then I should be careful. We should all be careful!
If Hank Green can get hooked, any of us can get hooked.
My current approach is evolving and multi-faceted. I am trying to be transparent about when, where, and why I use LLM tools. I try to keep unvetted LLM-modified / LLM-generated materials in a "sandbox" and only move them out once I have reviewed and tested them. I keep everything in a public GitHub repository, so that anyone can see the various human decisions that went into the making the materials (and/or propose their own changes).
The Hank Green "situation" has reminded me to stay vigilant on this front. I have definitely had days where it feels like "I" am producing more and more and more... copy, paste paste paste paste. Copy. Paste paste paste paste paste paste. Copy. V paste paste...
I was fortunate enough to have a teacher in fifth grade who was paying attention to students and recognized my dishonesty. It was not even for a school assignment, of course, but maybe that makes my dishonesty all the more condemnable.
Did I violate all that's ever sacred between people? My teacher had been generous enough to exchange emails. A piece of "electronic mail" should have been an offering of peace - a message from one soul to another.
By automating part of my message (command-v, command-v, command-c), I made it fraudulent. I was not sincere. The whole purpose of the exchange was to have a real human sender and a real human recipient, and my message broke that purpose.
Email is long gone as a useful means of communication.
But what about code?
Personally, I never thought I would write important messages in code, in programming languages -- certainly not in javascript. I have had seasons of work where I felt proficient in various facets coding, or specific programming languages, and that can certainly be satisfying. But the point was always to make something, to create something - to express something?
I know people who deeply enjoy programming, who find beauty and joy in it. If I can be vulnerable... yes, I work in a computer science department, and no, I do not consistently find beauty and joy in computing. Maybe in data, but not in computing.
For me, the computing/automation was always a means to an end. Do you know what I mean? And that is why I hopefully will not feel the pressure to simply create a larger volume of work. I actually want to keep my volume rather low, to stay focused, to move slowly and fix things.
Now, I will save this markdown file, and run it through a script to convert it to html, and then commit the changes in git, then push those changes to GitHub, where the html will be copied around their servers, and perhaps - perhaps! - one day be accessed by you, fellow human.