The pros and cons of DennisGPT
Some more thoughts on the encroachment of AI!
There are so many benefits to writing, and I find myself appreciating my two blogs more and more each day. Writing pieces for them requires me to be creative, think critically, and creates a wonderful archive of my thought process and ideas through time.
And quite recently, a very fun side effect has emerged: the creation of DennisGPT!
Not too long ago, I learned that you can make a custom GPT through ChatGPT.
Basically, this means that you can create your own model to do something you’d like, and all you need is a sufficient amount of training data to do so. I found out about this GPT-creation thing from my friend Nathan, who’s building one for UCLA Admissions to share with college applicants.
The general idea was that if you could feed this GPT all the available information about UCLA undergraduate admissions, it could create a custom chatbot that’s best suited to answering questions from prospective and current applicants. He fed in all the data he could find on UCLA’s website about admissions and different academic departments, and I provided the tour guide script (all 37 pages!) for further refinement. This was a cool project, and something I was happy to contribute to.
But I was curious about just how well this GPT-creation thing worked, so I decided to try and create one myself. I fed in all of my blog articles (all ~80 of them between Interosity and Adventure with Dennis), and asked it to create a model that copied my writing style.
Boom! Just like that, DennisGPT is born!
And the crazy thing is, since I had a sufficiently large training corpus (translation: I’ve written lots of articles), the model was pretty good at emulating my style. We’re talking random anecdotes everywhere (albeit hallucinated), floating conjunctions all over the place, and a complete disregard for normal formatting or citation standards.
Sound familiar? That’s me!
Funnily enough, I didn’t find many uses for DennisGPT, besides copy-pasting stuff and having it rewrite as if I wrote it myself. And there wasn’t any reason to do that, except for the completely hypothetical situation of my friend sending me his CS homework and DennisGPT going to town on it before I submitted it (enough said on that topic).
But the thing is, while DennisGPT can write like me, it can’t think like me. It can’t create anything new, and is stuck with rewriting existing work. It’s not me.
It’s a shadow of me.
At the start, I shared many benefits to writing, one of which is that it forces me to think critically. And that’s because writing well is actually really fucking hard.
Writing a good article involves consolidating a vast amount of information, coming up with a thoughtful analysis, and making it interesting to your audience. Gurwinder explains just why this is so hard: “Writing requires you to filter out bad information because you have a duty to your readers to not be full of shit.”
And writing without being full of shit is really, really hard to do.
But it’s also incredibly rewarding, and there are few things like it. The feeling I get when I read through a polished article draft is best described as the feeling you get when you’ve been debugging your code for hours, and it finally works.
Like a legend! Like the best that’s ever lived!
Dare I even say… godlike?
For an example of this writing process, let’s just look at how I wrote my most recent article, “Some thoughts on the encroachment of AI”. During the course of the piece, I share a few different anecdotes, each from a different place:
An intro about AI phone calls, which I read on someone else’s Substack (I forget exactly whose)
An anecdote about AI image creation, which I read about on LinkedIn
Harry Dry’s YouTube video analysis from his wonderful marketing newsletter
Another anecdote about my writing class and my professor’s grading
One more anecdote about keeping all handwritten notes I receive
Taking those different things and making an Interosity article out of them required me to sit there and ponder how to make it flow together well, which again, is really fucking hard. As Paul Graham says, “The easy, conversational tone of good writing comes only on the eighth rewrite.”
This conversational tone is something that I’ve gotten much better at in the months since I started writing (rereading some of my earliest articles makes me cringe), and something I’m extremely proud of. And unless DennisGPT can somehow download the entirety of my brain (unlikely), it’ll never be able to write these original pieces like I do.
Jack Raines describes this best:
The problem with taking an AI-first approach to tasks is that it robs you of everything that you would have gained by doing the work yourself.
I don’t write to simply generate a 1,200 word output. I consider writing to be an extension of my curiosity, and the writing process itself is what turns a rough idea into a finished product. I begin with a vague idea based on some observation of the world, and I put that on paper. As I’m writing that idea, two distant synapses in my brain connect, bridging seemingly-unrelated ideas. Maybe an anecdote from my time playing football relates to risk-taking in financial markets. Maybe a conversation I had at the bar the previous weekend sends me in a new direction entirely. As I continue down this path, the story evolves until it hardly resembles the original idea. Writing is a metamorphosis that turns vague abstractions into novel ideas, but you have to go through the writing process to connect the various points along the way.
And that’s basically how all of my articles are written. lol
Long story short, I’ve never written any blog post with DennisGPT, and never will. Because at the end of the day, DennisGPT is no match for Dennis himself.
Best,
Dennis 1.0 (me) :)
P.S. I was telling my friend Alex about my DennisGPT creation, and he told me that he wished he had an AlexGPT equivalent. The funny thing is, I had recommended that he start writing when he studied abroad in Barcelona, but he “didn’t have the time” and wished that ChatGPT could write the blogs for him. That means that unfortunately, he doesn’t have any training data to create an AlexGPT with, since you can’t train a model of your writing style with writing that ChatGPT wrote itself. Lol 🤦
Big yay for the value of writing, even (especially?) imperfectly! Writing is thinking! Sometimes, when I write, I spend so much time wondering if I'm being too dramatic or selfish or including too much bullshit my readers don't care about. But at the end of the day, our readers read because of the human at the keyboard, version 1.0 :)