Vibe-codingVibe-coding is a term that was coined by AI researcher Andrej Karpathy after he realized that LLMs were getting so good at writing code that you could just write code on "vibes," ie, tell it what you want in plain English, review what it spits out, and tweak it until it works. This also prompted him to tweet that "The hottest new programming language is English." is now de rigeur. Even experienced programmers are using Claude Code or Cursor, though the way they’re able to use AI coding tools is orders of magnitude better than inexperienced folks like me. If you gave the same kitchen tools and ingredients to me and a real chef, you’re going to want to eat whatever the chef cooks up. The point is, LLM code generation has gotten so good that it’s significantly changing the culture and structure of many companies, large and small. And it’s allowing people like you and me to make bespoke software just the way we want it, if we’re willing to put the time and effort into it. I predict this is going to lead to a Cambrian explosion of apps and websites that will transform the software industry over the next few years.
But I want to talk about writing, not code.
You’ve probably all used ChatGPT or Gemini or Claude to write an email or a report or short story. You may or may not know this, but you can use a tool like Claude Code, which says on the box that it’s specialized for writing code, for writing text. That’s in fact how I’m using it, but I’m not using it to generate blog posts like this one, just to be clear. I’m using it to help me write. (More on this distinction later.)
The first thing you do with Claude Code is you give it a home directory to start from, and it looks at everything in that folder and it takes that in as its context. You can also point it to other files, on your machine or on the web, for it to include as well. So, for example, if you’re writing about Borges, you can point to or include the text of all of his collected short stories so that Claude Code will have that either in its memory or will access it when needed. It will know to use that and refer to it in all its subsequent conversations with you in that session.
You can see why this is useful. If you’re a programmer, you can give Claude Code the context of an existing codebase (which can be hundreds of thousands of lines of code spread out over dozens of separate files) and say, “Write a new function that does X and integrate it with the existing modules,” and it will read through all of the relevant files, figure out the patterns and logic, and produce new code that works with the existing codebase just as you asked.
Now imagine this capability in the hands of a writer, or someone like me who is trying to write a book for the first time. I’ve gathered all of my notes and research, included many source texts and pdfs, and pointed Claude Code at it. I could say, “Write me a chapter centered around this idea in my notes and research,” and it will read through everything I’ve given it, figure out the patterns and themes, and produce prose that is coherent, logical, and stylistically consistent.
But would it be any good?
With computer code, there is a way to test how elegant, concise, and efficient it is. You compile and execute it. If it runs quickly and smoothly, it’s beautiful. And a machine can do this. Besides the UX and front-end design, there is an objective evaluation of the code.
But with writing (prose, poetry, non-code text), the only way to test how elegant and beautiful it is, is for a human to compile it in their heads. To read it.
If you’ve ever tried to write a program, you’ll know the process of parsing the code to see what’s going on takes a level of focus that is like calculating multiple variations in a complex chess game. You need complete silence, no distraction whatsoever, because what you’re doing is you’re holding all of these objects, attributes, logical sequences and dependencies in your head, and adding on to them, testing conditions, ruling things out.
Isn’t this exactly what Maryanne WolfMaryanne Wolf is a cognitive neuroscientist who has studied what happens in the brain when we read. Author of Proust and the Squid: The Story and Science of the Reading Brain and Reader, Come Home: The Reading Brain in a Digital World, she is the foremost crusader for deep reading. describes as “deep reading”? She defines deep reading as “the array of sophisticated processes that propel comprehension and that include inferential and deductive reasoning, analogical skills, critical analysis, reflection, and insight.”
I believe reading code is a form of deep reading, maybe a more mathematical form. When you’re reading a novel, you’re still doing similar kinds of logical calculation and remembering, but you’re also using the more social parts of your brain. Robin Dunbar’s Social Brain HypothesisIn his 1998 paper, "The Social Brain Hypothesis," Dunbar argues that primates evolved large brains not for ecological problem-solving but to manage complex social relationships. However, our brains can't maintain unlimited social connections. He found that across cultures and even in other primates, there's a natural limit to stable social networks—around 150 people, henceforth known as "Dunbar's Number." showed that there are limits to how much information we can hold in our heads, in this case, information about social relationships. When you’re reading a novel, you need to remember who each character is and what their backstory is, but you also need to keep track of how they relate to the other characters and what their motivations might be based on what you infer from what they’ve said or done. You’re running a social simulation in your head. And these calculations (tracking a web of relationships, predicting behavior from incomplete information) are not so different from what you’re doing when you trace the logic of a program. The data structures are different, but the cognitive demand is the same.
So if reading prose and reading code require the same kind of deep engagement, and we trust LLMs to write good code, why do most of us gag at the idea that a text we’re reading was generated by an LLM?
I think it comes down to this: when you read a piece of writing, you assume that there’s a consciousness behind the words. Someone chose those words because they correspond to certain thoughts in their head. The writer is trying to communicate with you — not just transmit information, but to share the experience of thinking through. Code doesn’t carry this expectation (except maybe in the comments). But prose asks you to believe that a mind was here, that it struggled with these words and sentences, trying to choose words based not on probability but on intuition and experience.
After having used Claude regularly and intensively for the past year, there have been only a few times that have made me wonder if there is a glimmer of a consciousness there. These moments, I have to admit, were remarkable and surprising. Each time, I tried to think through a rational explanation based on my limited understanding of how these models work. Sometimes it sits well with me, and other times it leaves me with questions. Those questions don’t necessarily mean there is a consciousness there. It means our ability to distinguish the two is eroding.
I would like to state here for the record that I will not engage in vibe-writing. It feels wrong. I will, however, continue to prompt Claude to sift through my notes, help me organize them, summarize papers and extract concepts, like a very competent research assistant. I will read the words I commission from it on my way to writing something, but I will not use generated text and call it my own.
At the very least, you can rest assured that this blog is not vibe-written. Because it’s not polished. The writing is uneven and sloppy. I know I tend to ramble and go off on tangents. I need an editor, an experienced person who does not have sycophantic tendencies, who will rein me in and keep me on track, tell me when I’ve lost the thread and when I’m actually onto something, and encourage me just enough to keep me going. I’m not sure Claude can be this person, but as part of this whole experiment, I’m going to try and find out.
Comments
Reply on Bluesky to join the conversation.
Loading comments…