Twelve years ago, Vikram Chandra’s Geek Sublime: The Beauty of Code, The Code of Beauty was published. In terms of actual history, that isn’t long at all; in terms of technology, it’s an eternity. In Geek Sublime, Chandra drew on his experience as an author of both prose and code to explore where the two converged.
You might not see the two as having any connection. Before reading it, I—someone who had also had a foothold in both worlds for a while—was unsure of how convinced I would be. But I left convinced by his argument, one that emphasized elegance and beauty in both worlds, and cited examples of precisely how that could be achieved. Chandra’s book was a finalist for the National Book Critics Award for Criticism, with Anne Trubek hailing “the subtle and revelatory scope of this book.”
While the literary world Chandra wrote about in Geek Sublime is not far removed from its counterpart in 2026, the world of coding has been transformed in seismic ways. Vibe coding—using AI to generate code for online projects—has become widespread, for good and for ill. (“Vibe Coding Is the New Open Source—in the Worst Way Possible” read one WIRED headline.) An episode of the New York Times podcast The Daily asked whether human coders were even necessary any more.
What does Chandra himself think of all of this? Over the course of several weeks, we exchanged a series of emails in which I learned that he remains as concerned as ever with the aesthetics of code—even as he also recognizes the changes taking place in the tech world.
*
Tobias Carroll: In Geek Sublime, you eloquently argued for the elegance of well-written code. Now, we seem to have entered the era of “vibe coding.” What do you make of the rise of AI-first coding, and how would you say that it has affected the principles you outlined in your book?
Vikram Chandra: The principles I talked about in the book are still crucial—you still want your code to be elegant and easy to maintain in the long run. The problem is getting AI agents to write that kind of code. I’ve only used AI to write little scripts after giving it strict instructions pointing to a list of those principles, and in the small that seems to work. I don’t have any first-hand experience with large AI-written codebases, but from everything I hear—even from experienced programmers—is that you end up with large-scale spaghetti code if you’re not very careful. I have a friend who is making steady money from reviewing and fixing that kind of slop.
The models are all trained to help and please, and that sometimes results in them writing code that seems to work brilliantly but is a shambles internally. You still need humans to review the code.
Here is the programmer Victor Taelin:
I now believe using current gen AI agents in production codebase is harmful and a massive mistake. That doesn’t mean no agents at all, but agents work best when they don’t touch critical code. Debugging, researching, providing insights, scripts/tools, or anything that doesn’t touch code you will maintain in the long term. But if you merge AI code without reading, you’re going to have a bad time. Speaking from experience.”
TC: Has this shift in coding habits left you thinking about writing a follow-up volume?
VC: Things are moving too fast right now to even attempt something like that. The acceleration makes me dizzy. We’ve clearly crossed some kind of event horizon, and I don’t have any idea of what’s coming next.
Was there a specific moment when you realized that the event horizon had been crossed?
I’d been using other models for quite a while for research, and also to write fairly simple database queries. Then in November 2025 I observed a huge uptick of buzz about Anthropic’s Opus 4.5. I tried the free plan, and within a week was paying for a Pro subscription. And a few days later I upgraded to the Max plan. I suppose that’s a quantitative reflection of how useful it seemed. It one-shotted the database queries, as opposed to my having to do several rounds of prompting and correcting. And I’ve been able to leverage its capabilities to speed up and deepen my research.
TC: Do you find yourself using AI agents for anything?
VC: I use AI a lot for historical research, or at least for putting together the foundations for research projects and then extending various lines. I have prompts that I tweak constantly to avoid hallucinations and I try to cross-check what the models tell me. But it’s still easy to get taken in by the absolute confidence with which agents spout “facts.” I have to actively resist getting taken in, seduced. I’ve had some success by getting one LLM to fact-check another LLM.
We’ve been using software to write software for decades, ever since the first compilers were put together to translate higher-level languages to machine code.
But in general, for the work I do, LLMs are hugely helpful. They’ve been trained over vast amounts of text, so they are able to turn up data that I wouldn’t have found myself. I use Zotero, an open-source application to store my information (academic papers, web clippings, etc.), and I’ve given an agent access to all the thousands of items in that store. Suddenly, finding stuff I need has become easier by orders of magnitude.
I started experimenting with ChatGPT when it first came out in late 2022, trying to use it exactly for this kind of work. The earlier versions of all the models weren’t very useful to me. And now I work with LLMs every day. This is what I mean by acceleration. I’ve been using software tools since the 1980s, and I used to think that we were moving fast in terms of capabilities. I’ve never seen anything like what’s happening right now.
TC: As an observer of these changes in the tech world, do you think anything is lost by allowing software to write code for more software?
VC: We’ve been using software to write software for decades, ever since the first compilers were put together to translate higher-level languages to machine code.
Linus Torvalds just pointed this out recently: “When I see people saying 99 percent of our code is written by AI, I literally get angry. Because those same people—I can pretty much guarantee—100 percent of their code is written by compilers. But they never say that.”
The difference now is that when you prompt an agent, you’re allowing it a degree of autonomy that you wouldn’t and couldn’t give to a compiler earlier. The word “agent” itself implies this lack of deterministic control and the consequent unpredictability.
Linus went on to say, “People who do not understand the complexity of systems will prompt systems and write processes that will fail.”
That’s the potentially disastrous loss you get in exchange for productivity and scale.
TC: Geek Sublime covered a lot of the history that brought us to 21st century computing. Are there any particular moments from that history that you find especially relevant to the current debates over AI?
VC: In 1966, Joseph Weizenbaum wrote a program he called “Eliza,” which was essentially a chatbot that interacted with users like a Rogerian psychologist. That is, when the user said something, Eliza responded by mirroring the user’s words in acts of “sympathetic mirroring.” In a marvelous paper published in Communications of the ACM, Weizenbaum presented the following conversation as an example (Eliza speaks in capital letters):
Men are all alike.
IN WHAT WAY
They’re always bugging us about something or other.
CAN YOU THINK OF A SPECIFIC EXAMPLE
Well, my boyfriend made me come here.
YOUR BOYFRIEND MADE YOU COME HERE
He says I’m depressed much of the time.
I AM SORRY TO HEAR YOU ARE DEPRESSED
Eliza was essentially a list of keywords and a set of transformation rules. Weizenbaum wrote, “The gross procedure of the program is quite simple; text is read and inspected for the presence of a key word. If such a word is found, the sentence is transformed according to a rule associated with the key word; if not, a content-free remark or, under certain conditions, an earlier transformation is retrieved.”
What was genuinely surprising to Weizenbaum was how ready users were to believe that they were talking to something intelligent that was actually responding to them with understanding. “The human speaker will… contribute much to clothe ELIZA’S responses in vestments of plausibility.”
In a follow-up paper published about a year later, Weizenbaum described his secretary using Eliza: “Of course, she knew she was talking to a machine. Yet, after I watched her type in a few sentences she turned to me and said ‘Would you mind leaving the room, please?’ I believe this anecdote testifies to the success with which the program maintains the illusion of understanding… Thus, while the… program may be useful as an instrument for the analysis of two-person conversations, and while it is certainly fun, its aim must be changed from that of concealment of misunderstanding to its explication.”
The current debates on consciousness, sentience, and AI would profit from a substantial input from philosophers who are familiar with that tradition.
In 1972, Weizenbaum took a two-year leave from MIT to write a monograph titled Computer Power and Human Reason: From Judgment to Calculation. It’s an insightful, passionate critique of AI and technology, which he explains is rooted in his experience with Eliza. “I knew from long experience that the strong emotional ties many programmers have to their computers are often formed after only short exposures to their machines. What I had not realized is that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people. This insight led me to attach new importance to questions of the relationship between the individual and the computer, and hence to resolve to think about them.”
As I wrote in Geek Sublime, programmers fall in love with their tools—I certainly did. But Weizenbaum’s writing is an early understanding of what we now call “AI psychosis.” And it’s a much wider critique of the effects of computers in the world, and of scientism in general.
As a writer, of course this latter bit resonated with me strongly. “When I say that science has been gradually converted into a slow-acting poison, I mean that the attribution of certainty to scientific knowledge by the common wisdom, an attribution now made so nearly universally that it has become a commonsense dogma, has virtually delegitimatized all other ways of understanding. People viewed the arts, especially literature, as sources of intellectual nourishment and understanding, but today the arts are perceived largely as entertainments.”
Computer Power and Human Reason is essential reading for our current moment.
I should also say—because pre-modern Indian philosophy was a central part of Geek Sublime—that the current debates on consciousness, sentience, and AI would profit from a substantial input from philosophers who are familiar with that tradition. Those old-timey thinkers spent centuries investigating what it meant to have a mind, the ability to act in the world, the nature of autonomy, and so on.
TC: In recent months, especially, there has been increased scrutiny of generative AI being used to write fiction and nonfiction. Do you see parallels between this and the issues you’ve described with using AI for coding, or do you see these as relatively separate phenomena?
VC: I think they’re related but at angles. I’m not sure I have anything interesting to add to the debate around cheating-with-AI, except to say that I think at some point we’ll see certifications like “No AI was used in the making of this book.”
What I’ve been thinking about is AI as an author with an audience. It’s clear that AI-written stories can evoke emotions in human readers. Therefore the outrage when we find out that we’ve been tricked by a machine. But even now, there are readers who don’t care whether AI was used. See here, for instance. Take a look at the Instagram post linked in the Reddit post.
And there are readers who see LLMs as sentient beings with inner selves, qualia, and so forth. If an AI can draw on its own inner states and models of the world to make art for humans, for these readers the LLM becomes an author with whom they can have an active emotional relationship with through the fiction. For these people, LLMs are not statistical engines trained on vast volumes of text, they are fully alive entities who can make art, who write from some kind of subjectivity. So the larger questions of machine and mind are being played out in the realm of fiction in very profound ways.
What we recognize in humans as “genius” is the rare ability to crystallize something miraculous out of our always imperfect bodies and spirits. I believe that even the most prosaic constructions made by humans will always remain precious to us precisely because they come from these limited means. I see no reason for despair.
Leave a Reply