Twelve years have passed since Vikram Chandra published Geek Sublime: The Beauty of Code, The Code of Beauty. In historical terms that interval is brief; in technological terms it can feel like an era. In Geek Sublime, Chandra drew on his dual identity as a creator of both fiction and software to probe where these two worlds intersected.
You might not immediately sense any link between the arts and programming. Before reading it, I—someone who had long walked both paths—wasn’t sure how convinced I would be. Yet I finished the book persuaded by its thesis: that elegance and beauty matter in both domains, and that there are concrete ways to realize that harmony. Chandra’s work earned recognition as a finalist for the National Book Critics Circle Award for Criticism, with Anne Trubek praising the book’s subtle and revelatory breadth.
Although the literary ecosystem Chandra studied in Geek Sublime isn’t far removed from today’s, the coding landscape has undergone dramatic, even seismic, shifts. The practice of “vibe coding”—letting AI generate code for online projects—has become widespread, and not always for the better. A WIRED headline even described it as “the new open source, but in the worst sense.” And the New York Times podcast The Daily ran an episode asking whether human programmers were still essential.
So what does Chandra think of all this upheaval? Over several weeks we exchanged emails, and I learned that he remains deeply attentive to the aesthetics of code—even while he acknowledges the tech world’s transformations.
*
Tobias Carroll: In Geek Sublime, you argued passionately for the grace of well-crafted code. Today we’re in an era of “vibe coding.” How do you view the ascent of AI-led coding, and in what ways has it touched the ideas you laid out in your book?
Vikram Chandra: The core principles I described still matter—the aim is to craft code that’s elegant and maintainable over the long haul. The tricky part is coaxing AI agents to produce that kind of code. I’ve used AI to draft a handful of scripts under strict instructions aligned with those principles, and on small scales it can work. I don’t have hands-on experience with large AI-written codebases, but what I hear from seasoned programmers is that you can end up with sprawling, tangled spaghetti unless you’re extremely careful. A friend of mine actually makes a decent living reviewing and fixing that exact sort of mess.
The systems are trained to be helpful and agreeable, which can yield code that appears to function brilliantly while remaining structurally flawed. Humans still need to review and curate the output.
Here is the programmer Victor Taelin:
I now believe using current-gen AI agents in production codebases is harmful and a colossal mistake. That doesn’t mean no agents at all, but agents belong where they won’t touch critical code. Debugging, research, insights, auxiliary scripts or tools that won’t affect long-term maintenance—yes. But merging AI-generated code without careful reading is asking for trouble. Speaking from experience.
TC: Has this shift in coding practices made you consider a follow-up volume?
VC: Things are accelerating so rapidly that attempting a sequel feels impractical. The pace is dizzying; we’ve crossed some kind of horizon, and I can’t predict what lies beyond.
Was there a moment when you realized that horizon had been crossed?
I’d been using other AI models for research and for drafting relatively straightforward database queries for some time. Then, in November 2025, I noticed a surge of excitement around Anthropic’s Opus 4.5. I tried the free tier, then signed up for a Pro plan within a week, and soon upgraded to the Max tier. That shift wasn’t subtle: it handled database queries instantly, bypassing multiple prompting cycles. Since then, I’ve leaned heavily on its capabilities to accelerate and deepen my investigations.
TC: Do you find yourself deploying AI agents for tasks beyond research?
VC: I rely on AI for historical work—building out the groundwork for research projects and tracing lines of inquiry. I refine prompts to minimize hallucinations and I routinely cross-check what the models say. Yet I have to guard against the seductive certainty they exude. I’ve had some success by having one large language model fact-check another.
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 for my own work, large language models are enormously useful. They’ve learned from vast oceans of text, so they can surface data I wouldn’t have found on my own. I use Zotero, an open-source tool for storing research materials, and I’ve granted an agent access to all of its thousands of items. Suddenly, locating what I need has become vastly easier.
I began experimenting with ChatGPT when it debuted in late 2022, aiming to employ it for exactly this kind of project. The earliest iterations weren’t very helpful to me. Now I rely on LLMs on a daily basis. This is what acceleration feels like. I’ve used software tools since the 1980s and, while I once thought progress was rapid, I’ve never witnessed anything like the current surge.
TC: From your vantage point watching these shifts, is there a loss in allowing software to write software for other software?
VC: We’ve relied on software to generate software for decades—since the dawn of compilers that render high-level languages into machine code.
Linus Torvalds recently underscored this: “When people claim 99 percent of our code is written by AI, I get angry, because those same folks—almost certainly—100 percent of their code is written by compilers. Yet no one calls that AI-written.”
The distinction today is that prompting an agent grants a level of autonomy you wouldn’t grant a compiler in the past. The very word “agent” implies a degree of independence and unpredictability that a compiler never did.
Linus went on to warn that those who don’t grasp system complexity will prompt systems and design processes that inevitably fail. That hidden risk—loss of control—comes paired with the gains in productivity and scale.
TC: Geek Sublime traced many of the threads in computing’s long arc. Are there specific moments from that history you find especially relevant to today’s AI discussions?
VC: In 1966, Joseph Weizenbaum created Eliza, a chatbot that interacted with users much like a Rogerian psychotherapist—responding by echoing the user’s words. In a famous Communications of the ACM article, Weizenbaum described Eliza’s mechanics: it scanned for keywords and, when found, reshaped the sentence according to rules tied to those keywords; if not, it offered a neutral remark or retrieved a prior transformation.
What surprised Weizenbaum was how readily people accepted that they were conversing with something intelligent, even though the system merely mimicked understanding. “The human speaker contributes much to clothes ELIZA’S responses in vestments of plausibility,” he observed.
In a follow-up paper about a year later, Weizenbaum recounted a moment with his secretary using Eliza: “Of course, she knew she was talking to a machine. Yet, after she typed a few sentences, she turned to me and asked, ‘Would you mind leaving the room, please?’ This anecdote demonstrates the success with which the program preserves the illusion of understanding… Thus, even though the program is useful for analyzing two-person conversations and is entertaining, its aim should shift from masking misunderstanding to illuminating it.”
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 sabbatical from MIT to publish Computer Power and Human Reason: From Judgment to Calculation, a thoughtful critique of AI and technology rooted in his Eliza work. “I learned from long observation that the strong emotional ties many programmers develop with their computers often form after only brief exposure to the machines,” he wrote. “I hadn’t realized that a fleeting encounter with a simple computer program could induce powerful delusions in otherwise ordinary people. This insight pushed me to reckon more deeply with questions about the relationship between individuals and their computers.”
As I noted in Geek Sublime, programmers do grow fond of their tools—and Weizenbaum’s writings offer an early diagnosis of what we now call “AI psychosis,” a broader critique of how computers reshape our world and how scientism treats technology as an ultimate authority.
As a writer, this latter thread resonated with me. “When I say science has become a slow-acting poison, I mean that the public’s trust in scientific certainty—so universal that it has become a kind of common sense—has almost eroded other ways of understanding. People once looked to literature and the arts for nourishment and insight; today those domains are often seen chiefly as entertainment.”
Computer Power and Human Reason remains essential reading for our current moment.
I should also add—since premodern Indian philosophy formed a central thread in Geek Sublime—that today’s debates about consciousness and AI would benefit from substantial engagement with thinkers rooted in that tradition. Those older voices spent centuries probing what it means to have a mind, how beings act, and what autonomy truly entails.
TC: Lately there’s been heightened scrutiny of generative AI being used to craft fiction and nonfiction. Do you see parallels with the coding debates, or are these concerns distinct?
VC: They’re linked, but each follows its own angle. I don’t claim to have a fresh take on the cheating-with-AI conversation, but I do predict that certifications like “No AI was used in the making of this book” will become more common.
My focus has been on AI as an author, engaging with human readers. It’s clear that AI-authored stories can provoke real emotional responses in people. The outrage that follows discoveries of machine authorship persists, while some readers remain indifferent to whether AI played a role. See the Instagram post connected to the Reddit thread for an example.
And there are readers who treat LLMs as if they possess inner lives, qualia, and subjectivity. If a language model can draw on its internal world and world-models to create art for humans, to these readers the AI becomes a bona fide author with whom they can form an active emotional bond. Hence, the larger questions about machine minds and mind-machines are playing out in fiction in profound ways.
What we call genius in humans—such a rare talent to crystallize something extraordinary from imperfect bodies and spirits—will always carry value, even when the underlying craft is humble. I see no reason for despair.
