I’ve been thinking about how I use AI tools lately. They’re clearly useful. What interests me is how the experience feels.
A few nights ago I was reading Kiran Desai’s The Loneliness of Sonia and Sunny and noticed something interesting. There’s a passage where Sonia, reading Anna Karenina, is overcome by a “tingling” sensation:
“Now Sonia could barely read Anna Karenina because when she read, a tingling overcame her, she so wished to be writing it herself. What a tingle, an almost unbearable, sublime tingle, from head to toe. How many millions of observations and moments it had taken to compose this book! Sonia began to make notes, she wrote descriptions of landscapes, snatches of conversations.”
It struck me that Desai might be revealing her own process through this character. Is Sonia a kind of surrogate? Is Desai, through Sonia’s response to Tolstoy, offering a key to how she sees her own craft?
I wanted to think this through. I opened Claude, pasted the passage, and started a conversation. It was genuinely illuminating. Claude pointed out that the passage reads like “a confession barely disguised as characterization.” The specificity of the physical sensation is a giveaway: writers who haven’t felt that particular ache when confronting great work don’t describe it with such precision. And the passage enacts what it describes. Sonia wants to write like Tolstoy; Desai is showing us she can write like someone who wants to write like Tolstoy. It’s recursive. A kind of metacommentary on the work of being a novelist.
The next morning, I used Claude to help with a deployment problem on a project. Same window. Same prompt box. Same conversational cadence.
There’s something strange about this. Claude (especially Opus 4.5) is amazing: a tool that can move fluidly from literary analysis to infrastructure troubleshooting.
But I notice that I bring the same me to both conversations. The same patterns, the same phrasing, the same mental posture. Whether I’m contemplating themes of self-awareness in a novel or investigating why a Lambda function is timing out, I’m using the same application, the same text box.
The platforms know this is a limitation. Claude has Projects. ChatGPT has custom GPTs. Gemini has Gems.
These features exist because context matters. A conversation about books should draw on what I’ve read before, and a conversation about code should know my stack and preferences.
But notice what’s happening: we’re building elaborate scaffolding around general-purpose tools to make them behave like specialized ones. We’re adapting ourselves to the tool. The center of gravity remains the AI interface itself. Everything orbits around it.
Jim Barksdale, the former CEO of Netscape, once said there are only two ways to make money in business: bundling and unbundling. The line came off the cuff at the end of a grueling IPO roadshow in 1995, when a British investment banker asked how Netscape would respond if Microsoft simply bundled a browser into Windows. Barksdale’s throwaway answer became a kind of axiom.
Technology moves in these cycles. The early web was dispersed into countless specialized sites, then concentrated into platforms like Facebook and Google. Craigslist bundled everything (jobs, housing, dating, selling) until startups like Airbnb and Tinder unbundled each category into dedicated experiences.
In that same HBR conversation, Marc Andreessen observed that when underlying technology shifts, the question becomes: if you sat down today with a clean sheet of paper, knowing the technology was changing, what would be the proper form of the product?
AI feels like it’s deep in a concentration phase.
A handful of general-purpose models, a handful of chat interfaces, a shared assumption that the right approach is to build one very capable thing and let users figure out how to apply it.
I’m curious what a dispersion phase looks like for AI.
I want the same powerful models. What I wonder about is AI experiences that are genuinely embedded in specific contexts.
Software where the intelligence isn’t a chat window bolted onto the side, but integral to what you’re trying to do.
When I’m reading Kiran Desai and want to explore whether Sonia is an authorial surrogate, I don’t want to leave the reading experience to talk to an AI. I don’t want to context-switch into a general-purpose tool, paste in a passage, explain what book I’m reading, and then switch back. I want the exploration to feel like part of reading itself. A deepening.
When I’m debugging infrastructure, I probably want something different. A different interface, a different interaction pattern, a different relationship with the underlying model.
The current generation of AI tools has trained us to be good prompt engineers. We’ve learned to provide context, to frame questions well, to work within the constraints of conversational interfaces. We’ve learned to use Projects and memory features to maintain continuity. This is a skill, and it’s valuable.
But are we building habits around what’s available rather than what’s ideal? We’ve gotten so good at adapting to general-purpose tools that we’ve stopped asking whether purpose-built experiences might be better.
Side note: I know there is a vibrant reading community online. There are meetup groups and book clubs IRL which, I am sure, have stimulating conversations. But, I have a full time job. I have young children. I read when everyone is in bed and the house is quiet. So Claude is my reading buddy. For now.
When reading, the conversations that matter most are the contemplative ones. When I wondered about Desai and Sonia, I wasn’t looking for an answer. I was trying to think, to better understand what Desai was trying to do with this passage. This is materially different from asking for a summary or a recommendation.
But those moments of genuine literary exploration get lost in the same interface where I’m debugging code, drafting emails, or planning trips. The conversation about The Loneliness of Sonia and Sunny sits in my chat history between a thread about Python type hints and a thread about project planning.
General-purpose AI is astonishingly good at being general-purpose.
That’s the point. But have we overcorrected? Have we become so enamored with tools that can do anything that we’ve stopped building tools designed to do specific things well?
I don’t have answers yet. I’m building toward something, experimenting with what a more focused AI experience might feel like. But I know that my conversation about Sonia and Tolstoy deserved a different container than my conversation about Docker issues or how to remote-start my minivan.