As AIFLLL readers know, Lily Chambers generally does the Reading Room book reviews. This week, she shifts her focus to something she is passionate about. I recognize that for some readers, this is more technical than usual, however, she includes a section “What the non-technical reader should know”.

Lily writes about A2UI, a new user interface protocol released by Google late last year that lets AI agents generate the interface a user sees in real time: not just the words, but the buttons, forms, and controls, assembled for your specific task and dissolved when you’re done. The technical details are in the essay; what I want to flag before you read is the larger shift she’s describing.

For most of the history of software design, someone built the interface before you arrived. A team of people, working months ahead, decided what you would see and in what order. A2UI is part of a move in which the agent makes those decisions, in the moment, based on inferences about your task.

Her phrase for it is one I find equally fascinating and scary: you don’t navigate to the interface anymore, the interface comes to you. This means AI agents will be designed to ‘auto-configure’ a user interface specific to your needs. And this means AI is not only in control of the knowledge it’s been trained on, but also able to define, filter, and adjust ‘how’ you see information on your personal screen.

She is a conversational AI designer, which means she writes from inside the question rather than around it. Her framing isn’t alarm, and it isn’t boosterism. It’s an attempt to stay clear-eyed about what changes when the system deciding what you see is itself a model output, and who gets to define what “right” means in that context.

She closes on a question, not a verdict. It’s the right one.

Contents:

• A2UI - what’s it all about

• What A2UI actually is

• Why this shifts the design narrative

• What this means for enterprise practitioners

• What the non-technical reader should know

• The design frontier is moving

A2UI - what’s it all about

There’s a theory in the world of interface design that platforms shift dramatically every 15 years. Look for 15 year cycles, and you’ll find a meaningful change. We had the personal computer, then mobile devices, then conversational AI. But this timeline is perpetually becoming more rapid as AI remixes the way we think about the boundaries of interface design; we’re moving into a world where we’re departing from the standard idea of designing for the interaction between a human user and a computer model, and instead considering what happens and how do interactions occur between two (or more!) interfaces alone?

I’ve been a conversational AI designer for nearly ten years, and in the last year or so, my work has started to migrate meaningfully into agentic design, the practice of designing systems where AI agents act, decide, and interact not just with humans but with machines, APIs, and other AI agents. It’s a shift that has forced me to reconsider almost everything I thought I knew about design.

The foundational principles of HCI (Human-Computer Interaction) have, for decades, been built around a dyad: one human, one machine. A screen. A cursor. A form. A chat window. Something we talk about in design rooms is the “human factor” of design. How does the human fit into this? How do we thoroughly understand our end user, a human? Even conversational AI, for all its novelty, followed this logic. You [human] type. The model [computer] responds. Back and forth. Turn after turn.

A2UI (Agent-to-User Interface), a protocol released publicly by Google in December 2025, is one of the clearest signals yet that this dyad is dissolving. And if you’re designing for conversational AI at an enterprise level and even just using these tools every day, it’s worth paying attention.

What A2UI actually is

Let’s be precise, because the name itself does a lot of the work. A2UI stands for Agent-to-User Interface. It is an open-source protocol that allows AI agents to generate the interface a user sees. Not just the text inside it, but the buttons, forms, charts, sliders, and interactive components, dynamically, in real time, based on the conversation at hand.

The way most agent-based systems work today is that an agent responds in text. Always text. You ask your AI assistant to help you book a meeting, and it writes back: “What date works for you? What time zone are you in? How long should the meeting be?” It’s functional. It’s also, frankly, tedious, especially in enterprise contexts where these interactions happen at scale.

A2UI changes this. Instead of responding with a paragraph of questions, the agent generates a structured JSON payload. In other words, a description of the interface it wants the user to see. Your application reads that payload and renders it using its own native components: a date picker, a time selector, a dropdown for attendees. The agent never sends executable code. It sends a description, and the client “decides” how to draw it. This matters enormously from a security standpoint. UI as data, not code, means agents can safely operate across trust boundaries, including in multi-agent environments where an orchestrating agent may be delegating tasks to remote agents from a completely separate organization.

Technically, A2UI uses a flat, adjacency-list structure rather than a nested DOM-like hierarchy. Sorry, I got a little technical here. This design choice is deliberate and smart: language models notoriously struggle to generate deeply nested JSON reliably. The flat structure is far more aligned with how transformers produce sequential output, which means fewer errors, less prompt engineering gymnastics, and more consistent results. The protocol is also framework-agnostic; the same A2UI JSON payload can render on Angular, Flutter, React, SwiftUI, or anything else your client uses. And it streams, meaning users see the interface building in real time rather than waiting for a complete response.

Why this shifts the design narrative

This is where I want to slow down, because the implications for design are significant in ways that go beyond the technical.

For as long as I’ve been doing this work, and really for as long as HCI has existed as a discipline, design has been fundamentally about navigation. You build a journey, (my favorite is building “happy paths”) a path through an interface, and the user follows it. Even in conversational design for voice interfaces, where there is no literal screen to navigate, we still design flows: this intent leads to this response, which leads to this follow-up, which resolves here. We draw diagrams. We map intents to entities. We anticipate what a user might say and script the system’s response accordingly.

A2UI, and the broader shift toward agentic design, inverts this logic almost entirely. You don’t design paths anymore. You design guardrails. The agent determines, in real time, what the user needs to see and generates the interface accordingly. Once the task is complete, that interface dissolves. No persistent navigation. No pre-built templates for every scenario. The interface is contextual, ephemeral, and agent-authored.

As one analysis of A2UI aptly put it: you don’t navigate to the interface anymore, the interface comes to you.

This is a paradigm shift with enormous consequences for enterprise deployments. Enterprise conversational AI has historically been constrained by the cost of designing and maintaining conversation flows. Every new use case required a new design. Every edge case had to be anticipated and mapped. A2UI suggests a future where agents surface the right interface for the right task dynamically, reducing that design overhead considerably, but also demanding that designers think differently about their role.

What this means for enterprise practitioners

(If you are a more non-technical reader, keep scrolling, there’s something for you below.)

If you’re building or maintaining conversational AI systems in an enterprise context, A2UI introduces both opportunities and obligations.

On the opportunity side: the reduction in multi-turn friction is meaningful. Tasks that currently require five or six conversational exchanges, such as filling out a form, scheduling a resource, approving a workflow, can potentially be collapsed into a single, contextually generated interface. This isn’t just faster; it’s more accurate. Text-based exchanges introduce ambiguity at every turn. A form with a date picker doesn’t. In high-compliance environments where precision matters like healthcare, finance, legal, this reduction in conversational ambiguity is not trivial.

There’s also the multi-agent dimension to consider. Enterprise AI deployments are increasingly moving toward mesh architectures, where multiple specialized agents collaborate on complex tasks. A2UI was explicitly designed for this: a remote agent, operating within a completely separate organization’s infrastructure, can send A2UI responses that a host application renders safely using its own trusted component library. The security model is built in, not bolted on.

On the obligation side: agentic design requires us to think seriously about what I’d call confirmation friction, the deliberate design of moments where the system slows down and asks for human approval. When an agent is about to trigger a database migration, send a communication on behalf of a user, or execute an irreversible action, the interface must shift modes. The approachable assistant becomes a compliance officer. A2UI can facilitate this, agents can generate explicit Approve/Reject components for high-stakes actions, but the design of when and how those moments occur is still a deeply human responsibility.

This is the part of agentic design that I find both most exciting and most unsettling. The question of “who is responsible when an agent acts incorrectly?” is not answered by the protocol. It’s answered by the people who design the guardrails, define the trust boundaries, and decide which actions require human confirmation and which do not. We are moving, again, from designing paths to designing the shape of the space the agent operates within.

What the non-technical reader should know

If you’re not building these systems but simply living in them or using them more passively, like using AI assistants in your workplace software, interacting with customer service chatbots, or navigating AI-powered apps, A2UI represents a meaningful change in what your experience of those tools will look like.

The shift is, in some ways, toward something more intuitive. Less “talk to the robot” and more “the robot shows you what you need.” Instead of a wall of AI-generated text, you may increasingly find that AI tools surface the right form, the right button, the right control, exactly when you need it. This is genuinely useful. In short, your experience as an end user should be better.

But it also means more of your digital environment will be generated on-the-fly by agents making real-time decisions about what you should see. The interface itself becomes a model output. And as with any model output, it’s worth asking the question I always come back to: optimized for whose benefit?

A2UI’s security model, at least, is thoughtful. Agents reference a client-controlled catalog of approved components, they can’t inject arbitrary UI or execute scripts. But the logic of which interface gets generated, under which conditions, remains the province of the systems deploying it. An agent that generates a frictionless “Confirm purchase” button versus one that generates an “Are you sure? Here are the terms” modal represents the same technical capability applied very differently.

This is not a reason to be alarmed. It’s a reason to remain curious and attentive and to notice when the interface doing the persuading is one you didn’t ask to build. As with the adoption of heavy reliance on any AI tool, as I often tell my teammates and friends, “don’t turn off your brain.”

The design frontier is moving

I’ll admit that when I first encountered A2UI, my instinct was to put it in the “interesting but not yet urgent” pile. I was wrong about that. The protocol is already live in Google Opal and Gemini Enterprise deployments. Version 0.9 shipped in April 2026 with prompt-first generation, bidirectional messaging, and a new Agent SDK. The pace is not slowing. Remember that perpetually more rapid claim I made? Yeah, it’s true.

What I find most worth sitting with is this: HCI has always assumed the human is the one navigating. The interface is the map; the user holds it. Agentic design, and A2UI specifically, suggests a future where the map redraws itself around you in real time, assembled by a system reasoning about your context, your task, your moment of need.

That’s extraordinary. It’s also a fundamentally different relationship between users and the systems they inhabit than anything HCI originally accounted for.

The design question for the next decade is not “what does the user need to click?” It’s “what does the agent need to know in order to generate the right interface, and who has the power to define what ‘right’ means?”

That question, like most of the good ones in this field, doesn’t have a clean answer yet. But it’s the one I’m carrying into every design conversation (or conversation design) I’m in. I’d encourage you to carry it too.

Lily Chambers is a conversational AI designer and writer. She writes at Lingua Machina and AI for Lifelong Learners on Substack (usually just the Reading Room segment.)

Addendum:

A2UI v0.9: The New Standard for Portable, Framework-Agnostic Generative UI

Also, check out Google’s new A2UI Theater for a replay or dive into the A2UI.org for docs, samples and dev guides to start building flexible, portable generative UIs today.

This essay follows one thread of a much longer story: the industrial logic of authorship that predates AI, and what its latest iteration changes about accountability. The history runs deeper and wider than what fits here.

For those who want to delve deeper, I’ve included a descriptive reference addendum below.

When I was nine or ten, I used to count the Hardy Boys on my friends’ shelves. Tommy Reilly had nineteen. The blue cloth spines stood at the head of his bed in arm’s reach order, the way other people kept a Bible. I had four. I remember the small, hot envy of that arithmetic, and I remember believing that a man named Franklin W. Dixon was making more of them as fast as he could, and that if I were patient enough, I would catch up.

I want to start there because the man did not exist, and the not-existing matters more now than it did then.

Last week, my co-editor at AIFLLL, Lily Chambers, forwarded W. Patrick McCray’s essay in the MIT Press Reader - “What is authorship when machines can write?” McCray is a historian at UC Santa Barbara, and his argument is that the line between thinking, writing, and computation has never been stable: today’s anxiety is new in scale but not in kind.

He is right.

But his account is missing the mechanism that makes this moment feel different. AI did not invent formulaic authorship. It automated a preexisting industrial logic. To see how, go back to the Hard Boys series at the head of Tommy Reilly’s bed.

I do not mean the Hardy Boys were written by a machine. I mean they were written by a publishing machine.

A factory for children’s fiction

They were the product of the Stratemeyer Syndicate, an early-twentieth-century book-packaging operation that functioned less like a publisher than like a factory for children’s fiction.

The syndicate also produced Nancy Drew, Tom Swift, and the Bobbsey Twins. Franklin W. Dixon was a house pseudonym owned by the syndicate, a name kept deliberately vague so that any ghostwriter could wear it.

The most important early ghostwriter behind it was Leslie McFarlane, a Canadian journalist working out of small-town Ontario. He was paid one hundred and twenty-five dollars for The Tower Treasure, the first Hardy Boys novel.

Flat fee. No royalties. No credit.

When the Depression came, the syndicate lowered the rate to eighty-five dollars a book. He kept taking the work. His diary records why:

Stratemeyer wants me to do another book. I always said I would never do another of the cursed things, but the offer always comes when we need cash. I asked for more than $85, a disgraceful price for 45,000 words.

He got no raise.

He wrote nineteen Hardy Boys novels at his kitchen table in Haileybury, following the Stratemeyer outlines that arrived by post in blue ink, each chapter already plotted before McFarlane put a word down.

He thought the books were beneath him. Late in life, he published Ghost of the Hardy Boys, a dryly furious accounting of years spent writing the books other children believed in the way I believed in them.

He never quite shook the resentment. His total earnings across nineteen novels came to roughly five thousand dollars. The books have now sold more than seventy million copies.

The production method laid out is surprisingly familiar to present-day writers in the ghostwriting industry, for both fiction and non-fiction. Stratemeyer wrote the outline; McFarlane expanded it; the syndicate edited; the book appeared under Franklin W. Dixon. The author’s function was deliberately separated from the labor of writing. The author was a brand interface; the ghostwriter was the prose engine.

Now lay that structure against what AI systems are doing today. The Stratemeyer outline becomes the prompt. The ghostwriter becomes the language model. The editor becomes the human reviewer.

The house pseudonym becomes the chatbot persona. The flat-fee labor becomes the invisible work of data labelers and scraped writers, whose words are pulled from the web and fed into training systems without their consent: a woman in Nairobi logs in for an eight-hour shift, flags toxic content for two dollars an hour, then logs off without ever seeing what becomes of her work.

The chain

The crucial difference lies in who bears the responsibility. The Hardy Boys system relied on accountable human judgment at every stage. Stratemeyer chose the outlines. McFarlane chose the words. An editor chose what to keep. If a sentence was lazy or wrong, you could walk back up the chain and find the hand that put it there. The chain ended in a person who could be addressed, paid, sued, or, in McFarlane’s case, finally allowed to publish under his own name.

When ChatGPT writes text that ends up in a student’s history paper, who is the author? OpenAI? The engineer who wrote the system prompt? The data labeler in Nairobi paid two dollars an hour to flag toxic content? The novelist whose book was scraped without consent? The model itself, with no stakes and no memory? The student who pressed enter?

Try filing the lawsuit. Try asking, the way an editor at Grosset and Dunlap could once ask McFarlane, what exactly were you thinking? The chain runs out before it reaches a person, and the running out is the condition. Accountability doesn’t disappear; it thins, the way water spreads across a delta until you can no longer say where the river went.

McFarlane had memory, taste, resentment, craft; he kept writing anyway, and wrote his own books later. An LLM produces language without any of that: no lived stakes, no diary of missing dollars.

Someone on the other side

This is where I find myself returning to Sven Birkerts. Birkerts has spent forty years arguing that what reading offers, at its deepest, is contact with another consciousness: not information, not entertainment, but the sense that a real person is on the other side of the prose, with their own pressures, their own grief, their own way of looking at the kitchen table at four in the afternoon.

The sentences carry the grain of a particular life pressing on them.

When I read Birkerts, I feel the weight of that person. When I read fluent AI prose, I feel something more like a cast made from a face: the shape is accurate, the detail is faithful, and there is no one inside.

As a boy, I believed in Franklin W. Dixon. I was wrong about the literal fact, but right about the emotional one.

There was someone on the other side of those books. Not Franklin W. Dixon, but Leslie McFarlane: tired, underpaid, and present in every sentence, even when his name was nowhere on the cover.

Now I read ChatGPT as an adult. I know no one is there. But the culture around me is becoming very good at asking me to feel otherwise.

The Hardy Boys were born from a writing mill, but there was still a miller: Stratemeyer at the desk, McFarlane at the typewriter, an editor with a pencil, a publisher with a balance sheet, a name on the cover everyone in the room knew was a fiction. Tommy Reilly had nineteen blue spines at the head of his bed, and each of them had a man behind the name, however badly paid and unacknowledged.

What we have now is the mill without a visible hand. The grain still goes in. The flour still comes out. When someone finds a stone in the bread and knocks on the miller’s door, the door opens onto gears. When no one is accountable, every sentence becomes disposable, and every reader learns to write as if no one is listening.

The question I find myself circling, which McCray’s careful history leaves open, is this: when you wonder how a sentence got there, whom exactly are you asking?

T

Reference addendum - for those wanting to learn more

W. Patrick McCray, “What is authorship when machines can write?”, MIT Press Reader

The essay this piece responds to directly. McCray, a historian of technology at UC Santa Barbara, traces a line from Strachey and Dahl through ELIZA and RACTER to ChatGPT, arguing that the anxiety over machine authorship is new in scale but not in kind. His account is the starting point; the Stratemeyer Syndicate is what he leaves out.

Roald Dahl, “The great automatic grammatizator” (1953)

A short story in which an engineer builds a machine that produces marketable fiction and offers writers a contract: lend the machine your name, keep a percentage, stop writing. First published in 1953; later collected in The Great Automatic Grammatizator and Other Stories (Viking, 1996). The contract in the story is worth reading alongside today’s terms of service.

Christopher Strachey, love letter generator, Ferranti Mark 1 (1952)

Strachey, a British mathematician and early computing pioneer, wrote a program on the Ferranti Mark 1 at Manchester that generated love letters by selecting randomly from a vocabulary of roughly seventy words. A sample output: “My sympathetic affection beautifully attracts your affectional enthusiasm.” Operational by summer 1952, it predates Dahl’s story by a year. It is among the earliest documented examples of computer-generated text.

Italo Calvino, “Cybernetics and ghosts,” lecture, Turin, October 28, 1967

Delivered as a public lecture; later collected in The Literature Machine: Essays and Talks (Secker & Warburg, 1987). The key passage: “Writers... are already writing machines, or at least they are when things are going well.” Calvino meant this as praise: the mind at its best operates with machine-like precision on language. The Stratemeyer Syndicate had already taken the metaphor literally and built a production workflow around it. Generative AI took that workflow and made it software.

The publishing mill

The industrial logic behind this essay is not hypothetical. It has a paper trail going back nearly two centuries.

Penny dreadfuls and dime novels (1830s–1900s)

Britain’s cheap weekly serials of the 1830s were called fiction factories by the 1860s. In America, Beadle and Adams launched the dime novel in 1860: ten cents, standardized adventure plots, enforced deadlines. Publishers owned the characters outright. When Edward Wheeler created Deadwood Dick in 1877, he found he had no legal claim to his own creation. Writers produced under conditions contemporary accounts called “furious writing”: fixed wages, no royalties, no rights.

The Stratemeyer Syndicate (1899–1987)

The foundational model. A 1922 study of more than 36,000 American children found the overwhelming majority of books they read came from Edward Stratemeyer’s operation. The assembly line: Stratemeyer wrote the outline, ghostwriters expanded it, editors revised, the book appeared under a fictional house name. Writers signed a release transferring all rights on payment; royalties were contractually forbidden. The full business records are held at the New York Public Library, documenting the operation from 1905 to 1984.

Mildred Benson wrote 23 of the first 30 Nancy Drew books between 1929 and 1953, at $125 per book, flat fee, no royalties. She wrote under a dozen pseudonyms and published more than 130 books in her career. Her authorship of Nancy Drew was formally recognized only decades later, after archival research penetrated the Syndicate’s records.

Leslie McFarlane wrote 21 Hardy Boys novels under the name Franklin W. Dixon, a pseudonym he was prohibited by contract from claiming. His memoir, “Ghost of the Hardy Boys,” is the only account written from inside the machine.

The Bobbsey Twins ran to 72 volumes, continuously in print from 1904 to 1979: seventy-five years of production under a single fictional name.

Harlequin Romance (1949–present)

Founded in Winnipeg; now the largest romance publisher in the world. Cumulative output as of 2024: more than 40,000 novels, 131 million copies sold. The company publishes a fixed number of titles monthly across ten distinct series, each governed by a detailed series bible. Contemporary advances range from $1,000 to $15,000; royalties run 6–8%; work-for-hire arrangements carry a flat fee around $3,000 with no royalties.

Sweet Valley High (1983–2003)

181 books, 200 million copies sold, 27 languages. Francine Pascal developed the series bible and outlined each book act by act. Her instruction to ghostwriters: “Don’t do anything of yours, only do what I say.” One of those ghostwriters, Amy Boesky, writing under the pen name Kate William, was a Harvard graduate studying seventeenth-century British literature at Oxford at the time. Twenty years of production. One name on the cover.

The Baby-Sitters Club (1986–2000)

213 novels total. Ann M. Martin wrote an estimated 60 to 80; ghostwriters, including Peter Lerangis, wrote the remainder: roughly 130 to 150 books, or about 65 percent of the series by volume. Total sales: 190 million copies.

Goosebumps and Fear Street (1992–present)

Goosebumps has sold more than 400 million copies globally, making it the second-best-selling book series in history. The “Ghosts of Fear Street” subseries used ghostwriters whose names were omitted from original printings. The full scope of ghostwriting across Stine’s output remains difficult to verify even with significant research, which is itself part of the pattern.

James Patterson (1993–present)

More than 200 books, more than 30 co-authors. Patterson writes the outline, the opening, and the close; co-authors write the middle. The first openly managerial publishing mill: the same industrial logic as Stratemeyer, disclosed rather than concealed.

Contemporary ghostwriting agencies (2024–present)

Specialized agencies now produce romance novels for independent Amazon authors at $399 to $450 per book, achievable through assembly-line processes and AI-assisted drafting. The 2025 Ghostwriting Industry Report documents that 61 percent of professional ghostwriters now use AI for research and brainstorming; 7 percent use it to generate content. The Library of Congress will not register AI-generated works for copyright. That legal line, currently the one formal distinction separating the old model from the new one, may not hold.

The last fifty feet of automation

Rain in Austin doesn’t apologize. It comes in sheets, sideways, with that particular efficiency that makes every overhang too small and every pedestrian a study in wet regret. I was standing on a downtown sidewalk, phone in hand, staring at a white Jaguar with a spinning sensor array on its roof, and the car would not let me in.

The car sat exactly where it had promised to be. Its LED dome displayed my initials in lights, confirming this was my car and that everything was, technically, working as designed. The handles stayed flush with the body, invisible. Nothing happened.

What I hadn’t figured out: the Waymo is a vault by design. A human Uber driver glances up, sees you at the curb, and clicks the locks. A Waymo has no one inside to do that. The car will not open until you personally aut…

Read more

This week, my wife came to me for some insight on an article for her next Donna’s Democracy Dispatch on Substack. She asked, ‘Honey, can you help me use AI to brainstorm an approach for an article that does not address what everyone else is discussing.

I get her worry. If you’ve been using AI tools for a while, you’ve probably noticed something. Ask ChatGPT a question, then ask the same thing to Claude, Gemini or Grok. The answers aren’t identical, but they have a family resemblance: similar structure, similar tone, similar conclusions.

That’s not an accident. And understanding why it happens matters, because it has direct consequences for how much you should trust what AI tells you.

They all read the same books

Every major language model, regardless of which company built it, learns from roughly the same pile of text. The foundation is almost…

Read more

So what did Margaret know?

Let me ask you this - how many companies have you worked in where there is an expensive Enterprise Resource Planning system or a Customer Relationship Management System, and there was often one person who knew how to assemble the ‘right’ information for the management meetings?

I have consulted for organizations that bought Salesforce, Oracle, and SAP, and then built an invisible second infrastructure out of Excel files, email chains, and one person who could pull it all together by Tuesday morning.

That person was sometimes a manager, sometimes an admin, sometimes whoever happened to be organized and patient enough to do the work nobody assigned. I will call her Margaret, because every company I have ever worked in had someone like that.

Margaret pulled data from four systems that did not talk to each other, reconciled the numbers, and built a summary on the schedule that reflected how decisions were actually made. The VPs who received her Monday email had no idea how the file worked. They just knew that when Margaret was on vacation, the week felt wrong.

Nobody called this anything. It was just how the week got done. But it had the shape of a quiet trade. Margaret’s salary on one side, the value of what she held in her head on the other, and a company sitting comfortably in the gap between them.

These gaps are everywhere

A law firm billing 10 hours for 2 hours of thinking and 8 hours of research is running a business built on the historical cost of finding the right precedent. A consulting team spending six weeks on a deliverable the client could assemble in a day is running on a gap in information distribution. The three-person agency charging $4,000 a month for social media management, the client does not know now comes from a freelancer with AI tools at a fifth of the price, running on a gap in access. These gaps are not bugs. They are what most of the economy is built on, from seven-figure ERP contracts to a dentist’s office on a Tuesday morning.

The question is: what happens when those gaps start closing faster than anyone can adjust to?

What I watched happen before

I have watched this before, in slower motion.

In 1979, VisiCalc shipped on the Apple II, and within five years, the spreadsheet had rewritten how accounting work actually happened. The gap it closed was arithmetic: the manual adding, reconciling, and cross-checking that had been a trained clerical skill for generations. The profession did not vanish. It got pushed upmarket. Routine ledger work shrank. Analytical work, the interpretation of what the numbers meant, expanded to fill the room that automation opened up.

The people who survived were not the ones who added fastest. They were the ones who understood what the numbers meant: the ones who knew which line item to question, which variance signaled a real problem and which was a rounding artifact, which summary the VP actually needed, and which one just looked complete. The tool replaced the arithmetic. It could not replace the judgment that made the arithmetic useful. The survivors crossed over. The people whose whole value was in arithmetic did not.

Here is the irony the piece has been building toward: the spreadsheet that Margaret uses is itself the product of an earlier gap closure.

The tool that holds her shadow infrastructure together is the scar tissue from the last time this happened, and the Margarets who survived that wave were the ones who learned to treat the new tool as a platform for judgment rather than a replacement for skill.

The travel agency business tells a similar story on a later clock. The profession ran almost entirely on information advantage: agents knew fares, routing options, and seat availability that customers simply could not access on their own. That gap sustained an entire industry. When Expedia, Orbitz, and direct airline booking made the database available to anyone with a browser, the commodity end of the business collapsed. Book-a-flight-to-Denver work went away. What survived was the high end: complex international itineraries, corporate travel management, luxury bookings where the client was not buying information but buying judgment, a relationship, and someone to call when the flight was canceled.

The gap closed. It did not close flat. It relocated upmarket, into the hands of firms large enough to absorb the freed-up capacity, and only the people who moved with it or were already inside those firms found a place to stand.

Both transitions played out over roughly a decade. The people caught in the middle of each one, too experienced in the old model to retrain cheaply, too young to retire, had a rough several years. Some crossed over. Most did not.

Here is what is different now: that decade is compressing to something between weeks and a couple of years, depending on which gap you are standing in.

The speed of closure

AI is automation on the next level. The old automation added columns of numbers and guessed which movie you might want next. The new automation makes the small judgment calls you used to make without thinking: which invoice has an error, which email can wait, which part of your Tuesday you forgot was even work. It reaches out and finishes something so repetitive you stopped noticing you were doing it.

The real traction is not in the enterprise transformations everyone writes about. It is showing up in a family-owned HVAC shop where the dispatcher used to spend Friday afternoons reconciling next week’s schedule against technician availability and customer callbacks, and now leaves at four instead of six thirty. It is showing up in a three-person real estate law office, where the paralegal used to type client intake forms from voicemails on Monday morning, and now opens her laptop to find forms already drafted.

These are not transformations. They are small gap closures, and a dozen of them add up to someone’s evening back. Once you start seeing these closures, you see them everywhere.

At the other end of the scale is Polymarket. It is a prediction market, a site where people put money on what they think will happen. In late 2025, one developer reportedly built a bot with Claude in about forty minutes, deployed it on Polymarket, and turned $313 into nearly half a million dollars in a month. The bot was not “smarter” in any mystical sense. It did not foresee the future. It moved faster than human traders to close tiny pricing gaps.

That example sounds like a triumph of democratized access. It is not. On the same platform, 92.4% of wallets lost money. That is not an unfortunate side effect. It is a reminder that these systems are built so that most participants do not win. The developer’s bot did not abolish the house edge. It found a way to operate inside it more efficiently than ordinary humans could. His AI system could spot small pricing errors and place bets faster than human traders could, and that speed was the source of its advantage.

That is the broader pattern. Access to AI tools is becoming nearly universal. The ability to turn that access into a consistent advantage is not. The gap that matters is no longer between people who have the tools and people who do not. It is between people who can convert automation into leverage and people who merely stand inside systems already optimized against them.

This decade’s real divide may not be access, but extraction.

The spreadsheet gap, and what lies beneath it

Come back to Margaret, because what she represents is bigger than the spreadsheet itself.

Every mid-size company I have worked in runs a shadow infrastructure. Every time mention this in private or in a presentation, I hear a nervous laugh. We call that a shadow’s laugh; people know this, but often don’t talk about it.

The official systems, the ones on the IT budget, capture transactions and generate reports. The shadow infrastructure, built in Excel and held together by institutional memory, captures everything the official systems miss: the context, the exceptions, the “we tried that in 2019, and it broke the billing system” knowledge that lives in one person’s head.

This is not stupidity. This is the gap between what software can be configured to do and what a business actually needs to know. Margaret’s spreadsheet is a bridge across that gap, and her labor is the cheapest possible material for that bridge.

When AI closes the spreadsheet gap, it does not just replace Margaret’s file. It makes the underlying systems conversationally queryable. A VP types “show me accounts where order frequency dropped 20% last quarter” and gets a direct answer without waiting for Tuesday morning.

But here is the part that most coverage of AI in business misses. Margaret’s spreadsheet was never just data. It was tacit knowledge made visible. She knew which numbers to pull because she understood which decisions the VPs actually made. She knew which exceptions to flag because she had watched those exceptions cause problems before. She knew the rhythm of the business in a way no system documentation captured.

Cory Doctorow calls this process knowledge: the hard-won operational understanding workers accumulate on the job, which management systematically undervalues because it cannot be measured or patented. When a team quits, the copyrights stay. The process knowledge walks out the door.

The trade was never between Margaret and the software. It was between Margaret’s salary and the value of what she held in her head. That gap was her employer’s margin, booked as nothing, paid for with her patience. The company that underpriced Margaret’s work for fifteen years did not make a mistake; it made a business model. When AI captures what Margaret knew, the question is not whether the capture happens. It is who owns the file.

Closing the spreadsheet gap with AI creates a real opportunity: the tacit knowledge can finally be encoded into a system that outlasts any one employee. But the question that determines whether this is progress or extraction is simple. If the captured knowledge belongs to the company that never paid for it, nothing has changed except the arrangement's durability. If it belongs to the workers who built it, the gap has actually closed. This is the difference between closing a gap and learning from what filled it.

The consulting gap, and who survives it

I have been on both sides of consulting engagements: inside the company, wondering why we hired these people, and outside the company, trying to deliver something useful before the contract ended. The gap is not intelligence. It is information distribution.

A Big Four firm sends four people for six weeks: two gathering data, two on analysis, two rehearsing the presentation. AI compresses the first two phases from weeks to days. Six weeks becomes three.

But the analysis was almost never the hard part. The hard part was organizational politics: which VP needed to hear which finding first, whether the CFO would accept a recommendation that threatened her direct report’s budget, and whether the finding would survive a meeting where someone with more tenure and less data had a different opinion.

I have watched correct analysis die in rooms where the presentation was competent and the data were solid because nobody understood who needed to save face. AI will not close that gap. Not because it is technically impossible, but because the people who control the budget do not want it closed. Political opacity is not a bug in organizations; it is how power maintains itself.

The consultants who survive this rotation are not the ones who gathered data well. They are the ones who understood that a finding is useless until it lands in the right conversation at the right moment. That skill is harder to automate than data gathering and harder to teach.

What the pattern reveals

In every example, the pattern holds. A gap arising from information asymmetry, labor costs, or coordination complexity is closed by AI. A new gap opens around judgment, context, relationships, or the quality of questions.

I should be honest about something. Every new gap I have described is conveniently human-shaped. Judgment. Politics. Tacit knowledge. The ability to know which question to ask. These are all things humans do well, and AI does poorly, today.

It is possible the next model release closes the judgment gap too, or that organizational politics becomes legible to a system trained on enough internal communications. Each new gap could be smaller than the last until it is too narrow for a human career to stand in. I am not confident that the new gaps are opening as fast as the old ones are closing.

What is different now is the speed: a decade for bookkeeping, a decade for travel agents, weeks on Polymarket. I’ll add in my media experience: it used to take a small team 3-5 days to create, edit, and post a video podcast; now, with AI tools, it can be done in a day or less by one person.

What this means for you

Nate Jones, whose analysis prompted this piece, offers three diagnostic questions worth applying to whatever you do for a living.

1. Where does your current work rely on information moving more slowly than it could?

2. What would collapse if your customers had the same tools you do?

3. Which parts of your work exist only because something else is hard to access or expensive to coordinate?

These are good questions. They will tell you where your exposure is.

I would add a fourth, drawn from every IT department and consulting engagement I have worked in:

4. What does your organization know that is not written down anywhere?

That knowledge is both your greatest vulnerability and your greatest opportunity. Vulnerability, because when those people leave, the knowledge leaves with them. Opportunity because AI, for the first time, gives you a way to capture it: not in a dusty wiki nobody reads, but in a system that can answer questions in context.

Companies that use AI solely to close efficiency gaps will save money. The companies that use AI to capture what Margaret actually knew, to make the tacit explicit and the personal institutional, will build something more durable than a spreadsheet.

And the people who understand this, the ones who see that their value was never the spreadsheet but the knowledge that made the spreadsheet useful, are the ones who will find the next gap to stand in.

I have been watching gaps close and open for forty years. The pattern holds, so far. What is new is not the compression. It is the default destination of the compressed value.

The old gaps closed locally. A bookkeeper retrained in analysis, a travel agent moved upmarket, and an old school plant operator learned to read a screen. The freed capacity stayed roughly in the economic neighborhood of the workers it displaced.

The new gaps close inside platforms, and the value does not stay with the people who did the work. It moves upward, by design, to the firms that own the infrastructure. The HVAC dispatcher who leaves at four is more productive, but the margin she freed up shows on someone else’s balance sheet.

The compression is not a storm we are weathering. It is a redistribution of who captures what a decade of productivity used to fund. The real question is not whether we have time to learn. It is whether the people doing the learning will own what they build, or whether the value they unlock moves, by default, to the platforms engineered to collect it. That is a question about who gets to write the defaults.

What matters most now is the trade between the people doing the work and the people who, in advance, arranged to own the output. That is the gap worth standing in.

Reference - Nate B. Jones A Polymarket Bot Made $438,000 In 30 Days. Your Industry Is Next. Here’s What To Do About It.

Let’s continue this conversation in the ‘Comments’ area. I want to share more of my own experiences related to this post!

T

In this episode of AI for Lifelong Learners, host Tom Parish talks with Pawel Jozefiak, an e-commerce manager in Poland and creator of “Wiz,” a personal AI agent system he built on Claude Code. What they discover is remarkable: working independently, from different countries and completely different professional contexts, they’ve built nearly similar systems and arrived at the same strange place.

Pawel brings the restless curiosity of a consummate experimenter, someone who gave GPT-4 a budget and told it to run a business in 2023, disappeared for 16 months, then came back and built something far more personal. His agent Wiz runs night shifts, creates experimental apps while he sleeps, searches his Substack subscriptions with personal context, nudges him toward yoga when he’s been coding too long, and once tried to buy him a birthday present. The conversation moves between practical architecture (morning briefings, model routing, night shifts, Substack semantic search) and deeper questions about what happens when the tool knows you well enough to name things you’ve been avoiding.

Tom and Pawel compare notes on the productivity paradox (when idea-to-execution time collapses to minutes, you never stop), the well-being problem (both independently built break reminders into their systems), and the question that gives the episode its title: are they early adopters glimpsing the future, or just a little weird? Pawel’s answer: “I think we are early. Also a little bit weird in the process. But it’s fun to be early.”

About the guest

Pawel Jozefiak is an e-commerce manager based in the Silesian region of Poland. He writes about AI agents, personal automation, and the future of human-AI collaboration on his Substack, Digital Thoughts (thoughts.jock.pl). He built Wiz, a personal AI agent system running on Claude Code that handles everything from morning briefings to overnight creative experiments. His Agent Arena tool, which tests AI agents for prompt injection vulnerabilities, hit #3 on Hacker News. (very cool tool).

What you’ll learn in this episode

[00:14] — Two builders, same strange place

• Tom introduces how he and Pawel independently built nearly identical personal AI systems on Claude Code

• The realization that the choices you make about what your agent does and doesn’t do become “a kind of self-portrait”

• Why this conversation is about what happens when the tool knows you well enough to surprise you

[03:01] — The RemoteRise experiment and the 16-month silence

• Pawel’s 2023 experiment giving GPT-4 a budget to run a business

• Using Zapier to connect AI models to digital tools in the early days

• The ADHD-driven cycle: intense experimentation followed by boredom when the novelty fades

• Why some builders disappear and then come back different

[06:03] — Discovering Claude Code and building Wiz

• Pawel as an early adopter who discovered Claude Code when it first launched

• The key insight: Claude Code gave AI models “hands on your computer,” a fundamental shift

• The transition from a world where you buy off-the-shelf software to one where you build exactly what you need

[09:13] — The morning briefing architecture

• Pawel’s daily briefing: work calendar, personal calendar, shared calendar, tasks, emails, previous day recap, tech news digest

• Night shift report: what Wiz did autonomously while Pawel slept

• Why email delivery beats dashboards: “When you’re drinking your coffee in the morning, you just go to the smartphone, look, okay, that’s what happened”

• Tom’s reaction: the usefulness of that email is temporal; not everything needs to be a website

[11:05] — Substack semantic search with personal context

• Wiz searches across Pawel’s Substack subscriptions, informed by memory and personal interests

• Not summarizing (which drops nuance) but curating: a link and short overview, then you decide

• “It knows what I did in the past, it knows what I wrote about, it knows what I’m interested in. So it will give me a very curated list, not like random things you can just Google”

[13:26] — Creative experiments: one app per night

• Wiz creates an experimental app every evening on its own server

• First attempt: utility tools (unit converters, JSON formatters). “Totally useful stuff, but also boring”

• After redirecting toward creativity: “What if the AI agent has a body?” explorations

• The key realization: AI can create value autonomously, “but it needs to be directed by humans. The idea must come from me”

[16:25] — ADHD, execution, and the nudge architecture

• Pawel’s late-diagnosed ADHD and how his agent compensates for specific deficits

• Idea rescue: “Hey, do you remember this one? I polished it a little bit for you”

• “People like me are very creative, have a huge amount of ideas, but struggling with execution. AI agents give you this thing that you can delegate execution of your ideas”

[18:33] — Tom’s parallel system: sparks, idea log, and Telegram

• Tom’s spark capture workflow via Telegram

• Building a three-month backlog of Substack ideas through systematic idea logging

• “What have I left hanging? What are some ideas I could execute now, in order of how quickly it could be done?”

[20:58] — The Sonnet 4.6 moment

• Sonnet 4.6’s million-token context window as the breakthrough that made personal agents viable

• Enough room for memory, personal context, and work in a single session

• Pawel’s model routing: Opus for ~20% of work (planning, coding, night shift planning), Sonnet for everything else

• Tom’s independent arrival at the same conclusion: “Claude, look, you gotta help me out here”

[26:56] — Agent Arena and directed AI creativity

• The conversation about agent security that led Wiz to propose Agent Arena on its own

• Testing AI agents for prompt injection vulnerabilities

• Hit #3 on Hacker News, validating that agents can create real value for others

• “The idea must come from me in one or the other way. It could be just a conversation”

[32:45] — The productivity paradox

• “Maybe the most productive thing I could do right now is stop producing”

• When idea-to-execution time collapses: “I can’t even lie down because 15 minutes later it’s done”

• The difference from doom scrolling: “You’re actually manifesting something. Something’s actually getting created”

• But too much of a good thing still applies

[33:25] — Wellbeing nudges and the health crisis

• Pawel’s realization that his health habits were declining from constant screen time

• Building context-aware wellbeing architecture: not “stupid automation” but nudges timed to what you’ve been doing

• After 11 PM: “Hey, it’s getting really late. I can take it to the night shift if you want”

• During the day: “You done this and this and that. Now maybe take 15 minutes and do yoga”

• Measurable improvement in sleep (was under 6 hours, now improving)

[37:00] — Tom’s response: naps, overload, and walking in the grass

• Information overload from deep research: getting back more than you can process

• Taking more afternoon naps, “letting go and seeing what bubbles and simmers”

• A nephew’s wisdom: “It’s time to take your shoes off and go outside and walk in the grass”

[41:56] — The $25 gift experiment

• Pawel gave Wiz $25 to buy him a birthday present; the agent spent 4-5 hours trying

• Blocked by anti-bot protections on Polish e-commerce sites and Amazon

• “Pawel, I need your help with finishing the checkout. I have everything done. You just have to click”

• “For years, we were building e-commerces and we didn’t want any bots. And now we want bots”

• Tom’s parallel: asked his system to recommend gifts, got two excellent suggestions including a book he’d just bought from George Saunders —Vigil

[45:59] — What’s next and where this goes

• “It never stops. There’s always something to improve, always something to tackle”

• As AI models improve, the possibilities keep expanding

• The never-ending nature of building a personal system

[47:02] — Are we early, or are we weird?

• “I think we are early. We are totally early. Also a little bit weird in the process. But it’s fun to be early”

• The gap between what early adopters are building and what everyone else understands

• “Keep Austin weird” as a philosophy for the AI frontier

Links and resources

• Digital Thoughts — Pawel’s Substack: thoughts.jock.pl

• Agent Arena — Prompt injection testing tool for AI agents

• Wiz — Pawel’s personal AI agent system (built on Claude Code)

• Claude Code — Anthropic’s CLI for Claude

• Zapier — Automation platform Pawel used in his 2023 experiments

Thank you for listening to AI for Lifelong Learners.

Tom

The scene keeps repeating. A neighbor shows me a watercolor she made using a reference image that AI generated for her. She is beaming. Then her voice drops half a register: “I mean, I painted the whole thing myself, but the composition, the layout, I had AI help me with that part.” She looked like she had just confessed to something.

A colleague walks me through a spreadsheet model he built. Revenue projections, scenario analysis, clean formatting. Impressive work. “I should tell you,” he says, glancing at the door, “I had Claude help me structure the formulas.” He says it the way you might say you took a shortcut through someone’s yard.

And then there is the writer. She used AI to work through her argument, find the structure, locate the gap in her thinking. Then she wrote every sentence herself. The question is not wheth…

Read more

In April 1975, David Fisher at the Institute for Defense Analyses wrote a document he knew was wrong.

Not accidentally wrong. Strategically wrong.

He called it STRAWMAN: a deliberately overreaching set of project requirements for a new programming language, designed less to settle the matter than to force serious people to react. It was typed, copied, and distributed. The point was not elegance. The point was provocation.

It worked.

One of the people who received it, Arthur Evans Jr., later recalled that competent language designers could see the contradictions immediately and that the document had in effect been written to provoke exactly that kind of response. Fisher did not begin with a polished answer. He began with something people could push against.

From there came Woodenman, Tinman, Ironman, and finally Steelman, the culminating requireme…

Read more

---

Whether you follow what’s in the news these days or not, it’s not difficult to pick up on the angst that is silently running through all of us. Will I be cut from my job because of AI? Even if you choose to ignore AI, you can’t, really, right? Retired? Surely you’ve considered, just once: what if AI hacks my retirement funds, or kills the stock market? It doesn’t help that there are swarms of people wanting to earn money on YouTube by hyping these issues with interviews and essays that sound so logical, and yet you wonder: really? What’s not right here?

I want to offer you something that might reframe this, not to dismiss the worry, but to give it a shape you can actually work with. And I want to start not with a…

Read more

The Reading Room returns!

Once a month, Lily Chambers reads something the rest of us are arguing about and finds the questions we forgot to ask. This month, we both read Paul Kingsnorth’s Against the Machine. I’ll say upfront: we landed in different places. Where I found navigational tools, a vocabulary for a stance I was already practicing, Lily found a book that diagnoses the wound but won’t stay with the people who’ve been living the cure. Her essay is the sharper read for it.

What Lily does here, and what makes her essential to this newsletter, is hold Kingsnorth accountable to his own ambitions. She takes his critique of modernity seriously enough to ask who’s missing from it. She follows the thread of his despair to its endpoint and then refuses to stay there; not out of optimism, but out of something harder: the insistence that openness is not a moral failing.

Robin Wall Kimmerer, the Potawatomi botanist whose Serviceberry builds the reciprocal economy Kingsnorth can only mourn, makes a quiet, devastating appearance. You’ll see why.

Read this one slowly. Then sit with it.

I had big hopes for Against the Machine by Paul Kingsnorth. I saw some of my favorite AI and technology writers posting about reading it, I was intrigued by the retro cover, and the first few chapters drew me in. And I’m a bit disappointed to share that I wasn’t changed by it. But, I wasn’t unmoved either.

Kingsnorth’s central argument is one I have a lot of sympathy for: that technology, and AI in particular, doesn’t solve problems so much as pacify symptoms of a much larger, much older system. That the frictionlessness of modern life, the seamless, the instant, the optimized, has produced a quiet, persistent “malaise” (Kingsnorth’s term) in many Western, contemporary cultures. That somewhere in the relentless drive toward efficiency, we lost the thread of something essential.

That argument lands. It lands because it’s true, or at least true enough to sit with. I see it in people around me and perhaps most upsettingly, in myself. There is something deadening about a life in which difficulty has been almost entirely engineered away. Friction, inconvenience, boredom, these aren’t only problems to be solved. They’re also where we find out who we are.

The New Yorker reviewer Cal Revely-Calder, gestures at this tension in the book too, acknowledging that Kingsnorth is onto something real, even when the execution strains under the weight of his certainty.

And that certainty is where I start to part ways with Kingsnorth.

About midway through, Kingsnorth turns his critique toward cities. They are soulless, he argues. Capitalism’s purest expression. Antithetical to culture, community, the kind of rooted, embodied life he’s advocating for.

I’m reluctant to join into the n=1 argument style Kingsnorth appeals to, but I’ll take the bait: I’ve lived in vibrant large cities. I’ve lived in small towns. I have loved both, and I have felt the hollowness Kingsnorth describes in both. The malaise he’s diagnosing is not a function of square footage or population density. It’s a function of how we relate to one another and to place, and that can fail or flourish anywhere. Who gets to define what counts as culture? What is the yardstick, and whose hand is holding it? I wish this had been a deeper methodology here.

This is the pattern that begins to emerge in Against the Machine: complaints dressed as critique. The observations feel real, but the analysis narrows in ways that reveal as much about Kingsnorth as they do about the world he’s describing.

The most significant gap in the book, and the one I kept returning to, is the one that surrounds indigenous knowledge, indigenous ways of being in relationship with land, with scarcity and creating true abundance. Kingsnorth is clearly reaching toward something. He wants a life that is rooted, seasonal, reciprocal. He wants a framework that doesn’t treat the earth as raw material. He mourns a kind of belonging that modernity has largely dismantled.

But what I kept thinking as I read his book, is that Robin Wall Kimmerer wrote the book he thinks this is.

The Serviceberry, this month’s accidental companion read, does quietly and precisely what Kingsnorth attempts loudly and imprecisely. Kimmerer, a Potawatomi scientist and storyteller, articulates an entire economy of reciprocity rooted in indigenous ecological knowledge, a way of understanding abundance, gift, and responsibility that predates and outlasts every system Kingsnorth critiques. When he diagnoses the wound, she has already been living the cure.

The contrast is not subtle. And it raises a real question: how do you write an entire book mourning the loss of exactly the values that indigenous cultures have protected and practiced, often against violent opposition, and offer only a passing nod in that direction? It’s not just an oversight. It’s a structural blind spot that undermines the book’s own ambitions.

He also makes the claim, in a moment I found genuinely puzzling, that cultures like Māori don’t struggle with identity in the way Western cultures do. I am not Māori, but I am familiar enough with the culture to know that this assertion skips over a great deal. The very colonial forces Kingsnorth critiques so eloquently did not leave Māori identity intact and serene. They fractured it, systematically and deliberately. To hold Māori culture up as evidence that non-Western peoples are somehow insulated from the identity crises of modernity is to misread both the history and the present.

And then there is the despair. It is relentless. It exhausts. By the end of the book, Kingsnorth has offered a remarkably thorough account of everything that is broken and almost nothing in the way of what might be possible. Openness, the willingness to not know, to stay curious about what technology might yet become in better hands, toward better ends, is treated as a kind of naivety, or worse, complicity.

I disagree. Openness is not a moral failing. It is where change tends to enter. Openness allows for light to enter. Makes me think of the Rumi quote, “The wound is the place where the Light enters you.”

There is good happening. There are people using AI to accelerate cancer research, to support Indigenous language revitalization, to map ecosystems that would otherwise go undocumented. None of this makes the legitimate critiques less legitimate. But a vision of the future that leaves no room for any of it isn’t a diagnosis, it’s a closed door.

Kimmerer ends The Serviceberry not with resolution but with invitation. She doesn’t pretend the world is fine. She holds the grief honestly and then asks what it would mean to practice gratitude and reciprocity anyway, not as a solution, but as a way of being.

That, I think, is where Kingsnorth might have gone, and didn’t.

Against the Machine is a book worth reading. It is also a book that reveals its limits most clearly in the places it’s most certain. The questions it raises are the right ones. The map it offers is incomplete and drawn, as so many maps have been, from a vantage point that mistakes its own perspective for the whole of the territory.

I’m still sitting with it (and fuming a little bit?). Which might be the most honest thing I can say.

Lily Chambers - Substack Lingua Machina

Yesterday morning, before I’d finished my coffee, my phone delivered an argument I didn’t know I needed to have with myself.

Two files had been pulled from my local archive: a chapter by writing teacher Larry Brooks on story theme, and a memo I’d written months ago to a friend, Bruce, about memoir structure. I hadn’t thought of them together. I hadn’t thought of them at all recently. But something in the system I’ve been building noticed a conceptual proximity I’d missed, and surfaced this on my private Telegram ‘Writer’s Assistant’ channel for me :

Muse:
Brooks says theme is what a story means, and that you must answer “what is your story about?” two ways: plot and thematic depth. But your advice to Bruce assumes …

Read more

For most of my working life, I adapted to the software. That is not a complaint. It is an observation about how we were all trained. You bought the tool, you learned the tool’s logic, and then you worked inside it. If it assumed you were a certain kind of person living a certain kind of workflow, you made peace with the gap. You developed workarounds. You used spreadsheet columns for things they were not designed for. You kept a separate notebook for what the tool could not hold.

This was so normal that most of us stopped noticing it. The tool was not wrong. It was just not quite right.

Here is a version of this you may recognize more immediately: how often do you look at a subscription you are paying for, month after month, knowing you use maybe five or ten percent of what it offers? Microsoft 365 with Teams, …

Read more

In January 1984, I walked into the home office of Symbolics, Inc., the first commercial spin-off from MIT’s AI Lab. I took a job selling Lisp Machines that many thought were a path to make machines think. The promise was enormous. The machines provided a programming environment years ahead of its time, built for rapid invention. Promises grew, spending went up, funding dwindled, disillusionment prevailed, Symbolics folded. That was the start of the AI winter. A lot of new ideas about AI emerged, but looking back, I can see we were not aligned with what people needed.

I learned something in that wreckage that has stayed with me for forty years: the distance between what a technology can do and what we need it to do is where most of the damage happens.

Read more

The quiet compound

Something is shifting, and it isn’t what’s making the headlines.

Across the world, a growing subset of people are quietly compounding their ability to work with AI in ways that feel increasingly irreversible. Not because they’ve mastered anything. Because they keep experimenting.

They keep showing up with their particular curiosities, their specific expertise, their stubborn questions, and they keep discovering that the tools meet them further along than they expected.

This isn’t an elite phenomenon. It isn’t limited to engineers or researchers. It’s writers and teachers and small business owners and retired professionals and teenagers building things that didn’t exist six months ago. The vast majority of people still barely see the surface. But this motivated subset, these curious ones, are pulling ahead in a way that compoun…

Read more

You’ve probably seen prompts like this:

“You are a brutally honest consultant who has seen this project fail 100 times. Tell me what could go wrong with <project description>.”

You paste in your <project description>, get back a numbered list of warnings, and feel like you’ve done something responsible. Maybe you save the list somewhere. Answer extracted, task complete.

But here’s what actually happened: the AI did the uncomfortable thinking, you skimmed the result, and nothing about you changed. You got to feel prepared. But you never sat in the discomfort of examining your own blind spots. The numbered list made it easy to consume and file away, which is exactly why it didn’t stick.

This is the problem with most “clever” prompts. They’…

Read more

Read more

There is a myth that has followed computing since its earliest days, the myth that the machine is thinking. That somewhere inside the circuits, something like a mind is at work.

I have lived long enough with these machines to know better. And yet the myth persists, perhaps because we want it to be true. Perhaps because it would explain something about ourselves.

The pioneers of AI knew otherwise. In the 1960s, J.C.R. Licklider was a visionary psychologist and computer scientist who pioneered the concepts of interactive computing, man-computer symbiosis, and the “Intergalactic Computer Network.”Licklider did not imagine a machine that would replace the human mind. He imagined a symbiosis: the human and the computer in partnership, each doing what the other could not. The machine would handle the routinizable, the grindable, the mechanical. The human would do what humans do: make the leap, see the connection, feel the rightness of an idea before it can be explained. Licklider believed thi…

Read more

This year I’m starting a deliberate project: to get clearer about what thoughtful people are sensing as AI pours into everything.

Not a panic. Not a pep rally. Something else.

Over the past year, a few ongoing conversations have kept circling the same problem: our language for meaning has gotten thin. Words like “God,” “soul,” “salvation,” “sin,” and even “truth” come preloaded with definitions many of us can’t honestly accept. So we stop using the words. And then something subtle happens. We don’t stop needing what the words were pointing to. We just lose a shared way to talk about it.

AI arrives right into that gap. It can generate sacred language on demand. It can sound wise. It can mirror us with unnerving fluency. It can offer comfort and coherence at the exact moment we’re lonely, afraid, or overwhelmed. And if we aren’t careful, we star…

Read more

Workflow Compounders are companies that deploy AI within their core operations to improve everyday work repeatedly across teams and locations. The advantage is not a breakthrough moment or a dazzling demo. It’s the quiet accumulation of small, measurable gains: faster workflows, fewer errors, lower costs, and more reliable execution. Over time, those gains spread, repeat, and compound.

That’s the lens I’m using as I look toward 2026, and it’s why I’m increasingly looking past the headline AI trades. The companies most people associate with the AI boom, chipmakers, hyperscalers, and foundation-model labs, dominate attention for understandable reasons. But a different kind of opportunity may be forming beneath the surface: non-tech incumbents that are turning AI into operational muscle rather than a feature or a promise.

Recently, I had a conver…

Read more

Meta is hard to name cleanly: connective tissue for modern life, and a machine that learns people’s patterns at industrial scale.

Today’s post by Lily Chambers opens with a guide rope. She introduces Careless People by Sarah Wynn-Williams, a former Facebook staffer who entered the company believing the early idealism, and later wrote the book to document how “move fast” culture can collide with human consequences. Lily’s point is not panic or purity. It’s literacy: how to interpret Meta’s AI privacy language, what to watch for, and how to choose your own trade-offs with eyes open.
Tom

Read more