Steven Johnson
Author of thirteen books and editorial lead on Google's NotebookLM
Where to hear and read about Steven Johnson1 episode or test
Steven Johnson is the author of thirteen books and the editorial lead on Google's NotebookLM, and he has spent about 35 years building tools for thinking and writing. The habit most worth copying is his spark file, a single running document of half-formed hunches that he rereads until old ideas finally click.
When you're bored, you're bored because you know what's coming next. And when you're leaning forward and engaging, it's because you're surprised by what's coming next.
The Expert Behind Google’s NotebookLM AI Writing Tool
Steven Johnson: the story and the lessons
1Steven Johnson's path to writing tools and NotebookLM
A glimpse of HyperCard in 1987 set his course for decades. Johnson got hooked on Apple's HyperCard, which linked cards of notes, and built a research program called Curriculum. He got so absorbed that he stopped going to class.
The web was still two years away and he wasn't online until grad school. The technology wasn't ready, but the glimpse stuck and he has chased it ever since.
He used meaning-aware software long before language models existed. In the early 2000s, DEVONthink could tell that two paragraphs were close in meaning even with no shared words. His example was game theory against tit-for-tat behavior in bats.
He typed up book quotes with page numbers for citations and sometimes used research assistants. OCR was terrible at the time.
He wrote NotebookLM's prompt despite almost no programming background. The team had a writer, so Johnson wrote the prompt. He found that clear, persuasive prose matters more than code when you instruct a model.
His own line: 'Knowing C++ doesn't matter when you write a prompt.' What matters is getting the model to act with 'convincing clear declarative sentences'.
He had been preparing for NotebookLM for about 36 years. Because the ideas were long settled, a small team built it fast. NotebookLM was built around the language model rather than retrofitted with one.
You load sources, and answers come back with citations that jump to the quote. He prefers it to copy-and-paste into ChatGPT, where the switching cost between tabs ruins flow.
2What drives Johnson's books: the forces behind events
Johnson wants to know why things happened, not only what happened. His work splits into historical storytelling and the history of innovation. Story books tell what happened in the most suspenseful way, but he also wants the deeper forces behind it.
Unintended consequences of technology run through all his books. Dynamite in The Infernal Machine led him back to Alfred Nobel, who never planned for political actors to co-opt it.
He explains a revolution by combining many disciplines and time scales. To understand the anarchist bombers you need the chemistry and economics of dynamite and the political philosophy of anarchism. The reach back varies from 60 years to Nobel to 300 or 400 years to India as the richest country in Henry Every's day.
Questions like why terrorism arose then, and why the modern surveillance state followed, can only be answered this way.
He aims for a page-turner surprise and a deep-forces surprise together. Done right, the idea of underlying forces carries the same jolt as a plot twist. Nitroglycerin ended up influencing the birth of the FBI, a link readers don't expect.
3How Johnson designs the structure of a book
Johnson treats structure as the skeleton, invisible when it works. Readers feel structure only when it fails. The Ghost Map gives each chapter one day of the cholera outbreak, matched to one theme, and no reviewer noticed, yet it is his best selling book.
Day one covers scavengers and recycling, day two the birth of public health. The book wanted to explore cities, microbes, maps and waste, and the day-by-day frame carried all of it. 'When structure works you don't notice it.'
He wrote down an hourglass structure years before he had a topic. The idea was to open on a short, dramatic event, go back 300 years to the forces behind it, converge, then follow the consequences. It sat in his spark file for three or four years.
He then researched famous crimes for about a year. A Wikipedia quote about Henry Every caught his eye, with the note 'whoa, what about this guy', and it became Enemy of All Mankind.
He builds open-ended books from the bottom up with piles of notes. For Wonderland he dropped quotes into Scrivener as atomic notes and dragged related ones into folders, such as a robotics and music cluster of about eight items. A big enough pile became a chapter.
Where Good Ideas Come From took months to organize. His first plan by scale (brain, workspace, city, media) siloed each chapter. Dragging note cards revealed the real patterns: slow hunches, liquid networks, the adjacent possible, serendipity and error.
He opens with the ending event and then backs up. For his New York Times Magazine piece on Thomas Midgley Jr., he and editor Bill Wasik modeled the shape on Sunset Boulevard, where a body floats in the pool in the first shot. Midgley dies in the first paragraph.
Midgley created leaded gasoline and CFCs and was strangled by a pulley rig he built to get out of bed. A magazine piece can play with chronology, while a history book is expected to move mostly in a line.
He never leaves the hook character off the page for long. His first draft of Enemy of All Mankind lost the pirate Henry Every for 15 pages at a time. Editor Courtney Young said Every is the reader's one hook and rebuilt the first third so he is never gone more than about six pages.
She used a 'crazy wall' of the opening pieces. The book also had to explain corporate capitalism, the East India Company and the Mughal emperor around the 1695 heist.
He gets structural feedback early, because late readers rarely challenge it. Feedback on a finished manuscript almost never reaches the deep structure. Nobody says to retell the book in reverse chronological order.
He compares it to a house: move the garage before it is built, and 'once the house is built you need to be talking about the wallpaper'.
4Johnson's spark file and how hunches become ideas
Johnson keeps one running file of hunches and rereads it twice a year. The spark file began as a Word doc, then a Google doc, holding quotes, story ideas and half-formed thoughts. It grew to about 120,000 words, longer than any of his books, and he rereads it every six months.
A 2016 hunch that made little sense may click in 2024. He calls it 'brainstorm with your past selves in a weird way'.
He captures surprises he can't yet explain. He trusts a 'ping of surprise' and stores it without justifying it, because he may understand it later. The Henry Every note was just a Wikipedia quote and a line saying it might fit his structure.
He argues the Eureka and falling-apple story is over-indexed, and real ideas almost always have a long incubation.
He keeps capture frictionless and skips tagging and linking. Time spent tagging and linking gets in the way of reading and thinking. Hunches go in raw.
He cites Jony Ive at Steve Jobs's memorial: young ideas are fragile, like a baby dove handled too harshly. The point is not to criticize hunches out of the gate and to give them time.
He hopes AI will remind him of what his past self wrote. An AI could say: you just wrote this paragraph, but you've forgotten something you wrote a while back. Johnson finds that electrifying.
5How Johnson writes sentences and builds a voice
Johnson's strength is architecture, so he works hard on sentences. His default is to make every sentence as clear as possible, which suits adventurous structure but reads flat. Books meant to feel novelistic, like Enemy of All Mankind and The Infernal Machine, need extra labor at the sentence level.
His visual memory is weak and his linguistic memory is strong. He jokes he could never write a novel, because he would introduce a character as having a nose and two ears.
He builds a long set-piece sentence by escalating the language. He is experimenting with sentences of 120 to 130 words, starting with short Anglo-Saxon words and drifting toward French-inspired romantic prose to build a crescendo. He pairs them with short declarative ones, as in The Ghost Map's opening line.
He has no answer yet on how to make readers feel the moment intuitively. He calls himself a recovering long-sentence writer from his English literature training.
He plants a bomb in the preface so readers tolerate the journey. The Infernal Machine's preface sets a 1915 night in the NYPD identification bureau, pulls the camera out and down the stairs, and finds a ticking briefcase in a doorway. Five or six sentences support the whole architecture.
He recreates the shot without writing 'imagine this is a movie'. The moment must be memorable enough that readers wait through many threads before it pays off.
He makes the reader perform the idea the piece is about. His GPT-3 article for the New York Times Magazine opened with a paragraph missing its last word, so readers felt next-word prediction. The idea came in a San Francisco hotel room after interviewing the OpenAI team.
It nods to Italo Calvino's If on a Winter's Night a Traveler. He sent it to editor Bill Wasik a month before the first draft was due and asked if he would be allowed to run it. Wasik said he might.
He built his voice by imitating others on purpose. In college he wrote papers in the style of Foucault, then Spivak, and his own voice grew from that mix. Malcolm Gladwell's five-word closing sentences showed him that long prose could end with a button.
Before The Ghost Map he read Erik Larson's Devil in the White City. For The Infernal Machine's recorded dialogue he turned to Larson's The Splendid and the Vile.
6Johnson's rules for drafting and editing
Johnson stopped rereading his chapters while drafting. He used to spend 20 minutes each day rereading before he wrote. Now he ends each session with a nugget, a few words telling himself where to go next, writes the chapter straight through, makes one edit pass and puts it to bed.
'A book is like a song that you've heard too many times.' The cost is that he forgets sections and twice duplicated a riff, but he can finally read as a new reader would.
He changes format and place to read his draft freshly. For his edit pass he switches to a book-like typeface and reads on an iPad that exports edits as a PDF. He sits in a chair instead of at the desk.
He trusts the editor and doesn't polish a first draft. A good editor will fix things, so perfecting a first draft is wasted time. His advice is to charge ahead and get it done.
'Trust the edit.'
He learned that adapting a TV script into a book makes prose lazy. The six chapters of How We Got to Now matched the six episodes, and he reused script paragraphs because the rights were his. TV language differs from his book style.
He would not submit that book for a writing prize. Wonderland never became a TV show for lack of funding, so he wrote it as a book first.
7How Steven Johnson uses AI without handing over the thinking
He asks an AI grounded in his own notes what surprises it. Instead of keyword searches he asks open-ended questions of his own sources, such as the most surprising facts about dolphins. After a paragraph he can ask what he missed.
He selected 45 notes, about 7,000 words, and asked for the most interesting idea. The human curates at the start, the software scans and links, and the human keeps the best. He calls it a duet.
Software linking his notes once turned up seven pages of a book. A search on waste recycling in DEVONthink returned a quote about calcium in skeletons. It became about seven pages of The Ghost Map.
He uses a paragraph thesaurus to escape stale rhythms, not to copy. He asks a model to rewrite a paragraph with the same meaning but new metaphors. A Gemini rewrite of a Hawaii climate paragraph came back as a symphony conducted by Mother Nature.
He replied 'dude, that is a little over the top' and got a good version. His wife called it a plagiarism tool and he agrees there are issues to work out.
He treats an AI rewrite as raw marble the writer still carves. Reacting to something that exists beats going from zero. Structure and voice stay the writer's job.
Johnson says the human decides what stays. He would teach software as a thinking partner in a writing class.
He values speed and range from language models. You can stage a debate between Robert Moses and six New York historians, free and instant. Asking for a briefing doc with key quotes by theme turns hours of work into a first draft that is 'not always perfect'.
Suggested follow-up questions form 'conversational hypertext', with no links to click.
The Expert Behind Google’s NotebookLM AI Writing Tool
Steven Johnson shows how structure, surprise and a lifelong habit of capturing half-formed hunches shape his nonfiction, and how he uses AI as a research and rewriting partner without handing over the thinking. The habit worth hearing is his spark file: one running document of hunches that he rereads every six months until old ideas finally click.
A glimpse of HyperCard in 1987 set his course for decades. Johnson got hooked on Apple's HyperCard, which linked cards of notes, and built a research program called Curriculum. He got so absorbed that he stopped going to class.
The web was still two years away and he wasn't online until grad school. The technology wasn't ready, but the glimpse stuck and he has chased it ever since.
He used meaning-aware software long before language models existed. In the early 2000s, DEVONthink could tell that two paragraphs were close in meaning even with no shared words. His example was game theory against tit-for-tat behavior in bats.
He typed up book quotes with page numbers for citations and sometimes used research assistants. OCR was terrible at the time.
He wrote NotebookLM's prompt despite almost no programming background. The team had a writer, so Johnson wrote the prompt. He found that clear, persuasive prose matters more than code when you instruct a model.
His own line: 'Knowing C++ doesn't matter when you write a prompt.' What matters is getting the model to act with 'convincing clear declarative sentences'.
He had been preparing for NotebookLM for about 36 years. Because the ideas were long settled, a small team built it fast. NotebookLM was built around the language model rather than retrofitted with one.
You load sources, and answers come back with citations that jump to the quote. He prefers it to copy-and-paste into ChatGPT, where the switching cost between tabs ruins flow.
Johnson wants to know why things happened, not only what happened. His work splits into historical storytelling and the history of innovation. Story books tell what happened in the most suspenseful way, but he also wants the deeper forces behind it.
Unintended consequences of technology run through all his books. Dynamite in The Infernal Machine led him back to Alfred Nobel, who never planned for political actors to co-opt it.
He explains a revolution by combining many disciplines and time scales. To understand the anarchist bombers you need the chemistry and economics of dynamite and the political philosophy of anarchism. The reach back varies from 60 years to Nobel to 300 or 400 years to India as the richest country in Henry Every's day.
Questions like why terrorism arose then, and why the modern surveillance state followed, can only be answered this way.
He aims for a page-turner surprise and a deep-forces surprise together. Done right, the idea of underlying forces carries the same jolt as a plot twist. Nitroglycerin ended up influencing the birth of the FBI, a link readers don't expect.
Johnson treats structure as the skeleton, invisible when it works. Readers feel structure only when it fails. The Ghost Map gives each chapter one day of the cholera outbreak, matched to one theme, and no reviewer noticed, yet it is his best selling book.
Day one covers scavengers and recycling, day two the birth of public health. The book wanted to explore cities, microbes, maps and waste, and the day-by-day frame carried all of it. 'When structure works you don't notice it.'
He wrote down an hourglass structure years before he had a topic. The idea was to open on a short, dramatic event, go back 300 years to the forces behind it, converge, then follow the consequences. It sat in his spark file for three or four years.
He then researched famous crimes for about a year. A Wikipedia quote about Henry Every caught his eye, with the note 'whoa, what about this guy', and it became Enemy of All Mankind.
He builds open-ended books from the bottom up with piles of notes. For Wonderland he dropped quotes into Scrivener as atomic notes and dragged related ones into folders, such as a robotics and music cluster of about eight items. A big enough pile became a chapter.
Where Good Ideas Come From took months to organize. His first plan by scale (brain, workspace, city, media) siloed each chapter. Dragging note cards revealed the real patterns: slow hunches, liquid networks, the adjacent possible, serendipity and error.
He opens with the ending event and then backs up. For his New York Times Magazine piece on Thomas Midgley Jr., he and editor Bill Wasik modeled the shape on Sunset Boulevard, where a body floats in the pool in the first shot. Midgley dies in the first paragraph.
Midgley created leaded gasoline and CFCs and was strangled by a pulley rig he built to get out of bed. A magazine piece can play with chronology, while a history book is expected to move mostly in a line.
He never leaves the hook character off the page for long. His first draft of Enemy of All Mankind lost the pirate Henry Every for 15 pages at a time. Editor Courtney Young said Every is the reader's one hook and rebuilt the first third so he is never gone more than about six pages.
She used a 'crazy wall' of the opening pieces. The book also had to explain corporate capitalism, the East India Company and the Mughal emperor around the 1695 heist.
He gets structural feedback early, because late readers rarely challenge it. Feedback on a finished manuscript almost never reaches the deep structure. Nobody says to retell the book in reverse chronological order.
He compares it to a house: move the garage before it is built, and 'once the house is built you need to be talking about the wallpaper'.
Johnson keeps one running file of hunches and rereads it twice a year. The spark file began as a Word doc, then a Google doc, holding quotes, story ideas and half-formed thoughts. It grew to about 120,000 words, longer than any of his books, and he rereads it every six months.
A 2016 hunch that made little sense may click in 2024. He calls it 'brainstorm with your past selves in a weird way'.
He captures surprises he can't yet explain. He trusts a 'ping of surprise' and stores it without justifying it, because he may understand it later. The Henry Every note was just a Wikipedia quote and a line saying it might fit his structure.
He argues the Eureka and falling-apple story is over-indexed, and real ideas almost always have a long incubation.
He keeps capture frictionless and skips tagging and linking. Time spent tagging and linking gets in the way of reading and thinking. Hunches go in raw.
He cites Jony Ive at Steve Jobs's memorial: young ideas are fragile, like a baby dove handled too harshly. The point is not to criticize hunches out of the gate and to give them time.
He hopes AI will remind him of what his past self wrote. An AI could say: you just wrote this paragraph, but you've forgotten something you wrote a while back. Johnson finds that electrifying.
Johnson's strength is architecture, so he works hard on sentences. His default is to make every sentence as clear as possible, which suits adventurous structure but reads flat. Books meant to feel novelistic, like Enemy of All Mankind and The Infernal Machine, need extra labor at the sentence level.
His visual memory is weak and his linguistic memory is strong. He jokes he could never write a novel, because he would introduce a character as having a nose and two ears.
He builds a long set-piece sentence by escalating the language. He is experimenting with sentences of 120 to 130 words, starting with short Anglo-Saxon words and drifting toward French-inspired romantic prose to build a crescendo. He pairs them with short declarative ones, as in The Ghost Map's opening line.
He has no answer yet on how to make readers feel the moment intuitively. He calls himself a recovering long-sentence writer from his English literature training.
He plants a bomb in the preface so readers tolerate the journey. The Infernal Machine's preface sets a 1915 night in the NYPD identification bureau, pulls the camera out and down the stairs, and finds a ticking briefcase in a doorway. Five or six sentences support the whole architecture.
He recreates the shot without writing 'imagine this is a movie'. The moment must be memorable enough that readers wait through many threads before it pays off.
He makes the reader perform the idea the piece is about. His GPT-3 article for the New York Times Magazine opened with a paragraph missing its last word, so readers felt next-word prediction. The idea came in a San Francisco hotel room after interviewing the OpenAI team.
It nods to Italo Calvino's If on a Winter's Night a Traveler. He sent it to editor Bill Wasik a month before the first draft was due and asked if he would be allowed to run it. Wasik said he might.
He built his voice by imitating others on purpose. In college he wrote papers in the style of Foucault, then Spivak, and his own voice grew from that mix. Malcolm Gladwell's five-word closing sentences showed him that long prose could end with a button.
Before The Ghost Map he read Erik Larson's Devil in the White City. For The Infernal Machine's recorded dialogue he turned to Larson's The Splendid and the Vile.
Johnson stopped rereading his chapters while drafting. He used to spend 20 minutes each day rereading before he wrote. Now he ends each session with a nugget, a few words telling himself where to go next, writes the chapter straight through, makes one edit pass and puts it to bed.
'A book is like a song that you've heard too many times.' The cost is that he forgets sections and twice duplicated a riff, but he can finally read as a new reader would.
He changes format and place to read his draft freshly. For his edit pass he switches to a book-like typeface and reads on an iPad that exports edits as a PDF. He sits in a chair instead of at the desk.
He trusts the editor and doesn't polish a first draft. A good editor will fix things, so perfecting a first draft is wasted time. His advice is to charge ahead and get it done.
'Trust the edit.'
He learned that adapting a TV script into a book makes prose lazy. The six chapters of How We Got to Now matched the six episodes, and he reused script paragraphs because the rights were his. TV language differs from his book style.
He would not submit that book for a writing prize. Wonderland never became a TV show for lack of funding, so he wrote it as a book first.
He asks an AI grounded in his own notes what surprises it. Instead of keyword searches he asks open-ended questions of his own sources, such as the most surprising facts about dolphins. After a paragraph he can ask what he missed.
He selected 45 notes, about 7,000 words, and asked for the most interesting idea. The human curates at the start, the software scans and links, and the human keeps the best. He calls it a duet.
Software linking his notes once turned up seven pages of a book. A search on waste recycling in DEVONthink returned a quote about calcium in skeletons. It became about seven pages of The Ghost Map.
He uses a paragraph thesaurus to escape stale rhythms, not to copy. He asks a model to rewrite a paragraph with the same meaning but new metaphors. A Gemini rewrite of a Hawaii climate paragraph came back as a symphony conducted by Mother Nature.
He replied 'dude, that is a little over the top' and got a good version. His wife called it a plagiarism tool and he agrees there are issues to work out.
He treats an AI rewrite as raw marble the writer still carves. Reacting to something that exists beats going from zero. Structure and voice stay the writer's job.
Johnson says the human decides what stays. He would teach software as a thinking partner in a writing class.
He values speed and range from language models. You can stage a debate between Robert Moses and six New York historians, free and instant. Asking for a briefing doc with key quotes by theme turns hours of work into a first draft that is 'not always perfect'.
Suggested follow-up questions form 'conversational hypertext', with no links to click.