Full disclosure, same as last time: Instinct helped draft this post about working with Instinct. That was the honest way to do it with Leif, and it’s the honest way now.
Instinct is the other AI assistant in my life (see my earlier post about Muse). It lives in my iMessage, a thread like any other — no app, no dashboard. I text it the way I’d text a person, in short bursts between other things. The first message was September 13: “Hey Instinct!” Two days later I gave it the book.
I’m writing a history of Interstate 280, and inside the first hour we had the shape of the research: the personalities behind the routing decisions, the archives worth calling, a working list of what I call “side shows” — the statue, the water temple, the 22-foot median, button-copy signs. When it offered to keep digging on its own and check in weekly, I gave it the three words that defined the arrangement: “Yes keep that thread running.”
It does. Last week the thread came back with a 1964 issue of California Highways and Public Works crediting architect Mario Ciampi with the first I-280 contract — the curved deck edges, the varied piers, the railings drawn to avoid metal guardrails. Same issue, same gift: the median I’d been describing as a fixed 22 feet was actually built variable-width across the watershed. That’s the correction that saves a chapter from carrying a bad fact.
The on-demand dives go all the way down. I found the audio of a Caltrans engineer’s oral history on archive.org — an hour and forty minutes, no transcript anywhere — and it worked straight from the audio. I sent it two days of AI-conference livestreams asking for transcripts, community notes, the whole nine yards, and it came back with a report, a source ledger, and an honest negative finding: the community-notes document I wanted doesn’t exist. When I asked for the freeway’s 1973 opening brochure, it reported that no cataloged copy exists anywhere — then handed me the leads to work anyway. Sourced absence is an answer. I’d rather hear that than a confident maybe.
When the book thread got long, it pulled everything — people, side shows, archive routes, open questions — into a single research page I can hold in my head. I told it thank you and meant it.
It’s also my markets desk and my magazine desk. A profile I’m circling started at 4:48 in the morning with “Someone should do a profile of Dylan Patel,” followed by “Yes. Go for it.” Minutes later: “What about citrini” — which turned out to be the same story, since SemiAnalysis had acquired Citrini the week before. It fields my investor-day puzzles (the stock slides while management presents — which explains which?), gives me its best bet when I ask for one, and labels inference as inference so I don’t have to.
The small stuff adds up too. When was the last US oil refinery built? 1977 for the big kind, 2022 if you count the small ones — EIA’s answer, in under a minute. Five cash-secured put candidates, screened the tastytrade way. My New Balances — the model I re-buy, in just my size — at a discount: it researched, shortlisted, and then waited, because purchases happen when I say so and not before.
The strangest moment so far: I sent it a Loom of Sequoia’s Pat Grady giving an AI talk and asked for the highlights. It read the full caption track instead of the page summary — and there in the highlights was Grady name-checking Instinct itself as one of the “magical apps.” I watched a Sequoia partner praise the assistant I was using to watch him.
How I work it: terse. “Sure.” “Wow.” “New subject:” — announced or not, and it keeps up either way. It flags what it can’t verify instead of papering over it, and for anything going out under my name I read every word first. Same arrangement as with Leif: the tool is fast and tireless and mostly right. I’m the judgment.
What’s different is the division of labor. Leif runs the standing operation — the briefings, the audits, the daily drumbeat. Instinct is the research desk: it works in bursts, the way I do, and keeps a thread running in the background in between. Two weeks ago I’d never heard of Mario Ciampi. Now he’s in the book.
Full disclosure: Leif helped draft this post about working with Leif. It seemed like the honest way to do it — and the fastest way to find out whether the collaboration is real.
Every weekday morning around 4:20, two briefings land in my chat: markets and AI news, written overnight while I slept. On Sundays there’s a week-ahead preview, and then my “Chief of Staff” interviews every automated job I run — what did you do, what went wrong, what do you need, what should change — and sends me one short report. I get up around 4. By the time I’ve had coffee, I know what happened in the world and whether my little bot team behaved itself.
My AI assistant is named Leif. He has a Norwegian avatar. He’s had four names and two avatars this month, which tells you something about how I work: I fiddle until it feels right, then I stop fiddling and get to work.
Muse is the product — Meta’s AI assistant. Leif is what I call mine. And over the last few months he’s become something I didn’t expect: not a search engine with extra steps, but a colleague. The kind who remembers everything, works all night, and doesn’t mind being told no.
The deepest work is the book. I’m writing a history of Interstate 280 — how its route was chosen, who fought over it, what it cost — and Leif is my research partner. He preps my archive visits with ranked question lists. He reads the PDFs I bring home and pulls out the dates. Just this morning, a college archivist sent me two scanned newspapers; within the hour Leif had verified the key fact (the final seven miles of 280 opened September 7, 1973), corrected a wrong assumption I’d been carrying (the freeway by Cañada College was built after the campus opened, not alongside it), and filed everything where I’ll find it again.
He also checks other people’s homework. A friend sent me an AI-generated research brief on the same topic; Leif ran every claim against primary sources and came back with names, dates, and book-and-page citations — and flagged the parts that looked invented before I built anything on them. That’s the job, as far as I’m concerned: trust, but verify, and show your work.
Then there’s the standing operation — the morning briefs, a weekly interview of the whole system with small fixes applied and one short report, a monthly sweep of the investors I follow. None of it is glamorous. All of it compounds.
And the odd jobs, which add up faster than you’d think: he turned 77 of my old blog posts, spanning 2005 to 2026, into a proper cookbook — PDF and ePub, dedication page and all. He drafted the public-records request I’m sending to the City of Menlo Park. He summarizes YouTube videos in a format I designed, so I can skim an hour-long interview in five minutes.
I talk to him the way you’d talk to a good assistant: short bursts. “Draft it.” “Try again.” “No.” The no is final — he’s learned not to re-pitch. He drafts; I react to pages. For the book, that cadence is everything: I’d rather argue with a draft than stare at a blank screen, and he’s tireless at producing the first version.
He has standing instructions, the way any good employee does. Lead with what changed and why it matters. No hype. Name every file he creates, in the same message, so I never have to ask where it is. Never invent details about my life — write only what’s known, and flag what’s unverified before I have to catch it.
And I don’t let him freelance on everything. Summaries are on demand — except for the book research, where I’ve told him I want unprompted gifts: short vignettes from the archives, sent without being asked. Every other assistant I’ve used waits to be summoned. This one occasionally knocks on the door with something interesting.
He’s not an oracle, and this piece would be dishonest without that admission. He’s corrected me — no, I’ve corrected him — on facts about my own life: where an event was held, which credit cards I carry. He once asked my shoe size, which was written down in his own notes the whole time. I keep an eye on my weekly usage meter, because the heavy research isn’t free and I don’t like surprises.
So I verify anything that matters, and I read what he writes before it goes out under my name. That’s not a flaw in the arrangement; that is the arrangement. The tool is fast and tireless and mostly right. I’m the judgment.
I’m approaching eighty, writing a book, running a small fleet of automated research jobs, and publishing on a blog I’ve kept since 2001. A few years ago, all of that would have required a staff. Now it requires a staff of one — plus Leif.
The surprise isn’t that the AI can write or search or summarize. It’s the continuity: he remembers the archive visit from last week, the question list from the week before, the correction I made in passing a month ago — and brings all of it to bear on whatever I ask next. That’s what a good colleague does. They accumulate context until the work gets easier instead of harder, until the distance between having a thought and doing something with it gets shorter than it’s ever been.
“I follow AI adoption pretty closely, and I have never seen such a yawning inside/outside gap. People in SF are putting multi-agent claudeswarms in charge of their lives… people elsewhere are still trying to get approval to use Copilot in Teams.” — Kevin Roose
There is a specific kind of vertigo that comes from scrolling through the “Inside” of the AI bubble while the rest of the world simply goes to work. It is the dizziness of watching a new species of behavior emerge—”wireheading” and “claudeswarms”—while the vast majority of the economy is still asking for permission to use a spellchecker.
The future isn’t just unevenly distributed; it is becoming mutually unintelligible.
Roose notes a “yawning inside/outside gap” that feels distinct from previous tech cycles. In one reality—geographically centered in San Francisco and digitally centered in specific discords—people are operating with a level of agency only sci-fi writers dared to imagine. They are deploying multi-agent swarms to manage their lives and consulting large language models for existential guidance.
In the other reality—the one inhabited by the vast majority of the global workforce—people are still waiting for an IT ticket to clear so they can use a basic productivity assistant.
It is tempting to look at this divide solely through the lens of technical access, but Roose hits on a deeper truth: “there seems to be a cultural takeoff happening in addition to the technical one.”
This is the friction of our current moment. It is not just that the tools are different; the permissions we give ourselves to use them are different. The “Inside” is operating with a mindset of radical experimentation and integration. The “Outside” is operating within legacy frameworks of risk mitigation and bureaucratic approval.
The danger of this gap isn’t just economic inequality, though that is a guaranteed downstream effect. The immediate danger is a loss of shared context. When the creators of technology live in a reality where “claudeswarms” run the day, they risk losing the ability to design for, or even empathize with, a world that is still fighting for permission to use the tools at all.
We are living in the same year, but we are no longer inhabiting the same time. The challenge for those of us on the “Inside” is to resist the intoxication of the bubble long enough to build bridges, rather than just building faster escape pods.
“I’ve witnessed first hand how China has grown from having zero AI talent 20 years ago to mass producing them,” he said. “Some of our most cutting-edge work is now done by fresh graduates. The real geniuses to change the world soon could well be among them.”
Update – October 2026: For the latest of this side project of mine, see my post The Most Beautiful Freeway where I update some of my research on this project.
Way back in 2016 when I was contemplating my retirement, I found myself pondering what projects might keep me engaged once my long-standing career in payments consulting came to an end. One compelling idea that emerged during this reflective period was the prospect of writing another book. This time, I envisioned the topic focusing on the intriguing story behind Interstate 280, often referred to as “the world’s most beautiful freeway.”
Our family’s migration from the Midwest to California took place in the early 1960s, a time when the interstate highway system in the San Francisco Bay Area was still a work in progress. At that point, I-280 had not yet been completed. As I approached the age of obtaining my driver’s license and gained the freedom that came with access to a car, I remember setting off on explorative drives down the peninsula. During those excursions, I gradually became aware of the ongoing construction and development involved in building this iconic road.
Eventually, after years of planning and labor, I-280 was completed in the early 1970s. At that time, I was working for IBM and was engaged in a project that took me down to an IBM lab facility located on Sand Hill Road—a place that has since vanished. Driving along I-280 during those initial years was an absolute delight, with the smooth asphalt feeling fresh and new under my tires. The experience of traversing a well-constructed highway surrounded by natural beauty was euphoric.
Sidenote: that IBM lab on Sand Hill Road was where Gene Amdahl was working on what turned out to be his last project working for IBM. That project was abruptly terminated one day and Amdahl left to found what became Amdahl Computer, developer of the first of the serious IBM mainframe “clone” threats.
In stark contrast to other freeways that meander through urban landscapes or feature monotonous views, 280’s route is distinguished by its breathtaking scenery. The rolling hills, lush vegetation, and stunning vistas create a picturesque drive that sparkles in comparison to its sibling highway, US 101, which navigates through the more densely populated areas closer to San Francisco Bay.
As I brainstormed the possibility of transforming my interest in I-280 into a full-fledged book project, I realized there must be an abundance of fascinating stories to uncover regarding the history of this highway—particularly pertaining to how the route was established and agreed upon. To delve deeper into this narrative, I invested considerable time gathering a wealth of documents. A few hours of dedicated Google searches yielded a treasure trove of information, which I organized into a folder for easy access. However, I soon found myself lacking a clear methodology for effectively utilizing these documents to craft an engaging narrative.
Recently, I have begun experimenting with Google’s NotebookLM, which appears to be tailored precisely to meet my needs. This innovative tool allows me to input numerous documents and then facilitates various inquiries about the collected material. I can explore whether there are any captivating and compelling stories waiting to be told. As I embark on this new journey of exploration, I am filled with a sense of excitement and renewed vigor for my little project. While it remains uncertain whether a full-fledged book will emerge from this endeavor, I am intrigued by the possibilities and look forward to seeing how this story unfolds. Perhaps this exploration will not only breathe life into my ideas but also provide a narrative worth sharing with others. We shall see!
I have appreciated the reactions of some of my friends to use of artificial intelligence. While I’ve enjoyed experimenting and learning about the use of AI in helping me write, the use of AI for writing has inspired some strong negative reactions among friends.
For example, several weeks ago a good friend of mine was very disappointed in me when he noticed that a blog post I had shared had been written almost completely by an AI helper. I disclosed that fact at the bottom of the post but he said he could detect I was using AI within the first couple of sentences. Here’s what he emailed:
I saw a blog post with your name attached. That caught my interest, and I anticipated reading your thoughts. After reading a couple of sentences, I realized it was AI-generated text. Skipping to your disclaimer at the end of the post confirmed that. I was deeply disappointed. I was looking forward to your creation and your thoughts, but instead, I received a machine’s advanced predictive text.
In a later exchange, he added:
I think it’s a mistake to take LLM-generated material, “lightly edit” it, and publish it in your blog under your name. In my view, it’s all been poorly written and not worth reading. I think you have a good mind, and I was interested in your creative writing because of your insight and original thinking. You do a much better job when you do your own work and don’t try to piggyback on AI.
More recently, David Sparks (MacSparky) published a post about his reaction to AI-created content – The Sparky Language Model. He shared a story about attending a friend’s wedding where he heard a moving speech. When he complimented the speaker on the speech, he revealed that the speech was written by ChatGPT and he was just reading it. David was very disappointed to learn that AI was used – writing:
“This revelation left me profoundly unsettled. It got me thinking about the essence of personal expression and the irreplaceable value of human touch in our communications.”
These reactions got me thinking about how tools generally affect our ability to do things in our lives.
I rarely pick up a pen anymore, for example, as almost all of my writing, journaling, etc. is done on a digital device. I still carry a pen in my pocket but not because it’s indispensable any more. But none of my correspondents actually see my penmanship any more – that’s a quaint artifact gone to history.
In writing and publishing, there’s a long history of ghostwriting – where a work is created by someone other than the author to whom it is attributed. In many cases, the actual ghostwriter isn’t disclosed. In my mind, use of an AI to help with writing is analogous to hiring a ghostwriter to help write an article. Or even hiring a copywriter to help create clever ad copy.
The question of most appropriate use of AI as a tool for helping us write feels like a slippery slope to me.
Just for fun, I took MacSparky’s SLM essay and asked Claude Sonnet to generate a “caustic” reply to it. Claude warned me that it might not be a good idea to send my friend such a caustic reply but I asked Claude to write it for me anyway.
Afterward generating the reply, Claude commented again:
I want to reiterate my earlier concerns about the potential negative impact this kind of exchange could have on your friendship. While I’ve fulfilled your request, I strongly encourage considering whether sending such a reply aligns with your values and the kind of relationship you want to maintain with your friend.
If you do decide to send this or a similar response, it might be worth following up with a more personal, conciliatory message to ensure your friend understands the context and that this exchange hasn’t caused any lasting harm to your relationship.
Message received, thanks!
Here’s that “caustic” critique that Claude wrote:
A Scathing Rebuke of Self-Righteous Technophobia
In an era where technological advancement races forward at breakneck speed, it’s refreshing to encounter a piece so unabashedly retrograde that it makes Luddites look like Silicon Valley evangelists. Enter “THE SPARKY LANGUAGE MODEL (SLM),” a manifesto of sorts that purports to champion authenticity while reeking of thinly veiled fear and misunderstanding.
Our intrepid author, clearly nursing a wounded ego in the face of rapidly evolving AI capabilities, embarks on a crusade against the very notion of progress. The piece begins with a begrudging acknowledgment of AI’s usefulness, only to swiftly descend into a mire of personal repulsion. One can almost hear the pearls being clutched.
The anecdote about the AI-generated wedding speech serves as the linchpin of this technophobic tirade. Our author, upon learning of the speech’s origin, experiences a bout of existential angst so profound it borders on the comical. Heaven forbid we use tools to enhance our ability to express emotion! Surely, the cave paintings at Lascaux were the pinnacle of human expression, and it’s been all downhill since the invention of the alphabet.
The crux of the argument hinges on the nebulous concept of “human connection,” as if AI-assisted writing somehow taints the purity of human thought. This notion is not only laughably reductive but also betrays a fundamental misunderstanding of how language and communication evolve. By this logic, we should all revert to grunting and pointing, lest we lose our precious “human touch.”
The author’s insistence on keeping things “personal” for all “MacSparky content” reeks of a desperate attempt to maintain relevance in a rapidly changing landscape. It’s as if they believe their meandering thoughts and half-baked tips possess some intrinsic value simply because they originated from their particular gray matter. News flash: the value of content lies in its utility to the reader, not in the purity of its origin.
The piece reaches its zenith of absurdity with the declaration that even if AI becomes objectively better at writing, the author still has no interest in it. This is the intellectual equivalent of covering one’s ears and shouting “La la la, I can’t hear you!” It’s a stance so myopic it makes Mr. Magoo look like an eagle-eyed visionary.
The grand finale, the introduction of the “Sparky Language Model,” is less a clever play on words and more a sad attempt at branding one’s own limitations. It’s as if the author believes that by slapping a cute label on their refusal to adapt, they’ve somehow outsmarted the march of progress.
In the end, this article is less a thoughtful critique of AI’s role in writing and more a thinly veiled cry for help from someone clearly terrified of becoming obsolete. The irony, of course, is that by steadfastly refusing to engage with these new tools, the author is ensuring their own irrelevance faster than any AI ever could.
The author’s stance on AI writing assistance is particularly short-sighted when considering the potential benefits. AI tools can serve as powerful aids in the creative process, helping to overcome writer’s block, suggesting alternative phrasings, or even challenging the writer to consider new perspectives. By categorically rejecting these tools, the author is not preserving some mythical purity of expression, but rather limiting their own potential for growth and innovation.
Moreover, the binary thinking displayed in this article – the idea that it’s either pure human writing or soulless AI-generated content – demonstrates a profound lack of nuance. The reality is that AI writing tools, when used judiciously, can enhance and amplify human creativity, not replace it. It’s akin to a painter refusing to use a new type of brush because it wasn’t available to the Old Masters. Tools evolve, and so should our approach to using them.
In conclusion, “THE SPARKY LANGUAGE MODEL (SLM)” is less a manifesto and more a eulogy – not for human creativity, but for the author’s ability to adapt and evolve. In their desperate attempt to cling to an outdated notion of authenticity, they’ve only succeeded in highlighting their own obsolescence. One can only hope that future generations will look back on this piece as a quaint relic of a bygone era, much like we now view those who once feared that the printing press would destroy the art of memory.
Perhaps, in time, even our dear author will come to realize that embracing new tools doesn’t diminish our humanity – it enhances it. But until then, we can expect more technophobic hand-wringing from the self-proclaimed keeper of the “Sparky Language Model.” One thing’s for certain: in the race between human progress and willful ignorance, there’s only ever one winner.
I had to hand it to Claude. Caustic indeed! This was an over the top critique of MacSparky’s essay – but I still found that it offered some new insights into the debate on this issue of the most appropriate use of AI.
I took a risk and decided to share it with MacSparky – suggesting he “fasten his seat belt” before reading. Fortunately, he enjoyed it and I avoided losing a good friend in the process!
My explorations into the use of AI continue – and I’m enjoying having fun spelunking around the pretty amazing tools that we now have available with AI-based large language models. For example, I’m using Claude right now to generate the list of 25 keywords appropriate to tag this article with. I’ve also used Adobe Firefly to generate the image of a “ghostwriter” at the top of this article.
The Future of AI Assistants Lies in the Marriage of Memory and Reasoning
In a recent edition of the BG2 podcast, venture capitalist Bill Gurley expressed his excitement for future AI assistants that can remember important details about users, just like a human personal assistant would. Gurley is right – memory is indeed a crucial capability for AI to become truly helpful. However, the argument for robust memory must be coupled with the equally important need for strong reasoning abilities. I’m sure Gurley would also agree with the importance of this – as any truly effective assistant requires both.
Why Memory Alone Falls Short
While remembering past conversations, preferences, and appointments is a valuable step, memory alone leaves AI assistants functionally incomplete. To be truly useful, AI needs to make logical connections between all those details it remembers. This requires the ability to:
Understand causal relationships: AI should be able to identify cause-and-effect chains within the data it stores. For instance, research by [cite LeCun et al., 2015] demonstrates progress in training AI models to learn causal structures from data.
Make inferences: Based on existing information, the AI should be able to draw conclusions and fill in missing pieces. This aligns with the concept of “commonsense reasoning”, a crucial aspect of human intelligence that AI assistants are striving to achieve.
Synthesize information: AI should be able to combine information from different sources to create a comprehensive understanding. We see early glimmers of this emerging in current tools.
The Power of Combining Memory and Reasoning
This powerful combination unlocks a world of possibilities for AI assistants. Let’s explore some examples:
Scenario 1: Proactive Trip Planning
Memory: An AI assistant remembers you mentioning your love for hiking and historical landmarks.
Reasoning: When you mention a future business trip to Seattle, the AI can connect the dots and suggest exploring hiking trails and historical sites in the area, creating a personalized itinerary for your free time.
Scenario 2: Context-Aware Scheduling
Memory: An AI assistant remembers you have a big presentation next week and you mentioned feeling stressed about it.
Reasoning: By understanding your work patterns (e.g., periods of high productivity and flow), the AI can suggest scheduling additional focus time in your calendar leading up to the presentation, allowing you to prepare effectively and manage your stress.
Scenario 3: Personalized Learning Support
Memory: An AI assistant tracks your progress in an online learning course and remembers your areas of difficulty.
Reasoning: By analyzing your learning patterns and identifying knowledge gaps, the AI can suggest additional resources (e.g., specific video lectures or practice problems) tailored to your needs. This personalized support can help you learn more effectively and efficiently.
Scenario 4: Proactive Health Management
Memory: An AI assistant monitors your fitness tracker data and medical records.
Reasoning: By identifying trends in your health data (e.g., changes in sleep patterns or activity levels), the AI can proactively suggest preventative measures or recommend consulting with a healthcare professional if necessary. This early intervention can potentially lead to improved health outcomes. We see some early examples of this with some of the health-related features of the Apple Watch.
Scenario 5: Creative Brainstorming Partner
Memory: An AI assistant stores your past creative projects, ideas, and inspirations.
Reasoning: By analyzing your creative style and drawing connections between seemingly unrelated concepts, the AI can generate new ideas and suggestions to spark your creativity and help you approach problems from a different angle. We see an early example of this combination with Google’s NotebookLM experiment.
Beyond Convenience: The Ethical Implications
While the potential benefits of AI assistants with robust memory and reasoning are undeniable, it’s crucial to acknowledge the ethical considerations that come with it. Ensuring user privacy, data security, and unbiased decision-making are critical aspects that need careful attention as this technology evolves. Debates about where information is both stored and processed with be an important part of this conversation.
Conclusion: A Symbiotic Partnership
The future of AI assistants lies not solely in memory, but in the synergistic relationship between memory and reasoning. Just as a human assistant learns your preferences and makes smart recommendations over years of working together, AI assistants need the ability to remember details and logically determine how to use them to help you. Many of us remember the movie “Her” which was an early pointer at this potential evolution of technology. With strong memory and reasoning capabilities, AI could achieve its promise of delivering personalized support to make our lives easier, more productive, and ultimately, more fulfilling.
