• I was making below minimum wage. You have to be a special type of worker to be fired when making $3/hr.

• My cousin fired me, because our mutual boss was too conflict-avoidant to do it herself.

My first job was making sandwiches at a Subway in Beltsville, MD. (Not the current one; the original has long since been demolished.) I worked the 4-7pm shift after school, having gotten a work permit the moment it was allowed (which in MD is at 14 years of age).

This was way before Jared. Not just before Jared was busted — I mean before we even knew about a guy named Jared. 1990 was, in fact, a time when Subway debuted the inane practice of canoe-cutting the bread. That’s right: instead of cutting a piece of bread the way any human being would, Subway decided at the time that its clever differentiator would be to cut a loaf of bread the way you might hollow out a canoe from a large tree trunk, carving a V-shaped trench into the loaf, leaving the top piece comically small like an inadequate carbohydrate toupée barely covering the filling.

I could make sandwiches like the dickens. I was a veritable maven of meats, industriously assembling foot-longs like Henry Ford himself with adept snaps of the wrist. No, I wasn’t fired for my sandwich-making ability. I was fired because I stood behind the counter waiting for customers.

My boss had a (totally reasonable) expectation that I would clean the tables or mop the floor in between customers, especially during the more idle times around 4-5:30pm. But the key is that she never once told me that was an expectation. Sure, as a 50 year old, it’s crystal clear to me that:

• when you’re on the boss’s time and you’re idle, you should find useful things to do

• you should ask how you’re doing in the job every once in a while

But 14-year-old Philip didn’t know any of this. 14-year-old Philip felt that standing at attention behind the counter in between customers was the best representation an employee could make for the Subway franchise prior to Jared holding large pants open on national television. And his manager never once asked him to do any differently. She just one day decided it was no longer worth her money to pay me below minimum wage, and asked my cousin to fire me.

Paper Folly

Did you know the newspaper lobby convinced Congress to pass a special law in 1949 exempting them from child labor wage regulations when delivering newspapers (ironically called the “Fair Labor Standards Amendments”)? This is how I ended up getting paid $0.07 per Washington Post delivered in the late 1980’s.

My first job, technically, came before being fired by Subway. It was delivering then then-ubiquitously-popular Washington Post newspaper. These were the glory days of tree cutting, back when Sunday editions could easily hit seven pounds each (not a joke — fully documented), with issues regularly cresting over 500 pages (!). You could in no way ride a bicycle delivering 54 issues (378 lbs.) on my route unless your “bicycle” was a Honda Valkyrie towing a trailer.

I was thirteen years old at the time, alongside my brother pushing a shopping cart filled with several hundred pounds of WaPo every morning before dawn through eastern University Park, Maryland. This was no “sunlit newspaper boy with a small bike basket tossing light rolls of newspapers onto porches” deal. This was walking up to the doors of 54 houses down twelve streets every morning in the dark, through rain and snow. The newspapers were too heavy to throw, so you had to walk up to each door individually.

Well, you didn’t have to. I tried at first to throw a few, which seemed like much hilarity until one went clear through the glass front door of a house on Pineway (in fact — and here I must insist I’m not making this up — this very house, which used to have a slatted glass door until I arrived on the scene). Turns out, despite paying its child workers below minimum wage, Washington Post had a slush budget set aside for this very thing… which was a relief, because replacing the two broken panes of glass cost $110.

Which was, if you’ve been paying attention, a month’s wages. Because delivering the post was equivalent to walking every house in the neighborhood in the dark collecting seven pennies from every porch.

My brother and I took the job in January in hopes of lasting until Christmas, when some families would add bonuses to the monthly bill — bonuses which, when you’re making seven pennies a delivery, matter a whole lot. But the job was much harder than I had realized, and I argued a bunch with my brother until we quit in April.

I’m not sure there’s a moral to this story. I’ve felt guilty for a few decades that I disappointed my brother by quitting before Christmas, since a several-hundred-pound shopping cart is clearly not navigable by one teenager alone through rain and snow. At the same time, I can’t say that I would, in hindsight, rather have stuck with the job all the way through the entire year, because I’m thankful I didn’t spend an hour every morning for an entire year to collect what was likely an extra $100-200 that Christmas.

It’s Academic

My job during college was to code for $8/hour at Academic Data Systems, a department at the University of Maryland focused on mainframe data processing. At the time, all coding was done on IBM MVS running System/370 using a language called IDEAL, which was, I assure you, very far from ideal. For instance, IDEAL had no local variables, such that 30 kLOC programs worked from only global variables. You can imagine the nightmare of maintaining such systems.

I remember two embarrassing incidents from this time:

1. I was happily listening to thumping techno music through my wired headphones in the middle of our cubicle farm, but a bit confused why everyone walking was staring at me. Nobody said anything, but it turns out my headphones weren’t plugged in all the way, so the music was blaring throughout the office.

2. I had created VENUS, a course scheduler where you could see different layouts of possible schedules. This was at a time (~1995) when cookies weren’t even standardized on the internet, so it was a lot more innovative than it now sounds. The program was a big success at the university, so I wrote several other universities asking whether they’d want to buy it. I had proposed and single-handedly created the entire VENUS system, so I naively thought it was mine to sell. I still can’t tell if I was just atypically clueless about this type of thing, or whether many 19-year-olds would feel the same. In any case, one university contacted the head of Academic Data Systems and I got a stern talking-to. Only after that discussion did I understand that, when creating things on someone else’s dime, all proceeds from those creations are theirs to keep.

Early Microsoft Days

I did many embarrassing things in my first year at Microsoft, but one that stands out as unusually foolish was suspending a CRT monitor over my head. I had been experiencing a bit of back pain from long hours, and wanted to try working from a prone position, lying down on the floor with my monitor suspended over my head between two tables nudged just closely enough to support the bezel without obscuring the screen.

At the time, 19” CRT monitors were typically 55-70 lbs. If the tables shifted apart any more than about an inch, the monitor would fall directly on my face from two feet above me. After operating like this for a few days, I had the presence of mind to secure the base of the monitor with a metal chain I attached to a leg of the table. But in thinking about this even now, I’m pretty sure my setup would never have passed OSHA and I could have died a gratuitously-self-induced death for the sake of ergonomics.

But 22 years old is a silly time. A coworker of mine at the time, who’s now an SVP somewhere and had also headed major teams at Google at Meta, asked his manager whether he could take a longer vacation by taking on a bunch of sick days. Another coworker would frequently take home backpacks full of sodas, an action he fondly called “JanSporting it.”

These follies all seem ridiculous now, so much so that it’s hard to believe there was a time when they all seemed totally fine and natural. But I guess that’s the nature of growth and perspective — things we once believed and took for granted as acceptable sometimes seem foolish in hindsight.

Listening to village elders made sense for eons because things didn’t change decade to decade or even generation to generation. Despite the modern myth that you stand a better chance to escape a crocodile by running in an arc (because crocodiles run much faster in a straight line), empiric evidence from experts is that sprinting in a straight line affords you the best odds of survival because crocodiles tire easily and will only give chase a handful of yards.1

Your great-grandmother would surely have known this if you lived with her several millennia ago. When things don’t change, you get wiser as you age simply by remembering lessons from your past. Oral traditions persist those lessons across generations — lessons which, in a world of stasis, remain perpetually relevant.

But is that still our world? Or do things now change so quickly that past lessons are oft outmoded by the time you tell the next generation?

If I’m honest, it hurt me just to type that previous paragraph. Veneration of filial piety is at the absolute core of the Confucian ethic which I and many other East Asians were brought up with. So it’s difficult to even permit the possibility of questioning the wisdom of elders.

Perhaps it’s easier if we look at ourselves.

Is Self-Improvement a Myth?

Most musicians’ biggest hits were made in their 20’s. Einstein’s world-changing thoughts on special relativity were published at age 26 while working as a patent clerk. Infinite Jest (33), Pilgrim at Tinker Creek (29), Jane Eyre (31), Wuthering Heights (29), Frankenstein (20), The Great Gatsby (28), The Sun Also Rises (27), The Red Badge of Courage (24). Even The Hitchhiker’s Guide to the Galaxy (27)!

If we get better with age, where’s the evidence?

Is self-improvement a real thing, or merely fixity of thought as we age? Are we simply convincing ourselves we’ve become better year over year? If we truly improve as we age — and most people seem to believe this — why is it that we don’t think of most older people in our society as complete mensches? Why is it, in fact, that the average person’s assessment of old people, other than their own grandparents, is slightly (or even greatly) negative?

Despite this, why is it that, other than physical ability and the priceless luxury of yet-unlived years, most people wouldn’t want to revert to their younger selves if given the chance? As we age, we come to see some of our past convictions as foolish, some of our past behaviors as regrettable. I’d like to think this is proof positive we improve with age.

And yet, look around. I don’t think anyone’s making a real argument that older people are in net better than the young. In fact, it’s clear most people also lose things — very positive things — as they age. Optimism is replaced by cynicism. Creativity by structure. Spontaneity by habit.

All old people think their generation was better, that the world has gone downhill since their youth. It’s not that they’re improving in comparison with younger people. They’re actually talking about older times being better. This, at least, is clearly a myth, since this perspective is shared across every generation as they age. When we complain about “kids these days,” we aren’t making an insightful observation about the degradation of society. We’re making a declarative statement about ourselves, about our subjectively favorable opinion of our own youth.

Or Is It Just Me?

Maybe you feel old people are clearly better than the young, that our rapidly changing society hasn’t come to favor youth if not simply for the fact they’re more adaptable. Perhaps you talk about all the wisdom bombs that the elderly are dropping on you all the time. Mayhap you consult the aged, use their products, follow their example, consume their creations.

In fact, is this very article proof-positive that I’m the one, having turned 50 this year, who has now become more crufty — doubtful the world is as good as I thought when younger, eager to point out that perhaps, after all, personal “growth” is an illusion belied by the simple act of looking around? Perhaps mine is a minority view, tainted by my own cynicism.

It’s hard to remember that every divorce started as a wedding. That if I live to be 85, I yet have as much time and possibility ahead of me as Mozart’s entire life, birth to death. That it’s possible still to look in a baby’s eyes and not just lament the suffering, hurt, and despair it will no doubt one day experience, but also see all the joy and wonder. That change is possible — even now, even still, even yet.

(If that last paragraph resonates with you, I highly recommend watching the short film Trading Cards by Radheya Jung, which I saw at this year’s Palm Springs ShortFest. It does a great job elegizing the many positive things we lose as we grow into adulthood. I hope it’ll be available for general viewing once its film festival rounds are complete.)

Age Favors the Unchanging

Maybe the right way to see things is that age favors the unchanging.

Eating the wrong mushroom has always killed in the past and will likely continue to kill well into the future. So it’s helpful to be old enough to have witnessed one or two people dying after eating certain types of mushrooms so you can warn gastronomically wayward younglings.

Human nature hasn’t changed through the ages. You get a sense of this when reading how the Stoics describe the human drama of their day. It’s all recognizable, as fresh as if it were all written yesterday. I feel similarly when I hear Oprah advising young women to “ignore what men say; just judge their supposed love by what they do.” That’s real wisdom, and it’s unchanging.

The question is how much in our modern world remains unchanging, and how much is shifting so radically as to make the opinions of the elderly less relevant — or in some cases, downright counterproductive. This intuition is what causes the common experience of young people rolling their eyes as an elder talks about things “back in the day.” Half the times, of course, the young eye-roller one day comes to see things the exact same way as that elder, and recognizes they were too naïve when younger to see the wisdom in what was said. But at other times, there’s justified exasperation at the speaker’s lack of insight into how the world is fundamentally different from the one in which they first derived their “wisdom.”

What’s tricky is that every old person thinks their advice falls exclusively in the former category, while most young people are convinced the same advice falls mostly in the latter. This tension will only increase as our world changes even more quickly. Once everything is exponential, you begin to wonder whose linear extrapolations from past experience remain prescient.

I go to old people for advice all the time, so perhaps the entire premise of this article is hypocritical. I guess what I’m saying is that, as I myself become older, I begin to see both the ways in which the elderly can hold on to some now-irrelevant views, some of which clearly mislead or hold back society, and the way in which I likely need to temper my convictions about my own insights even as the world changes all around us.

I’d like to think I’ve improved in my life. I’m just less positive about that than I used to be.

At the time, Facebook was 500 engineers in total, small enough for all the senior people in every major product area to know each other. I was 34 years old at the time, and remained the oldest person in the Seattle office for its first six months. This was before Google+, before The Social Network. MySpace was rumored to be coming back. People forget there was a time when Facebook’s future was far from certain.

I had joined Facebook after working twelve years at Microsoft. I was excited to move back into an IC role where I could focus purely on shipping things, at a company where the speed of shipping reigned supreme. I had no idea the company even had career ladder levels; it’d be an entire year before I discovered I was an IC7.

I chose to work as the only full time engineer on video calling back in 2010, but was eventually joined by some true superstars through V1 and V2:

• Thiago Hirai, who later led projects at Snap and OpenAI.

• Denise Noyes, who rose to be a Distinguished Engineer at Meta.

• Vijaye Raji, the current CTO of Applications at OpenAI.

I knew they were amazing even at the time. But there’s a way in which, in hindsight, one is astounded at the company one kept, at the tremendous serendipity of it all.

The Months Leading Up

Having come from Microsoft, which was at the time perhaps 140,000 employees, it was refreshing to be on a tiny, tiny team trying to ship Facebook’s first version of video calling. We had only a fractional PM and designer, and for a while, I was the only full-time developer. I’d soon be, thankfully, joined by Denise, Thiago, and Vijaye.

The company and the team were so small at the time that we did product reviews directly with Mark Zuckerberg. It seemed nothing but natural at the time, and I appreciated how he pushed us to test the extremes. For instance, Mark once suggested we not require people to explicitly answer a call, but instead for calls to immediately be connected. The idea was to simulate aspects of real life (e.g. a coworker can walk right up to you). We sat aghast for a moment at his suggestion, after which Mark revealed he wasn’t really suggesting that approach, but instead just testing the limits to which we would explore the idea of immediacy and in-person interactions.

Video calling was much tricker at the time because WebRTC didn’t have widespread adoption. Chrome and Firefox weren’t even the dominant browsers. So establishing a call wasn’t anywhere near as trivial as it is today. We had negotiated a contract with Skype, prior to it being acquired by Microsoft, so that the underlying video technology would be provided by their client-side binaries, which then had to be installed on people’s computers before calls could be made.

Skype is proof positive that even a terrible product can be sold three times to different buyers (eBay, Silver Lake, Microsoft). I had been a casual user of Skype until that point, but working directly with Skype in order to ship video calling exposed me to the dark underbelly of terrible technology. At the time, I was told that Estonians working in Skype were local superheroes, treated pretty much as Ajay Bhatt entering any arena, complete with wind-swept hair.

But the code itself performed terribly. Skype engineers were incredulous — positively gobsmacked — when I showed them data suggesting call-establishment success rates of 82%. (Video calls take a few back-and-forth messages before a connection can be fully established. At the time, 18% of attempted call establishments through Skype ended in failure.) When I couldn’t convince anyone at Skype that the connection success rate was so bad, I asked them to instead show me their own data about their call establishment success rates. The answer at the time was — and I insist I am not making this up — “We don’t have those numbers.” The logic, as it was explained to me, was that the only truly useful metric was user-reported call satisfaction. Never mind the Skype UI didn’t ever ask you how your call was if it never got connected in the first place. It was a philosophical argument akin to trees falling in forests.

A few individuals I worked with at Skype were stellar. Perhaps the lesson is that even a few stellar folks can be trapped within a subpar organization to produce terrible products. Or perhaps it’s a big-fish/small-pond lesson around being skeptical when people treat you as a rockstar, to look around at the broader world to be sure you’re as amazing as people say you are.

Turning the eye back on ourselves, a big benefit of using Facebook Messenger’s data channel for initial call establishment was confronting the fact of Facebook’s own poor delivery reliability back in 2010. Through video calling’s logs, the team got much stronger evidence around the percentage of messages which simply went undelivered. And, just as the Bluetooth stack on iPhones had poor reliability until AirPods were released, video calling gave Facebook a strong reason to improve underlying infrastructure in order to make a feature possible.

Factoid: the ring tone of Facebook’s first video calling feature was F-A-C-E, a tidbit which those with perfect pitch no doubt noticed upon its release. This stroke of brilliance was the brainchild of an audio designer who I don’t remember meeting in person, but whose cleverness lives on.

Shipping the Product

Facebook was so small at the time that I was one of the key people meeting with the press to launch the feature. It felt a bit unnatural to get “press-trained” by a member of the PR team prior to being allowed to speak with reporters, but I can see why that’s considered a required step at most companies (prior to setting The Crazy Engineers™ loose to say whatever it is they want).

The six weeks after launch were the hardest sustained number of hours I’ve ever worked in my career. I moved our young family down to Palo Alto because I knew work would be crazy. Right after launch, all the bugs started surfacing, requiring hotfix after hotfix. I worked literally all my waking hours, starting around 8am and ending around midnight every day. My wife remembers it as an idyllic time for our toddlers. They visited places like the Palo Alto Junior Museum & Zoo and the local library almost daily.

By the time the post-ship crunch was over, the product was in much better shape, connecting more reliably and delivering a new experience many people hadn’t had with their loved ones. Facebook video calling was the first time I had spoken with my parents over video. It felt like magic to bring people closer with a free, easy-to-use form of video calling.

Sixteen years later, after the multi-year crunch mode of COVID, video calling is a staple in people’s lives. WebRTC has made it so that even small startups can release products which include video connectivity. Everybody is used to reminding others to unmute themselves. But there was a time sixteen years ago where none of that was a given, and it was wonderful to experience the baptism of fire that was my induction into the first version of Facebook Video Calling.

Life as a Whole

• Four Thousand Weeks, by Oliver Burkeman. A reminder of how precious little time we have, and the importance of treating every single week as a substantial block of our entire lives.

• In a similar vein, Tim Urban’s The End of Life was a revelation on how little time we have with our parents. By the time you leave home, you will have spent more than 90% of the time you’ll ever spend with your parents even if you visit them 10 days a year. Two years ago, I moved my parents from Maryland to our home in Seattle when my mom was diagnosed with advanced liver cancer. It was invaluable to spend those final 9 months of her life in person with her.

• The Practicing Stoic by Ward Farnsworth. This makes the otherwise disorganized work of famous Stoics like Seneca, Epictetus, and Marcus Aurelius readable. Focus on what is within your control. What is said isn’t itself offensive; your choice to be offended turns it into an offense. Was your broken leg a travesty, or perhaps the lucky happenstance that saved you from subsequent devastating loss? By realizing our responses to events dictate our experience of life, we turn almost everything around us into something within our control. I’ve read this book multiple times in my life and find it profoundly helpful each time.

• Meditations by Marcus Aurelius is one of the classic Stoics texts, straight from the horse’s mouth. These diaries were were published posthumously without its author ever meaning for them to be read by the public. Knowing this makes it even more remarkable how wise an emperor Marcus Aurelius was. The book makes concrete the daily struggles of living well, and reminds us how people even thousands of years ago were pretty much the same as today.

• Man’s Search for Meaning, by Viktor Frankl. Freud, with nothing but untested philosophical principle, claimed he could predict the behavior of any man by first reducing him to starvation. Frankl, having lived through a WWII concentration camp, proffers empiric evidence that people behave very differently from each other — even when starved equally near to death — based on their individual sense of morality. This book is a reminder of the depth of human suffering as well as the existential importance of hope as a motivator of our lives. Frankl feels that asking the meaning of life while in the midst of living it is like asking the meaning of a movie in the middle of watching it. We can only truly say what the meaning of any life is once we see the entirety of its arc.

• I used to recommend Delivering Happiness (by Tony Hsieh) as a guide on how the interests of employees, shareholders, and customers can all be satisfied if we see the world as positive-sum. Given Hsieh’s subsequent apparent suicide after years of struggling with drug addiction, I’m now more likely than not to instead share Jewel’s letter pleading with Hsieh to, in a way, save his own soul. Delivering Happiness is nowadays a reminder to me that there’s a vast canyon between understanding what’s good for our lives and living it.

Systems Thinking

• Meditations on Moloch, by Scott Alexander, for which this Substack is named. Everything is determined by systems — which, if you understand them, predestine outcomes. Everyone can live in a little house in the neighborhood if only one parent works. Once a few families start earning dual incomes, the price of housing will quickly require both parents of all families in that neighborhood to work in order to live in those exact same houses. If you bail out banks each time they fail, you should not be surprised when they take increasingly larger risks that cause even more failure. We debate the outcomes of systems more often than addressing the design of the systems themselves. This one insight, if applied to human governance systems, represents the greatest single improvement we could make to global happiness and wellbeing.

• Sapiens, by Yuval Noah Harari, is literally epic in scope, covering the greater arc of humanity at a scale I had not contemplated before. Fire, religion, capitalism, enlightenment, corporate personhood, democracy — there are a handful of key ideas that additively changed our collective fates. If anything, Sapiens recounts how the greatest of human systems (e.g. the cooking of foods, the advent of agriculture) have shaped the lives of everyone today.

Money and Finances

• The Intelligent Investor by Benjamin Graham and the annual letters of Warren Buffett for Berkshire Hathaway. “Can one of the most successful investment firms of all time have a website that still looks hosted by GeoCities?” Yes — most definitely, yes. When you are as successful as Buffett and Munger have been, you can afford to have a website that looks like Mosaic rendered it in 1994. The biggest takeaways from Graham and Buffett are the seemingly simplest principles: stocks are the fractional ownership of companies; the market is a voting machine in the short term but a weighing machine in the long; whenever you buy a stock at a certain price you think is a good deal, someone is selling it to you because they think the exact opposite; we’ll know who’s swimming naked when the tide goes out; buy great companies at good prices.

• The Truth About Money by Ric Edelman. There are far more books with everyday money advice nowadays, like Rich Dad Poor Dad, and I’ve also not reread this book since it came out over two decades ago so I hesitate to vouch for it. But The Truth About Money influenced me a lot in my early 20’s, just as I started to earn my own money, by encouraging two things: a) saving when you’re young and investing in a diversified portfolio, and b) treating your house as a place to live, not as “your biggest investment.”

• The Psychology of Money, by Morgan Housel, is a contemporary book which I do vouch for and highly recommend. It’s the one book I purchased for each of our kids when they began to earn their own incomes. Far more important than sage investment advice is understanding how your psychology dictates financial outcomes far more than any clever strategy. This book is modern wisdom in the guise of financial advice.

• Die With Zero, by Bill Perkins, is an amazing book saddled by a terrible title. Did you know that the average American receives their inheritance in their mid 60’s when their parents pass away? How much do you think you’ll need your parents’ money by the time you’re in your 60’s? This book single-handedly changed the way we think about giving to our kids. Our plan is now to give them their inheritance in their 20’s to early 30’s — an age when our own lives would have been most improved by an injection of money. My other takeaway is that you’ll die a hundred deaths before your final, physical death. There will be a last time you lift your child up in your arms; a last time you ski; a last time you ride a boat; a last time you travel internationally. The insight is that most of those times, you don’t realize it’s your last time. Knowing this, you should arrange bucket lists for every 5 years in your life as opposed to having just one big Lifelong Bucket List, front-loading activities you may not have the interest or ability to do later in life.

The Self

• Solitude and Leadership by William Deresiewicz, who also wrote Excellent Sheep: The Miseducation of the American Elite and the Way to a Meaningful Life. As Susan Cain wrote in Quiet, America values the loud and brash leader who stands out front. But there’s value in the contemplative leader who in fact speaks not as a way to garner attention but as the inevitable outcome of having thought through and reached an important conclusion. Solitude and Leadership was Deresiewicz’s address to the cadets at West Point to remind them that true leadership cannot just be the remixing of thoughts whipped together from 8 hours of nonstop meetings every day of the week. True leadership requires the space and solitude that nurtures novel insights.

• Citizenship in a Republic, Theodore Roosevelt’s 1910 address at the Sorbonne. “It is not the critic who counts; not the man who points out how the strong man stumbles, or where the doer of deeds could have done them better. The credit belongs to the man who is actually in the arena, whose face is marred by dust and sweat and blood; who strives valiantly; who errs, who comes short again and again, because there is no effort without error and shortcoming; but who does actually strive to do the deeds; who knows great enthusiasms, the great devotions; who spends himself in a worthy cause; who at the best knows in the end the triumph of high achievement, and who at the worst, if he fails, at least fails while daring greatly, so that his place shall never be with those cold and timid souls who neither know victory nor defeat.” If you judge all past writings through the lens of modern conventions, you’ll pass a kidney stone at Roosevelt talking about women cooking. But an inability to see past that would blind you to a host of wisdom around our higher calling to make for us a great democracy, and to reflect on the type of men and women required to sustain it.

• The short stories of Alice Munro. I didn’t want to just focus on Hateship, Friendship, Courtship, Loveship, Marriage: Stories — winner of the Man Booker Prize — or just on Alice Munro, for that matter. I wanted some representative of the importance of fiction in shaping one the most important aspects of my life, which is how I think of my relationship to others. The Great Gatsby, A Separate Peace, Of Mice and Men, To the Lighthouse, The Lord of the Rings, The Fault in Our Stars, Les Miserables… there are just too many to name. One of the greatest strengths of fiction is to remind us what we already know but have been blinded to by the barrage of everyday life. Munro’s writings consistently do that by exploring the space between us.

I’m sure I’m missing plenty from this list, but it’s a start. I’d be interested to hear some of yours in the comments!

• ReachVIP (https://reach.vip). You should be able to reach anyone in the world, no matter how famous, if they name their price. ReachVIP lets any expert create a profile and set prices for what they’d charge in order to answer your question.

  • Luminaries wouldn’t do this, you say? A friend of mine received a birthday video from Kirsten Bell via Cameo. Kenny G sent a birthday video to my wife. Most people are happy to do a little something helpful in their day if they’re compensated for something straightforward.

  • You can already book 1:1 videos with people like the founder of Reddit with services like https://intro.co. Video coordination is hard, though, for both the expert and the client. ReachVIP instead offers text-based inquiries that don’t need schedule coordination. It’s like being able to email any expert in the world by paying for a thoughtful response.

• Thankverse (https://thankverse.org). Thank anyone in the world with a thoughtful note online which they get to keep forever. In high school, I went to a camp where, on the last night, everyone wrote everyone else a note of appreciation. Those notes were treasured by me for decades. I wanted to create the same experience for people, but in a digital world. There’ll never be ads on Thankverse; it’s meant to be something positive for humanity.

Each of these went from ideation to launch within one weekend, which opens the door to several observations.

The Consequences of Easy

The ways in which AI supercharges creation are obvious:

• Non-coders can now make their ideas a reality. It’s going to be a PM and Designer heyday.

• Individuals and small companies can accomplish heretofore unimaginable amounts of work.

• Deep expertise in all areas is no longer a prerequisite to prototyping. I built ReachVIP’s Stripe integration without ever having looked at the APIs.

But the consequences may not be as obvious at first blush.

Cheap Things Are Wasted

Book publishing used to exert a strong selection bias toward quality writing because printing physical books was both expensive and time consuming. Nowadays anyone can put something on Amazon via CreateSpace, uploading a PDF for free and collecting royalties only if the book sells. You can bet the average quality of books has gone way down.

Similarly, just because you can build AcmeSite.ai doesn’t mean you should. Back when developers were expensive and prototypes took weeks, you’d think long and hard before commissioning the work. These days, the whole Lean Startup mindset has gone to extremes as more and more vaporware / abandonware is released online.

Discovery Becomes Impossible

Some time in the late ‘90’s, there was a comedic trope about having “200 channels of cable but nothing to watch.” There was CourtTV with nonstop gavel banging. The Box, where you paid by telephone to queue music videos. Prevue Channel, where the only thing happening was a scrolling schedule of other channels’ content.

You might suppose moving from three national TV stations to 200 cable channels would be an unequivocal boon to viewers worldwide, but that’s not what happened. Instead, the relative ease with which you could create new channels led to a morass of slop.

If you think the App Store’s 2 million apps are already a problem, brace yourself for the world of 160 million apps.

The crappy cable problem was solved over 20 years through a combination of:

• Quality aggregators. When you tuned in to HBO, you were paying for discernment and taste.

• Personalization at scale. The promise of personalization was hyped even in the year 2000 by Pandora, though its algorithms were bad enough nobody subscribed. Nowadays, however, YouTube and TikTok recommendations have gotten so good that global tastemakers are no longer necessary. You can now have a channel which you always love.

We’ll no doubt also experience this transition in the apps/sites domain over the next few years, but in the meantime, buckle your seatbelts for a barrage of bad apps.

Copying Goes Thermonuclear

Copying an app has always been far easier than building it in the first place because the initial inventor threw away 90% of various experimental features until the core value prop was discovered. Copying requires that you only do the 10% which survived, with much less market risk at that.

AI takes this to a different plane altogether. If you’ve never pointed AI at a web app and asked it to simply replicate its entire functionality, you ought to experience such magic at least once. There’s little doubt we’ll soon be able to replicate mobile apps just as easily.

Many traditional software businesses were built on the premise that software is hard, and that the code itself forms a strong barrier against competition. With AI, you’ll need to rely on other moat-building principles, like branding, network effects, or scale efficiencies.

Calendly is a great example. Anyone can replicate its functionality in a weekend. In fact, Google Calendar now offers the same functionality for free. Calendly’s business is 100% based on the brand it established via rapid adoption shortly before the AI revolution.

The Timeliness of Now

If all of this didn’t discourage you from building, I’ll add another major problem: the breakneck speed of improvements in AI exerts deflationary pressure on the value of your time.

Is an hour of your time building with AI more valuable now, or six months from now? If you wait six months, you’ll likely be able to build whatever you wanted to build today in 25% the time.

Most of us have only ever experienced inflation so we tend not to have strong intuitions about human behavior under deflation (as there was in Japan between the ‘90’s and ~2012). If you knew a dollar would buy more next month than this month, you’re going to put off most purchases until you absolutely need them.

This is the sense in which improvements in AI cause deflation in the value of your time. You would understandably be tempted, with that logic, to defer building things.

I’ll here make the argument you probably should build now anyway:

• First movers often still have the advantage. See Calendly above. You could invent and define a new category of app. Your app’s name could become a verb.

• There’s joy in the building. Overcoming difficulties is part of the fun of it all. Games are, in a way, putting up deliberate obstacles for ourselves in order to feel the sense of accomplishment once overcome.

• And of course, life isn’t forever. During deflation, a dollar might buy more for you tomorrow than today, but it certainly buys you nothing 90 years hence.

I built ReachVIP with a talented Nigerian designer, Daniel Afolabi. The interpersonal experience of collaborating with Daniel was truly enjoyable. But if I’m honest, it may be the last time I work with a designer to prototype ideas. AI is getting to the point where the price differential (several orders of magnitude) and speed differential (waiting several days for a design vs. getting an updated design within minutes) no longer justify the cost of human design labor for small firms or solopreneurs. The import of hiring human coworkers is fast becoming a luxury belief which only the wealthy, or those who don’t sign the fronts of paychecks, can afford.

With that said, I’d especially encourage you to build now, not later, if you have one or two coworkers to brainstorm alongside. Such opportunities will likely become rarer in the near future.

• Anaximander (~570 BCE): The world can be explained by natural causes instead of myth. The author, Rovelli, treats Anaximander as the start of science.

• Democritus (~400 BCE): Matter as discrete particles (“atomism”).

• Aristotle (~350 BCE): The four elements (earth, wind, fire, water) each have a “natural place” toward which they return. Not a correct theory, but one that dominated Western thinking for over a millennium.

• Copernicus (1543): The Sun, not Earth, is at the center of our then-comprehensible universe.

• Galileo (1632):

  • Things fall at the same speed (absent air resistance)

  • “Relativity of motion”: there’s no experiment which could tell you whether you’re moving or still within a smoothly-moving closed system. This, coming nearly 90 years after Copernicus, finally answered one objection of heliocentrism: why don’t things fly all over the place if we’re circling the Sun?

• Newton (1687):

  • Everything attracts everything else with calculable force (i.e. “universal gravitation”).

  • Things exist in an absolute space & time. Even if the world had only one object in it, you could state “where” that object is.

• Faraday and Maxwell (1800’s): “Fields” like electromagnetism govern forces and give structure even to “empty” space.

• Einstein (early 1900’s):

  • Special relativity makes absolute simultaneity nonsensical. Light is quantized as “photons.” There is no absolute space & time.

  • General relativity posits that gravity is not a force but a distortion in space and time.

What strikes me isn’t just how fundamentally mind-shifting each of these advances are, but that they’re separated so widely in time. Look at those gaps! What was happening all those years?

Humanity Beyond Physics

As important and life-changing as the above ideas are, physics is just a small slice of the total space of ideas impacting humanity. If we take a much broader view of key advancements in human thought, the picture is even more dramatic.

With a nod to books like Sapiens, here are some of the ideas that have fundamentally changed the arc of humanity:

• Fire (~400,000 years ago). Warmth is just one aspect. The other is the “cooking hypothesis”: the idea that fire not only eliminates pathogens, but also changes starches and proteins to make caloric absorption much more efficient. It’s unclear whether our brains would have ever developed their modern capabilities without this.

• Agriculture (~10,000 years ago). Towns and cities are impossible without agriculture because we otherwise could not control the proximity of food. Most subsequent major intellectual advancements required the density of cities in order for ideas to intermix sufficiently.

• Writing (~3200 BCE). We’re the only species capable of conveying complex knowledge across generations, allowing humanity to reach heights no single human could. Writing is superior to oral communication because it transports information error-free and survives even the death of all who knew its knowledge (as we’ve seen many times with ancient texts rediscovered by other civilizations hundreds of years later).

• Democracy (~500 BCE in Athens/Rome/India, then dead over 1000 years until it was retried in Europe). The unlock of human rights, self-governance, peaceful transition of power, and resistance to idea lock-in under authoritarian governments made many subsequent human advancements possible.

• Free speech (~400 BCE with Greek parrhesia, “frank speech”). Though sometimes linked with ideas around democracy, free speech as a concept isn’t constrained to a form of government. Free speech is a prerequisite for structural change to improve governance and science. Restrictions here have suppressed many intellectual advancements which humanity could arguably have made sooner.

• Free markets (1776). Adam Smith’s The Wealth of Nations might be an agreed starting point of systematic free market thinking, but various forms of free trade have existed all the way back to prehistoric bartering. It’s become trendy in some American circles to question capitalism, but there’s no doubt that price signals without central direction have unlocked tremendous social benefit in recent centuries.

• Corporate legal personhood (1500’s). Even the Romans had universitas, the idea of legal entities distinct from their members. But I’m specifically interested in the idea of corporations being able to fail without bankrupting their owners. This sparked ventures like the Dutch East India Company: for the first time in history, people could attempt brave products and endeavors without fearing personal bankruptcy. The import of this on human flourishing is grossly underestimated.

• The Enlightenment (1700’s). These ideas didn’t all debut only during this period, but gathering them together under one moniker further emphasizes the heavily-quantized nature of human advancement:

  • Reason. Not authority, tradition, or religion.

  • Empiricism. Things should be measured and testable.

  • Universalism and natural rights. People deserve things beyond whatever their current government happens to give them. These things are shared across all humanity.

I’ll here make a few observations:

1. Each of these ideas probably seems natural and obvious to you… but all of them, other than fire, are new — brand new — in the long-arc history of humanity.

2. Remember that at some point in the past, every one of these ideas was considered not only novel, but something which had never been thought of before. Ideas go from never-having-existed to permanently-world-changing.

3. See, again, how far these advancements are separated from each other. Ideas don’t require a minimum bake time before subsequent ideas can be had. The gaps of hundreds (or thousands) of years between these ideas implies people were stuck with the same notions all those centuries whilst knowing full well all the ideas prerequisite to the next innovation. Those ideas just didn’t occur to them.

4. And I’ve not even covered domains like philosophy — which impacts human thought itself — or medicine, which alone has also revolutionized the human experience. You’ll find the theme is the same in those domains: a few key ideas punctuate hundreds of otherwise eventless generations.

You (Or, “The Individual”)

Every human being experiences the same types of quantum leaps in the lifelong development of their own mindset. Common changes include:

• Object permanence (4-8 months). What? That box doesn’t cease to exist when it goes behind the curtain?

• Self-recognition via mirror test (15-24 months). Hey, that’s me! Not just some random, amazingly cute baby.

• Theory of mind (3-5 years). I bet she wouldn’t know I ate her candy when she left the room.

• Death is permanent (4-6 years). Spot’s never coming back, even if I bury him in a pet sematary? (sic)

• Personal & parental mortality (5-9 years). You’re telling me my parents won’t be here forever? I’ll one day die too?

• Moral reasoning shift (7-10 years). What’s right and wrong might not be fixed, but instead dependent on social and personal context.

• Risk estimation maturation (20+ years). Maybe my homemade bungee cord isn’t the best idea.

• Religious & philosophical shifts (anytime after childhood, if at all). I’m not sure Xenu, the galactic ruler who killed billions of people near volcanoes with hydrogen bombs, is a real thing.

Many of these shifts can only happen once the brain has developed certain capabilities, but there are a host of other life-changing ideas people only discover decades into adulthood which, in theory, can be discovered earlier. The point is that our own lives, much like human history, might mostly consist long periods of stasis punctuated by rare, revolutionary epiphanies which we were never, after all, blocked from experiencing earlier.

The question is how to spark these quantum leaps, to deliberately advance ourselves and our societies as soon as we possess the requisite insights instead of decades (or even millennia) later.

Frustratingly, good ideas don’t just arrive and stay forever. The British first discovered the cure for scurvy in 1617 (via ship’s surgeon John Woodall), then promptly forgot the cure for 130 years before James Lind of the Royal Navy rediscovered it in 1747. Shockingly, the Brits forgot it again in the transition to West Indian limes (1860’s) only to re-rediscover scurvy’s cure 70 years later in the form of vitamin C.

And this, at last, is why I wrote this article — not just to ask how we can surface these insights sooner in our societies and our selves, but also to remind us of some key ideas which have fundamentally improved the human condition.

The good news, if you look back at these lists, is that it only takes a handful of ideas to completely change how we and our societies work. Across Warren Buffett’s 60+ years of investing, he says Berkshire Hathaway’s phenomenal success has “been the product of about a dozen truly good decisions – that would be about one every five years.”

It’s not about having a ton of ideas. It’s about having:

• … the highest impact ideas

• … sooner

• … and keeping from regressing back to the way things were.

Godspeed!

This week, Mariah passed along an insightful article, 5 Rules of AI Writing, which we wanted to use as a basis for back-and-forth discussion. The rest of this post will likely make more sense if you take a moment to first read that article.

Mariah

Thanks, Philip! I’ve been such a fan of reading your thoughtful Substack articles, especially the ones about the very rapidly changing technology environment we’re currently in. I’m excited to explore this particular topic today, because so much of what has been written about AI ethics was written in academic journals before these new LLM models caused a Cambrian explosion of AI use cases. The Rachman article you mentioned above isn't an academic white paper written from an earlier time that posits hypotheticals; it really brings us into the messy ethical scenarios we’re finding ourselves in right now, truly for the first time. In this article Rachman delineates the current tension that exists around AI transparency and the possible structural responsibility needed for AI use in writing so that we don’t end up “downgrading our species’ intellectual conversation into a feedback loop of artifice, in which we are mindless mouthpieces of literate machines.” With that being said, I’m curious how you’re currently thinking about AI transparency and which parts still feel like unknown unknowns?

Philip

I’m so glad you shared the article because it made me consider what sorts of guidelines we’d want to encourage around AI-generated and AI-assisted prose content. For instance, Rule #2 seemed eminently reasonable:

2: Test whether you’re going too far by asking: Would it be wrong to accept this help from a human without crediting them?

But what started out simple got more complicated for me as I considered several analogies further. To wit:

• Most copyeditors are never credited. This is pretty much accepted by everyone, perhaps because copyediting leans more toward what Word’s red and green squiggles do. It’s a form of correction of mistakes, never meant to change the “tone” or “voice” of the author.

• Magazine and newspaper editors are credited once, though it’s understood their fingerprints likely fall throughout the publication’s many articles. In this case, editors are exercising at least two types of judgment for which we feel they deserve recognition:

  • deciding which articles even belong within an issue, and

  • giving broad feedback to authors which can significantly reshape submissions.

• Ghostwriters are never credited. When The Art of the Deal is billed as being written “by Donald Trump with Tony Schwartz,” everybody understands it to mean “by Schwarz based on interviews with Trump.” But true ghostwriters are never credited. The whole point is to pretend the celebrity wrote it themselves, to sell to the type of person who’d only buy a book because it purports to be written by that celebrity.

• And then there’s outright deception. There’s a type which most people don’t judge, like Mary Ann Evans publishing as George Eliot in order to be taken seriously, or the Brontë sisters using male pseudonyms for their first books. But then there’s outright lying, like The Hitler Diaries and Go Ask Alice, where the whole point is to mislead you about authorship.

I got this far in my thinking before ethical nuances started to feel overwhelming. What do you think of the spectrum of credit-giving when it comes to AI? Nobody faults an author for using spell-check. Some authors judge those who use a thesaurus, just like some people judge the difference between genetically bestowed beauty vs. that which was surgically achieved. Then there’s how the public assesses people who have flat abs due to disciplined workouts vs. those who got liposuction. In a way, do you think it’s ultimately about the amount of effort or time a person spends? Do we basically want to know whether a person has sweat as hard as their output suggests?

Mariah

These are indeed tough questions to answer because, as you outline above, a spectrum of assisted authorship already exists. The new AI spectrum of writing now seems to range from 100 percent AI-generated, to varying degrees of AI assistance, to no AI being used. Today, there are already laws and company policies being put in place to transparently identify the 100 percent AI-generated content. New York State put a law in place in June requiring all businesses creating AI advertising to disclose to consumers what is synthetic. Similarly, platforms like Meta and TikTok are showing AI information badges on synthetic images. Spotify is working on how to identify and share what music is AI-generated. These systems aren’t perfect yet, but I imagine the laws and identifiers will get better and be more common. It seems we have instinctively come to a conclusion as a society that content that is fully AI should be marked as such.

Ethics around transparency become murkier in the areas that relate most to the nuances you call out above that already exist with copywriters, ghostwriters, and even cosmetic modifications. With AI, the circumstances where we can no longer tell what is attributable to one human being become hard to navigate. As your questions also suggest, higher tension seems to exist in domains where people spend significant time and effort, making shortcuts to achieve similar or better results morally loaded.

When pondering your question, “Do we basically want to know whether a person has sweat as hard as their output suggests?”, I was thinking about a recent experience I had with a woman I hadn’t seen in six months. When I saw her again, she was in the best fitness shape I’ve seen her in. My naive first thought was that maybe she had a great Pilates studio I needed to find out about. I finally asked, “You look great, what have you been doing?”, to which she responded that she had gotten on a low dose of GLP-1s. Finding that out certainly satisfied my curiosity, but never finding out wouldn’t have caused me any harm. In other words, of course we want to know how much effort someone has put into their output, but I don’t know if we have a right to know in many situations. However, I did think of a few situations where I do think we have a right to know.

• We are in direct competition. The clearest example is anti-doping in competitive sports, where transparency ensures fair competition.

• Society or individuals are caused direct harm by not knowing. The immediate example that came to mind is doctors being transparent about their training before operating on patients.

I’m wondering what your thoughts are on whether applying these standards to transparency in AI-assisted writing is appropriate, or not strict enough? Are there other qualifications where transparency is needed rather than being a nice-to-have?

Philip

Your framing around competition and harm is insightful. I’d like to posit an example or two in each category to further test the framework’s boundaries.

Competition

You’re right we want disclosures to ensure fair evaluation when competing with others. If there were a digital illustration contest, contestants would rightfully protest if some entries were secretly AI-enhanced (or even AI-produced) unless rules explicitly allowed for it. The latter case is interestingly what has happened to Meta’s software development interviews: Meta started not only allowing AI assistance during interviews, but redesigning the entire interview process to expect AI assistance as a concession to AI’s increasing dominance in software jobs.

A grayer area for me is when competition isn’t official, or cases where not everyone agrees there’s even competition. An example might be the appearance-enhancing photo filters used by many influencers. When this trend first started, many might have found it unethical; but as usage became widespread, it became not only expected, but perhaps even required, in order to maintain one’s followers. But the appearance of one’s social media photos is a case where not everyone might perceive it as a competition… in which case it’s unclear to me whether we should force AI attribution in any and all cases where even some small subset of people perceive some form of competition.

Your GLP-1 example highlights this well. Some people might take GLP-1s purely for personal health reasons, but others in the dating market might feel GLP-1s convey an unfair advantage. Where society seems to have fallen on this, at least in the domain of cosmetic surgery, is that people aren’t forced to declare their use of it (even if it conveys an “unfair” advantage as perceived by some). Increasingly, society has also become more accepting and less judgey of those who do get cosmetic surgery due to its rising popularity.

I wonder if the same will happen for AI as usage increases towards ubiquity. Perhaps at some point we won’t even ask for declarations anymore because it’ll be assumed or understood that most people are getting AI help. The alternative world, where eventually nearly everything is marked AI-Enhanced or AI-Created, might be analogous to the EU cookie banner, which started with similarly good intentions but ended up so omnipresent as to be ignored to the point of irritation. It’s even imaginable we might end up the same place we have with #nofilter on Instagram, where the rare 100%-human produced prose is what gets the badging. But in the meantime, I think you’re spot-on that people will want badges that disclose when AI has been used in competitive environments. Questions remaining for me include:

• How much AI participation warrants disclosure? To me, AI copyediting feels the same as Word’s underline squiggles, which nobody declares. And when AI writes enough that, were it human, we’d feel ethically bound to disclose their involvement, I suspect many people would agree that disclosure is necessary. But there’s so much in the middle that feels unclear.

• Who adjudicates which environments are “competitive” enough to warrant disclosures? I attended a talk by Fei-Fei Li where she said she loves that her grad students can now use AI to author scientific papers, because, in her mind, the “real science” is already done by the time the paper needs authoring, and many of her international students otherwise struggle to communicate clearly in English. While most PhD candidates would say that publishing papers is a highly competitive environment, I’m guessing most would say what’s being competed isn’t the strength of a paper’s prose but instead its scientific novelty.

Harm

Your second category, that of causing harm, made me think a lot. At first, I wondered what sort of harm to another person could come from not disclosing AI-generated prose, but then things started getting clearer once I considered deception as a sort of “intellectual harm” one might inflict.

When The Daily Show puts up all sorts of hilarious graphics, there’s no need to tell you they’re Photoshopped because: a) they’re obviously so fake that no one’s fooled, and b) they’re meant to make you laugh. Essentially, no one is deceived.

It’s not the same when you submit a college application essay edited by AI, since admissions officers have a reasonable expectation you’ve written your own essay. (Actually, is that still a reasonable expectation? I’d imagine admissions officers have recalibrated to assuming essays are now heavily edited by AIs… at which point is there still an ethical issue? In fact, might admissions essays go the way of AP Calculus exams, where it’s assumed everyone’s using a calculator?)

When I applied to college, it never crossed my mind to have anyone read my essays prior to submission. I don’t know if I was atypical, but by the time my kids applied to college, it was commonplace for students to ask friends, family, or counselors to give feedback on essays. I imagine this feedback leads to radical changes in some essays — changes which, if the essay were submitted to a magazine, might warrant crediting those who gave input.

All of these gray areas have me wondering whether it’s better to abide by Rule #5 in the article as a way to capture the need for exercising judgment:

If you—or the reader—would care who the named author is, don’t generate.

Do you think it’d resolve some of these difficulties if we instead tweaked this recommendation to something like, “If you—or the reader—would care who the named author is, acknowledge AI’s contribution explicitly?”

I feel like our conversation has done a good job opening the aperture of considerations and highlighting several nuances. Are there ways you might now recommend for us to narrow in on some concrete recommendations for how we can incorporate AI ethically into prose composition? Are there additions or changes you might suggest to the original article’s five rules?

Mariah

When contemplating your question, “How much AI participation warrants disclosure?” I became increasingly unsatisfied with the idea that the varying amount of AI used should determine whether disclosure is necessary. The “how much” framework quickly broke down when I considered several of the examples discussed above, such as copyediting, ghostwriting, and even the example of Fei-Fei Li’s expectations for her class. Instead, I began to think disclosure might be better understood through the lens of what is being evaluated and audience expectations. The consistent factor in the cases above for disclosure was not how much assistance was received, but whether nondisclosure would mislead readers about what they were evaluating or who was responsible for the work.

For example, current expectations when submitting a term paper are generally that students are expected to cite outside sources because the professor is evaluating the student’s own thinking and writing. Failing to disclose those sources misrepresents what is being assessed. The audience, in this case the professor, also expects that the student is the sole author of their own paper unless otherwise noted through citations. Conversely, in the Fei-Fei Li example, the professor is assessing the science of the paper, not the writing. Fei-Fei Li has also set the expectation for her class that AI use is okay. In this scenario, it seems okay that disclosure within her classroom is not needed given the set expectations. If her students were submitting to an academic journal, which would likely have different expectations for evaluation and authorship, disclosure makes sense to me.

Another example of where we are seeing varying degrees of AI use undisclosed is in the field of software engineering. Today, AI assistance in writing code is increasingly common and, as far as I have seen, software engineers are not disclosing during code review which lines were written by AI versus by the engineer. Not disclosing this assistance is not perceived as deceptive because reviewers are evaluating whether the code is correct, maintainable, and accomplishes its intended purpose, not evaluating specific code choices of an individual engineer. Additionally, there is already an understanding in this setting that AI is being used heavily to improve productivity.

This leads me to think that disclosure should be guided less by how much AI contributes and more by whether nondisclosure would mislead readers about what they are evaluating and who was responsible for the authorship. Viewed through this more contextualized and setting specific lens, I imagine there will be many sources (e.g. universities, governments, businesses) that adjudicate when disclosure is needed. As we navigate the unconsolidated spaces where AI disclosure is unclear, I’d ask myself what readers are evaluating and what the general expectation is for authorship in that specific setting. While I don’t think Rachman’s five rules specifically identify where disclosure is warranted, they certainly can be used as a starting point to reflect more on what is being evaluated and what the audience’s expectations might be. We’ve now navigated several edge cases and gray areas above, I’m curious if you have any final thoughts about how to navigate this evolving space?

Philip

You’ve convinced me that I was overly focused on the mechanics of exactly how much an AI has contributed to a piece of writing rather than questions you mentioned, like:

1. Whether the reader cares about authorship. There are many cases where people don’t actually care about the “true” authorship but instead just want the information conveyed by the writing.

2. Whether knowledge of authorship changes the reader’s evaluation of the work. This even includes social cases like an email between friends, where many people would, understandably, care whether AI composed some or all of an exchange, because part of what’s being evaluated — even if subconsciously — is the amount of care a friend has for you based on the effort they put into their email to you.

And ultimately, as you mention, institutions in various domains will likely need to each articulate what level of disclosure is ethical when writing with AI.

Thanks, Mariah, for helping us all explore the space of ethics of AI-assisted writing! You said a lot of things that made me think more deeply about what I actually believe. We live in interesting times.

You can find more of Mariah’s writings at her Substack, Mental Loose Threads. For instance, one recent article I liked was about AI, Authenticity, and Imperfection, which posits that we may come to appreciate imperfections once AI improves the production quality of many things but makes them uniform in much the same way that the Industrial Revolution improved the production of clothes yet eliminated the imperfections that reveal our humanness.

Here’s photographic evidence:

“Wait!” I hear you saying, “Aren’t those windows I see in dark brown vertical columns arrayed four times down the length of the school?” And you’d be right — those are indeed windows. But the few narrow columns of windows you see are only found in emergency fire escape stairwells leading out of the school, presumably for rescue staff to extract a few lucky students during infernos.

Can a school really insist that three thousand students don’t see the sun from 8:30 a.m. to 3:10 p.m. every weekday of the year?

Yes, yes it can. And it’s not alone.

Windowless Schools

ERHS was built during the tail end of the “windowless school” movement. (Yes, there was an entire movement!) Opening in 1976, ERHS caught the tail end of a movement which lasted from the 1950’s to the 1970’s. The lack of windows at ERHS, thought by students to be an unfortunate oddity, turns out to be by design. It was intentional.

Windowless schools became a thing in the second half of the 20th century because the prevailing theory, supported by a shocking paucity of evidence, was that windows created distractions which detract from learning efficacy. Never mind that the same theory was rarely applied to colleges, much less workplaces. People in the ‘50’s through the ‘70’s liked windows in their own workplaces (e.g. the concept of the corner window office remained an elite perk for careerists), but they knew — just knew — windows couldn’t possibly be good for students.

New Mexico was the trailblazer in windowless school maxxing, though it was soon outdone by California’s enthusiasm in constructing the most windowless schools in the nation. By 1966, over $100M of windowless schools had already been built, and there were still another 10+ years of construction to come. (That’s over $1B in today’s dollars.)

Another reason commonly cited during that period in defense of windowless schools was — wait for it — vandalism. That’s right. Having windows makes you susceptible for those same said windows to be broken by hooligans and pranksters. By the same reasoning, I’m mildly surprised exterior walls weren’t also eliminated, given their siren call to the likes of Banksy.

2 Brutal 2 Last

Thankfully after three decades of enthusiasm behind locking kids into windowless schools, even education architects began to rethink their theories. An entire generation raised in windowless schools did not outperform previous sun-exposed millennia of students. And that generation, once adults, started speaking out against the windowless movement much as I am today.

Perhaps most amusingly, Windex (that’s right — the window cleaning spray company) has an entire page devoted to these mistakes of yesteryear, featuring modern schools which have repented of their windowless ways.

More disturbingly, New Mexico, that bastion of windowless school “innovation,” continued to publish papers even as late as 1975 reassuring us that, despite many other studies showing the positive benefits of windows on both students and teachers alike, you can design studies where it’s still a toss-up of an open question. Never mind that New Mexico’s own state capitol features many (no doubt endlessly distracting) windows:

“What’s that?” you say. “Do mine eyes deceive me, or is that a statue of children, outside, exposed to the many distractions of daylight?” And you’d be right. The artist, Glenna Goodacre (yes, of Sacagawea dollar obverse fame), meant for it to evoke, amongst other things, openness.

ERHS Today

Despite all the depressing memories, there were silver linings to the Eleanor Roosevelt High School story even during my time (1990-94), one of which was the fact the school was woefully under-built for the 3000+ students it housed in the ‘90’s.

As a result, more than 20 portable classroom units (which we called “temps”) sit, pretty much permanently, throughout the parking lot like a Walmart full of RVs:

I couldn’t find a great photo, but it’s still clear the temps each had windows. Though back in the day, those temps were all uniformly beige like the color of all PCs in that pre-iMac era.

The main building of ERHS, fifty years in, remains windowless despite decades of research indicating it’s a bad idea. Yet it continues to be the top high school in PG County, leaving the open question of how very much better its students could have done over the past half century if it hadn’t bought into the trendy, misguided windowless movement of the mid 20th century.

It managed to produce Sergei Brin. Martin Lawrence, the comedian who one day drove up in a brand new Lexus to gift to a teacher who always supported his class-clowning days. And Mýa, the Grammy-winning R&B singer.

During my time, it also featured principal Gerald Boarman, who amongst other things was later accused of financial impropriety by the Washington Post based on unusual spending patterns benefiting general counsel Natasha Nazareth, who followed him from a school he led in North Carolina back to Bullis School in Maryland.

But never mind that. Far more important is the fact that you, too, can now own this amazing t-shirt designed by yours truly:

Buy yours today! Also available as mugs, mousepads, hats, aprons, and other configurations.

If your school had windows, you escaped decades of seriously misguided architecture. If you didn’t, my sincere hope is that college and adult life have allowed for windows, regardless of their supposed vandalism risk.

Writing has always been a source of joy and flow. It’s an act of thinking. For this reason, I have never used AI to generate any English prose — not in Molochinations, not in emails. I even gave up dictating Molochinations after trying it for a few paragraphs because the slow act of typing enables a form of parallel contemplation I can get no other way. It’s like the way many insights come to us when washing the dishes or showering, instead of when laying still or focusing exclusively on a thought. The idle physical engagement allows the mind to roam semi-free as the body keeps it on task through rhythmic, near-autonomic action.

So here we are a year later.

Reader Favorites

From the past year, here are the most viewed articles:

• Insta-Sell All RSUs, Always (24k). Most people have correlated risk bound with their employer by holding RSUs (not to mention owning a house in their employer’s city). You shouldn’t hold what you wouldn’t buy. I continue to abide by this, but made an exception last year when selling OpenAI shares because I was persuaded by Steve Na’s rebuttal to keep a small portion of the stock as a way to support my former coworkers.

• How I Got a Job at OpenAI (22k). I’m surprised how little people prepare for interviews when they dictate so much of our future lives. The majority of getting a job has already happened by the time you interview, based on a series of past experiences which should hopefully build on each other. But you can significantly improve your odds of getting an offer by intensely preparing for interviews.

• No More Code Reviews: Lights-Out Codebases Ahead (20k). It’s self-evident that humans can’t possibly keep up with code reviewing every line of AI-generated code. We need to build systems for this new reality. And there will definitely come a day — I’d argue soon — where it’ll be a liability to your company’s reputation to let human beings touch code. My convictions have only gotten stronger since authoring this piece given interim developments in AI.

• AI Killed the Individual Contributor (10k). Developer jobs are becoming far more like manager jobs and PM jobs. The age of a lone developer thinking hard on implementation details is over. This may not be good news for many people as AI erodes craft/flow by requiring we split our attention amongst many parallel agents for the sake of maximum efficiency.

• You Should Change Companies (9k). Companies, like countries, can give us a mistaken sense we know how best to organize and behave in groups. Only by switching companies can you truly experience different cultures and draw insights otherwise not possible about yourself and about how to build successfully.

• Waxing Asymptotic in Career Velocity (9k). Everybody has a terminal level, a maximum they’ll reach in their careers. Acknowledging this may help reduce the urgency to speed-run the entire experience. It may also ameliorate feelings of incompetence or languishing if we accept our limits. That said, most of us probably underestimate the levels we’re truly capable of.

• Sometimes a Thing Is Just Expensive (8k). You often get better quality by paying more, but many items are Veblen goods, where demand increases as price increases (i.e. where the point is paying more). I’m not sure why this article resonated so well with many people. Maybe we like hearing that expensive things aren’t always what they’re made out to be. Maybe it feels like pragmatic advice.

Note that power laws seem to apply: AI Killed the Individual Contributor had half the views of the three articles above it.

The things that are popular tend either to be controversial (sell everything! no code reviews!) or self-improvement-y (i.e. of potential career benefit to the reader). Though this is not surprising, it’s clear why these dynamics would cause more people to post extreme claims in order to get more viewership.

If your favorite article wasn’t listed above, I’d love to hear about it in the comments.

The Duds

You might be wondering which articles did the most poorly. Here they are:

• The Design of Everyday Things. I just got back from a fancy conference center where I watched many people struggle with push doors which had pull handles. The constant violation of simple design principles seems to be a pastime of sorts, so the fact that this article was the least read of the 53 that have been published bodes poorly.

• The Morass of Idea Blindness. To paraphrase William Gibson, “Great ideas are already here — they’re just not evenly distributed.” We seem willfully blind to applying well-understood, time-tested ideas to different geographies or domains. Digital technologies will hopefully hasten the spread of good ideas as it becomes easier than ever to learn what others are inventing and doing.

• Whither Titlelage? Whether to have publicly visible titles in a software organization is often debated in small companies, whereas the same question rarely comes up in other fields. It’s probably because software organizations often have meritocratic ideals which bristle at the notion of title-based authority. I’ve seen hidden titles work well in smaller organizations but have never seen them scale.

• Form Line! Form Line! Serpentine lines (i.e. lines which bifurcate into multiple lanes at the last possible moment) have been shown to not only be more fair but also to reduce average wait times. Yet we see them in places like Europe and the Far East but rarely in the US. The time for serpentine lines has come.

• Technological Progress Leaves the Dregs of Employment. Is there a chance that technical advancement brings with it the degradation of job satisfaction? We saw this during the Industrial Revolution as well as the introduction of Ford’s assembly line, but might it be a general principle that new technologies sometimes hollow out the most enjoyable parts of a job, leaving behind only the worst parts in the name of efficiency?

Each had under 2k views. Three of the articles are thoughts around design, so perhaps lack of interest reflects a readership more interested in technology and career topics.

But I’m fascinated by the impact design has on our everyday lives, so brace yourself for more unpopular posts about this essential but neglected topic. I’m willing to lose readers over this. 😆 (I’m told by youngsters that the use of laughing emojis alone reveals my age; it’s apparently so passé.)

The Year Ahead

I’m partly unsure whether to continue Molochinations. As you probably know, this publication is named after the concepts in Meditations on Moloch, reflecting my belief that systems determine outcomes. My hope is that some of these articles help us consider the broader systems around us, and to focus more on changing broken systems as opposed to merely addressing the bad outcomes they predestine.

But the other part of me wants to simply enjoy the act of writing and not be captured by my audience to produce only what’s popular. It’s a selfish act to foist writings about random unpopular things on people… but at the same time, I’m trying not to get lost in chasing the big clickbaity topics du jour.

Please comment below if you have thoughts on how where we could take this publication. I’m hoping for a year ahead where I can continue enjoying the act of writing while at the same time producing things which are worth reading, at least to some niche audience.

Thanks for your support of Molochinations over the past year!

About a decade ago, a friend of mine said he started watching TV shows at 2x. This seemed to me ridiculous at the time, and if I’m honest, continues to go against the spirit of enjoying television. I watch many videos online at 2-3x, but those videos are mostly informational and have little meaningful emotional content. I still have a hard time imagining enjoying a TV drama at 2x, though.

Then again, I listen to podcasts at 3.7x. It’s taken me around 15 years to get up to that speed, starting from 1.25x and slowly ratcheting up through the years. My habits here are a huge part of why Superphonic: a) is the only player that goes up to 5x, and b) allows 0.1x increments because that’s what you need to smoothly ratchet your listening speed. When I first started listening at 2x and beyond, jokes weren’t funny because the timing was off. Nowadays, I laugh at jokes even at 3.7x. The mind is malleable.

Years ago I started flipping to Page 3 of online recipes because SEO incentivized recipe authors to insert multipage preambles about how this particular lasagne was something their paternal grandmother used to cook for sick members of the extended family shortly after the war using only what scrap foods were available at the time. SEO is why no recipe online ever starts on the first page.

One in four American adults don’t read even part of any book in an entire year. As reading online becomes increasingly popular, so has the skimming of articles — the visual equivalent of watching TV at 2x.

The death of word-for-word long form reading is a travesty.

Deluge-Induced Collapse of Reading

Traditional barriers to publication were also served as a form of selection bias pressure towards higher quality. Back when printing presses were rare, you had to decide very carefully what to print. When the self-publishing of books used to be expensive, literary agents and publishers acted as helpful quality filters when deciding what to read.

Nowadays anyone can publish a Substack, including yours truly. Amazon’s CreateSpace allows anyone to upload a PDF (for free!) to sell both Kindle and print forms of their book. When the cost of something goes to zero, you can bet you’ll get a lot more of it.

We are being veritably inundated by words. This is causing us to skim everything, looking for unexpected words and pausing occasionally to consume a whole sentence or perchance even an entire paragraph.

The long-form essay is dead.

• There’s loss of trust around authorship. As AI is increasingly used by even well-known writers, I’m less and less convinced each word was specifically intended by the author. An author spending less time on writing makes me want to spend less time on reading.

• Content queues have become infinite. I used to save stuff to my Watch Later list on YouTube and my Save for Later list on Facebook. But after a few years, I’ve come to realize they’re basically guilt-free trash cans, not actually repositories, because I’ve rarely ever consumed content once I mark it for “later.” This is mostly because the amount of new content streaming at our faces each day is already so overwhelming that it’s hard to imagine setting aside even more time in order to catch up on past items.

• Summarizers will be needed just to survive the flood. As our content worlds get flooded with ever more AI-generated and AI-augmented content, we’ll inevitably need summarizers just to say acceptably on top of our reading lists.

I now find myself skimming articles these days, even ones written by friends, because I’m not sure how much is AI generated. (By the way, this publication is 100% human generated — not just generated, but typed instead of dictated because I find the sentences come out very differently when I type them instead of dictating them. Using AI on any Molochinations article would defeat the primary point of this publication, which is that I enjoy the act and art of writing itself.)

Musicians are intentional about the whole of their work. For most of recorded audio history, albums were meant to be listened to as a whole. Artists would agonize over whether a particular song belonged on a specific album because they wanted the whole work to mean something. Listeners respected this intention and listened to albums cover-to-cover. It took much effort to skip individual tracks on records and cassettes, which further led to the standard practice of viewing albums holistically. CDs made skipping easier, and then streaming completely demolished the idea of albums as unitary pieces of art. These days, it’s all about individual tracks.

Relatedly, I’ve never believed in abridged versions of novels. With perhaps the exception of most translators relegating Hugo’s 60-page treatise on the Paris sewer system, with nary a mention of a single character from Les Miserables, to an appendix. When David Foster Wallace was asked by an audience member to cut more out of Infinite Jest, his 1,079-page tour de force, he replied that he’d not cut a single word. As with musicians, meticulous authors intend for every iota in the work to be there.

We now skim and skip. We retweet and restack selective quotes the way private equity firms break up entire companies to sell individual parts. Don’t get me wrong — there’s great value in amplifying a good article by advertising one of its gems. But these days, far more people just gorge on gems and hardly ever read an entire piece.

In Defense of Mindful Reading

I’ll here make a defense of word-for-word reading much the same way some espouse slow food.

Long-form reading is a different type of thinking. Sustained attention to one topic feels qualitatively unlike running your mental life as a slideshow of snippets. It’s the difference between experiencing deep flow versus being mesmerized at a kaleidoscope of disjointed sights and sounds. This is why meditation is both harder and better in the modern era than it’s ever been — because our lives have so little sustained attention time these days, unlike when your 10-hour day was basically driving oxen to furrow a field.

Reading is not just extracting information like a doctor reading a medical chart. It’s a pleasurable way to spend your time, the way children dwell in rich worlds of imagination. Long-form reading, perhaps even once in a while putting down the book or article to contemplate a profound sentence, is the only true way to change your mind.

You can read all you want that “Money doesn’t buy happiness.” There’s no way you’ll be convinced. You need to read a long-form essay about how that manifested in one person’s life in order to truly buy into it.

Some arguments — many arguments — cannot be properly mounted, much less concluded, in 120 characters. What you get in snippets are clapbacks and witticisms, equivalent to stuffing your mind with junk food all day. And, as with food, each piece of junk feels much better individually, releasing all the right neurotransmitters, while at the same time it’s sustained healthy eating which leads to more enduring life satisfaction.

The reason I hand-type every single word in this publication is because the point of my writing is not to convey information or tidbits to readers — it’s to enjoy a weekly act of creation where I try to make sense of this complex, uncertain, and often confusing world.

Find a long-form article to read this week in its entirety without skipping paragraphs once their first few words don’t sound new. If you find the right things to read and follow, I believe you’ll experience a deeper sort of enjoyment that our modern world is increasingly making extinct.

He didn’t mean his other colleagues. By “everybody,” Fermi meant all the alien civilizations we should have heard from by now.

• There are hundreds of billions of stars in our galaxy.

• Many stars have planets. In fact, we’ve found that planets are extremely common.

• Even if intelligent civilizations are rare, there should still have been many over the Milky Way’s history.

• A civilization with modest interstellar travel capabilities could potentially spread throughout the galaxy in a few million to tens of millions of years — a short time compared to the galaxy’s age (~13 billion years).

• Yet we see no clear evidence of alien civilizations.

His lunchtime question has since come to be known as the Fermi Paradox: given the reasoning above, it’s surprising we’ve not yet encountered aliens.

Whereas I think the answer is obvious.

Common Explanations

The Drake Equation, developed by Frank Drake a decade after Fermi’s question, attempts to formalize the locus of our uncertainty around why we haven’t encountered aliens. I’ll here summarize common explanations:

• They don’t exist. Perhaps you don’t believe in evolution. Or you believe intelligent life is so incredibly improbable that it has only happened once.

• They live too far. Did you know there are places in the universe which we will never know about, not because we don’t have the technology but because it’s literally impossible by the laws of physics? Space is expanding faster than the speed of light. So if you believe light speed is insurmountable, there are places in the universe we’ll never be able to know about. Perhaps there’s intelligent life there.

• They’ve chosen to be quiet. Several fun possibilities include:

  • The “zoo hypothesis,” where aliens know about us but are trying to keep us unadulterated like some planet-sized nature preserve.

  • Or, more credibly to me, the “dark forest hypothesis” recently popularized by Three Body Problem. If it’s reasonable to expect that some alien life is hostile and likely much more advanced than your civilization, it’s wise to stay undetectable lest you be destroyed.

  • Or perhaps they live amongst us, quietly blending in as pretend humans or microscopically undetectable beings.

• They’ve all destroyed themselves. This seems to me like the obvious answer, and it’s frankly surprising so few people agree.

Veritably Guaranteed Annihilation

Witness the following:

• There’s an asymmetry to what it takes to create versus destroy. It is far easier to topple someone’s Jenga than build your own.

  • Aircraft carriers, for decades the strongest way to project military might, are now called into question by advancements in drone technologies. Cheap drones now regularly destroy far more expensive equipment.

  • Warships were themselves outmoded by the invention of the aircraft carrier. In the same way that Mongols beat Europeans because their bows shot farther, thus allowing them to remain barely out of reach of their enemys’ arrows, so too did naval warfare used to comprise ever larger cannons. The height of this was the mighty Yamato Cannon during WWII, capable of firing a bullet the weight of a Honda Accord at targets 26 miles away. Yamato class ships seemed impossible to beat prior to the advent of the aircraft carrier, whose planes could fly near and sink huge ships with cheap armaments.

• The capability of unilateral action continues accelerating.

  • A millennia ago, if you didn’t like a village, you could knife some villagers or shoot a few of them with a bow prior to being yourself killed.

  • Once you graduate to revolvers and eventually assault rifles, it’s possible for one deranged person to, say, kill 60 and wound 413 before being stopped.

  • You can now buy DNA synthesis machines at reasonable prices. It won’t be long before someone can create viruses on their own.

  • Even the miniaturization of nukes is a threat. As far back as the 1960’s, the M28/M29 Davy Crockett was invented as a 50-pound nuke. That was 60 years ago.

• The acceleration of urbanization concentrates single points of failure. The world continues urbanizing because of the lifestyle and work benefits that accrue from more efficient human interactions. When you destroy a city nowadays, you’re not just destroying national infrastructure (e.g. datacenters and power stations), you’re destroying high concentrations of intellectual know-how — the same people you may well need to rebuild that city.

If it’s far easier to destroy than to create, and we’ve concentrated our civilization into dense cities, the acceleration of destructive capabilities subject to unilateral action by one deranged person or group is only going to get worse. Take the ability to genetically alter viruses or the fact that smallpox can be re-released into the world today.

I’m not even talking about nations destroying each other. The world has enough nukes to flatten the Earth several times over. I’d remind us of the many times we nearly destroyed the world during the Cold War, when the policy on all sides was “mutually assured destruction”:

• During the Cuban Missile Crisis, Soviet submarine B-59 was being attacked by US depth charges. Their crew believed the war might have already started. Of the three senior officers required to launch a nuke, only one, Vasily Arkhipov, dissented.

• In September 1983, Soviet satellites reported the US had launched first one nuclear missile, then several more. The duty officer Stanislav Petrov thankfully reasoned the US would likely launch far more missiles to start a real war, so reported the alarm as false.

• Just two months later, NATO decided to do a series of military exercises simulating the escalation of war, called Able Archer 83. Given the Soviet Union already feared a surprise NATO attack, these exercises were interpreted by the Soviets as a potential ruse to disguise preparations for a real war. The Soviet 4th Air Army began loading nuclear warheads as units were placed in East Germany and Poland for heightened readiness.

It seems to me obvious that the asymmetry between creation and destruction ultimately favors entropy. It’s a minor miracle we haven’t destroyed ourselves already.

Any sufficiently advanced society would suffer the same fate because the dynamics above are independent of humanity and Earth — they are universal. The answer to Fermi’s question is that unfettered technological advancement inevitably leads civilizations to destroy themselves. Far before you can create a rocket that travels a century near light speed, you would have invented technologies so powerful that a single person or a small group of people could destroy everything unilaterally.

Rage Against the Machine

Ok, well if that’s true, what should we do about it?

I’ll first say it’s not so bad if you believe in the ultimate heat death of the universe. As with the saying, “Everybody dies — it’s just a matter of when and how,” I feel similarly about civilization. We’ll definitely destroy ourselves at some point, but let’s not rush into it.

Or, much better said by Dylan Thomas:

Do not go gentle into that good night
Rage, rage against the dying of the light

Getting to Mars is one step, but insufficient. If someone releases an amped-up version of smallpox on Earth, having a set of people on Mars would at least give us an extra shot at continued life. But even the colony on Mars is subject to the same dynamics we’ve outlined previously, because those dynamics are universal. In fact, I’d speculate the colony on Mars is likely to be far more fragile for many decades than any on Earth.

Going interplanetary is about placing more bets. We could also strengthen our societal defenses by instituting a police state. The premise is this: since individuals are becoming unilaterally more dangerous as technologies advance, the only way to protect us is to strongly spy on and control everyone. This approach has obvious undesirable side effects, but I must admit I like visiting Singapore and would gladly trade individual freedoms for collective safety if I could be assured that central powers wouldn’t be abused. But alas, they always are.

Conflict between nations also seems a likely cause for how we might destroy ourselves. The Earth and its nations is like a playground with no teacher. Might ultimately makes right. The pen is easily destroyed by the sword. This feels to me a distinctly suboptimal way to run the world, but it’s unclear how we’d keep a universal government well-functioning over time… though a brief interlude with a benevolent dictator is imaginable. Ultimately, we may just have to embrace the existential uncertainty which independently sovereign nations guarantee.

Or perhaps we could genetically modify the human race to not allow for the penchant to inflict mass destruction. I’ve not thought about this seriously because it seems impracticable: as soon as some subset of humanity has aggressiveness bred out of them, it seems all too likely the rest of society will just take advantage of them and destroy them entirely.

If the answer to Fermi’s question is that all sufficiently advanced civilizations destroy themselves, there’s a beauty to the resulting epistemologically unanswerable question of whether there exist intelligent civilizations beyond ourselves. We get to choose our preferred answer to this question knowing that the true answer is ultimately unknowable.

There’s a way in which everything about your first job feels normal, a matter of course, because you have nothing to compare your experiences against. This was the case for me when it came to employee generosity at Microsoft, my first job out of college.

I didn’t realize that anything was different about the generosity of Microsoft employees when it came to charitable giving until I worked at several other companies. I now look back on my Microsoft days as anomalous. It’s not that other companies I’ve worked for don’t have charitable employees; it’s that Microsoft employees, more than any other set of folks I’ve worked with, are incredibly generous in their charitable giving.

I think this is a result of corporate culture being shaped from the very top downward. After working in industry 28 years, I’d say the biggest thing I miss about working at Microsoft was its widespread culture around generous support of charitable causes.

The Giving Campaign

The very first Microsoft Giving Campaign was in 1983, launched alongside United Way, raising about $17,000 from a couple of hundred employees. Even accounting for the fact a single dollar could by you much more than a coffee back then, this might sound like small change. But last year, for example, Microsoft’s Giving Campaign raised $255.6M, which is nothing to shake a stick at.

The annual Giving Campaign is a huge deal at Microsoft. I witnessed this even in my first first year, right out of college, when I joined in 1998. Every fall, Microsoft launches a charitable campaign lasting about a month, featuring things such as:

• General calls for generosity. I learned for the first time about many charities through employees talking about their favorite causes. At Microsoft, it’s normal for employees to boost charities they support, whereas I found it’s far less common to talk about charities you love in other companies where I’ve worked.

• A huge auction. Employees donate all sorts of unique experiences in the annual auction, some of which you can’t obtain through any other reasonable means. There are the common things you’d expect, like lunch with some VIP, but also whimsical experiences like the opportunity to put your exec in a dunk tank. One of the most sought after items each year is a privileged reserved parking space right next to your building’s entrance, lasting an entire blissful year.

• Corporate matching of all giving. Many companies do this to some extent, but Microsoft is over the top when it comes to matching donations.

• Dedicated charity events. Things like poker tournaments and bake-offs. Much hilarity and good spirits.

• Collective volunteering. Groups of employees or even entire teams volunteer together at local charities. In fact, Microsoft pays charities throughout the year for each hour that employees volunteer, making volunteering doubly attractive because your time also comes with a bundle of money.

An annual rhythm around the Giving Campaign set an expectation and anticipation around being generous alongside your coworkers. Each year featured much giving and much fun.

Founders Shape Everything

Apple’s secrecy, even amongst (and against) coworkers, was legendarily set from the top down. There’s a famous story about an intern at Apple who didn’t know how to answer when Steve Jobs turned around to her in the lunch line to ask what she worked on. She didn’t know whether it was a pop quiz to test her loyalties, carrying the risk of being instantly fired if she revealed her project’s existence.

Similarly, Microsoft’s great interest in charitable giving started with Bill Gates… or perhaps more accurately, his parents, who had a lifelong history of supporting United Way. This, for instance, is why Bill Gates Sr. was the co-chair of the Gates Foundation for many years.

Bill Gates’s strong support of the Microsoft’s Giving Campaign is what gave it the companywide impact it had. By setting an example from the top, fired by a genuine passion for the cause, Bill inspired everyone to dig deeper in support of charities.

I once attended a surprise talk given by Bill Gates prior to his quitting Microsoft to focus on the Gates Foundation full time. He apparently had sent a general invite to employees in a few buildings near Building 33 to talk to them about global health simply because he wanted a venue in which to talk about things which mattered to him a lot. I worked in Building 32 back then, and am so glad I attended that spontaneous talk because it was inspirational to personally witness the depth of Bill Gates’s obvious passion for solving problems in world health.

This sort of top-down inspiration is also what motivates VPs at Microsoft to often compete with each other around how much they can inspire their teams to give. These competitions have included celebrated stories which have become legends, like people shaving their heads or otherwise making spectacles of themselves all in the name of charitable causes.

A Continued Legacy

Through a combination of my positive experiences at Microsoft and the legacy Christian faith of my youth, our family continues to give 10% of our income every year to charitable causes. It’s been a rewarding experience to support a variety of causes through the years and to see individual charities grow from startups to maturity.

One organization we supported in the past few years is GiveDirectly, which operates on the thesis of books like Give People Money and the many papers showing that just giving people cash is often the best way to help them (as opposed to money earmarked for specific uses). Economists would agree it’s often the most efficient way to provide direct benefit to those in need.

GiveDirectly will even organize one-time gifts from you to an entire village to fundamentally change real people’s lives. Here are a collection of stories from several villages we’ve contributed to. You can see the wide variety of uses to which people put their money, and the long-lasting sort of benefits that are possible if you give people a leg up.

I miss the way Microsoft coworkers regularly talked with each other about causes they support. A thing I thought was common in every workplace turns out to be relatively rare, at least in my experience.

Some studies show that lower income households give more than higher income households, as a percentage of their total income. More recent better-controlled studies show that families on average give around 1.4 - 2.0% of their income regardless of income bracket.

In either case, it’s statistically unlikely you’ll give more once you have more. If you feel it in your heart, give now. There’s no better time, and you’re unlikely to find yourself any more generous if you suddenly get a windfall cash infusion or start making boatloads more money.

Now’s your chance. Now’s as good of a time as any. Thank you for your generosity!

Mid-Career Dominates Long-Term Compensation

I lead a seminar every year in University of Washington’s computer science program around career topics. Each year, understandably, seniors ask me a lot of questions as they mull over how to respond to job offers. Compensation differences between offers often cause them uncertainty.

My biggest suggestion to students contemplating different compensation packages is not to overthink them, because mid-career compensation far outweighs early-career compensation. If you want to be compensated well, the most productive thing you can do is build yourself into an amazing mid-career engineer.

This is the spiritual opposite of a child star going nowhere in their adult acting career. You’re like a child star when you get hired out of college: people haven’t seen much of you yet, but they think you have promise. By the time you become an adult, the best actors have been separated from many who never progressed beyond their childhood potential.

How dominant is mid-career compensation over early career? In my case:

• Started at Microsoft at $42k a year, with perhaps an additional $12k RSUs a year.

• Ten years in, went from L66 → L67, where base pay was around $150k with around $50k RSUs.

• Fifteen years in, my pay at Facebook was ~$250k with $1M RSUs. Note this is after taking a pay and RSU cut to join Facebook from Microsoft.

• Twenty-five years into my career, at age 48, my base pay at OpenAI was around $450k with $2M RSUs.

Several things might stand out:

1. The magnitude of difference in what you might earn in your 40’s is hugely different from your first job out of college. In looking at the numbers above, it’s clear that optimizing for maximal mid-career compensation is the wisest strategy while nearly disregarding early offers. Remember: Microsoft was my second lowest offer, but I took it believing it would build skills which would make me more valuable later — which it definitely did.

2. Companies compensate differently. OpenAI has a different comp philosophy from Microsoft. As a gross oversimplification, harder-to-get jobs tend to compensate more because they generate higher applicant demand. Getting yourself that harder-to-get job is all about what skills you’ve built earlier in career.

3. Your compensation growth will be lumpy. It’s more a punctuated equilibrium with large discontinuities than some sort of monotonically rising line graph. Expect years where nothing changes much. In fact, expect to sometimes take jobs offering less in order to invest in yourself. And if you work at startups, of course, pay cuts are a real thing. I’ve omitted experiences from the above timeline, like my three years starting a nonprofit where I was paid less than a tenth what I made the years before.

4. Compensation growth is geometric. Students in my seminar have a hard time intuiting that compensation in their 40’s will dwarf compensation in their 20’s because they don’t quite see that it grows geometrically. For instance, companies tend to increase pay by fixed percentages (e.g. “Everyone who gets promoted this year gets a 2% raise.”). Since those percentages apply on top of whatever your compensation happens to be, growth over time is geometric.

5. Some promotions have outsized consequences. Back when I worked at Microsoft, a promotion from L65 → L66 had a fixed percentage increase in RSUs, but going from L66 → L67 doubled your RSUs. There are similarly large discontinuities when going from L67 → L68 (“Partner” or E8). This tends to be because the market of qualified candidates is uneven between certain levels. At Microsoft, something like 1% of employees can be Partners; the discontinuity in pay reflects demand chasing much smaller pools of capable individuals.

6. Your mileage may vary. Compensation is ultimately aligned with how rare and valuable your skills are. I grew from an E3 to an E9, so my compensation changed very significantly between college and midlife. The average engineer progresses to E5 or E6 and stays at that level for the rest of their careers. Note, however, that even E6 compensation is dramatically higher than E3, so everything said here still applies.

The biggest thing you can do to increase your compensation is to focus on growing your skills — not negotiating your offers (which you still should) or choosing higher offers. The goal if you’re income-maxxing is to become the very best version of yourself by the time you’re 35-40 years old, which will be the prime years of much higher income if you’re in tech. So choose jobs and teams where your skills will grow the most.

Be Best

In Outliers, Malcolm Gladwell discusses the effect of birthdays on children becoming professional hockey players. Children born the right time of year end up bigger than their peers in children’s hockey leagues, so they perform better. And better performing children are given more opportunities where they further bloom. The advantage of being better than your peers compounds.

At most major tech companies, compensation bands at different levels overlap (a lot). So, for instance, an exceptional E4 often out-earns a below-average E5. Several reasons that compensation bands overlap so much:

• It gives companies the flexibility to hire individuals at equivalent levels without distorting companywide fairness. For instance, a staff engineer at Hooli might only be willing to go to Vandelay Industries if they can stay a staff engineer. Yet their performance during the interviews relative to Vandelay’s staff engineers was not crushingly terrible, but definitely below par. In this case, Vandelay can make an offer with a Staff title but compensate at the lower end of that level in order to keep compensation fair relative to others in the company.

• It gives room for rewarding exceptional performance without requiring promotions. Companies can give an E4 a meaningful pay bump even if they aren’t quite performing at E5 yet. Conversely, and perhaps equally importantly, poor performers’ compensation can drift to below average as overall compensation bands are increased over the years (i.e. by not giving a below-average IC4 a raise while the overall IC4 compensation increases, that IC4 naturally gets paid commensurate to their sub-par contributions over time).

Knowing that compensation bands overlap a lot leads to a consequence many people don’t internalize: it is often better, compensation-wise, to consistently exceed the expectations of your level than it is to just meet the expectations of your level.

Concretely: Bob and Sally start their careers in the same company at the same time. They grow equally fast in skills over the years. Bob constantly pesters his boss to promote him as soon as possible. He goes from E3 → E5 in the space of 4 years, but has rarely exceeded the expectations of his level since he always gets promoted as soon as he’s performing on-expectations at the next level. Sally works on the same team as Bob, but her manager only promotes her when he feels she’s greatly exceeding the expectations of her level. Sally goes from E3 → E5 in 6 years, even though she had the same skills as Bob along the way. The main difference between them is that Bob’s promotions happened a bit before he was fully ready while Sally’s happened a bit after.

While Sally gets to E5 more slowly, she has exceeded expectations in every performance review along the way. Whereas Bob mostly met the expectations of his level except for the two times leading to promotion. In most tech companies, Sally would have earned more (sometimes much more) than Bob even though she reached E5 later because her performance was always stronger than her peers. Most companies give outsized rewards to people who greatly exceed the expectations of their level; for many years at Microsoft, merely meeting expectations would lead to no bonus and no raise — you were essentially “doing your job.”

There’s a subtler long-term benefit to only being promoted when you’re fully ready to excel in the next level: you tend to be given the best opportunities for growth. Managers tend to remember outperformers and give them first dibs on new projects. When asked whether you’d give an attractive project to someone who’s always delivered beyond expectations or to someone who’s never done so, managers understandably tend to bias toward those who’ve always done a great job. This exposure to better opportunities compounds over the course of a career, leading to far more meaningful compensation differences in the long run.

I once sensed I was about to go from L64 → L65 at Microsoft and asked my manager to only promote me once he felt I was going to do well relative to peers at L65. Saying this delayed my promotion by six months, but it was well worth it to continue a history of overdelivering rather than to arrive at a new level early and struggle to tread water.

I’m not suggesting that everyone redshirt themselves and deliberately slow-play well-deserved promotions. I’m instead arguing two things:

• Focus on building the skills that will make you great mid-career. Building these skills happens largely independently of when you get which promotions.

• Realize you’ll usually earn more by consistently delivering beyond expectations at a lower level rather than always barely meeting expectations at a higher level.

[Warning: This essay contains spoilers for several TV shows. But the shows I discuss are all old, so I’m unlikely ruining anyone’s enjoyment.]

I’ve never liked a TV show more after rewatching it years later, though I’ve often had that experience when rereading a book or rewatching a movie. I’m not sure why this is.

For about a decade, I had been telling people my favorite show of all time was Breaking Bad. Recently, upon the tenth anniversary of my having first watched the series, I binged all five seasons over two weeks. I can now say definitively that it’s no longer my favorite show of all time, for sure, but it’s not because I found a new favorite in the intervening years. It’s because Breaking Bad fell significantly in my estimation after the second watching.

There are a few classes of changes which could account for the phenomenon whereby TV shows have never improved upon rewatching, while many get worse when revisited.

Perspectives Change

“No man steps into the same river twice — because it is not the same river, and he is not same man.”

— Heraclitus.

It’s one thing to like Voltron when you’re twelve. But nearly no one likes it when they’re 50, because their perspective changes. Every episode of Voltron starts with the heroes battling the villain as five separate lions. They always do terribly and are nearly overcome by the time they finally decide to join into Voltron (“I’ll form the head!”). The fight still doesn’t go too well, at which point they pull out the Mondo Sword™, which, without fail, always kills the enemy in one slash. Never mind why they don’t just always start the battle with that sword, given history and all.

A-Team falls into the same category for me:

• They arrive into some new town where a local shop owner is being harassed by hooligans.

  • The owner herself, or perhaps her daughter, is attracted to Face.

  • Murdoch performs general shenanigans.

  • B.A. calls Murdoch a fool.

• The team welds, screws, and duct tapes together some number of weapons and machines to counteract the hooligans.

• They save the shop, and Hannibal says, “I love it when a plan comes together.”

I loved A-Team as a kid. But rewatching it as an adult is a huge mistake. Blues Clues revolutionized children’s television with the insight that very young children prefer the same exact episode airing five days a week. Anticipating what you know will happen is its own form of fun. It’s predictable and safe. You love reliving the feelings. This, it turns out, is part of what was great about Voltron and A-Team. Its predictability was a key part of its value proposition.

Our perspective also changes because life experiences help us identify more strongly with certain stories over time. For reasons unclear to me, my high school required us to read The Sun Also Rises, Hemingway’s story about a man’s struggle with alcohol addiction and impotence — two things that I had absolutely no understanding of as a 16 year old.

Given that, you’d think I’d identify more strongly with Walter White upon rewatching Breaking Bad this year: Walt turns 50 in the first episode, and I’ve just turned 50; my mom passed away last year from cancer, and Walt is diagnosed with cancer in the first episode. But despite my new perspective, I strangely didn’t find the show more compelling for reasons I’ll get into.

Expectations Change

I watched Rashomon and Seven Samurai, both legendary Kurosawa films, late in adulthood and found them completely unremarkable. I only learned later that this was because I’ve taken for granted several common occurrences in film which those movies invented:

• Rashomon was the first time a movie showed the same story looking very different depending on whose perspective you’re showing.

• Seven Samurai invented the thing where one enemy crests into view over a hill, followed shortly by hundreds of his compatriots cresting all at once.

Similarly, if you’ve ever rewatched the original Jurassic Park, you’ll likely feel unimpressed. Back in 1993, when seeing it for the first time in theaters, audiences positively lost their minds when the brontosauruses (or are they now apatosauruses?) first came into view in the grassy valley. People went nuts because they looked so real.

We laugh when we hear stories of people diving out of the way of trains driving straight towards the camera in black and white movies… but the locomotives really seemed that real to them. The original Jurassic Park dinosaurs inspired that same incredulous awe. However, when you now go back and watch that exact same scene which in 1993 caused you to nearly lose continence, you’ll find plenty of flaws in those dinosaurs. The lighting… the textures… it all looks as real as a black and white locomotive coming toward the screen looks threatening — which is to say, not at all.

Friends and I rewatched an episode of MacGyver about fifteen years ago after reminiscing about how great the show was in our childhood. Big mistake. Its production had such low standards that a car crashing into the living room window of a house was shown with the stuntman’s plywood launch ramp still visible in the final edit. V: The Original Miniseries has the same flaws. A show I loved as a kid now looks low budget and melodramatic.

Sometimes, a show is just relevant to its time period. I’d put Twin Peaks in that category: perfect for 1990, completely non-sequitur and inexplicable in 2026.

One last major shift in expectations is around the norms of storytelling and pacing. The Wire was a groundbreaking, magnificent show, loved by critics and fans alike. But anyone watching it for the first time in 2026 will likely need to amp the playback speed to 2x because shows like Breaking Bad have completely changed the pace at which we expect things to happen in TV dramas. Somewhere in Season 1 of The Wire, you’re likely to drop off entirely the way most people do when J. R. R. Tolkien goes into Yet Another Five Page Poem About Lothlorien. In fact, The Wire doesn’t even really take off as a story until Season 2.

Details Matter

Rereading a book and rewatching a movie both give opportunities for you to notice details you missed the first time. When you rewatch The Usual Suspects or Sixth Sense, you’ll likely notice far more details which increase your appreciation for their writers’ craft.

I’ve not found this same dynamic in rewatching TV shows, though, and I’m not sure why. You for sure notice more details. But seeing more details in television shows often leads to more disappointment, not more appreciation.

In my second watching of Breaking Bad, I noticed signs of Walt’s ego and his selfish behaviors very early in Season 1. Whereas my previous memory of the show was that Walt was good for a while and eventually became beat down by circumstances and corrupted by his work. With my rewatching, I now recognize Walt was bad from the very beginning, and thus felt a lot less sympathetic towards him because he committed so many selfish acts at the behest of his insatiable ego.

Whereas in rewatching shows like Twin Peaks, you end up noticing far more inconsistencies where the story doesn’t hold together, or where actors are shown in obviously different positions via the two cameras in a scene. Idle cycles of the mind, formerly spent on tracking the plot, can now roam the scene picking up all sorts of distracting, unhelpful details that were mistakes or omissions.

I still can’t explain why all of this is any different between TV shows, movies, and books. While I’ve often enjoyed rewatching movies and rereading books, I’ve never never had a better experience with a TV show the second time around.

I’ll leave you with one last thought. It often pays to watch TV shows for the first time many years after they debut, for the same reason as reading older books: survivorship bias. There’s a reason we’re still talking about some books centuries after they were written. Star Trek and The Office were both much more popular in syndication than they ever were when new. And if you’re lucky, as with me watching The Sopranos a decade after it first aired, nobody will have inadvertently spoiled the ending.

The world was my oyster.† It felt a good time to buy something representing the fact I had arrived. It was time to buy a pen.

And I don’t mean a fountain pen, or even the Sylvester Stallone Chaos, which I’m told had gem-encrusted versions selling for $50k each. I was under no illusion that those pens could possibly justify their cost. And I always knew fountain pens were more trouble than they’re worth, largely a nod to robber barons and the Gilded Age, a sort of declaration your name was probably “William Wentworth IV” of the landed gentry.

I wanted the best pen you could buy if money was no object. So I relied on New York Magazine’s review of 100 best pens, and bought its #1 pick: the Baron Fig. Machined out of block aluminum, weighted a certain way, blah blah blah. Ignoring price, one $55 pen seemed a reasonable nod to the fact my career was going well. A worthwhile indulgence, a bit of a splurge.

Never mind I rarely wrote anything. In fact, I’ve written so little in the past two decades that my handwriting now looks like a child’s, and each time I write feels like the start of physical therapy, with my inept fingers shakily forming letters from days of yore. But I wanted The Best Pen because I Had Made It.

I unboxed my newly arrived pen with a giddy anticipation and a warm satisfaction that, yes, all the years of hard work were worth it. Now I would write like few have written before, luxuriating in a joy The New Yorker Magazine assured me topped 99 alternatives.

This glow faded within a month, by which point I was rarely, if ever, using the pen. Turns out a heavy, perfectly cylindrical pen is almost never what you want, unless you only ever work on professionally-leveled tabletops. The weight, which reviewers praise, is a drawback — in fact, the same drawback as the heavy smooth-metal chopsticks Koreans often use, which constantly slip out of my unpracticed hands onto whatever plate I’m pointing them towards. Twist-actuating a pen always requires two hands, unlike click-activated pens. The idea that I owned An Expensive Pen™ was quickly overtaken by the sheepish admission that about the only thing the Baron Fig was good for was the meager satisfaction I could afford its $55 price tag.

The joke was on me.

Know What You’re Paying For

Not everything works this way, of course. Oftentimes you get what you pay for. Then again, it’s important to be clear about exactly what you’re paying for.

Some things are positional goods, whose value derives purely from the fact only a limited number of people can have them. In the UK, a funny-to-an-American positional good is the desire to have the equivalent of low-number plates. Vanity plates with only two alphanumerics are especially rare and thus coveted almost exclusively for their exclusivity. (Though the funniest case of this I observed in London was a house with two Aston Martins out front: “2B” and “NOT 2B,” cars which could only ever be legitimately parked relative to each other).

Then there are Veblen goods, things whose values increase precisely because of the amount paid. Two examples:

• Mooncakes are often a formal gift amongst the Chinese during the Mid-Autumn Festival. Because of their value as gifts, there are a variety of ridiculously priced mooncakes in superb packaging where the entire point of those mooncakes is to convey to the recipient just how much you’re willing to spend on them. Their expense is the point.

• In a more humorous vein, Microsoft SQL Server 6.x used to cost several hundred dollars a pop, when Oracle already cost 5-digit sums. In a brilliant move, SQL Server 7 was priced nearly two orders of magnitude more, at which point customers took the product much more seriously instead of viewing it as a toy database. The surprise wasn’t that SQL Server 7 made more revenue; the hilarity was that the much more expensive version sold more total units. The much higher pricing itself did most of the marketing.

Yet sometimes you’re truly paying more for quality. Or perhaps the thing you’re buying is an experience, something memorable to share amongst people you love. An example is when our family bought, for the only time in our lives, Suites tickets on Singapore Airlines. This was a huge splurge, something not to be repeated — but sharing this experience as a family was a worthwhile one-time expense. What we were buying was not the suite itself, which, judged on its own merits (e.g. by the square footage of plane it was taking up or the opportunity cost of other tickets to be sold in the same space), likely doesn’t pencil. What we were buying was the memory of living like sultans, if only for a day.

Oftentimes, Cheaper Is Better

I’m a Costco shopper, so you probably could guess my opinion on price vs. value. But I’ll ground it with some examples.

Another thing I bought for myself which I now feel foolish for is a Toto Washlet. Google employees rave about how they’d never poop anywhere else. Oh, the luxury. Oh, the amazing feelings of cleanliness. So many of my techie friends swore by their Totos that I ended up buying one myself for $2,000.

As with the Baron Fig, I had high expectations of a scatological life revolutionized. I don’t know if I’ve ever been as excited to poop as on the debut day of the Toto Washlet, which I even had to rewire my bathroom for in order to operate.

Over the past few years, it has malfunctioned many times. I’ve replaced the entire unit once, and refuse to a second time. It makes noises at night while you’re sleeping, like a dog who loudly yawns awake to remind you he’s restless. It causes even more cleaning headaches because of — putting it delicately — collateral splash damage.

Turns out flushable wipes have none of the drawbacks and most of the advantages, coming perhaps 100x cheaper. I had not known this when starting this whole Washlet adventure, but suffice it to say that I now Subscribe and Save™ on Amazon.

Or take Lamborghinis. I owned this Diablo SV many years ago, repo’d in Las Vegas and sold to me for less than the price of a Porsche:

Changing the clutch was nothing like a Honda, requiring the wholesale removal of its engine and transmission in order to replace the $2,000 clutch yourself (photos here). The Diablo came with a switch you had to actuate whenever parking it for more than a few days because they were manufactured with so many electrical leaks that the makers simply gave buyers the ability to disconnect the entire electrical system, passing this ridiculous responsibility to owners.

I’ve since discovered you’d much rather drive a Lotus Exige Cup R, at perhaps a quarter the price, at a place like Spa Francorchamps. It’s lighter and has better road feel with much cheaper consumables. Repairs are easy. It’s better in almost every way if you’re interested in racing. But the Lambo is better if you’ve bought BTC and want the world to know about your diamond hands.

It comes down to purpose and your ability to use things. I once saw someone at Sonoma Raceway driving a stock Honda S2000, lapping people in much more expensive Porsches and Aston Martins.

Oftentimes cheaper really can be better. You need to know why you’re actually buying something.

Asymptotic Limits of Performance

“A billionaire can’t buy a better Coca-Cola than the bum on the corner.”

— Warren Buffett

Similarly, people have expressed ideas like the fact that Bill Gates can’t buy a better iPhone than you can. It might have been John Steinbeck who said something to the effect that “no man can eat more than two eggs for breakfast.”

I’m not saying expensive things are crap, or that they’re all a scam. What I’m suggesting is to:

1. Know why you’re buying what you’re buying

2. Decouple price paid from value received

3. Realize there’s no end to the ability of positional and Veblen goods to separate you from your money

After several false starts, my favorite pen is actually the Pilot G2, which writes easily and smoothly with a gel that doesn’t bleed when wet. It click-retracts and doesn’t require capping / uncapping. It doesn’t roll on tables. And it’s $1.91 on Amazon.

Not everything turns out this way, of course. But far more things do than I ever thought did whilst I was less well off.

“I think everybody should get rich and famous and do everything they ever dreamed of so they can see that it’s not the answer.”

— Jim Carrey

† This, as with many other phrases (and words like “manager”), comes from Shakespeare. In The Merry Wives of Windsor: “Why then the world’s mine oyster, which I with sword will open.”

If the past has taught us anything, it’s that predicting the future is hard because the machine of the world is much more complicated than it seems, especially to those who confidently prognosticate about the future.

Then again, those who cannot remember the past are condemned to repeat it. And though history doesn’t repeat itself, it rhymes.

We basically have a world where half of the wise say to study past patterns in order to predict the future while the other half bring up innumerable examples where past performance is not indicative of future results.

I’ll now wade into speculation about the societal impacts of AI in full awareness of the folly of doing so.

Tooling Universality

I have a friend who owns a mechanical orange juicer with a huge two-foot arm for maximum leverage. It squeezes oranges better and faster than anything else I’ve used, including motorized juicers. But it also weighs about 20 lbs and displaces all sorts of other equipment on their countertop.

Specialized tools outperform generalized ones but come with significant tradeoffs. Whereas tools are most broadly impactful when they’re generalized, even if they can’t perform any one specific function better than custom tools.

Take the opposable thumb, whose only biological value (and a hugely significant one at that) is its generalizability. It’s credited with enabling humans to far exceed other species in tool use because the opposable thumb is a meta-tool: a tool enabling you to use many other tools, a sort of second-order generalization.

CPUs are this. Steve Jobs put it best when he called the personal computer a bicycle for the mind.

As late as the 1970’s, large corporations usually had an entire floor of secretaries (the “typing pool”) whose only job was to type memos and documents dictated by others. The adoption of “word processors” in the ‘80’s eliminated these dedicated cohorts of typists. Though these full-time typists created documents faster than the workers who replaced them, efficiency favored the integration of typing into every office worker’s job because it afforded them the flexibility and immediacy which dedicated floors of much faster typists did not. Generalization won the day.

AI, similarly, is a meta-tool that will allow anyone to harness the power of customizing software for themselves, of building personal tools. Just as you used to need dedicated floors of software engineers to make your company run, everyone will soon instead pilot the creation of their own software at their desk, just as workers in the ‘80’s harnessed word processors for themselves.

Hard Fork had a great episode recently where non-coders sent in their stories of using AI to create software for themselves. A plumber created his own customer scheduling software. A father created a leaderboard for his children’s chores. Everyday people were making their lives easier through the power of AI.

This trend will only grow, quite rapidly, as more people realize the power of creating software tools to increase their efficiency. But I’m not sure it’s an unalloyed good.

Indie Films vs. WALL-E

Of many possible futures if we were to adopt universal basic income, I find the two extremes most thought-provoking:

• Proliferation of indie films and performance art. There’s a certain sort of person — perhaps a journalist who graduated from Cornell’s philosophy program or a progressive community organizer — who advocates UBI because they imagine a world free of work constraints, a world where people can finally have the time to create indie films and stage life-changing immersive experiences for others. A sort of perpetual Burning Man.

• Then there’s WALL-E. A world where everyone’s overweight, scooting around on mobility machines while watching Netflix and sipping gigantic sodas.

The real future is unlikely to be as extreme as either of these two poles, but I’ll just make one observation about the indie film crowd: they imagine that future because they themselves, no doubt raised in upper-middle-class neighborhoods and Dead Poets Society-style private schools, would make indie films and coordinate flash mobs if they didn’t have to work for a living. But what they don’t perhaps realize is that their personal initiative and bent towards creative endeavors or artistic expression isn’t common.

In fact, it’s downright atypical. Employment of men in their early 20’s has dropped in the last decade. What they’ve done with that time, empirically, isn’t to create a heretofore unseen explosion in the arts; what they’re doing is playing 50% more video games. More WALL-E, less Burning Man.

And here’s where I loop it back to AI: the type of enterprising reader who follows this Substack is likely the same type of person who, given extra time or money, would love to create something novel or impactful in the world. You, reader, are atypical. The survivorship bias alone of making it to this paragraph is indubitably strong given the meandering nature of the previous dozen paragraphs. You would create things if given the opportunity.

But empirically, most people don’t. WALL-E is the far stronger pull than Burning Man for most of the population.

The Road Less Traveled

AI will exacerbate this bifurcation because it is a meta-tool that accelerates the utility of all other tools. The type of person with strong individual initiative will greatly benefit from the sudden increase in their personal agency, their ability to make their visions a reality. For them, the AI future enables a Diamond Age where their ability to reshape the world is limited only by their creativity and drive.

But what happens to everyone else? One type of person loves an art class where they’re given a blank canvas. Another is paralyzed by the open, undirected possibilities that same blank canvas creates. Sometimes it’s nice to show up and have the art instructor tell you what to draw instead of having to always have an idea yourself… much less have your income be dependent on your ability to generate new ideas all the time instead of doing whatever the world tells you needs doing.

There’s a kind of career-related stress, common nowadays, which is quite historically unusual. For nearly all of history, most of humanity didn’t need to make “career decisions.” If your dad was a blacksmith, you’d naturally be a blacksmith. There weren’t ever decisions about promotions or relocating to Blacksmith Headquarters™ to get better hammering opportunities. Life was a straight line with few forks in the road.

We now have a sort of paradox of choice when it comes to work and life. While it might seem like an unalloyed good for many people to have a variety of possible paths for career and life to choose from, have we considered the possibility that a panoply of choice might actually be overwhelming, if not downright paralyzing, for many people?

The advance of AI greatly accelerates this bifurcation of society. People with plenty of innate initiative and ideas will find themselves veritable wizards of possibility. But for the majority of humanity, who until only the most recent period of history were able to survive simply by doing whatever circumstances required of them, things are about to get much worse as those with opposite personalities accelerate way past them.

If you’re lucky enough to be blessed with plenty of creativity and personal initiative, AI will soon amplify your strengths toward even more differentiated outcomes from the average person. But if you’re the average person, now is a critical time to reevaluate whether you can lean into what creativity and initiative you can muster before it’s too late.

The road forks before us.

The problem is this: just because you see where the road is going doesn’t mean you can change course. Usually, knowledge is power. But sometimes, knowledge is just you cosplaying Cassandra, the Trojan princess beloved by Apollo, who at first bestowed upon her the ability to see the future, but then, upon being rebuffed, added the devastating provision that no one would ever believe her dire warnings. There’s a world of difference between seeing the future and preventing it.

It’s true you can only get senior engineers by growing them from junior ones. Making that observation is the easy (I’d say obvious) part. But who shall hire said junior engineers, and pay the cost of growing them into senior ones? Who’s going to take the hit?

The far smarter move is to let your competitors hire junior engineers and for you to simply poach them once they’ve grown with experience. In fact, a cynic would say it’s smart business to convince all your top competitors to hire junior engineers, loading them down with the tremendous cost of training while you feast on the resulting bounty.

The average tenure in Mag 7 companies is four years. Conveniently, this is also just around the amount of time it takes to grow a star junior engineer into a senior one. In other words, that superstar is likely to leave right when you’ve finished growing them into someone your competitor would love to hire.

Collective Action Problems

While everyone acknowledges we won’t have future senior engineers if no one hires junior engineers, who’ll be the sucker to take the hit? Every (smart) employer is waiting for others to pay the price.

This dynamic falls into the general class of collective action problems, of which the most famous is the prisoner’s dilemma. The specific case of junior vs. senior engineers is a public goods dilemma, where everyone benefits if some members contribute something, but the greatest benefits accrue to “free riders” who simply mooch off others’ largesse.

More dramatically, we all need a few good men:

Son, we live in a world that has walls, and those walls have to be guarded by men with guns. Who’s gonna do it? …You weep for Santiago, and you curse the Marines. You have that luxury. You have the luxury of not knowing what I know -- that Santiago’s death, while tragic, probably saved lives; and my existence, while grotesque and incomprehensible to you, saves lives.

You don’t want the truth because deep down in places you don’t talk about at parties, you want me on that wall -- you need me on that wall.

Colonel Jessep passionately highlights a public goods dilemma. Society benefits from the public good of national defense, but from Jessep’s perspective, producing that good requires paying the cost of moral ambiguity during training, a cost most of us are unwilling to pay as we free ride on the liberties those good men provide.

So we’re all wise to the need to produce future engineers. It’s just that no half-thinking CEO wants to pay the cost.

Meditating on Moloch

More broadly, the whole reason for this publication’s name is Slate Star Codex’s Meditations on Moloch, an article that opened my eyes to the idea that systems dictate outcomes.

If you put two prisoners into the classic dilemma, you should only be surprised if they don’t choose the worst outcome. If you allow dual-income families into your neighborhood, you should not be surprised at all to find in several decades that the only way to afford those same said houses is for all your neighbors to become dual-income. Every family will work double just to remain in the exact same house because you’ve designed a system whose outcome was preordained.

If we lived in the 60’s, when men worked at GM for 40 years with GM guaranteeing their retirements, during an age where loyalty between employer and employee ran both ways, we would fully expect GM to be plenty willing to pay the cost of training junior employees because they stood to reap the benefits for decades.

Conversely, in the modern era where employers cut employees to juice quarterly earnings, and employees leave employers at the drop of a hat when greener pastures come calling, we should be not at all surprised when few employers are willing to hire and train junior engineers.

Systems preordain outcomes.

In an era where the average software engineer’s tenure is four years per company, we should rightfully expect collective inaction when it comes to the hiring of junior engineers. Everyone is waiting for someone else to pay the price. Everyone will free ride. And they’ll do this, just as our two prisoners, with full understanding of exactly what harm comes to everyone, including themselves, through their inaction.

Despite this, act they will not. The system guarantees this.

Meditating on Missiles

There’s a way in which concerns about AI-driven warfare can be seen as a collective action problem. In fact, you could say from a guns vs. butter perspective that whether to develop AI warbots is a prisoner’s dilemma.

Whether nations develop AI-driven war machines will be dictated entirely by the systems of international relations we design. If we let today’s system stand as-is, we should fully expect all nations to develop AI-driven weapons of war with an urgency bordering on the existential because our current system is exactly a prisoner’s dilemma. Every nation knows the ideal outcome for all would be for none to develop AI warbots. Yet the first nation to develop superior AI warbots makes suckers (or worse yet, servants) of everyone else. Since everyone understands this, Skynet is bound to come online.

Unless we change the system.

Nuclear non-proliferation and mutual deescalation treaties were able to change the international system enough to reduce stockpiles, at least for a few decades. We’ll likely need something similar if we don’t want escalation of AI-driven weapons of war.

The Yearn to Earn

“I don’t much care about the threat of global AI-driven apocalypse. I just want to hear how we’re going to restore jobs for junior engineers. Because income.”

You’ve come to the right place. Maybe.

It’s one thing to identify a problem (i.e. our current system punishes companies for hiring junior engineers). It’s a whole other thing to posit viable solutions. I’ll here share just a few observations:

• Junior hiring won’t ever go to zero. Just as there are still farmers today, or, even more rarely, there are still people hand-weaving rugs, there will always be appetite to hire a small number of junior engineers because the very best ones are miles better than mediocre senior engineers.

• High retention organizations are still incentivized. Tech companies with long employee tenures will still find the amortized cost of training junior engineers worthwhile. No top tech company fits this category, but many other organizations do. Academic Data Systems, the organization at the University of Maryland which gave me my first coding job, had people who worked there for decades. A place like that won’t hesitate to hire junior engineers.

• Industry coalitions are possible. SPF / DKIM / DMARC arose out of the obvious benefit to all email providers for more reliable spam filtering. It’s not unimaginable that the Mag 7 might agree to solve their collective action problem by doing OPEC-style quotas where each company agrees to hire a certain percentage of junior engineers yearly. In such a system where everyone pays the cost of training engineers equally, everyone benefits.

While none of these in and of themselves will solve the whole problem, my aim with this essay is to at least convince you that all the people smugly wondering where we’ll get senior engineers from if we don’t hire junior engineers haven’t actually articulated a solution. They’ve only described a problem.

Calling the emperor naked doesn’t make you smart when it’s already obvious to everyone. Worse yet, it doesn’t clothe the emperor. When will someone actually propose a sartorial solution?

To gauge the extent of damage, doctors needed to take a liver biopsy. The procedure involves local anesthetic, followed by a 6” needle (that looks like a slim straw) being stuck into you, between rib bones, like a climate scientist taking a core sample of Antarctic ice. The main risk is hitting an artery — a risk, if you can believe it, modern medicine mitigates by… actually, they don’t. It’s Bleed Out Roulette Night at Seattle General, where any spin can win!

The risk of bleeding out and requiring emergency surgery was on my mind as they swabbed the right side of my rib cage with iodine. I felt the anesthetic jab, which, compared to the rather bigger, disturbingly hollow needle on the nearby tray, seemed no more bothersome than a mosquito bite. My anxiety peaked as the surgeon picked up the big needle and asked me to hold my breath.

At that moment of highest intensity there on the operating table, I felt the warmth of someone’s hand grabbing mine, that of the nurse standing next to me. The comfort and relief that flooded into me at that moment was like a light being turned on in an otherwise pitch black basement. I welled up with gratefulness.

The procedure was over in seconds, and she released my hand silently. Though we didn’t acknowledge anything had happened, it continues to be one of the kindest things anyone has ever done for me. She took a risk by reaching out at a critical moment to do the human thing, and I’m so grateful she did.

All-American Rejects

You’ve no doubt on occasion felt the instinct to say or do something that could be helpful, wonderful, or amazing. But I bet if you’re anything like me, you resist nine out of ten times. After all, there’s the distinct possibility of rejection, or of the other person mistaking your good intentions.

Once while navigating a Tube station in London, I saw an old couple standing at the top of a long flight of stairs staring at their two big suitcases. A swirl of busy commuters eddied past them. I went up and offered to help lift their bags down the steps. The wife was just beginning to smile when her husband interrupted with an angry, “No! We don’t need help!” as he swatted at me like the biggest Asian fly you’ve ever seen. I felt embarrassed even though I had done nothing wrong. Perhaps I looked too much like the type of guy who would “wash” your windshield using brown water at a stoplight and then demand you pay for the “service.”

But here’s the thing: I’ve helped many people with their luggage down Tube station steps. And even more frequently, I’ve helped parents by lifting the lower half of the stroller while they held the handles, all the while hoping I didn’t turn the endeavor into some slow-motion baby-stroller-bounces-down-Union-Station-staircase homage. Each and every time, the people were thankful for the help and happy I had offered. In fact, in my five years riding the Tube in London, there was only that one man who ever responded angrily.

Most times, people are happy someone tried to help them. Most times, you’re thanked, not shouted at.

And each time you do such a thing, it forms a distinct, positive memory that punctuates the otherwise mundane, always identical other 1,249 days you commuted by Tube while living in London. Our memories comprise the unexpected.

Risks I’ve Taken

My therapist of many years — let’s call her Carly — had been her brother’s caregiver as he underwent increasingly worsening dementia. Any of you who’ve taken care of someone with dementia will know just how difficult and stressful it is to always be on call trying to ensure the safety of your loved one. There was even one video therapy session during covid when, as I talked, Carly’s brother wandered into the background and began rifling through various items on the credenza behind her. I had been totally ok with this, but Carly was mortified when she noticed her brother in frame.

In the midst of those months, as things got steadily worse with her brother, there was one session where I noticed Carly seemed not quite herself. She looked almost teary during much of our conversation. As we continued talking, I kept flipping back and forth wondering whether I should check on how she’s doing. Doing so felt like it would violate a sort of social expectation around our roles. But after some internal debate, I decided to just ask Carly how she was doing, mentioning I was concerned she looked very sad.

She broke out in tears and explained that things with her brother had reached a breaking point of her ability to continue helping him. As with most caregivers, I’m sure she also felt guilt around the unavoidability of institutionalization. She said she was embarrassed and mumbled something about her being unprofessional, but the sentiment she expressed most was her thankfulness I had noticed and asked. It was one of the most human moments of our work together over many years.

In a much riskier vein, I once woke up around 3 or 4 a.m. and couldn’t stop thinking about a good friend of mine, so much so that I couldn’t fall back asleep. Perhaps I had watched way too many ads for Time-Life’s Mysteries of the Unknown as a child, but I kept thinking I should call him. This was back before smartphones, so there was no way to just send a message; you actually had to risk having a rotary phone clang repeatedly in someone’s bedroom.

Ninety-nine times out of a hundred when you call a friend out of the blue at 3 a.m., they’re going to have been sound asleep. The remaining one out of a hundred times, they’re awake but completely baffled why you’d call. You’d end up muttering something around being concerned for them, etc, but instead just sound kooky.

Well, in this one particular case, which has never happened before or since in my life, this friend of mine was awake when I called and indeed had wanted to talk. I drove over and we had a very meaningful conversation. I am so glad I called despite how unreasonable and embarrassing it all seemed before I did.

The Road Not Traveled

It’s popular these days to repeat the aphorism that people tend to regret not the things they did, but the things they didn’t do. This is certainly true of me for one seemingly tiny event in my life.

One night in London, some very, very late hour, I was walking through a mostly-empty Leicester Square. (For Americans out there, that has less syllables than you think.) A guitarist had set up an amp and microphone — for who knows what reason, since there wasn’t an audience beyond his four friends in a hundred meters — and began playing Under the Bridge by Red Hot Chili Peppers.

It was a song I had played many times on guitar, one of my favorites. After the first few bars, I expected him to start singing into the mic… but he didn’t. He just kept playing. I looked around at his friends, but they all simply watched him. It was unclear what the microphone was even doing there on its stand.

I had such a compelling urge to go up to sing. He was playing so well; the frisson’s siren call was nearly irresistible. I took one step forward but stopped. Wouldn’t it be so embarrassing? I knew I wouldn’t be ruining any busking prospects since it was way after midnight and nobody was around. But sing? In front of five strangers?

I wavered in indecision as the song wound on, all the way until the guitarist landed its last melancholy broken chord. There were a few scattered claps. He turned off the amp and unplugged all the gear. Apparently the goal was just to play that one song to his four friends standing in an empty Leicester Square. I walked on, pensive the rest of the way home.

I don’t know why I still think about that initial moment, that one step I took toward nearly singing that night. It would have been beautiful. It would have been memorable, not just for me, but for everyone — not because I’m Pavarotti, but precisely because I was a stranger. A moment of human connection, a moment of delight and wonder. Maybe I think of this moment often, and of the decision I made not to participate, because it represents a greater pattern I see in myself, in my life. To have joined in that night would have been to live in the moment, to live with abandon. It would have been to, for just once in my life, dance like no one’s watching. It sounds trivial or perhaps nonsensical, but Leicester Square may well be the answer which comes to mind when someone asks what I regret in my life years from now.

Unlike, say, The Life of Chuck. If you haven’t seen this movie, you really must. It’s only “Stephen King” in the same way that Stand By Me or Shawshank Redemption are “Stephen King” — there’s nothing scary about it. The Life of Chuck is exactly what I’m talking about.

Reach out to friends. Say yes to more things. Speak when you’re thankful. Compliment someone when they do something well. Sing. Take the risk.

I, for one, will never shoo you away like some sort of do-good Jeff Goldblum.

The first time I ever heard of specs as a concept was nearly 30 years ago as an intern in Microsoft Word. The PM repeated the team’s common knowledge to me: “Words are cheap. Code is expensive.” That observation was the foundational justification of Microsoft’s waterfall days: it’s wiser to scrutinize a spec until you were absolutely sure everything was right before handing it off to developers. Measure twice, cut once — because cutting again was supremely time consuming.

But what if the economics of time shifts so that, with AI’s advancement, coding eventually approaches wish fulfillment?

One obvious consequence is that classic software moats become harder to maintain. In a world where we have what is essentially the Star Trek replicator for Earl Grey software, the cost of mimicry goes to zero. You’d need something else to sustain a business, like scale economies, network economies, or branding.

I’m more interested in the personal consequences. What happens to you, someone working in technology, as the speed of code production goes ludicrous?

From Founder to Billionaire

What are the odds that a young founder goes from ideas to billionaire without any issues? Sure, you can name a handful of twenty-something founders who successfully helmed their eventual unicorns — but I’d argue that’s survivor bias.

More often than not, and quite understandably, what got you here won’t get you there. The odds you’re a great engineer who solders your own hardware, as well as a brilliantly charismatic salesperson, nears zero. The world, mostly, comprises different talents in different people. It’s why the phrase “triple threat” even exists. Steve & Bill are the exceptions, not the norm.

The march from solopreneur to unicorn has some recognized stops along the way:

• When it’s just yourself, you need to:

  • have an idea or ideas

  • have a tolerance for personal risk

  • have strong self-confidence and self-reliance

  • be able to build prototypes well

  • be good at courting angels

• Once your team is a handful of people, you need to:

  • manage ICs well

  • be one of the highest technical output people on your team

  • be discerning when hiring ICs

• By the time you grow to 50 people, you need to:

  • set and reinforce a positive culture

  • manage managers well, which is quite different from managing ICs

  • manage people with entirely different competencies, like sales staff

• At 200 people, somewhere beyond Dunbar’s number, social currency is no longer enough to sustain healthy cross-group relationships. You now spend your time:

  • establishing scalable people infrastructure for performance evaluation, compensation, and inter-group collaboration

  • getting serious about pricing strategy and other long-arc business considerations

  • interfacing with your lead PMs and designers enough to make product decisions without you

• Things take off and you 10x suddenly to 2,000 employees. Along the way, you’ve had to:

  • be amazing at recruiting and hiring executives, which also includes the ability to discern excellence in domains you know little about

  • outmaneuver competitors

  • raise multiple rounds of investment under increasing amounts of scrutiny

  • be charismatic at every product launch and in every public appearance

Look at this (very incomplete) list of the things you’d have to do to build a successful company. The requisite skills at each stage are materially different, some dramatically so.

Almost everyone has a point in this journey from seed to exit where their talents and their growth max out and are no longer sufficient to run the company. In fact, most people can’t even take the first step of starting on their own.

AI Accelerates You Into Your Incompetence

There’s an analog to this pattern of ever-changing competencies required to stay successful in the world of AI.

There used to be a time when Microsoft talked about “journeyman developers” — people who reach L63/64 (Senior SDE, E5) — being solid lifetime employees. Every team needed engineers who could bang out code independently and reliably, at high quality, and think holistically about the architecture and infrastructure of the entire product while balancing technical realities with business needs.

But by making some things increasingly cheap and fast, like junior-level coding or exhaustive test coverage, individuals are quickly being pushed to their level of incompetence because many of the things they were good at, which used to take up entire sprints, are now commoditized and take minutes.

By creating something that spews code all night while we sleep, we’re left with the other parts, things like:

• Managing multiple agents and multiple development trees simultaneously. I wrote previously about how all of us are becoming managers. And many of us won’t naturally be good at managing.

• Deciding what features to build all the time. Can you imagine a PM’s day in a future world where entire features might be written in minutes once they’re articulated in words? Teams in large companies will need to shift to higher PM:dev ratios. Smaller companies may well ask their engineers to PM more of their own work, which isn’t something many engineer are naturally good at doing (or frankly enjoy doing).

• Coming up with business strategy. Sure, business strategy was always necessary, much like product design was always necessary even back in the 80286 days. But we’re talking about a vast difference in the quantity of those decisions which now need to be made on a day-to-day basis, whereas these key decisions used to bake for long periods, sometimes even over months, giving everyone a chance to rethink things to be sure what path to take. Quantity has a quality all its own.

• Organization, prioritization, scheduling, and multitasking. These, too, were always domains the strongest engineers already mastered. But if you have any weaknesses in these areas, the impact of those weaknesses are about to be greatly amplified.

When the largest part of a traditional job, the coding portion of software development, takes increasingly less time, the substance of one’s days becomes dominated by the remaining non-coding tasks.

Romance author Coral Hart released over 200 novels in 2025. She went from a traditional romance author into possibly the most AI-forward author of last year. Setting aside all the snarky, self-smug comments I’m bound to receive on this post about how “her books must be slop” — a premise, by the way, I’d challenge unless the commenter is truly well-read in not only the genre as a whole but also in Hart’s 2025 novels — the real point is that Hart took one aspect of a romance author’s job, inventing stories and plots, and made it the core of her work.

If she is successful, it’ll mean she’s very good at those aspects of her profession, and that, at least amongst her base of followers, the sentence-by-sentence act of writing is not central to her continued success. If she is successful, she will have made the transition to the AI-enhanced world of book authoring in a way that perhaps many authors can’t… or won’t.

What’s the equivalent in your work? Are you ready for your non-coding skills to be an increasingly large determinant of your future success?

A representative sample of this angst arrived by email this past week:

How do I position myself to survive the tide of AI without deprecating my technical competency as a junior engineer? There’s a large disparity in how senior engineers pilot AI to achieve their goals versus junior engineers like myself.

I am of the opinion that using AI and relying on AI will deprecate my technical competence. I’m not technically competent enough to say that I can completely understand every line of code that AI writes, to be frank.

Do you have any advice for me on how I can position myself to grow as an engineer without being left behind by agentic AI while maintaining the high output from using AI?

Junior Hiring Has Shifted

Numbers are all over the place when it comes to how US CS grads are doing in the job market.

• The NY Fed claims there was 7% unemployment for CS grads in 2024. Note that this rate is only topped by anthropology, fine arts, and performing arts.

• Furthermore, 19.1% of CS grads were underemployed, meaning they’re working in fields unrelated to their degree or involuntarily working less hours than they wanted.

So just a bit over 1 in 4 CS grads aren’t working in CS, which, for reference, is more than peak unemployment during the Great Depression. But the bulk of this is underemployment (e.g. Wendy’s or Allbirds before its AI transformation), not literal unemployment.

These numbers aren’t as bad as some outlets have claimed — I’ve seen breathless (and factless) declarations of 10% job placement in noisy outlets like Reddit — but they’re also from 2024, the last year the Fed has reliable data broken down by majors. My personal anecdotal data, however, feels worse.

Mix in the fact that my son was interviewed via voice by an AI for a possible internship… The entire landscape of junior hiring is changing rapidly.

To AI or Not to AI

Back to one of the questions posed at the top of this post: should junior engineers embrace AI? How do you build technical skills if you’re relying on AI all the time?

The asker was spot-on in observing that senior engineers pilot AI completely differently from how junior engineers do. One advantage of having a lot of coding experience is that you tend to develop a Spidey Sense™ for where the skeletons are buried. For instance, when building a website using AI recently, I suspected it was probably hardcoding a bunch of colors and other CSS values that should really be constants. I double-checked my suspicion with the AI, and indeed it turns out the entire codebase was littered with hardcoded values which I then asked it to extract into constants. Code architecture is another thing which I often have a sense that AI will do poorly, so I actively ask about whether it’s factored a certain component in a specific way (mostly to be told that, yes, it had written code in an unmaintainable fashion).

Experienced engineers have a sense of what types of shortcuts are taken because they’ve worked enough to have seen it all — either by reflecting on the growth of their own skills over time or by observing the types of errors which junior developers often make. Experienced engineers proactively pilot AIs to circumvent these pitfalls.

Junior engineers are caught in a bind. On one hand, it’s essential to embrace AIs (yes, multiple ones, so you know the relative strengths of all the major models) because future employers will likely expect (and eventually require) it. On the other hand, constantly relying on AI to jam out all the code eliminates the core opportunity by which you learn the insights that will grow you from a junior engineer into a senior one.

I’d recommend these principles when considering how to walk the fine line between being overly reliant on AI and being left behind by everyone:

1. Always build the first instance by hand. If it’s the first time you’re doing something new (e.g. using a NoSQL DB, implementing an advanced data structure), do it by hand. Nothing replaces the experiential learning of doing something yourself.

2. Steadfastly review all code produced by AI. You should carefully review all AI code even when doing things you’re already familiar with. This isn’t about catching mistakes (though you inevitably well) — it’s about learning two types of insights:

   1. Better approaches to specific problems. It’s a regular occurrence for me, even as an engineer with decades of experience, to learn new and better ways of solving specific coding problems by reviewing AI code. It’s like crowdsourcing approaches to expressing logic: you’re bound to learn alternatives to solving the same problem.

   2. Anti-patterns to avoid. One great part about reviewing current AI code is that it’s often poorly constructed. So poorly, in fact, you’re bound to learn anti-patterns simply by correcting it so often. Cleaner factoring. Maintainability. You know the saying that you never know something as well as after you’ve just taught someone else? You’ll have this experience all the time reviewing AI-generated code.

3. Always Be Coding. Although this was already to your advantage well before AI, the practice of constantly coding is super valuable now more than ever, if not simply for the fact you can compress years of classical coding experience into mere months by being an order of magnitude more prolific. The other reason, of course, is that tools and approaches are shifting so quickly that you’ll need lots of hands-on experience just to stay relevant.

Gain Real-World Experience

The job market has now become an employer’s market (a buyer’s market), especially since even senior engineers are regularly laid off. Employers can be, and have become, very choosy when it comes to which college grads to hire (if any at all).

In this environment, real-world experience counts far more than academic achievements. Having a set of internships each summer is an essential differentiator. Getting the first one is the hardest part, so you’ll need to pound the pavement and work much harder than your peers in applying to opportunities proactively. But once you have one internship, it’s much easier to parlay that experience into a better internship the following summer as long as you take each experience seriously by working above-and-beyond your fellow interns and by putting in extra time to achieve outsized results.

Remember that the tech world is small, and the best people intermix amongst top companies all the time. In each job you ever have, do the very best work you can. Your reputation carries with you.

Also remember that for the very best internships, you should be willing to pay them. Never turn down an internship due to compensation. The foremost consideration should always be, “Will this experience make me a stronger candidate for future jobs?” And, of course, if you have the luxury, “Will I enjoy the work?”

The Nxtern Revolution

Toward that end, Microsoft Corporate Vice President Ashok Kuppusamy and I are launching Nxtern (https://nxtern.com), a new augmentation to a traditional college education that emphasizes real-world experience and tangible skills.

College students can take a gap year to be a Nxtern. They’ll work three days a week at Microsoft, doing real work and shipping real features. The other two days are spent in-person in Seattle with their peers doing AI-driven learning that directly pertains to skills required to become great engineers.

Nxterns get paid for the year, even while gaining concrete skills and experience to make them even more valuable to future employers. They also meet a set of strong peers who can become allies in career growth.

If you’re interested in learning more about this, please reach out at https://nxtern.com. Please also pass this along to others you know! I’d love for Nxtern to flourish as a way to increase job opportunities for college students.