Modernity & The Mountaintop

On the eve of his assassination, Martin Luther King Jr. spoke of having been to the mountaintop, referencing a vantage point of great potential, with a likely bumpy path ahead:

Right before this section, which is the end of his speech, he (eerily and presciently) spoke about the previous assassination attempt where someone stabbed him in the chest. I strongly recommend listening, not just reading; the emotion is touching. That assassination attempt happened in New York City. While thinking about all he endured and worked for, his reflection of a little white girl writing him a letter, his willingness to call White people who wanted to kill him brothers, while riding my bike down the streets of Manhattan, I began to cry for the beauty, the potential and the dark sides of humanity. All history is cyclical and instructive to our modern times.

Measuring human progress

In my writing and thinking, I always try, admittedly at times for topics that are difficult to quantify or hold to a scientific rigor, to find patterns and norms that explain our shared (albeit, diverse) realities. Given the variety of belief systems and human experiences, there are many different kinds of frameworks that people have tried to use to measure human progress, assessing qualities like fulfillment, satisfaction, happiness.

Technology as a key driver

Technology has always undergirded at least a portion of that, in that what we have access to, or the status of technology, can have an important foundational element to our ability to flourish.

Rules & access matter

It's also the case that there are many dimensions to health and wellness and fulfillment, and another important component, as it pertains to modern times, is access to that technology. That's usually determined by governmental structures, rules, regulations, and laws that have an influence on how broadly adopted available technology is. There's no hard and fast rule, but in many cases, the technology can concentrate power. The market doesn't necessarily demand, in any way, shape, or form, that it be broadly shared. There are wildly broad strokes of advancement of technology, and human progress and fulfillment are definitely up and to the right on graphs and correlated. We have plenty of data around health outcomes, life expectancy, reduction of poverty over time, these sorts of things. Those are all true, and so you might say that, for the advancement of technology and innovation, many would cite that it can be neutral depending on how it's used.

Compounding pace and complexity

As we have moved through time, and technology compounds, so does its complexity. Many have written about this.²³⁴⁵ There are a couple of important factors:

• The technology itself is harder and harder to understand, therefore fewer and fewer people do understand it. In most modern frontiers, no one fully does.
• Combinatorial evolution, meaning modern technologies are built off of prior ones layered on top. There's just more to know and as a result,  fewer people understand. The burden of knowledge becomes problematic, or can be problematic.

There are downsides to this. The rapid pace and scale of innovation, and the expectation (and in some important cases, requirement... think Covid vaccines) of adoption require us all to trust others to use them or not. 

All "knowledge" is (or has been up until recently) social, you have to figure out who and which sources to trust and where to get your information from.

The digital age paradigm shift

The digital age is both a continuation of trends of combinatorial evolution playing itself out, while also, deservingly in a class of its own. Digital goods are basically marginally free to duplicate and distribute globally instantaneously, a previously unprecedented reach and scale capability. 

A lot of the infrastructure of the internet and modern computing was designed and built on public funding, with the advances freely shared to the world for smart people who can make something people want. In short, fewer people can have more impact than ever. 

Societal influence

This compounding aspect and scalability of digital does manifest in some real and unprecedented concentration of power. One instructive example is Meta and the history of acquisition of the now most valuable components of the overall business: 

We may see continued layoffs and downsizing while productivity or profits of the organization continue to rise, even at that, or more downsizing, given that they touch almost half of humanity on a daily basis. Meta reached roughly 3.58 billion daily users at the end of 2025 with about 78,000 employees, which works out to something on the order of 45,000 people touched for every single person it employs.⁶ That has to be unprecedented. For a relatively modern comparison: GM in 1955 was the largest company in the country and the first in history to clear $1 billion in after-tax profit, and it had almost 600,000 employees at that point.⁹ Meta, Nvidia, Apple, and Google have all left that kind of scale far behind on a per-employee basis.

Zooming out a bit wider (though including Meta), the concentration of the Magnificent Seven¹⁰ within the S&P 500 has risen to roughly a third of the entire index, up from about 12 percent just eight years ago.¹¹

Makers amass power, and, ergo, influence

Markets produce externalities as a feature of their operations: e.g. pollution alongside a mine, smell next to the landfill. Directors at U.S. corporations owe their duties to the corporation and its shareholders; there is no legal obligation to the public good.¹² Large companies do plenty outside their business model in the public interest, and while many people within those organizations truly care about those programs, I would argue they’re primarily driven by how public perception hits their bottom lines.

So we have created agencies, rules, taxes, licensing to address externalities, almost as a rule, to address a damage that has already been done. E.g., the FDA only gained the power to require proof of safety before a drug reached the market in 1938, after an untested sulfanilamide elixir killed just over a hundred people, many of them children.¹³ 

The public-interest bodies are often out-resourced. The FTC runs on roughly $384 million a year.¹⁴ Technology companies spent more than $100 million on federal lobbying in 2025, and Meta alone accounted for $26.29 million of it, more than any other company in any industry.¹⁵  That asymmetry concentrates power further, because it can be cheaper to shape a rule, or deal with noncompliance, than to behave in accordance.

There is a question underneath this about the people making the decisions, and I want to state the version of it I can defend. It is not that engineers are worse people. It is that neither their training nor their profession asks them to be accountable to anyone outside the company. Cech's longitudinal study found that engineering students' concern for public welfare measurably declines over the course of their education.¹⁶ She identifies two mechanisms, and both are recognizable to anyone who has worked in this field. The first is depoliticization: the belief that engineering can and should be separated from social and political questions, because considering them would introduce bias. Notice how respectable that sounds. It is not taught as ignore the public; it is taught as stay objective. The second is technical/social dualism: the ranked pairing in which the technical work is the real work, hard and legible, and the social part is soft, secondary, somebody else's department. Between them they produce a professional identity in which caring about downstream human effects reads as a lack of seriousness.

I should be honest about the size of that study. It followed 326 students at four schools, all of them in Massachusetts, so it is suggestive rather than conclusive, and I would not build a national claim on it. I include it because it matches what I have watched happen, and because the direction of the finding is hard to explain away: these were the same individuals measured over time, so it is not that the field attracts people who care less. It is that four years of training left them caring less than when they arrived.

The closest comparison is not medicine, it is architecture. A doctor or a lawyer harms one client at a time, through direct contact. An architect harms people who never met them, through a decision made years earlier, at scale and all at once. That is software's shape exactly, and software knows it, because it borrowed the title: we have software architects, systems architects, solutions architects. What it did not borrow was the licensure. Architects are licensed precisely because indirect structural harm is still harm. And even in the engineering disciplines that do license, the industrial exemption means most engineers working inside a company never need the credential at all, which removes it in exactly the setting where the work reaches the most people.¹⁷ We license roughly in proportion to how badly someone can hurt a stranger. A plumber needs a license. Someone shipping a ranking change to three billion people does not.

The most honest illustration I know comes from inside the industry. Y Combinator's founders were, by their own account, excellent judges of technical ideas and not good judges of character, and so the judging of people fell to Jessica Livingston, the co-founder they nicknamed the social radar.¹⁸ The most influential startup accelerator in the world understood that reading humans was a separate and necessary skill, one its technical founders did not have, and it named the person who did. That is the division of labor this whole section is about, stated plainly by the people it describes.

We likely don’t have the luxury to wait for damage and handle it retroactively going forward; we may need a new approach and even new institutions. But, we do have multiple precedents for asking more of private organizations: American corporations were originally chartered for public purposes: a bridge, a canal, a bank the state wanted built. The charter was a privilege granted in exchange for a public benefit, not a right, and that link was severed over the course of the nineteenth century.¹⁹ It has since been partly rebuilt. More than thirty states now recognize benefit corporations, which are permitted, and in some formulations required, to weigh public benefit alongside profit.²⁰ Anthropic, an organization on its way to being one of, if not the most influential one in the world, is formally incorporated as a public benefit corporation (PBC).

And it is not purely hypothetical. The first regulation I know of that tries to get ahead of the damage rather than answer it is already law. Since 2023, New York City has required an annual, independent bias audit of any automated tool used to make employment decisions in the city, before it is deployed and every year after, testing for disparate impact by race and sex.²¹ The European Union's AI Act requires a conformity assessment for high-risk systems before they reach the market at all.²² This is the FDA's 1938 lesson applied in advance instead of after a body count. It is also fragile: Colorado passed a comparable law and then repealed and replaced it with a substantially weaker one before the original ever took effect.²³ Which is both halves of my argument in a single example. We are capable of acting before the harm, and the same concentration of resources that makes acting necessary is very good at making sure we do not.

Smartphones and social media, lauded at first

The sentiment around social media and smart phones was euphoric at first. Naturally, both provide tremendous value, and it's clear because they are so broadly adopted.

But as the business models around them evolved, we began to see some serious potential downsides in their over-usage. The Attention Economy for an ad-funded business refers to the design goal of optimizing time spent on the platform. 

What holds attention, we've learned, turns out to be problematic. Out-group animosity, for example, drives engagement.²⁴ Every moral-emotional word added to a message raises its expected sharing rate by roughly twenty percent.²⁵ In headline tests on real readers, each additional negative word raises click-through by about two and a half percent.²⁶ Anger and sensationalism are not incidental to the feed; they are what it selects for, an unfortunate feature of human psychology.

Many will be familiar with the term: "doom-scrolling", which measurably lowers wellbeing, and a mechanism the researchers identify is envy: watching an edited version of everyone else's life.²⁷ Over the same two decades these products spread, in-person time with friends among fifteen to twenty-four year olds fell from roughly a hundred and fifty minutes a day in 2003 to about forty in 2020, a decline of nearly seventy percent.²⁸ Among American teenage girls, the share reporting persistent sadness or hopelessness rose from thirty-six percent in 2011 to fifty-seven percent in 2021.²⁹

I want to be careful with that last one, because it is the claim most often overstated. The best-powered analyses find screen time explains well under one percent of the variation in adolescent wellbeing, a far smaller effect than most headlines imply.³⁰ The honest position is that the timing is striking, the causal case is unproven, and the documentary evidence of what these companies knew internally is stronger than the correlational evidence about their users.

The results have not been good

The mental health numbers moved in one direction. The share of American high schoolers reporting persistent sadness or hopelessness went from thirty percent in 2013 to forty-two percent in 2021, and among girls to fifty-seven percent.²⁹ Among college students, the share who had ever seriously considered suicide rose from about twenty-four percent in 2011 to thirty-seven percent in 2019.³¹

But the more revealing pattern is what young people stopped doing. Across seven nationally representative surveys covering 8.44 million adolescents from 1976 to 2016, nearly every marker of independent life declined together: having sex, dating, drinking, working for pay, driving, and going out without a parent.³² The share of high schoolers who had ever had sex fell from fifty-four percent in 1991 to thirty percent in 2021.³³ Twelfth graders who had ever been on a date fell from eighty-seven percent in 1980 to fifty-eight percent in 2014.³² Teen substance use now sits at historic lows.³⁴

That last one is genuinely good news, and it is what makes the pattern legible. If drinking, sex, driving and holding a job all fall at once, health is not the common factor. Something more basic contracted: the amount of unsupervised life a young person gets to have. Risk is not only a cost. It is also how a person finds out what they can survive.

Which matters because the thing being lost is the thing that matters most. The Harvard Study of Adult Development has followed people since 1938, and relationship quality at fifty predicted physical health at eighty better than cholesterol did. A meta-analysis of 308,849 people found strong social ties predict about fifty percent better odds of survival, an effect comparable to quitting smoking.³⁵ We are social animals, and connection is not a nice-to-have on top of a good life. It is most of what a good life is made of.

The result reads like fragility, and I think it is more accurate to call it inexperience. If you have had fewer unsupervised hours, fewer awkward conversations, fewer small rejections survived in person, you arrive at adulthood with less evidence that you can handle things.

I want to be careful how I say that, because complaining about the young is the oldest reflex there is. It has also been studied: across five preregistered experiments, adults reliably judged today's youth as deficient, and specifically in whatever the adult happened to be good at. Well-read people thought the young read less; intelligent people thought the young were less intelligent. The mechanism is a memory bias that projects who you are now onto who you remember being.³⁶ So the reflex is real and it is a documented error.

The difference is that this is measured rather than remembered. And the young people in that data are not a worse generation. They are the same kind of people we were, handed an environment engineered by adults to hold their attention, and they are the ones who will shape everything that comes next. That seems like a reason for concern and effort on their behalf, not contempt.

Enter AI (the LLM era)

Now, enter AI. I deserve a purple heart for this being the first time in this article those letters were mentioned. Go ahead, CTRL+F it; I'll wait.

I'd be remiss to not recognize the field is both over 70 years old, and that Large Language Models (LLMs) really changed the game in 2022, largely made possible by the Attention Is All You Need research published by Google in 2017.³⁷ Math that can understand and replicate natural language to a degree that is either the same or imperceptibly different from how we produce language has been a game changer upon which modern commercial AI has promulgated.

Why AI is different in important ways

Combinatorial evolution and burden of knowledge naturally apply to AI, and it's of course a component of the digital/information paradigm. 

Capability: Every previous technology was a tool that did a defined job. Even computers and software were almost exclusively deterministic (had consistent, predictable, and programmable outcomes from inputs). AI is the first technology that learns from patterns it finds, not from directions we give it. The mathematics that do this can be explained, but the output can be indistinguishable from, and increasingly superior to, human intelligence. That is new(ish). 

Perception: Fifty-two percent of Americans now say they are more concerned than excited about AI in daily life, up from thirty-seven percent in 2021 (pre commercial LLMs), and among adults under thirty it is fifty-five percent, up from thirty-one.³⁸ Its unprecedented lack of popularity, and its continuing decline, are in lock step with its staggering pace of improvement. Importantly, most correctly feel they cannot opt out of it, which exacerbates the issue.

The concerns are varied and valid: 

Work: Employment for twenty-two to twenty-five year olds in the most AI-exposed occupations now sits roughly nineteen percent below where it would be if it had tracked their peers in less exposed work. The mechanism is reduced hiring rather than layoffs, which is why it is easy to miss.³⁹

Energy: Data centers account for close to half of all growth in US electricity demand between now and 2030, and are projected to consume between nine and seventeen percent of the country's electricity by then.⁴⁰

Fraud: Americans reported roughly $893 million in losses to AI-assisted scams in 2025 across more than 22,000 complaints, including cloned voices used for family-emergency calls.⁴¹

Children: More than seventy percent of American teenagers have used an AI companion, and about one in three has used one to discuss something serious instead of talking to a person.⁴²

Concentration: Everything in the earlier sections of this piece, except that the capital requirements are larger and the number of organizations that can compete at the frontier is smaller.

Worst case: AI is not the first innovation that some have speculated could end it all for the human experiment. Some project as low as a 1% chance⁴³, but to me, that's a non-trivial amount and should be taken seriously as a possibility that we continue to collectively discuss and do our best to avoid. 

We must shape the future together

Given the reach, potential and probable society-changing impact AI will continue to have, we have no choice as a society other than to face it head on and do our best to shape it to optimize the benefits to everyone. I offer practical thoughts on how we should do that individually and societally in part two.