~15 minute read

This is the second essay in a series on deliberate identity construction. The first, “The Story Has to Hold,” examined why most attempts at personal change fail. This one asks what comes before change: seeing the story you're already living inside.

There is a particular kind of discomfort that comes from being asked to describe yourself. Not your job or your life history, not the biographical facts that would appear on a form, but who you are as a person.

The question arrives occasionally at dinner parties, job interviews, sometimes in therapy, and almost never in the right moment. Most people, when confronted with it, offer a list of traits. I’m introverted. I’m competitive. I’m not great with money. I’m loyal, sometimes to a fault.

The statements come quickly and have the texture of observation…of things noticed and reported, the way one might say the sky is cloudy today or this shirt is grey.

But these are not observations. They are conclusions. And the difference matters more than most of us ever realize.

In a previous essay, I argued that durable identity change depends on narrative coherence. That the person you become must be someone the person you were could plausibly have become.

Most personal transformations fail because they attempt to install a new self without updating the story that makes it believable.

What I did not adequately address was a more foundational question that precedes any attempt at revision: What story are you telling right now?

You cannot edit a draft you have never read..

This is harder to answer than it appears. Most people have never articulated their identity narrative. They could not write it down if asked. And yet they live inside it every day. It shapes which opportunities feel plausible and which feel absurd. It determines which risks they perceive as courageous and which as reckless. It whispers, constantly and with brute authority, about what kind of person you are, and therefore, what kind of life you’re allowed to lead.

The narrative is there. It is just invisible to the person playing it out.

For most of my life, I believed I was an introvert. Not in the colloquial sense of being shy or socially anxious (I was neither) but in the more precise sense that solitude was where I drew my energy, and other people, in sufficient quantity, depleted it.

This felt like a fact about my constitution, as fixed and unchosen as my height. Some people are energized by crowds. I was not. That was simply how I was built.

As a child, my mother, a pediatric cardiologist, had needed to work constantly. My brother, four years older and by then a teenager, had little interest in spending time with me. Friends existed, but going to someone’s house was an occasional indulgence rather than a daily rhythm. And so I learned early on to be my own best companion.

I read books. I watched movies. I spent hours on the computer, in the early days of the internet, following threads of curiosity into corners no one had asked me to explore. I remember coming home from school and looking up animal species with the seriousness of a doctoral researcher: the Arabian sand fox, the snow leopard, creatures I would never see but whose existence I was fascinated by. The world was enormous and accessible from my desk, and I did not need anyone else to explore it.

This was not deprivation. It was, in many ways, a gift. I developed an interior life that was rich, self-sustaining, and deeply pleasurable. But it also laid down a template…a kind of default setting that would run, unchallenged, for decades. Solitude was comfortable. Groups were overwhelming. Large social gatherings produced a specific, low-grade exhaustion that I interpreted as evidence of my nature rather than as a conditioned response.

The conclusion was obvious and felt permanent: I am someone who thrives alone.

I did not experience this as a story. I experienced it as a fact.

The psychologist Jefferson Singer has spent much of his career studying what he calls “self-defining memories”. These are the specific episodes from our past that carry disproportionate weight in how we understand who we are.

Not all memories are created equal. Most of what we experience passes through us and leaves only a residue…a vague sense that something happened, like a general impression without specifics. We forget almost everything.

This is not a failure of the system; it is the system working as designed. A mind that remembered every moment with equal vividness would be paralyzed by its own archive.

But certain memories resist being forgotten. They persist with unusual clarity, return unbidden and, crucially, end with a conclusion about the self.

Singer has found that these self-defining memories share a cluster of characteristics: they are vivid, emotionally intense, frequently recalled, and linked to an enduring concern or unresolved theme in the person’s life. They are not random. They are the episodes the mind has selected, from all the available evidence, to serve as proof.

I think of them as anchor memories because they are the load-bearing elements of the identity structure and they tether us to a specific conception of the self. When a person says, with quiet certainty, “I’m not the kind of person who enjoys networking,” there is almost always, somewhere behind that claim, a specific scene. A crowded room. A conversation that went nowhere. The particular fatigue of performing interest in someone else’s small talk while calculating the earliest polite moment to leave.

That memory has been elevated above the fray. It has been rehearsed numerous times, sharpened, and compressed into a kind of thesis statement about the inherent dispositions of the self.

One episode, speaking for a lifetime.

The problem is not that anchor memories exist. We need them. A sense of self without specific evidentiary grounding would be unstable, like an assertion without supporting evidence, vulnerable to any contrary wind.

The problem is that most people have never examined the evidence their identity rests on. They have never asked whether the pillars deserve the load they’ve been assigned.

My anchor memories for introversion were not single dramatic episodes.

They were an accumulation: hundreds of afternoons alone in my room, the particular contentment of a book and silence, the contrast between that contentment and the mild overstimulation of a birthday party or a family gathering.

The sheer volume of these episodes formed a pattern. And patterns, by their nature, feel more like truth than interpretation.

But even patterns can be misleading when they are drawn from a biased sample. And this is where Martin Conway’s research on the working self becomes essential.

Conway describes the working self as an executive function that regulates which memories we can access and which we suppress. Its primary concern is coherence: ensuring that the memories we retrieve are consistent with our current understanding of who we are.

When a memory aligns with the prevailing self-concept, it is retrieved easily, vividly, and with a feeling of rightness. When a memory contradicts the self-concept, the system resists… not through erasure, but through reduced accessibility. The contradicting memory becomes harder to recall, less vivid when found, and easier to dismiss as an exception.

This creates a feedback loop of extraordinary power. The anchor memory supports the identity claim. The identity claim makes the anchor memory easier to retrieve. And the ease of retrieval reinforces the sense that the claim is simply true as opposed to just a conclusion drawn from selective evidence.

The circularity is invisible to the person inside it. It feels like self-knowledge. But it is, more precisely, a self-fulfilling prophecy.

The mechanism has a corollary that is less often discussed: suppression.

If the working self elevates memories that are coherent with the self-concept, it must also be suppressing memories that are not. And this is exactly what the research suggests.

For every anchor memory holding a self-narrative in place, there exists somewhere in the archive, unretrieved and narratively inert, a collection of counter-evidence. Episodes in which the person was exactly what they claim not to be. Moments of social ease from someone who “isn’t a people person.” Moments of courage from someone who “lacks confidence.”

These episodes are not imagined. They happened. They are stored. They are simply not being retrieved, because retrieving them would threaten the coherence the working self is designed to protect.

When I became a venture capitalist, the introvert narrative ran into a structural problem.

The profession does not accommodate solitude in the way that, say, writing or research does. A venture capitalist’s success depends almost entirely on relationships: the ability to meet founders, build trust, sustain networks, and show up, repeatedly, in rooms full of people. The job is essentially a social practice disguised as a financial one. And the story I had been telling myself, that I was constitutionally unsuited for sustained social interaction, was, for the first time, not merely limiting but professionally untenable.

What is interesting, in retrospect, is not that the story was challenged by external circumstances. Stories are challenged all the time and they usually survive.

What is interesting is what happened when, under the pressure of professional necessity, I began to look, really look, at the evidence the narrative had been suppressing.

There was, it turned out, quite a lot of it.

I had always been good at presentations. Not merely adequate but really good. In school, while other students dreaded oral assignments, I found that something activated when I stood in front of a room. The nervousness was there, but beneath it was a current of energy that I did not experience in other contexts. Public speaking came naturally to me in a way that my introvert narrative had never adequately explained. The working self had handled this contradiction by filing it under “performance”: a separate category, unrelated to the “real” self, and therefore not admissible as evidence.

There was more. As an entrepreneur, before venture capital, I had been great at closing a sale. I was good at reading people. I was good at the particular kind of attention that makes another person feel heard. These were not the skills of someone who was depleted by human contact. They were the skills of someone who was, under certain conditions, energized by it.

But the working self is a ruthless editor. These memories had never been granted narrative standing. They sat in the archive, tagged as exceptions and flukes of circumstance, irrelevant to the larger truth.

The story said introvert. The counter-evidence was inadmissible.

We do not experience ourselves as a collection of data points. We experience ourselves as characters in an ongoing story, and we evaluate our actions not against abstract criteria but against the internal logic of the plot.

This is why the audit cannot be conducted in purely rational terms. It is not enough to list counter-evidence and expect the narrative to update. The working self will resist. The anchor memory will reassert itself and the new evidence will be dismissed…unless it is allowed to become part of the story rather than an appendix slapped onto it.

The process begins with surfacing the anchor memories.

You identify the specific episodes carrying the weight of your self-concept. You ask: What is the memory I return to when I need to explain why I am this way? What is the scene that plays, automatically and without invitation, when someone suggests I might be different from who I believe myself to be?

Then you test the anchor. Not to destroy it, but to evaluate it honestly. Four questions are useful here, though they are not exhaustive.

1. Recency: When did this happen?

   If the anchor is rooted in childhood, for example in the specific circumstances of a mother forced to work long hours and a brother who had better things to do, you are building your present self-concept on the evidence of a child’s adaptation. The conditions that produced the pattern no longer exist. The pattern may have outlasted its cause.

2. Source: Who authored this interpretation?

   In my case, no one told me I was an introvert. The conclusion was self-generated, which gave it a feeling of authenticity that externally imposed labels sometimes lack. But “self-generated” does not mean “accurate.” A child who learns to entertain himself because no one else is available is drawing a reasonable conclusion from the data at hand. A 37-year-old who continues to live inside that conclusion, without ever revisiting it, is being governed by the reasoning of a child.

3. Representativeness: Is the evidence typical, or is it a biased sample?

   My archive of “introvert” memories was large, but it was drawn almost entirely from unstructured social situations: parties, conferences, large gatherings where the mode of interaction was surface-level and diffuse. It did not include the one-on-one conversations where I came alive or the negotiations where I was more present and more engaged than in any solitary hour. The sample was not representative. It was filtered.

4. Function: What is this memory doing for you?

   My introvert narrative was, among other things, a permission structure. It allowed me to leave events early, to decline invitations without guilt, to treat my preference for solitude as a necessity rather than a habit. The story protected me from the discomfort of situations I had not yet learned to navigate. It was a cage that felt like a home.

In my case, what happened upon that kind of narrative interrogation was not a dramatic reinvention. I did not declare myself an extrovert and begin working the room at every conference.

What I did was smaller and, precisely because of its modesty, more durable.

I started attending the events I had been avoiding. But I changed the terms of engagement. Rather than attempting to interact with everyone, which had always produced the overwhelm the narrative predicted, I would find one person, sometimes two, who seemed genuinely interesting.

Not interesting in the professional-networking sense of being strategically useful, but interesting in the human sense of having something to say that I wanted to hear. If those one or two conversations were all I managed in an evening, that was enough. It would be better than not going.

The results were not what my introvert narrative had prepared me for. The conversations went deep. I discovered that I had a natural curiosity about high performers…about the mechanics of how they thought, what drove them, what they had learned through difficulty.

This was not small talk. This was the kind of exchange that, far from depleting me, left me more energized than I had been when I arrived. Over months and then years, these encounters accumulated into relationships of remarkable depth. The kind of relationships that do not form at cocktail parties but might begin at one, if you’re willing to ignore the cocktail party’s rules.

What I had been allergic to, it turned out, was not people. It was a particular mode of interacting with them. The shallow, performative exchange in which eye contact is sparingly deployed and attention is always partly elsewhere, scanning the room for someone more important. The polite nod and the smile that means nothing. That was what overwhelmed me. And I had mistaken the specific for the general. I had concluded that I was not suited for human connection, when in fact I was not suited for its counterfeit.

Over time, and it was years, not months, I became someone I would not have recognized a decade earlier. In the right setting, with the right people, I was a source of energy rather than a drain on it. I could hold a room. I could draw people out. I was animated, engaged, and, most surprisingly, inexhaustible.

But the shift had a condition that I never abandoned: depth. I never learned to tolerate shallow discourse, and I stopped trying.

The story you are living in was written by someone. A younger you, operating on limited evidence, under circumstances that no longer obtain. It was reinforced through years of selective recall and never subjected to serious editorial scrutiny. It feels like fact when in fact it’s a first draft.

What does it feel like when the audit is complete?

In my experience it’s disorienting. The identity that felt solid reveals itself as fluid. The bedrock becomes a draft. You see the anchors for what they are: selected by you, not assigned by destiny.

You begin to see the suppressed evidence for what it is: real testimonies of character, not anomalous flukes. And you find yourself in a strange and unfamiliar position, standing outside a story you had been living inside, holding it at arm’s length, seeing its skeletal structure laid bare for the first time.

The comfort of a fixed identity, even a limiting one, is that it provides certainty. You know who you are. You know what to expect of yourself. You know, with depressing reliability, how the next chapter will go.

To see the story as just that, a story and not a fact, is to lose that certainty. It is to accept that the self is not a given but something that is available for manual control.

The question is no longer Who am I? That question accepts the premise that identity is fixed and discoverable, like a continent waiting to be mapped.

The better question, the one that makes revision possible, is What story am I telling, and does it still deserve to be told?

In a subsequent essay, I will take up the mechanics of revision: how to reframe an anchor memory, how writing encodes new meaning into the self-narrative, and how to borrow the structural tools of fiction to reshape the arc of a life.

These are specific, learnable skills. But they depend on a precondition that cannot be skipped.

You must first see the draft. You must read what was written before you can decide what to write next.

You can now get medical advice, legal analysis, architectural sketches or tax appraisals in seconds, courtesy of your favorite AI bot.

In most cases the output is directionally correct, and in many it is excellent.

But expertise is not the same thing as accountability.

When advice crosses into high-stakes territories like health, safety or financial exposure the question becomes: who bears liability if the model is wrong?

Under US common law, a professional must exercise the degree of skill and care ordinarily exercised by members of the profession under similar circumstances.

If a model misidentifies a toxic plant, overlooks a tumor, or drafts a flawed contract, someone must assume responsibility.

Disclaimers alone are unlikely to solve this.

Under well-established principles of US tort law, courts look to substance over labels. Where a company undertakes to provide advice in circumstances where reliance is foreseeable, it may owe a duty of reasonable care, regardless of boilerplate warnings.

Tort law does not recognize probabilistic brilliance.

This creates a need for a new category of AI-native companies that provide Accountability as a Service (yes, the acronym is doing it no favors).

In AI-Native Services, I argued that AI will enable a new breed of services companies to operate with software-like economics, replacing human labor with agentic infrastructure and selling outcomes rather than hours. One doctor oversees what previously required ten. One attorney reviews what once required a team.

Accountability as a Service is the mechanism that forces this convergence:

AI performs the cognitive work at near-zero marginal cost, while licensed professionals remain in the loop to validate, supervise and assume legal responsibility.

Because professional licensing is inherently local (think medical boards, bar associations, engineering certifications), this category will fragment by jurisdiction.

And that means the next wave of value creation may not concentrate in Silicon Valley. It may emerge through a Cambrian explosion of vertically integrated Accountability as a Service players embedded in local regulatory regimes.

For twenty years, SaaS pricing was built around a simple assumption: humans are the bottleneck of production.

Software amplified human productivity, but it did not remove the human as the unit of throughput.

Instead software gave humans access to insights from massive data sets + it translated that data into an interface they could navigate and use.

But most importantly, and this is the part that the whole AI killing SaaS debate largely misses, until now the interface was also a measuring device!

Logins were how you counted usage. And usage was how SaaS companies were able to charge. And so revenue scaled with the number of human users a company wanted to buy software access for.

The problem is that AI breaks the measuring device.

When an agent can perform the work of ten or one hundred knowledge workers, the relationship between number of seats and value collapses.

The constraint is no longer human attention. It is agent capacity...and that in turn is a function of compute capacity. Compute is variable, non-linear, and impossible to anchor to a fixed “seat”. Depending on the task, you might need the equivalent of 0.1 or 10 FTE’s .

What matters is not whether enterprises have fully deployed agents yet.

What matters is that they understand the implications of unbundling software functionality and interface.

If future workflows require fewer seats, pricing pressure starts now.

So what replaces the seat as the metering unit?

Three new pricing anchors are emerging:

1. Inputs: tokens, API calls, compute cycles.

You charge for usage. Like AWS. Scales with consumption, but weakly tied to business value. You charge for outcomes. Reduce churn by 15%. Close the books in 3 days. Pass the audit.

Each layer increases alignment with value but also increases the difficulty of delivery.

Input pricing is easy to meter but commoditizes quickly. If you charge per token or per API call, you’re charging for raw consumption.

But:

1,000 tokens could generate a trivial email.

1,000 tokens could generate a life-saving medical insight.

1,000 tokens could generate nonsense.

The price is the same.

That means pricing is disconnected from the customer’s economic gain.

When price ≠ value, customers optimize for lowest cost.

And that’s how commoditization starts.

2. Outputs: tasks completed.

You charge per contract reviewed, per invoice reconciled, per ticket resolved. Stronger alignment with value, but only works when tasks are standardized.

Output pricing requires domain control and repeatability. To charge per task, you must:

Clearly define what “done” means.

Ensure quality meets a professional threshold.

Handle edge cases.

Understand the workflow deeply.

You can’t charge per “contract reviewed” unless you understand:

What counts as a review?

What risk thresholds apply?

What clauses matter?

What happens when the contract is unusual?

If you don’t have the domain knowledge, you can’t reliably define the unit you’re charging for.

3. Intent: objectives achieved.

Intent pricing requires something else entirely:

the ability to interpret ambiguous enterprise goals, orchestrate execution, and absorb accountability for the result. That looks less like traditional software and more like vertically integrated accountability (more on that in a later post).

And so, the key question isn’t whether AI replaces SaaS.

It’s what will become the new meter...

Linda has been a 911 dispatcher for twenty-two years. She started in 1998. Back then, the dispatch room was loud: phones ringing, voices overlapping, the scratch of pencils on paper maps. She learned the city by feel: where traffic backed up at five o’clock, which neighborhoods required a second unit after dark. Her knowledge became tacit…accumulated through thousands of small decisions, none of which she could have fully explained if asked.

Now, in 2026, the room is quiet. The paper maps are gone. Calls route through a system that scores them before Linda even hears the end of the first sentence. Sentiment analysis, situational priority, estimated duration, optimal unit assignment, all calculated in milliseconds. Her job, most nights, is to review what the system recommends and click confirm.

She still overrides sometimes. A regular caller the system doesn’t recognize is fragile. A route that looks fine on the map but floods after rain. Each override is logged and reviewed to determine whether her intervention improved outcomes or degraded them. Linda has seen the reports. Her overrides help, on average, but the margin has been narrowing. Her “optimization rate” is now 5%, down from 9.7% the year before. The system learns from her corrections, incorporating them into its models. Each time she teaches it something, she makes herself slightly less necessary.

She tries not to dwell on this.

So what’s actually changing? It’s not that machines can perform work. That’s been true ever since machines started replacing workers on the assembly line.

It’s that machines can now manage.

Bessemer’s Systems of Action thesis correctly identified the first wave of AI automation: software that doesn’t merely assist human effort but replaces it. In other words, software as the worker, not software as the tool. That thesis is now the baseline.

What comes next is software as allocator: Systems of Planning that decide what work gets done, by whom, in what order, under what constraints.

Systems of Action automate work.

Systems of Planning automate the allocation of work.

The first displaces hands. The second displaces everyday judgment, historically the purview of middle management.

And that judgment is where leverage historically lived.

For most of the twentieth century, labor and capital existed in complementarity. Machines were productive only when combined with human work. As capital accumulated, it created demand for labor to operate it. Wages rose and returns on capital fell. For decades, roughly two-thirds of national income flowed to labor and one-third to capital.

This stability of the distribution of wealth was not a moral imperative. It was structural. It rested on a production system in which human judgment remained indispensable.

Frederick Winslow Taylor’s 1911 Principles of Scientific Management began extracting that judgment. Knowledge of how to do the job would no longer reside in the worker; it would be analyzed, codified, and placed in management. The worker executed. Management conceived.

Harry Braverman later named this for what it was: a mechanism of control. Separating conception from execution rendered workers interchangeable.

“Machinery,” Braverman wrote, “offers to management the opportunity to do by wholly mechanical means that which it had previously attempted to do by organizational and disciplinary means.”

And yet, despite this steady extraction, labor survived.

It survived because automation historically targeted action, not planning.

Machines replaced hands. The loom replaced the weaver. The assembly line replaced the artisan.

Later, machines replaced routine cognition: the filing clerk, the switchboard operator, the bookkeeper. But displaced workers did not simply disappear, instead they moved:

The craftsman became the foreman.

The clerk became the coordinator.

The driver became the dispatcher.

Labor survived mechanization by migrating upward: out of doing and into deciding.

Firms needed human planners because planning could not be mechanized. Data was sparse. Real-time allocation across many variables exceeded computational capacity. Information was local and incomplete. Exceptions dominated. Much of what mattered was tacit.

And so the equilibrium was maintained.

This pattern, “humans plan, machines execute”, held long enough for it to feel inevitable.

That assumption is now failing and the reason is straightforward: AI is starting to collapse the cost of planning.

AI systems now operate on continuous streams of real-time data. It’s a perpetual loop of sense, decide, act, observe, revise.

Machine-learning systems allocate under uncertainty, updating decisions as evidence arrives. Sequence models anticipate downstream consequences. Reinforcement systems adjust continuously based on outcomes.

As a result, exceptions which were long the refuge of human planners, are shrinking. In emergency response, companies like RapidSOS and Palantir already triage incidents, prioritize resources, and route responders across complex systems faster than human operators once could. In high-volume labor markets companies such as Fountain automate hiring and shift allocation at scale, determining who gets hired, who gets which shift, and who gets let go.

In each case, the human remains in the loop as an executor. But the planning authority of who does what, when, and in what sequence, has migrated into the system.

Execution may remain human (until it’s not). But planning, and therefore leverage, will increasingly be eaten by AI.

A manager whose judgment is required is hard to replace. A worker who simply follows instructions is not.

Ownership of planning infrastructure is valuable because it replaces the coordination that humans once provided.

This is the moat Systems of Planning create: control over how scarcity is arbitrated. The firms whether vertical SaaS companies, AI-native startups, or infrastructure players, are not competing for workflow efficiency. They are competing for the margin that once accrued to human judgment.

Systems of Action scaled productivity.

Systems of Planning scale power.

The first replaced workers.

The second will determine where the next compounding advantage will reside.

Next week in Systems of Planning Part II: the second and third-order effects, who captures the value, what happens to labor leverage, and why the political window is finite.

Every January, millions of people announce their intention to become someone else. New year, new me. Gym memberships spike. Alcohol sales dip. Willpower surges like a tide across the population, and then, quietly, it recedes. The failure rate of New Year’s resolutions is so well documented it has become its own genre of commentary. What is less often examined is why the failure feels, to the person experiencing it, less like defeat than inevitability.

The resolution does not usually collapse in a single dramatic moment. It fades. The person who swore to become disciplined finds himself, one Tuesday evening, back where he started, as if waking from a brief and unconvincing dream.

The cultural mythology of transformation favors the clean break. The narrative arc we are sold - by advertisements, by self-help literature, by the rhetoric of reinvention - implies that identity change is primarily an act of will. Decide who you want to be. Act accordingly. The old self is a costume to be discarded; the new self awaits, fully formed, on the other side of commitment.

But this model misunderstands something fundamental. Identity is not a costume. It is a story already in motion - open to revision, resistant to erasure.

The psychologist Martin Conway has spent decades studying the architecture of autobiographical memory: how we store, access, and organize our personal histories. His research suggests that identity functions as a “self-memory system,” a dynamic interplay among our goals, our self-images, and our recollections of the past. At the center sits what Conway calls the working self, an executive function that regulates which memories become accessible and which remain suppressed, depending on their coherence with the person’s current understanding of who he is.

The working self is not neutral. It prioritizes consistency. When we try to access memories that contradict our prevailing self-image, for example recollections of cowardice when we think of ourselves as brave or of stinginess when we see ourselves as generous, the system resists. Not through outright denial, but through filtering. The memories still exist, but they become harder to retrieve, less vivid, less narratively central. The past quietly reshapes itself to match the present.

This coherence-seeking is not a flaw. A stable sense of self allows us to plan, to commit, to act with confidence in an uncertain world. But it also means that change which contradicts the past is experienced as threat, an implausibility that must be resolved or rejected. When someone declares, “I am now a person who exercises every day,” and the memory system returns decades of evidence to the contrary, the claim does not feel heroic. It feels false. And identity claims that feel false rarely survive contact with stress.

Most attempts at transformation fail not because of weak will, but because of narrative incoherence. The person is trying to install a new identity without updating the story that makes it believable to his own mind.

Dan McAdams, a psychologist who has devoted his career to understanding how identity evolves, offers a different model, one that depends not on rupture, but on reinterpretation. Identity, in his account, is not a fixed set of traits or a checklist of behaviors. It is an internalized, evolving story that integrates a reconstructed past and an imagined future into something resembling coherence. We are, in a deep sense, the stories we tell about ourselves.

McAdams emphasizes a particular cognitive skill: autobiographical reasoning. This is the interpretive work through which we draw on memory to make sense of who we are and what our lives mean. It is not the raw material of memory that shapes identity, but the meaning we assign to it. A person might recall the same childhood episode at twenty and at forty and derive entirely different lessons. The facts remain; the significance changes.

We store the raw footage of our lives, the episodes, the scenes, the specific moments, in one cognitive register. We store our sense of who we are, our traits, our values, our self-concept, in another. These systems are related but not identical. A person may recall periods of sustained effort yet still describe himself as lazy, or remember instances of generosity while believing himself fundamentally selfish. What connects the two is narrative: the interpretive work of turning episodes into meaning.

This distinction, between the episodic content of memory and its narrative interpretation, is central to understanding durable identity change. The episodes of our lives are, in a sense, fixed. What happened, happened. But meaning can be revised. A humiliation can become a crucible. A failure can become the precondition for later success. The same autobiographical data can support very different stories, depending on how it is arranged.

A counterintuitive truth follows: people do not need new experiences to change who they are. They need new interpretations of the experiences they have already had.

For most of my adolescence and early adulthood, I was significantly overweight. Food was many things to me: comfort, reward, a reliable source of pleasure in a life defined largely by pressure and performance. I did not understand this at the time, but in retrospect food functioned as emotional regulation. It was the thing that worked, in the absence of other things that might have worked better.

I tried to lose weight many times. Diets, exercise programs, various frameworks of restriction. None lasted. The pattern was always the same: effortful compliance, gradual erosion, return to baseline. Each failure felt like proof of something, proof that I lacked willpower, that I was simply not the kind of person who could be fit. The identity I was trying to claim felt implausible. It contradicted too much of what I knew about myself.

What eventually allowed change to begin was not a new program or a stronger commitment. It was a different interpretation of the past. I stopped seeing my history with food as evidence of weakness and started seeing it as evidence of adaptation. Given the circumstances of my life, specifically the stress of performing at the highest level academically, the lack of alternative coping mechanisms, the environment in which I grew up, overeating made sense. It was not a character flaw. It was a solution to a problem I did not yet know how to solve any other way.

This reframing did not excuse anything. It did not make the past disappear. But it changed its meaning. And that shift made it possible to imagine a different future without having to repudiate who I had been. I was not replacing the old self with a new one. I was extending the story.

The changes that followed were slow. I built systems. I developed habits. I started moving my body in small ways, then larger ones. I began to think of myself not as someone trying to become disciplined, but as someone learning what discipline looked like for him. The shift was not a declaration; it was an accumulation. As the narrative updated, the changes compounded. What began as effortful became automatic. What had felt implausible began to feel true. Over several years, the overweight adolescent became a long-distance runner.

But narratives are not settled once and never revisited. A few years ago, within a short span, I lost both of my parents. Not long after, I found myself raising two small children. The stress was not abstract; it was physical, daily, cumulative. And slowly, almost imperceptibly, old patterns resurfaced. Food began functioning again as regulation. Clothes grew tighter. I noticed this the way one notices a tide coming in: not wave by wave, but all at once, when the water is already at your ankles.

What surprised me was not the return of the behavior, but my response to it. I did not experience the weight gain as failure, or as evidence that the transformation had been illusory. I understood it as a temporary, intelligible response to extraordinary circumstances. The coping mechanism I had once relied upon had returned because the conditions that originally called for it had, in a sense, returned as well.

This was not collapse. It was continuity of a different kind. That distinction mattered. It meant I did not have to fight my way back through self-condemnation or prove something to myself all over again. The story still held. The person who had learned to manage his relationship with food was the same person now navigating grief and exhaustion with imperfect tools.

What eventually became clear was that something else had quietly disappeared. Running, which had served as my primary release valve, had fallen by the wayside, crowded out by grief, exhaustion, and a sense that taking an hour for myself each morning felt like a small act of selfishness. In its absence, food resumed its prior role.

Recognizing this made the path forward legible. If I did not allow myself that daily, seemingly self-indulgent act of running, emotional regulation would take on a more corrosive form. The return to equilibrium was not a battle. It was a logical progression.

The injunction to “decide who you are” or to “fake it till you make it” assumes that identity is a matter of assertion. It implies a gap between who you are and who you claim to be. That gap is unsustainable. Under stress, under fatigue, under the ordinary pressures of life, people revert to the identities they actually hold, not the ones they have rhetorically adopted. The person declares himself transformed, but the memories remain, untouched by new interpretation. And because memory exerts a gravitational pull on identity, because the working self seeks coherence between what we remember and who we believe ourselves to be, the old story reasserts itself. The reinvention does not stick because it was never integrated.

The alternative is slower but more durable. The past is revisited, not rejected. The same episodes are arranged into a new narrative, one that allows for growth without requiring amnesia. The person who was once anxious becomes the person who learned to manage anxiety. The person who once failed becomes the person for whom failure proved instructive. The facts remain; the meaning evolves.

I do not want to suggest this process is easy, or that it can be accomplished through sheer intellectual effort. Reinterpretation often requires time, distance, sometimes therapeutic help. Some pasts are harder to integrate than others. Trauma, in particular, can resist narrative coherence in ways that make identity change extraordinarily difficult. But the principle holds: change that denies the past is fragile; change that incorporates it is not.

There is something humane in this understanding. It suggests that transformation does not require self-rejection. That who you were is not an obstacle to who you wish to become, but the material out of which that person must be built.

Every January, people resolve to become someone new. Most will not succeed. Not because they lack will, but because they are asking the wrong thing of themselves.

Identity is not interested in novelty. It is interested in coherence. The person you become must be someone the person you were could plausibly have become.

The story has to hold.

For a long time, I avoided SaaS businesses selling to SMEs like the plague.

Not because the entrepreneurs weren’t strong - often they were excellent.

But because the economics were almost always an uphill battle.

SME software budgets are structurally constrained.

Average contract values tend to cap out around $500–$1,500 per month. At the same time, these customers are often non-technical, which means you still need a real sales motion, sometimes even door-to-door, and meaningful customer success just to get people onboarded and using the product properly.

Put differently: you’re selling low-ticket software into a segment that still requires high-touch human labor. The CAC/LTV math rarely made sense.

A few years ago, embedded fintech started to shift this equation slightly. Vertical SaaS companies began embedding payments into their platforms and monetizing a percentage of GMV. This helped a bit since revenue became tied to customer activity but in most cases it only became meaningfully attractive at very large scale (think Toast).

What’s different now is AI.

The key shift isn’t that AI makes software cheaper or smarter.

It’s that it changes what you’re actually selling.

If you’re selling software, you’re competing against a tightly constrained software budget.

If you’re replacing labor, you’re competing against payroll.

That’s a very different pricing anchor.

When a product removes or materially reduces ongoing human work, whether that’s back office operations, admin, customer support, or workflow execution, the willingness to pay increases dramatically. You’re no longer asking an SME to justify another tool. You’re asking them to justify keeping headcount.

That doesn’t mean SME SaaS suddenly becomes easy or universally attractive. Distribution is still hard. Churn is still a nightmare. And many companies will overestimate how much labor they’re actually replacing.

But the equation itself has changed.

For the first time in a long while, SME-focused businesses can plausibly charge based on economic value delivered, not on what a small business thinks “software should cost.”

And that’s a subtle but important shift.

This is not a typical Shifting Moats post.

It’s not about markets, strategy, or business models.

It’s about something more personal - something that has quietly done more to change my trajectory than any framework or insight I’ve written about here.

Over the last few years, I’ve become convinced that identity is far more malleable than we tend to assume. That the story we tell ourselves about who we are isn’t just descriptive…it’s generative.

It shapes what we attempt, what feels possible, and what we instinctively rule out.

Exploring this idea, and applying it deliberately, has had a disproportionate impact on my own life: how I work, how I take risks, how I relate to effort and change. At some point, it felt wrong not to write about it.

So this post, and some that follow, are an experiment. I’m sharing something that worked for me, in case it’s useful to others who are curious about intentional self-reinvention. No broader claim than that.

Most people think their identity is shaped by circumstances: childhood, environment, early experiences. That once formed, it’s more or less fixed.

I’ve come to believe the opposite.

That identity, both the headline narrative and the traits beneath it, is far more fluid than we assume. That you can deliberately reinvent yourself, not once, but repeatedly.

I wanted to pressure-test that intuition. So I went down a rabbit hole of personality psychology research. I came across work by Dan McAdams, William Fleeson, and others that fundamentally changed how I think about the self.

Their research suggests that identity is not a static essence but rather a narrative or story we construct through what psychologists call autobiographical reasoning. We interpret our experiences, select which moments matter, and weave them into a coherent account of who we are.

Crucially, many of the things we think of as fixed “traits” aren’t fixed at all.

They’re better understood as mental state distributions. We’re not “disciplined” or “introverted” in some absolute sense. Instead, we’re a pattern of moments that we’ve averaged into a trait label.

Which means: your identity is largely the story you keep telling yourself.

The mechanism is straightforward and largely subconscious:

Your actions create episodes.

Episodes become memories.

Memories shape the story you tell yourself about who you are.

And that story determines which future actions feel “like you.”

It’s a compounding flywheel. And for most people, it’s running on autopilot.

Here’s the trap. We look at our past and project a “compatible” future. We stay loyal to who we’ve been instead of who we could become. Consistency becomes a constraint rather than a virtue.

Every small action that aligns with a target identity becomes evidence. Stack enough evidence and the brain is forced to update the story. The narrative shifts. And once the narrative shifts, decisions that once felt unnatural start to feel obvious. Cognitive dissonance, often experienced as imposter syndrome, fades, and the new identity begins to feel authoritative.

This isn’t just a personal development curiosity. It’s a pattern that shows up repeatedly at inflection points throughout history.

Many of history’s most consequential leaders, from Lincoln, Churchill, and Deng Xiaoping to Steve Jobs and Reed Hastings, weren’t defined by continuity. They stepped into identities their pasts did not predict and let disciplined behavior compound until the narrative became inevitable.

In that sense, identity is not a byproduct of success. It’s an input.

A strategic asset that can be deliberately constructed.

I’ll be unpacking the research and ideas behind this in a short series on deliberate identity construction.

I’ll be learning and writing in public. If that’s something you’d find useful, follow along.

(And yes, there’s a quiet nod to Friedrich Nietzsche in the title.)

For decades, software made humans more productive. AI replaces large part of the work humans do altogether.

In Systems of Action I wrote:

“Systems of Action execute entire workflows with humans stepping in only to approve exceptions. By replacing human labor, services businesses—historically seen as non-scalable—begin to exhibit product-like margins. Law, accounting, and property management firms may soon become scalable and investable businesses.”

Previous technology waves expanded software TAM to roughly $1 trillion by digitizing workflows and automating coordination. AI goes further.

By making human judgment a digital and scalable good, it expands the market from software spend to labor spend.

We are moving from roughly $300 billion in SaaS to the trillions spent on human labor.

As Jensen Huang put it at Sequoia’s AI Ascent: “This is the first time in technology we’re not creating a tool. We’re displacing labor budget.”

Lightspeed captures the same idea concretely: Salesforce helps manage sales workflows, but an AI sales agent can do the selling. Genesys runs contact centers, but an AI agent can resolve queries directly. Excel helps analysts build models, but an AI analyst can do the analysis.

The pattern holds wherever work is structured, repeatable, document-driven, and rule-based.

This shift is giving rise to AI-native services: companies that present as service providers but operate with software economics, selling outcomes rather than seats or billable hours.

Two paths lead here:

1. AI-native services built from scratch with humans in the loop

2. SaaS companies adding agents and forward-deployed engineers

This essay provides a framework for understanding this inflection point.

I. Why Now: The Enabling Technologies

AI-native services are enabled not by a single breakthrough, but by the convergence of capabilities that allow software to perform real work reliably.

First, general-purpose models have crossed a threshold in handling service workflows. Large language models now excel at unstructured document processing, data reconciliation, semantic search, multi-step reasoning, and tool use. This matters because most services work is interpretive rather than creative: reading messy inputs, matching them to rules and precedents, resolving inconsistencies, and producing a decision or action.

Second, voice AI has quietly become good enough to unlock entire categories of service work. Advances in speech synthesis and understanding mean agents can handle phone calls without breaking flow or sounding robotic. Tasks like collections, intake, scheduling, and customer support move from “human-only” to automatable.

Third, agents are no longer confined to single applications. Emerging browser and desktop automation allows AI systems to operate across heterogeneous software environments much like a human worker would. Agents can navigate legacy systems, fill forms, reconcile data across tools, and complete workflows end to end without bespoke integrations.

Together, these capabilities remove the historical bottlenecks that kept services labor-bound. Software can now read what humans read, speak where humans speak, and operate where humans operate. The remaining challenge is no longer model capability, but workflow redesign, which is why AI-native services are now viable rather than theoretical.

In practice, this approach works best in workflows that are document-heavy, rules-driven, and repeatable, with humans stepping in primarily for judgment and edge cases.

II. The Low-Hanging Fruit: BPO

The BPO market, projected to reach roughly $525 billion by 2030, illustrates the opportunity. As a16z notes in Unbundling the BPO, incumbents like Cognizant, Infosys, and Wipro each generate $10–20 billion handling work enterprises prefer not to do internally: customer support, claims processing, reconciliation, and IT operations.

The experience is often broken. Turnaround times are long, errors accumulate, and staff lack the context or authority to resolve edge cases. Enterprises tolerate this because building equivalent internal capacity is worse.

BPO is the low-hanging fruit for AI-native services for three reasons. First, the budget already exists, creating a PMF advantage and reducing adoption friction. Second, much BPO work is repetitive and rules-based, making it especially vulnerable to automation that improves accuracy and speed. Third, outcome-based pricing aligns incentives far better than billable hours or fixed retainers.

BPO incumbents are structurally unprepared for this shift. AI-native operations compress margins, cannibalize labor revenue, and require product and ML capabilities that outsourcing firms largely lack today.

III. The North Star: Revenue per Employee

If AI-native services truly behave more like product businesses, that should show up in the numbers. The clearest early signal is revenue per employee.

This metric reveals whether automation is compounding or whether the business is still scaling like a traditional service firm. In practice, it should be evaluated against service benchmarks in the same vertical, not against SaaS peers. Past waves of “tech-enabled services” often promised operating leverage without delivering it; applying an exacting standard here is essential to avoid mistaking hype for progress.

IV. Getting Automation Right

Full automation is not required at the outset. In practice, it is often wiser to “trade margin for moat” by automating where leverage is highest while keeping humans in the loop elsewhere.

This is the central operating challenge for AI-native services. More than model quality or interface design, how, when and where automation is applied, determines whether a company compounds into a defensible business or stalls as a labor-heavy service.

Done well, this approach allows companies to understand true complexity, refine workflows with real data, build proprietary training sets from human-handled cases, and identify which steps can be automated versus augmented at each stage of maturity.

Subscribe now

V. Beyond BPO: Property Management as a Case Study

Consider AI-native property management. The workflow spans tenant communication, maintenance coordination, lease administration, rent collection, and owner reporting. Some components are ready for automation immediately: tenant inquiries, maintenance triage, rent reminders, and reporting.

Others benefit from humans initially. Lease negotiations, complex maintenance decisions, disputes, and vendor management require judgment. Keeping humans involved while structuring every interaction generates training data for future automation.

Over time, this compounds. Eighteen months of human-led lease negotiations can yield tens of thousands of tagged transcripts, forming the basis for a negotiation agent competitors cannot easily replicate.

The margin trajectory reflects this sequencing:

• Year 1: ~30% gross margin

• Year 2: ~45% as core workflows automate

• Year 3: 60%+ as agents handle most cases

The winner is not the fastest automator, but the company that sequences automation to build defensible data assets.

VI. Will SaaS Companies Become Service Providers? The Rise of the Forward Deployed Engineer

AI also pulls SaaS companies toward services.

As agents move from assisting users to performing work, success depends on how well systems reflect real operational workflows. This is where forward-deployed engineers become central.

Forward-deployed engineers embed with customers to understand how work is done on the ground, then redesign workflows around AI capabilities. They map business logic, define where automation is safe, integrate fragmented systems, and turn edge cases into structured rules the product can learn from. In doing so, they translate operational reality into executable systems.

This model shares core characteristics with services businesses: deep operational involvement, humans in the loop, accountability for outcomes, and early economics shaped by human effort. Forward-deployed engineering becomes a bridge between SaaS and services.

Decagon illustrates this pattern. Its agent product managers work closely with customers to deploy AI support agents in production, integrating systems, adapting workflows, and handling early failures so agents can resolve requests reliably. An older precedent is Palantir, whose forward-deployed teams have long embedded inside customer organizations, co-running mission-critical workflows and redesigning processes as reality changes.

What we once called “implementation” has become an ongoing operational function. As software moves from enabling work to doing the work, that function becomes load-bearing.

VII. A Blurry Line

As AI capabilities advance, the line between AI-native services and SaaS with agents will blur.

The distinction will hinge on the answer to one question: who owns delivery?

AI-native services sell completed work and bear responsibility for execution, quality, and failure.

SaaS with agentic capabilities leaves accountability with the customer and prices access rather than outcomes.

As AI systems improve, companies will move along this spectrum.

Some SaaS companies will be pulled towards services and operational responsibility as customers demand outcomes rather than tools.

Similarly, some AI-native services will push towards software models as automation absorbs more of their workflows.

Conclusion

By automating jobs rather than tasks, AI-native services capture labor budgets and achieve ACVs SaaS alone cannot.

Over time, the market will sort companies not by how they describe themselves, but by who performs the work, how incentives are aligned, and how value accumulates as it scales.

Attribution

This essay synthesizes ideas from several original contributions on AI and services, including:

• Lightspeed, AI Will Eat Services

• Andreessen Horowitz, Unbundling the BPO; Trading Margin for Moat

• Point Nine Capital, AI-First Service Businesses

• Bessemer Venture Partners, AI Systems of Action

• Sequoia Capital, AI Ascent 2025

• Emergence Capital, forward-deployed engineers

Any interpretations or errors are my own.

Couple that with image and video generation capabilities that are already approaching studio-level quality, and you get the conditions for a shift from mass-produced entertainment to individually constructed narratives: a completely new kind of media primitive that adapts to the user rather than asking the user to adapt to it and is completely unconstrained by the economics of traditional production.

We are beginning to see this in early, almost trivial use cases: kids and parents bringing storybook characters to life, inserting the child as the protagonist who overcomes a specific challenge.

The technology is impressive, but it doesn’t scale culturally. Most people have little interest in watching AI-generated videos made by other users if the content is too specific, too personal. A video of your nephew defeating a dragon in your backyard has meaning to your family, but to almost no one else.

Now change one variable.

Imagine those AI-generated videos featured characters that were familiar to everyone.

With the appropriate licensing in place, a global army of creators, amateurs and professionals alike, could generate feature-length films using recognizable characters. Studios could monetize on their IP based on number of views with incentives aligned across creators and, where relevant, human likeness holders.

In that world, the constraint isn’t the technology. It’s the IP.

The real opportunity lies in indexing this AI-generated content and making it discoverable. A plausible outcome is that dominant AI platforms evolve into content portals where users can search for studio-grade material across thousands of parameters, resulting in experiences deeply tailored to individual preferences.

And if you don’t like part of the plot, you don’t abandon the film. You rewrite it.

This outcome is far from inevitable. The aggregation role for user-generated AI-content could still belong to Netflix, YouTube, TikTok or some new challenger. AI platforms only win if they become the primary interface for discovery and consumption, not just creation. But history suggests that whoever controls the creation layer has a least a path to controlling the consumption layer.

To understand where this leads, it helps to look backward.

In the streaming wars, aggregation platforms consistently gravitated toward vertical integration. Netflix didn’t start by making shows but rather it licensed them. Over time, it realized that owning content was the path to durable margins and defensibility. Disney, in turn, pulled its catalog from Netflix and launched Disney+, reclaiming control over its distribution.

In the next era, vertical integration will mean something different. The strategic asset will not be the exclusive rights to entire films or series, but rather the exclusive rights to repurpose IP at a more granular level: not the movie, but the character. Think Captain America rather than any given Avengers film.

The atomic unit of entertainment IP is shifting from the story to the character.

Two players enter this landscape with asymmetric advantages.

On the content side, Disney is uniquely positioned. Its vast IP library contains characters that are largely unconstrained by human likeness rights. Mickey Mouse, Elsa, and Simba do not require negotiating with actors’ estates or navigating consent and compensation complexities tied to real people. That is a structural advantage in a world where AI-generated content scales.

On the distribution side, the question is simpler: who has the largest footprint in consumer AI adoption? Today, the answer is undoubtedly OpenAI. ChatGPT has become the default interface for hundreds of millions of users interacting with AI.

Seen through this lens, the recent Disney–OpenAI partnership makes strategic sense. According to the announcement, OpenAI users gain access to over 200 Disney characters for one year, spanning Disney, Pixar, Marvel, and Star Wars, while Disney retains ownership of the underlying IP.

OpenAI receives differentiated, culturally resonant creative building blocks that deepen engagement and strengthen ChatGPT’s position as a creative interface. Disney, meanwhile, unlocks a new monetization channel for dormant IP with minimal incremental cost. More importantly, it gains early insight into how its characters perform in AI-native environments without ceding distribution or committing to a full platform strategy.

This is clearly v1, i.e. just a controlled experiment. But the implications are easy to extrapolate.

If a single AI platform were able to offer creators exclusive access to characters like Captain America and if audiences could only consume AI-generated content featuring those characters through that same interface, it would create a powerful new form of competitive advantage. This is precisely why such deals will be short-term until the dust settles. Exclusive character licensing is the next evolution of exclusive content licensing, and both sides know it.

The obvious counterargument is brand dilution. Disney has historically resisted remix culture. Allowing millions of users to generate their own Captain America stories risks narrative incoherence and loss of control.

But the alternative is worse. AI-generated content featuring beloved characters will exist regardless of studio participation. The only real choice is whether IP owners participate in the economics or watch from the sidelines. Controlled licensing, with parameterized creativity and guardrails, is a better outcome than uncontrolled proliferation.

The question is whether the same sequence that defined the streaming wars will repeat itself:

Content studios license to platforms (collaboration phase).

Platforms own demand and commoditize studios.

Studios consolidate and reclaim distribution.

Platforms move upstream into production.

The sum total of that war was that Disney and Netflix ended up controlling the lion’s share of the streaming market.

OpenAI may now be playing the role Netflix played a decade ago when it was the new streaming platform on the block, but with an important distinction.

Netflix aggregated demand within a category. OpenAI aggregates user intent across categories. The interface that becomes default for work, learning, and creation has a natural path to becoming default for entertainment, if the right strategic and design choices are made.

The Disney-OpenAI role may be a brief era of collaboration before competition hardens, reminiscent of the collaborative phase of the streaming wars. That likely explains the one-year term of the Disney–OpenAI deal. Disney saw this play out before…

Which brings us to Netflix.

If there is a credible counterweight to a Disney–OpenAI axis, it likely runs through Google. Gemini, paired with YouTube’s unmatched global distribution and Netflix’s scale in premium long-form storytelling, would combine frontier AI capability with the world’s largest video discovery engine and one of the deepest data-driven content operations ever built.

Regardless of how these partnerships ultimately shake out, Netflix must confront a future in which its distribution advantage is no longer exclusive. One where AI platforms, embedded upstream in how users search, create, and decide what to watch, begin competing for consumer mindshare long before the play button is ever pressed.

The streaming wars were fought over content libraries. The next war will be fought over characters and over the platforms that control how those characters are created, discovered, and experienced.

NB: This analysis is informed by Ben Thompson’s recent Stratechery piece Disney and OpenAI, Totems in an AI World, Google Versus the World, which frames the Disney–OpenAI partnership as both a testament to Disney’s enduring assets and a vector for competition with Google in the AI landscape.

My argument takes that framing a step further by applying aggregation theory to media economics: as AI advances, the strategic unit of value shifts from completed content to the characters themselves, and the platforms that control access to those characters, both for creation and discovery, will be the ones that capture long-term demand and monetization. This builds on Thompson’s platform perspective by extending it into the mechanics of AI-driven content personalization and IP repurposing, rather than focusing solely on corporate positioning within the AI ecosystem.

Every generation gets a technology that “changes everything.”

And every generation learns the same lesson: the technology can be real, and investors can still lose fortunes.

Railroads transformed America, but most railroad investors were wiped out.

The internet reshaped commerce, yet most dot-com era companies went under.

When people ask whether AI is a bubble, there is often an implicit assumption that an affirmative response would turn every AI breakthrough we have seen into nothing more than a fad.

But that is the wrong way to think about it.

Howard Marks, in a recent note drawing on the work of Carlota Perez and others, provided a useful distinction between two kinds of bubbles that behave very differently.

Mean-reversion bubbles

These bubbles, as the name implies, are financial excess without lasting productivity gains. After the crash, activity reverts to the baseline of where it was pre-bubble, and little remains.

An example of this is the subprime mortgage crisis. When housing prices declined, mortgage-backed securities deteriorated and credit default swaps written against them triggered cascading collateral calls. There were no new industries or transformative infrastructure. All that remained were widespread home foreclosures and broken balance sheets.

A simple test helps here: after the crash, is the world structurally different?

If not, you are likely looking at a mean-reversion bubble.

Inflection bubbles

These are bubbles driven by optimism about the transformative potential of a new technology. The excitement is well-founded and the direction of change tends to be broadly correct. But the assumption that capital deployed early and broadly will earn attractive returns is often wrong.

The railroad boom of the 19th century is a canonical example of this.

In the 1850s and again in the 1870s and 1880s, capital poured into rail construction at a rapid clip.

Competing companies often built parallel lines between the same cities, even though traffic did not justify it, because they believed demand would eventually catch up and wanted to get there first.

By the early 1870s, the United States had more rail mileage than it could profitably support at the time.

Freight rates collapsed and many railroad companies were unable to service their debt, triggering a wave of bankruptcies that dragged down banks, manufacturers, and land developers across the country.

But the physical infrastructure did not disappear.

Bankrupt railroads were reorganized, debts were written down, and ownership consolidated. Trains continued to move goods and people at low cost and ultimately enabled decades of economic expansion.

So while capital was largely destroyed, the benefits endured.

In these periods, when the future is unclear but the prize is large, capital funds many competing approaches at once. Carlota Perez calls this the Installation Phase, the frenzied build-out driven by speculative mania before anyone knows what works.

Then the bubble collapses. Prices are driven down faster than costs and before demand fully materializes.

But the collapse marks a transition to the Deployment Phase, when what survives becomes productive infrastructure and the foundation for real progress.

That pattern repeated itself with virtually every technological breakthrough in modern history: canals, automobiles, aviation, telecommunications, semiconductors just to name a few.

What This Means for AI

AI can be transformational and still be a bad investment.

To be clear, this is not an argument that AI is a mirage or that recent advances are illusory. Quite the opposite.

Like railroads or the internet before it, AI appears to be a genuine general-purpose technology with the potential to reshape large parts of the economy.

The risk lies elsewhere. If AI is an inflection bubble, history suggests that much of the capital being deployed today may fail to earn attractive returns.

That does not mean investing in AI is a mistake. It means understanding where we are in the cycle matters.

So the right question is not whether AI is a bubble. It is whether investors are being compensated for the risks they are taking at this stage of the cycle and whether they can survive the transition from speculation to consolidation.

If history is any guide, AI may well change everything.

It just will not make everyone rich along the way…

For years, the AI race was defined by a simple question: whose model is smarter?

Each new release brought measurable leaps in reasoning, coding, and general capability. GPT-3 to GPT-4 felt like a generational shift.

The assumption was that these gains would continue and that whoever stayed on the frontier would capture most of the value.

People went as far as to claim that models would “eat the application layer” and that anything that sat on top of them was just a “GPT-wrapper”.

That assumption is now under strain.

The Curve Is Flattening

Frontier model reasoning gains are flattening relative to cost and complexity.

This isn’t because progress has stopped, but because marginal gains now require disproportionate increases in data, compute, and capital.

When you look at model performance benchmarks on authoritative intelligence benchmarks, e.g. “The Last Test of Humanity”, you can clearly see a flattening of the curve and, its corollary, convergence of model performance across frontier companies.

As models converge, firms are forced to subsidize usage to win distribution, compressing margins even as absolute usage explodes.

Massive capital is being spent to push down the cost of a unit of reasoning, often at significant operating losses.

But economically, what this race produces is not durable competitive advantage but rather cheaper tokens.

The strategic consequence is that intelligence itself is becoming both abundant and cheap, i.e. a commodity.

What’s the clearest supporting evidence of this trend? It’s that leading companies are no longer attempting to differentiate via reasoning superiority but rather distribution, UX, and verticalization.

ChatGPT now unequivocally owns consumer while Anthropic is increasingly going after the Enterprise and “prosumer” markets.

To add fuel to the fire, Chinese companies are pioneering the open-source LLM movement.

DeepSeek shows that near-frontier reasoning can be delivered at a fraction of the token cost by optimizing for “good-enough” performance, compressing margins for everyone else and accelerating commoditization.

The implication is significant: if intelligence is commoditizing, the question is where does competitive advantage shift to next?

Embodied AI

The current generation of AI lives in the cloud.

You send a prompt, a data center processes it, and you get a response.

The constraints that matter are training compute, inference cost, and latency. These are software problems, and they favor companies with algorithmic expertise and access to GPUs.

But the next generation of AI will live in the physical world.

Self-driving cars that don’t just drive but converse, negotiate traffic, and coordinate behavior with other vehicles.

Home assistant robots that complete chores, coordinate deliveries, energy use, and security while adapting continuously to the people inside.

Personal drones that film, follow, and assist in outdoor activities.

Wearable exoskeletons that augment strength and reduce fatigue.

That’s to speak nothing of the many industrial and (sadly) military applications.

In each case, intelligence is no longer something you open on a screen, it’s something that moves alongside you in the physical world.

Those form factors are the body that AI will soon inhabit. Hence the term “Embodied AI”.

And what will Embodied AI depend on to act in the world? Energy, manufacturing, and supply chains.

The question that will define the next decade isn’t who builds the smartest reasoning model. It’s who can manufacture AI at scale in the physical world.

Seen through that lens, you can view the Chinese push to commoditize intelligence with open-source LLM’s that perform at near-frontier quality as a classic “commoditize your complements” strategy.

Microsoft executed the same playbook three decades ago.

By commoditizing PC hardware, it captured value at the operating system layer and dismantled IBM’s hardware-centric advantage.

Ironically, today, advantage is shifting back towards hardware.

Here’s why that shift matters.

A robot moving through a warehouse can’t rely on a data center thousands of miles away.

It needs to process information locally, in real-time, while managing its own power supply.

The constraints that matter are no longer just compute and algorithms.

They’re energy density, motor precision and manufacturing cost.

Packy McCormick and Sam D’Amico frame this as the “Electric Stack”: the four core technologies that convert electricity into mechanical action, namely lithium-ion batteries, electric motors, power electronics, and embedded compute.

Their argument hits home: If we want AI to play a role in our physical world, it needs a body.

And that body will be built on hardware that must be manufactured at scale.

Batteries Are the Rate-Limiting Factor

If embodied AI depends on the Electric Stack, which component of that stack matters most?

Compute, sensors, motors, and energy storage all matter, but only batteries simultaneously constrain mobility, endurance, unit economics, and manufacturability across every embodied AI category.

A robot’s operational range, a drone’s flight time, an EV’s miles per charge are all determined by battery energy density.

You can have the most sophisticated AI in the world, but if your robot runs out of power after two hours, its utility is capped.

Battery progress is governed by learning curves.

This is worth understanding because it explains why early advantages compound over time: scale drives cost down, lower costs unlock new products, and new products create more demand, pushing scale further.

When Sony commercialized lithium-ion batteries in 1991, a kilowatt-hour cost roughly $7,500 in today’s dollars.

Now we’re at $100-150…that’s a 99% decline over three decades, driven by manufacturing scale and process improvements.

Sony’s Handycam was the “alpha product” that pushed batteries down the curve. Then came laptops. Then smartphones.

And now, crucially, Electric Vehicles are the new frontier of battery technology.

Whoever controls the battery value chain, captures compounding advantages across every product built on top of it.

Right now, that’s China.

China’s Structural Advantage

Starting with the 2001-2005 Five-Year Plans, China made electric vehicles and battery technology a strategic priority.

It invested in the full stack.

Everything from mining rare earth minerals to refining materials, building cell manufacturing capacity, and integrating vertically into finished products.

BYD, China’s leading EV company, actually started out life as a battery manufacturer.

BYD controls the full stack, from raw materials to finished cars. Once you’re ahead on the learning curve, you’re producing at lower cost than competitors. Lower cost means you can price more aggressively, win more volume, and accelerate further down the curve. It’s a self-reinforcing advantage.

Meanwhile, Western battery companies stumbled.

A123 Systems is a good counterfactual to BYD.

The company focused on materials innovation, while leaving battery management to customers.

By not controlling the full system, they couldn’t optimize across it and eventually went bankrupt.

And who was there to snap up A123 Systems for the bargain-basement price of $257 million when they went under in 2012? None other than Chinese conglomerate Wanxiang, gaining not just assets but critical institutional knowledge in the process.

Today, China controls over 70% of global lithium-ion battery production.

This dominance didn’t happen by accident. It was built deliberately over decades.

1

Semiconductors: A Parallel Battleground

AI hardware depends on advanced semiconductors.

If batteries are the requirement for embodied AI to be mobile and energy-autonomous, then semiconductor chips are the brains or substrate on which that intelligence actually runs. In more strategic terms, they are the computational chokepoint that governs performance, latency, and cost at scale.

The problem is that global semi-conductor supply is concentrated in Taiwan (TSMC) and South Korea (Samsung). Both are within easy reach of Chinese missiles.

If a conflict disrupted access to these fabs, the U.S. would lose not just AI chips but the basic semiconductors that power essentially everything.

As Ben Thompson points out in his Stratechery newsletter “US Intel”, this explains why the U.S. government moved to bail out support Intel with a $10bn investment this year.

Without getting too into the weeds, Intel needed a credible guarantee of longevity before it could make the significant Capex investments necessary to remain competitive in the semiconductor space.

The US also doubled down with the CHIPS Act, export controls, fab subsidies and a whole slew of other measures.

But here’s the dichotomy: semiconductors have received this level of attention. Batteries have not.

If the logic for chips is that abandoning the leading edge means never regaining it, the same logic applies to batteries. Both are decades-long endeavors. Both require massive scale to be economically viable. Both involve learning curves that compound early advantages. Both are vulnerable to geopolitical disruption.

Yet the policy response is asymmetric.

The U.S. is actively trying to re-shore chip capacity while battery dominance slips away with far less urgency.

Thompson’s argument for chips is a template for batteries.

The question is whether policymakers will recognize this before the window closes.

Old-World Moats Reasserting Themselves

The new AI moats may turn out to be the old world industrial moats: manufacturing, supply chains and energy efficiency, rising once more to reassert themselves in a software-first world.

For the past few years, the dominant narrative was “software is eating the world”. It now seems like we’re in the midst of a giant reversal.

One might almost say, hardware is eating software.

NB: A note on sources: this essay borrows heavily from, and builds upon, the thinking of Packy McCormick’s Not Boring and Ben Thompson’s Stratechery. Both have been early and rigorous in framing AI not just as a software phenomenon, but as a shift with deep implications for hardware, industrial capacity, and geopolitical power. Any synthesis here stands on that foundation. Any errors or extrapolations are my own.

For decades, the conventional wisdom was that you couldn't directly challenge legacy ERP systems like SAP, Oracle or Workday.

Instead, the new generation of enterprise SaaS companies built on the margins: workflow tools that sat on top of the system of record, analytics layers that made the data more accessible, point solutions that handled specific tasks the ERP didn't do well.

They made the legacy system more useful, but they didn't replace it. The ERP remained the source of truth and the place where mission critical data and core business logic lived.

Even when a new software vendor promised enhanced utility with a reimagined workflow, the switching costs of ripping out legacy ERP were simply too high to justify any incremental improvements. Migrating decades of business data, rebuilding integrations and retraining employees simply never made sense.

But there’s a paradigm shift unfolding that’s wedging open the window of opportunity.

Bessemer Venture Partners calls it the shift from “systems of record” to “systems of action.” Sequoia frames it as “service-as-a-software,” where the addressable market isn’t software spend but labour spend, measured in the trillions. The distinction matters.

Systems of record store data and empower humans to become more efficient. Systems of action, on the other hand, execute entire workflows with humans only stepping in to approve exceptions and deviations.

A few examples of what this looks like in practice:

• Harvey (Sequoia) drafts legal documents and conducts due diligence that would take a junior associate days to complete

• Prepared (a16z): triage 911 calls in real time, routing genuine emergencies to dispatchers instantly while handling noise complaints autonomously

• Decagon (a16z): handles customer support, resolving 90%+ of inquiries without human involvement

• Factory (Sequoia, NEA): deploys autonomous “Droids” that review code, write tests, and debug across the entire software development lifecycle, saving customers thousands of hours of engineering time

• Tennr (a16z): reads incoming healthcare referrals, extracts patient data, and routes cases to the right specialist before a human touches the file

• Resolve AI (Greylock): monitors production incidents across a company’s entire software stack, escalating to engineers only when necessary

These aren’t copilots making humans faster. They’re systems doing the work.

The strategic implications are difficult to overstate: when the ROI of the new system finally justifies the pain of migration, the moat that protected incumbents for decades begins to erode.

This won’t happen overnight. But sectors that were previously unassailable suddenly can’t afford to ignore the gains in productivity at zero marginal cost. Couple that with a C-level mandate to do “something with AI” and you get the perfect storm.

Banking, healthcare, government, higher education...sleepy industries where you had legacy ERP players sell $5M - $10M software licenses in the late 90’s or early 2000’s and sit tight for a couple decades are now fertile selling ground.

The strategic question for incumbents is whether they can recast themselves as systems of action and not merely custodians of data. This will involve all kinds of trade-offs and value chain constraints a-la Clay Christensen’s Innovator’s Dilemma. One simple example is outcomes-based pricing: when the software can quantify the exact savings in labour cost or the incremental revenue it is generating for you at zero marginal cost, the next logical step is to only charge you for outcomes that are demonstrably better than your historical baseline. That is a paradigm shift compared to the traditional Enterprise SaaS model of seat-based pricing or generic annual licensing fees. And that’s just one example…

But before I wrap up this note, I should mention one other big change that these systems of action promise to herald…

By replacing human labor, entire business categories will be redefined. Services businesses, which we historically thought of as non-scalable (selling money for time), will all of a sudden exhibit product-like margins. Law, accounting and property management firms, just to name a few prominent examples, will become scalable and, soon perhaps, investable businesses.

That will likely lead to both VC and PE investors having a field day with consolidation platforms that can purchase individual firms at traditional multiples of 0.5-2x revenue / 3-5x EBITDA to something more like 5-6x revenue / 6-9x EBITDA. That’s not quite as high as software multiples but it’s quite the arbitrage opportunity.

Right now that’s a blue ocean but I sense that soon enough there will be blood in the water…

Imagine an online proxy for Benjamin Franklin's Junto Club…or an engaged community of experimental foodies, birdwatchers, emerging market VC’s or you name it.

Now imagine these online communities looked less like Reddit forums and more like Twitter meets Houseparty for x area of interest and y level of expertise.

That vector of engagement, connecting individuals with similar interest and knowledge graphs and facilitating meaningful connections online is still a nascent (albeit growing) category.

FB/Reddit groups and forums are horizontal and generic attempts to solve for this but they don't even come close to the experience dedicated companies leveraging curation, supply constraint driven network effects and vertically-tuned matching algorithms could produce.

These new kids on the block will represent a gradual unbundling of mainstream social networks and, unlike the old guard, could become catalysts for real human connection and self-actualization.

Coho Capital wrote a great memo on Spotify, demonstrating why the company’s stock is significantly undervalued. It’s a great example of thesis-driven thinking.

Here’s a quick summary of the 13-page memo:

1. Labels are Losing Leverage

The bear case on Spotify is that they will forever be captive to the music labels with 80% of current streaming content controlled by four labels. 

To know where we are going, it is helpful to know where we’ve been. Prior to streaming, the music industry endured fifteen years of decline with recorded music revenues dropping from $14.6 billion in 1999 to $6.7 billion in 2015. It was only once music streaming gained traction that the industry was able to return to growth.

Streaming now accounts for 50% of global music revenue with Spotify at almost 70% market share. 

Before streaming, it made sense for music labels to collect the lion’s share of profits. After all, they funded the retail network, oversaw the capital-intensive business of producing and distributing physical media and discovered and promoted stars. Many of those tasks have been rendered obsolete by music streaming.

The labels need Spotify for maximum distribution and while, in theory, they could threaten to pull their catalogues, that wouldn’t go down well with artists, whose earnings would see a sharp drop. 

Despite this shift in power dynamics, industry profit pool participation hasn’t really changed. There is little doubt that over time Spotify will use its leverage to capture better economics commensurate with its importance in the music industry value chain.

2. Discovery Flywheel

Discovery is Spotify’s superpower with over four billion playlists at the backbone of its discovery focused UX. By providing best-in-class discovery and personalization tools, Spotify creates a virtuous flywheel of demand: discovery drives engagement which feeds data algorithms further improving personalization. The success of playlists in keeping listeners engaged is reflected in listener data with a third of listening time spent on Spotify generated playlists and a third of listening time spent on user generated playlists.

3. Google of Audio

Daniel Ek once said “the market we’re going after is audio”. As such, Spotify is moving to capture several emerging audio categories outside of music, including podcasts, books, courses, sports, news, etc. 

By aggregating content and serving as a central depository of all things audio, Spotify can remove frictional search costs and become a one-stop-shop for audio content, a sort of Google for audio search.

4. Podcasts - the New Frontier

Podcasting is now a $1.3 billion market growing at 22% CAGR. Podcast listeners are a deeply engaged bunch, consuming six hours per week.

Spotify spent $600 million over the past year to acquire four podcasting companies and is clearly angling for vertical integration to both publish and distribute content. Given its large base of 300m users, the company has a real chance of becoming the preferred platform for podcast discovery.

Like Netflix, its scale gives it a significant competitive advantage both in aggregating non-exclusive third party content and in producing originals at lower cost than any competitor.  

The shift towards original content should help Spotify replace the variable costs paid to music labels with fixed costs that create operating leverage and yield better unit economics. 

Relative to consumption hours, podcasts are woefully under-monetized, promising significant future upside.

5. Keeping the Horsemen at Bay

Spotify has to contend with formidable opponents in the form of Apple, Amazon and Google. Each benefits from a significant advantage in customer acquisition and is happy to utilize music as a loss-leader to sell more phones (Apple), serve more ads (Google), or deepen overall commitment to its ecosystem (Amazon).

Yet despite their advantages, each of Spotify’s competitors’ ability to scale globally is constrained. For Apple, gains have been constrained by the company’s inability to gain significant traction outside of its iOS operating system, which currently hovers around 25% on a global basis.  Amazon is in 45 markets relative to the 92 markets in which Spotify has planted its flag. Last, while YouTube is everywhere, its model will always be subpar due to an artist payout ratio per stream only 1/6th that of Spotify

6. 2x Future Gross Profit

Coho sees Spotify growing its average revenue per user from $5 to $10. This may sound aggressive, but unlimited on-demand streaming of a catalog of 50 million songs across devices and without commercials for $9.99 a month is one of the best deals around. While the price remains the same, increased engagement means lower costs for each unit of content consumed (songs for a service such as Spotify and shows or movies for a service like Netflix), a sort of personal operating leverage for consumers. This means marginal costs for extra music consumption is zero. In economics, this is known as a “consumer surplus,” which reflects the difference between the price consumers are paying for a service and the price they are willing to pay. Once Spotify migrates beyond the landgrab stage it will not hesitate to press on the pricing lever with consumers. Similarly, margin extraction from labels in line with its central position in the music value chain is only a matter of time. 

The market for paid music streaming subscribers will grow to 1.2 billion by 2030. If Spotify nets out at 50% of the market revenue would jump to $72 billion a year. At Spotify’s gross margin guidance of 35%, the company would generate $25.2 billion in gross profits. Put another way, Spotify currently trades for roughly two times where gross profits are expected to be in a decade.

Read the full memo here: https://bit.ly/3js4Lzp

Subscribe now