The Scaling Laws
On artificial intelligence, a broken heart, and the measure that matters
A few months ago, I put ten years of silence into words.
I called the essay The Consequences of a Broken Heart. It was about cardiac arrhythmia, but not only the electrical kind. It was about what happens when the body lives under a threat nobody else can see. The constant alarms. The loss of control. The grief. The masking. The long negotiation between a heart that will not settle and a mind trying to convince the rest of the world that everything is fine.
Writing it settled something in me.
The response made me keep writing.
I wrote and published additional articles, each approaching the burden from a different angle. Across them, the replies accumulated.
One person sent an essay to his wife because he had never been able to explain what the condition felt like. Another saved one in a notes app and returned to it during a PVC and PAC storm. Someone described the impossibility of having a pleasant conversation in the kitchen while the house is on fire. Someone else said that even after an ablation brought relief, a part of the carefree person who entered the hospital had never made it back out.
There were people living with tens of thousands of extra beats a day and people living with only a few. There were people newly diagnosed and people who had carried the burden for more than thirty years. There were shocks, ablations, medications, emergency rooms, long drives to reach specialist care, and repeated assurances that the symptoms were benign or psychological.
The burdens differed.
The alarm was recognizable.
Again and again, people described the same invisible negotiation: the fear of the next episode, the fear that nobody believed them, the fear of exhausting the people they loved, the fear that speaking honestly would make them sound weak, dramatic, or ungrateful.
One reader was only a week into a diagnosis and already considering whether to speak with a mental-health professional. Another had found a therapist who specialized in cardiac patients after thirty-one years of arrhythmia. One said the writing gave her strength while she waited for an ablation. Another said it helped him show his wife what he could not make visible on his own.
That was the signal.
The literature can quantify parts of it. Published estimates vary by population and instrument, and they should not be collapsed into one universal number. Even so, the direction is difficult to ignore. Anxiety and depression have been reported in roughly 28 to 38 percent of people with atrial fibrillation. A meta-analysis covering 39,954 patients with implantable cardioverter-defibrillators estimated prevalence of 15 percent for depression, 23 percent for anxiety, and 12 percent for post-traumatic stress disorder. In one tertiary atrial-fibrillation population, 35 percent reported severe psychological distress and 20 percent reported suicidal ideation.
I do not see those statistics.
I feel them.
I feel the low-level torment one person described. I feel the dread washing over another when a missed beat announces what may be coming next. I feel the person who looks physically capable but has nothing left to offer the family member who thinks he simply did not show up. I feel the person afraid to visit family in the countryside because medical care is too far away. I feel the exhaustion of trying to explain an internal emergency to a world that sees an ordinary face.
Those feelings are not a substitute for evidence.
They are the reason I refuse to look away from it.
They are the vigilance.
The machine’s scaling laws
In artificial intelligence, scaling laws describe a striking empirical pattern: as model size, training data, and compute increase, performance often improves in ways that can be estimated. The relationship is not magic, and it is not without limits, but it has been reliable enough to shape the direction of an entire industry.
The scaling did not end with pretraining. More capability can also be produced at inference: giving a model more time, more attempts, better search, tools, verification, or specialized context when it works on a particular problem.
When that acceleration became impossible for me to ignore, I made the decision to lean hard into understanding it.
Vigilance requires action. When you see uncertainty with the capacity to produce consequences, ignorance is not a strategy. I did not need to know exactly where the technology would lead to know that falling inference costs, rapid specialization, expanding compute, and better ways of making models reason through difficult problems were going to change the texture of work.
I needed to understand the instrument.
The practical consequence is larger than a better chatbot.
Natural language is becoming a control surface for machine capability.
A person who can define a problem clearly can now move from prose to research, from research to a specification, from a specification to code, from code to tests, from tests to an operating system, and from an operating system back into language another person can understand. The borders between natural and machine language have not disappeared. But the cost of crossing them has fallen.
For most of modern work, productive reach was constrained by access to specialized execution. An idea might require a statistician, a software engineer, a designer, a researcher, an editor, infrastructure, and capital before it could become anything another human being could use.
Those skills still matter. Deep expertise may matter more when plausible answers are cheap, because someone must be able to distinguish a credible output from a confident imitation of one. AI does not make knowledge irrelevant. It makes unverified fluency dangerous.
What changes is the upper bound.
The market value of one narrowly scarce skill no longer has to define the total reach of one person’s work. A capable operator can orchestrate across disciplines, use different models for different roles, test outputs against evidence, bring in human experts where the boundary demands it, and build at a scale that would once have required an institution.
The center of gravity begins to move.
Away from:
What single task are you qualified to perform?
Toward:
What meaningful outcome can you responsibly produce?
That is not the erasure of skill.
It is the expansion of agency.
And it forces a harder question than whether the next model will be larger, faster, or cheaper.
What happens when human purpose begins to scale with it?
The fear
I see the fear people carry about artificial intelligence.
I understand it.
There are real risks. Work will change. Authority can become concentrated in a small number of companies and institutions. Models can fabricate facts, reproduce bias, scale persuasion, lower the cost of surveillance, and give weak judgment an appearance of technical legitimacy. A machine can make an error faster than a human. An organization can use that speed to spread the consequence across millions of people.
Nobody should be shamed for seeing danger in that.
Fear is often the nervous system identifying a variable before the conscious mind has words for it.
I learned that from my heart.
The alarms were not theoretical. A sensation in my chest could be nothing, or it could be the beginning of an event that changed the rest of the day, sent me to a hospital, triggered a shock, or reset the horizon of hope. The body becomes a prediction engine under those conditions. It absorbs priors. It updates from every episode. It notices small changes because small changes have carried large consequences before.
That kind of vigilance can protect you.
It can also consume you.
The lesson was never to stop feeling fear. The lesson was to stop allowing fear to make every decision.
Embrace it.
Interrogate it.
Ask what it noticed.
Then make it show its work.
What is the threat?
How likely is it?
What evidence would change the estimate?
What can be reversed?
What must be held until more is known?
Who carries the consequence if the calculation is wrong?
Fear becomes useful when it is translated into questions, boundaries, and action. Left unexamined, it becomes paralysis or panic. Dismissed entirely, it becomes exposure.
This is how I think about AI.
I will not tell frightened people that there is nothing to fear. That would be dishonest.
I will tell them that fear does not have to be the end of their agency.
Capability is not direction
The phrase AI for good is almost useless on its own.
Good for whom?
Defined by whom, measured where, over what time?
Who receives the benefit - and who carries the loss?
Artificial intelligence is a multiplier. It can scale the quality of a careful process, and it can scale the damage of a careless one. It can help a patient find language for an experience nobody around them understands. It can also bury that patient under an ocean of authoritative-sounding misinformation.
The machine does not resolve the moral question by becoming more capable.
It makes the moral question more consequential.
This is the missing variable in most conversations about scaling. We talk about how much compute enters the system and how much intelligence appears to come out. We talk less about the objective function that capability is serving.
If the system is optimizing engagement, it will become better at holding attention.
If it is optimizing labor reduction, it will become better at removing labor.
If it is optimizing revenue, it will become better at finding revenue.
If it is optimizing human capability, it can help a person do something that institutional access, money, illness, geography, or a broken trajectory once placed beyond reach.
AI does not choose among those purposes.
People do.
That is why I do not believe the future should be organized around artificial intelligence replacing human judgment. The humane use of this technology should make people more capable of exercising judgment - and more able to prove how they exercised it.
The model can propose. It can search, calculate, and write code. It can attack an argument, identify a contradiction, run another scenario, and return with a better question.
It cannot decide what another human being is worth.
It cannot decide which uncertainty a patient should be forced to carry.
It cannot decide whether a trade-off is acceptable to the person living with the irreversible consequence.
It cannot carry the liability for a claim simply because it generated the sentence.
The model can hallucinate and move on.
The human must not. Their name remains on the claim, and another human being may carry its consequence.
That boundary is not a limitation to engineer away.
It is the necessary human gate.
The human scaling laws
The machines have scaling laws.
So do we.
Human capability scales when the cost of translation falls.
One person’s lived experience can become a question. The question can become a research protocol. The protocol can become an analysis. The analysis can become code. The code can become an inspectable model. The model can be attacked by other models and reviewed by people with deeper domain expertise. The corrections can be preserved. The reasoning can be published. The work can reach someone who was never in the room where any of it was built.
Human judgment scales when it remains visible.
An answer without provenance asks for trust. A decision with its sources, assumptions, transformations, dissent, corrections, and approvals exposed can earn it.
Human dignity scales when capability is handed outward.
The point is not to build a more impressive machine beside the person who has been overlooked. The point is to put a rung within reach.
And impact scales when the work returns to the burden that created it.
That is the law I care about - not how many parameters a model contains, not how many tokens I can generate, not how many disciplines I can appear to cross.
How much human burden can be responsibly reduced?
How much uncertainty can be named without manufacturing false certainty?
How many people can become more capable without becoming more dependent?
How far can one person’s vigilance travel before it loses contact with the people it was meant to serve?
These questions are not soft. They are design requirements.
They determine the architecture.
The edge was lonely
There is a cost to climbing early.
I was living at two edges at once. One was the edge of machine capability - orchestrating models toward serious work while most of the software world still considered the whole category unserious. The other was the edge I was recovering from: the extreme end of arrhythmia burden, a body coming back from a decade of load, a mind discovering how thick the fog had been only because it was finally gone. Neither edge had shared language yet. Neither could be seen from the outside. And the people closest to me would have needed to understand both to understand me at all.
When I leaned into this technology, the environment was hostile. Much of the software world met AI-assisted building with something between dismissal and contempt - and I want to be fair about why, because the fairness is the analysis.
Start with the debt. The enterprise of software development built the world this essay travels through. Every system we orchestrate, every model we call, every network that carries these words exists because generations of engineers turned thought into instruction with a rigor the rest of us inherited for free. Nothing here is contempt for that enterprise. It is awe with a question attached.
Because code was never only a tool. It was a language - and for decades, fluency in that language decided who could write the rules of digital life. Lawrence Lessig saw it early: code is law. The architecture of software governs behavior as surely as statute does. What Lessig described as regulation, I lived as a border. The law of the digital world was written in a language most of humanity could not read, let alone write, and that asymmetry shielded the institutions that employed its fluent speakers. The priesthood was real. So was the moat.
The early models earned skepticism - they failed often and confidently, and an experienced engineer who tested one, watched it fabricate an interface, and walked away was making a rational call about their time. But underneath the rational part was something I recognized immediately, because I had spent my whole life on the receiving end of it: identity threat. When your language is the law, a translator is not a convenience. It is a constitutional event. And when something arrives that a person cannot categorize without redrawing their own borders, people do what people do.
They reach for labels.
I knew that gesture. I had been labeled by it my entire life.
Here is what I believe was actually arriving, underneath the arguments about tooling. Code is still law. The difference is that natural language can now enforce machine language. The borders between the two languages have not disappeared - but for the first time, the crossing runs in both directions, and the cost is collapsing. That is symmetry between languages, and symmetry between languages is symmetry of power. A person who carries a code of their own - values, intent, a law of conduct written in nothing but honest words - can now put it through a compiler and watch it become something a machine will faithfully execute and another human can inspect. The creative can enforce their human code.
That is not the death of software development. It is the enterprise’s own achievement completing itself: the translation layer engineers spent seventy years building finally reached all the way down to ordinary language. The people who built the bridge should not be surprised that so many are crossing it.
I kept building - with verification, with gates, with the discipline this essay describes - through a stretch when almost nobody in the room believed the room existed. It wasn’t until the tools crossed a threshold that could no longer be argued with that the conversation turned, and many of the same voices that had dismissed the category quietly picked it back up. I don’t hold that against them. Their skepticism was rational and identity-load-bearing at the same time, and nobody has to be the villain - not even my critics.
Towards the end of 2025, in the middle of that hostile stretch, an engineer from my hometown’s university found one of my articles on LinkedIn. He said it read like a PhD student defending a thesis.
I have no degree. I have a derailed academic track and a stack of labels older than my career. That one sentence, from a stranger carrying every credential I lack, meant more to me than he could have known.
And here is the part that still stings. At the same moment a stranger was reaching for a new category to describe my work, the people at home were reaching for the old ones - the labels that historically made sense. They could not see what I was building, and they could not see what I was recovering from. Two invisible frontiers, and both got filed under what was already believed about me.
The stranger updated on the evidence. The family updated on the history.
That is the loneliness of building on the edge of capability while healing from the edge of burden: the work arrives before the categories do, and the recovery runs deeper than the people watching can measure. The people closest to you hold the oldest priors, and the oldest priors move slowest. If you are early enough - at either edge - recognition will come from strangers before it comes from home, and you will have to decide, in that gap, whether the work and the healing are true anyway.
I decided they were.
From feeling to infrastructure
This is why the work extends beyond an essay.
The Consequences of a Broken Heart settled the direction. The aggregate response to that essay and the articles that followed made the first need unmistakable: people living with arrhythmia need language, confidence, connection, and psychological support that recognizes the cardiac reality of what they are carrying.
Confidence grounded in transparency.
The work must respect the clinical boundary. Lived experience gives me standing to ask the question. It does not give me authority to predetermine the answer. Artificial intelligence gives me additional capacity to investigate and build. It does not confer a medical credential or remove the need for expert review.
That is why OneRhythm begins with the human experience but cannot end with testimony alone.
It must connect the person to evidence, to other people, and to care without pretending to replace any of them.
It is also why Phial exists.
Phial began inside a consequential health-economic question, where evidence had to be selected, assumptions defended, models tested, errors corrected, and interpretation held behind explicit human gates. But the larger idea is not limited to economics.
Phial is being designed to preserve the space between machine output and human decision.
What evidence entered?
What was excluded?
Which arguments were considered?
Why was one accepted and another rejected?
Where did the model contribute?
Where did a human overrule it?
What changed after adversarial review?
Who approved the next step?
What uncertainty remained when the decision was made?
This is reasoning provenance.
The goal is not to automate judgment out of the system.
The goal is to make judgment inspectable enough to be challenged, taught, corrected, and trusted.
That is AI for good made operational.
Not a slogan.
An objective function.
A set of constraints.
A record.
A human being accountable at the gate.
What the replies proved
The replies did not prove that a series of essays can solve psychological distress in arrhythmia.
They proved something more useful.
Language can reduce isolation.
One person’s attempt to name an invisible burden can help another person explain it to a spouse. It can give someone words during a bad night. It can help a newly diagnosed patient recognize that mental-health support belongs inside the response to a cardiac condition. It can remind someone waiting for a procedure that fear is not evidence of weakness. It can tell a person who has been dismissed that psychological suffering does not make the physical experience imaginary.
The essays began with one person’s attempt to name the burden.
The recognition belonged to many.
That is scale.
Let the heart do what it wants
When the arrhythmia finally quieted, I expected peace.
What came first was clarity.
The burden had occupied so much of my internal world for so long that I could not see its full shape while I was inside it. Once the rhythm settled, I could see what remained: the vigilance, still running; the grief, still present; the mind, still scanning for the bear.
I could try to shut it all down.
Or I could give it somewhere to go.
The only way this heart stays calm is if I let it do what it wants.
It wants to work.
It wants to take fear and turn it into a question - the question into inquiry, inquiry into something another person can inspect, inspection into confidence, and confidence into agency.
It wants to build for the person whose pain is invisible, whose credentials do not tell the truth about their capacity, whose geography put expertise out of reach, whose illness interrupted the path, whose family does not understand, or whose fear of artificial intelligence is really the fear that the future no longer has a place for them.
I will not lie to that person.
The tools are imperfect. The risks are real. The work is difficult. Judgment cannot be delegated. Verification is not optional. And nothing about expanded capability absolves the human being using it from responsibility for the consequence.
But the ladder is real.
At a scale that was recently unavailable, a person without institutional reach can now build with something approaching it. A person fluent enough in the problem, disciplined enough in the process, and humble enough to expose the work to correction can move between natural language and machine language without asking permission at every border.
The machine’s scaling law asks how much capability emerges when we add more resources.
The human scaling law asks what happens when that capability is placed in the hands of someone who knows exactly who they are trying to serve.
My answer is still being built.
It is in the essays.
It is in OneRhythm.
It is in Phial.
It is in the record of every question, correction, and human gate.
It is in the refusal to let fear have the final word.
And it is in a heart that is calm only when the vigilance has somewhere useful to go.
That is the scaling law that matters:
Artificial intelligence can expand the reach of a human mind. Human judgment must decide the direction. Human consequence is the measure.
If this technology is going to scale, then so must our responsibility for what we ask it to amplify.
I know what I am asking it to amplify.
The people who answered those essays made certain of it.
There is a reason this essay exists at all, and it is the same reason I survived long enough to write it.
I have always maintained that words have more value than any currency the world has ever known. I believed it when language was the only asset I had - the thing that carried me into rooms my paperwork could not. I believed it when writing put ten years of silence into a shape a stranger could hold. And I believe it now, in an era when words have become the way humans direct machines.
Words have healed me.
They are my value.
And they are the most important asset I will ever own.
- Matthew J. Adams
ad astra per aspera