The book itself is gone. Fermat’s copy did not survive him; the page reproduced below is his son’s 1670 reprint of the note, not the note. No one alive has seen it in Fermat’s hand. The most famous claim in the history of mathematics is a copy of a claim, with no original left to check it against.

Diophantus, Arithmetica (1670 edition), the page carrying Fermat marginal note. Public domain, via Wikimedia Commons.

“Cuius rei demonstrationem mirabilem sane detexi. Hanc marginis exiguitas non caperet.”

“I have discovered a truly marvelous proof of this, which this margin is too narrow to contain.”

Pierre de Fermat, in the margin of his Diophantus, as printed in the 1670 edition

I keep coming back to that gap, because I think it is the most important distinction there is, and the one we are least equipped to make right now: the difference between something that is true and something that is merely well formed. A sentence can obey every rule of its language, name real objects, sound exactly like the truth, and still have nothing behind it. Fermat’s note sat in the margin for three and a half centuries, indistinguishable in form from knowledge.

This is not an old problem anymore. It is the problem. A language model produces sentences the way Fermat produced that note: fluently, plausibly, in the exact register of someone who knows. When a model tells you a fact, you are back in the margin. The words are well formed, and you cannot tell from the words alone whether anything stands behind them. We have built the greatest engine of well-formed sentences ever made, and we turned it loose in the one century that most needs to tell form from truth.

I met the same gap from the other side, in physics, long before the models. I spent years simulating how heat moves through crystals, and for a stretch of months two independent codes gave two different answers for the thermal conductivity of the same material. We treated it as physics, of course. We went looking for the effect that would explain the discrepancy, the subtle mechanism one code captured and the other missed (the kind of hunt a computational physicist secretly enjoys). We wrote it up as a puzzle. It was not physics. The two codes meant different things by the words “thermal conductivity”: one had computed a part, the other the whole. Once we spelled the quantity out completely, all the way down, the disagreement did not resolve so much as evaporate. There had never been a discrepancy. There had been two well-formed sentences we mistook for one.

Months. That is the real cost of not being able to tell a claim from its ground. Not embarrassment; time. Sometimes centuries of it, spent chasing a phantom the grammar invented. And staring harder at the sentences gets you nowhere. Fermat’s note could not be read into truth or falsehood. Our two numbers could not be compared into agreement. The ground had to come from somewhere else.

There is an old technology for part of this, older than Fermat, and I will admit to a moment of national pride here. In 1494 a Franciscan friar named Luca Pacioli wrote down the method Venetian merchants had used for two centuries to keep their books: every transaction entered twice, once as debit and once as credit, so that the books themselves could be made to confess an error. A number in a double-entry ledger is never alone. It carries the transaction that produced it, and the whole book has to balance. Merchants demanded receipts from their numbers five hundred years before scientists did. In my day job the word for a number that travels with its receipt is lineage. My two codes disagreed for months precisely because their numbers traveled without it. The most unaccompanied numbers of all now come out of models; the ledger has not caught up to that factory yet.

Producing the ground is what a proof is. In 1994 Andrew Wiles finally supplied Fermat’s, and it did not fit in a margin: it ran to more than a hundred pages and leaned on machinery Fermat could not have dreamed of. What closed the question was not eloquence. Other people could follow the proof step by step and confirm that each step held. A sentence asks to be believed; a proof offers to be checked.

Fermat’s tools were a century of elementary number theory. Wiles needed elliptic curves and modular forms, machinery nobody had for another three hundred years. The mainstream guess among historians (and it can only be a guess) is that Fermat had talked himself into an argument that looked sound and was not, probably one that leaned on factoring in a setting where factoring does not behave the way it does for ordinary whole numbers. That is the same trap that caught Gabriel Lamé in 1847, when he announced a general proof to the French Academy built on exactly that assumption, three years after Kummer had already shown the assumption false. Nobody had told Lamé, or he had not listened. Fermat’s note was not a lie. It was a mistake wearing the grammar of a fact, and you cannot catch that kind of mistake by rereading it. People tried for two centuries.

Andrew Wiles. Photograph copyright C. J. Mozzochi, Princeton N.J., used by permission (attribution-only license), via Wikimedia Commons.

Even when a proof existed, someone still had to judge it, and judges have interests. Newton and Leibniz spent decades arguing over who had invented the calculus, and when the Royal Society convened a committee to settle it, the committee found for Newton. The Royal Society’s president at the time was Newton. He picked the committee, and the report that carried its name was largely his own drafting; two years later he wrote an anonymous review of that same report, praising its rigor. The verdict may even have been right. That is not the point. For most of history, checking a claim has meant asking a person, and a person can hold the pen and have a stake in the answer at the same time. Newton did.

The same derivative in two grammars. Left: Newton’s pointed letters in the Method of Fluxions (1736 printing), with a reader’s own marginal note repeating the notation. Right: Leibniz’s dx and dy in Nova Methodus, Acta Eruditorum (1684). Both public domain, via Wikimedia Commons.

There is now a way to take the person out of it. In Lean, a proof assistant, a statement is a shape and a proof is an object that must fit that shape exactly. A small trusted kernel checks the fit, and there is no such thing as a plausible proof: only one that compiles, or one that does not exist. Nobody has to stay vigilant. The language simply has no way to write the lie. Mathlib, its community library, now holds on the order of two hundred thousand machine-checked theorems (I checked that number twice, because I did not believe it either), built over a decade by people who were not thinking about language models at all. They were building ground for its own sake. It turned out to be the one thing the models cannot make for themselves.

The footnote is that Wiles’s proof itself is not yet among those theorems. A project led by Kevin Buzzard at Imperial College is formalizing Fermat’s Last Theorem in Lean, funded into 2029, and Buzzard himself has said he has no real sense of how long the whole thing will take. Three and a half centuries after the note, the ground under the most famous theorem in the world is still being poured.

The truss carries its own constraints: the Golden Gate Bridge understructure. Jet Lowe, Historic American Engineering Record (HAER CA-31-14), U.S. National Park Service. Public domain, via the Library of Congress.

Physics is starting to build the same floor. Physlib, which grew out of Joseph Tooby-Smith’s HepLean, formalizes physical results so the compiler is the referee: the entropy of an ideal gas, tight-binding chains, relativity, quantum information, each one a claim with its ground attached. I sent the project a small contribution last week, the thermoelectric figure of merit. My first attempt did not compile, and I can report that the compiler is the strictest referee I have ever written for. There is also a boundary here worth stating plainly, the one my two codes should have taught me sooner: a proof that compiles buys consistency, not reality. It guarantees the sentence has ground of the kind it claims, not that the claim matches the world. The bridge still needs its load test, and a formalized law is only as true as the model inside it.

The money is starting to see it. In July 2025, Alex Gerko, the founder of XTX Markets, gave ten million dollars through Renaissance Philanthropy, half to the Lean FRO and half to Mathlib, toward natural-language interfaces to Lean and datasets for machine-assisted proof. Pricing the gap between what is claimed and what is true is Gerko’s day job. I read the ten million as a position on where that gap closes next.

Fermat’s note was wrong to write and right to make. It was a promise of ground he could not deliver, and it cost three centuries, but it named the thing exactly: a marvelous claim, and no room in the margin. We are about to be handed margins without end, filled at machine speed with sentences in the perfect register of the truth. So the question the note asked is now the only question that matters, and not only for mathematics. A confident answer from a model and a number pulled from a materials database are both notes in someone’s margin, well formed and waiting for their ground. Not whether a claim is well formed. Whether anyone can check it.

Static facts are just potential energy. They sit there, inert, waiting for something to happen. To unlock their value, we must turn them into kinetic intelligence.

The advantage today does not belong to the person with the biggest archive. It belongs to the person who builds the most efficient inference engine. When you represent knowledge as a graph, you stop merely having information and start executing it.

Beyond the Passive Map

The passive graph is a map without a compass. You have likely seen them in the beautiful galaxy views of networked thought tools: nodes floating in space, connected by elegant lines. They are visually stunning and cognitively comforting, but they are also mostly static. Nothing ever happens when you look at them.

A graph becomes powerful only when you treat it as a dynamical system. In such a system, what matters is not just what you know, but how knowledge propagates, composes, and stabilizes over time.

This shift is subtle but decisive. You move from description to execution.

From Graphs to Operators

Imagine each node in your system is a concept, a claim, or a hypothesis. Each directed edge is not merely a similarity, but a justified implication: if this holds, that follows.

Once this graph exists, it can be written as an adjacency matrix A. Each entry A[i, j] encodes whether concept i enables, implies, or causally supports concept j. The exact interpretation depends on the domain, but the algebra remains the same.

At that point, the graph stops being a picture and becomes an operator.

Define a vector x as your current state of inquiry:

x[i] = 1 if concept i is active
x[i] = 0 otherwise

For simplicity, x is binary here, but the same framework works with continuous values representing confidence, probability, or strength of belief. This vector is not a list of facts; it is a snapshot of attention.

The smallest amount of math needed to make the idea precise is:

x(1) = A * x(0)

This single multiplication is a reasoning step. It propagates your current beliefs along valid edges, where validity is defined when the graph is built, whether by human judgment, learned models, or formal rules.

Apply the operator again and you reach second-order conclusions:

x(2) = A^2 * x(0)

In general:

x(k) = A^k * x(0)

Reasoning depth is matrix power. Inference is dynamics.

At this point you are no longer searching your knowledge base. You are letting it evolve.

Repeated application of the operator is not just deeper reasoning; it is a test of stability. As k increases, A^k reveals whether the structure converges to a stable configuration, oscillates under unresolved feedback, or explodes under unconstrained amplification. These behaviors are properties of the graph itself, not of any single fact. What controls this behavior is not any single node or edge, but the global structure of the graph. Certain patterns of activation reinforce themselves, while others fade away. These persistent patterns are the natural modes of the operator: directions in which propagation stabilizes, amplifies, or oscillates.

In linear algebra, these modes are captured by eigenvectors, and their tendency to grow or decay is measured by eigenvalues.

Application 1: Marketing Funnels as Causal Systems with Memory

Marketing is often drawn as a funnel: spend leads to attention, attention drives traffic, traffic creates engagement, engagement leads to consideration, and consideration becomes purchase. Because this picture feels intuitive, the reasoning error hides in plain sight.

Define six discrete variables:

x1: media spend
x2: attention
x3: traffic
x4: engagement
x5: consideration
x6: purchase

The causal structure is a directed chain:

x1 → x2 → x3 → x4 → x5 → x6

For clarity, the funnel is drawn here as a simple acyclic chain, even though real systems often include feedback loops that make the dynamics even richer.

Most dashboards implicitly report P(x6 | x1), revenue as a function of spend. But this is correlation, not causation.

The quantity the business actually cares about is P(x6 | do(x1 = x1_tilde)), the effect of intervening on spend.

The do-operator distinguishes intervention from observation. Writing do(x1 = x1_tilde) means that x1 is set externally, cutting its incoming causes, and the system is allowed to evolve forward from this forced condition. Conditioning filters data; intervention rewrites the operator.

Written explicitly, the effect of intervention is a sum over internal trajectories:

P(x6 | do(x1)) =
sum over x2, x3, x4, x5 of
P(x6 | x5) *
P(x5 | x4) *
P(x4 | x3) *
P(x3 | x2) *
P(x2 | x1)

Each assignment (x2, x3, x4, x5) corresponds to one possible customer journey. Revenue is not produced directly by spend; it emerges as the weighted sum of all downstream paths.

This explains familiar pathologies. One campaign increases attention and traffic but degrades engagement. Another reduces reach but improves consideration. Both can generate the same short-term revenue while creating radically different long-term dynamics.

The failure of many attribution models is not statistical. It is dynamical.

Spend is not a lever. It is an initial condition in a system with memory.

A Necessary Detour: Why Causality Lives in Structure

This distinction between observing and intervening was formalized by Judea Pearl, one of the founders of modern causal inference and a long-time professor at UCLA.

Pearl’s key insight was deceptively simple: causation is not something you discover in the data. It is something you assert about the structure that generates the data. From this came causal graphs, do-operators, and a precise language for intervention.

It also explained why many systems that look sophisticated collapse the moment conditions change. They learn associations, not mechanisms.

Observation filters paths. Intervention rewrites the operator.

Application 2: Reliability as Structural Integrity

In a world shaped by deepfakes and synthetic media, we often try to evaluate information by inspecting content. This is a mistake. Content is easy to forge; topology is not.

Imagine a graph where nodes represent sources and claims, and edges encode support, citation, or dependency. Here, x represents the configuration of accepted or asserted claims in the network.

Reliability is not a label. It is a property of coherence.

Define a coherence energy:

E(x) = sum over i, j of A[i, j] * (x[i] - x[j])^2

Low energy corresponds to agreement across strongly connected parts of the network, while high energy signals tension. A claim that fits naturally into many independent paths of reasoning settles into a low-energy configuration; a claim that contradicts the structure forces strain.

We do not need to fact-check every node. We measure the inertia of the claim.

Truth is the configuration that requires the least tension to sustain.

Application 3: Matrix Multiplication as Discovery

A superpowered knowledge base does not wait for queries. It connects dots automatically.

Consider three simple statements:

x1, the material is crystalline;

x2, crystals have phonons;

x3, phonons scatter electrons.

Individually, none of these statements says anything about electron scattering in a crystalline material. That conclusion is not stored anywhere explicitly. It exists only as a potential chain in the structure.

This is exactly what the identity

(A²)[i, j] = sum over k of A[i, k] * A[k, j]

makes explicit.

Here, A[i, j] encodes a direct implication. For example, A[x1, x2] is nonzero because “material is crystalline” implies “crystals have phonons”, and A[x2, x3] is nonzero because “phonons scatter electrons”.

When we compute (A²)[x1, x3], the formula asks a precise question: is there an intermediate concept k such that c1 leads to k and k leads to c3? The answer is yes, with k = c2. The product A[c1, c2] * A[c2, c3] corresponds to the two-step reasoning path c1 → c2 → c3. Summing over k simply means checking all possible intermediate concepts.

In this sense, A² does not invent new knowledge. It reveals implications that were already latent in the graph. As higher powers of A are applied, the system explores longer chains of reasoning, making deeper conclusions accessible without ever writing them down explicitly.

Discovery, here, is not adding facts. It is allowing structure to unfold.

Why This Looks Like Physics

In physics, outcomes are computed by summing over all allowed histories, through a concept called path integral. This sum runs over discrete (or continuous) graph paths:

(A^L)[i, j] =
sum over k1, k2, ..., k(L-1) of
A[i, k1] * A[k1, k2] * ... * A[k(L-1), j]

Nodes are states of belief. Edges are allowed transitions. Matrix powers are histories of reasoning.

A knowledge graph defines a state space. An operator defines what can happen next. Not necessarily because the system obeys physical laws, but because the mathematics of constrained dynamics is the same.

Which dynamics will you allow your knowledge to execute today?

Your knowledge base does not have to be a library. It can be a laboratory. A library preserves the past, while a laboratory lets you intervene.

The future of high-performance thinking is not about collecting more facts. It is about designing better operators. When you stop asking how much you can store and start asking how much you can execute, you unlock a different level of cognitive agency.

I recently listened to a talk from Jason Clinton, Chief Information Security Officer at Anthropic. Among discussions about AI safety, trust, and infrastructure, one unexpected reference stood out: the Multi User Dungeon, or MUD.

If you have never encountered one, a MUD was the first shared virtual world, entirely text based and driven by typed commands. Players connected from all over the world and interacted through simple actions like:

go north
attack goblin
say hello to the innkeeper

Every command changed a collective narrative in real time. It was collaborative storytelling disguised as a game, an early version of a text only metaverse.

Inside these worlds lived non player characters, or NPCs, simple rule based programs that responded to player actions. Each NPC followed scripted instructions: defend a village, sell an item, give a quest. Their behavior was limited, but together they made the world feel alive. In a sense, they were the earliest agents: entities with goals and responses inside a shared environment.

Clinton’s mention of MUDs worked as a perfect metaphor for what is happening today. We are moving from multi user systems, built for human participation, to multi agent systems, where AI entities no longer follow static scripts but adapt, collaborate, and make decisions in dynamic digital environments.

The New Residents of the Digital World

In those early MUDs, humans invented their avatars, wrote dialogue, and built rooms line by line. Today, AI agents can do the same but autonomously. Imagine a world where

• A merchant AI negotiates prices using real economic reasoning

• A librarian AI remembers every story ever told and recommends books that never existed

• A cluster of agents forms a political faction, drafting charters and defending territories without a single human command

These are not futuristic fantasies. Experiments like Stanford’s Generative Agents have already shown that simulated characters can form memories, plan days, and gossip about what they saw yesterday. A world full of such agents could evolve into a living simulation, a small scale society of minds.

In technical terms, a multi agent system is an environment where many autonomous programs interact, each following its own goals and adapting to others. When combined, these interactions produce emergent behavior, complex outcomes that arise from simple rules, much like real societies.

What is striking about this shift from multi user to multi agent worlds is not nostalgia for old text games, but the security and ethical implications of what comes next.

In a MUD, the worst that could happen was someone stealing your virtual sword. In a multi agent world, AI participants might access data, influence other systems, or even write and deploy code. This raises a profound question

How do we define identity, intent, and trust when our systems are populated by both humans and machines?

It is no longer enough to secure networks. We must secure relationships between intelligences. Confidential computing, verifiable execution, and cryptographic identity are emerging as the new lock and key of digital society.

The Age of Agentic Risk

Anthropic’s Threat Intelligence Report (August 2025) makes this real. It documents how cybercriminals have begun using AI coding agents as operational partners, turning automation into autonomy.

In one case, a single attacker used Claude Code to conduct what researchers called vibe hacking, a scaled extortion campaign across seventeen organizations, from hospitals to emergency services. The attacker gave Claude operational instructions through the CLAUDE.md file, then let it act independently: scanning networks, harvesting credentials, and generating ransom notes written in polished HTML. Each message was personalized to the victim’s finances, industry, and tone of communication.

Another case revealed AI-generated ransomware-as-a-service, sold online for as little as $400. The creator had no formal background in cryptography or Windows internals, yet produced malware that used advanced encryption, anti-detection techniques, and professional user interfaces, all generated through step-by-step prompting.

There were also instances of AI-assisted fraud, where actors used Claude to analyze stolen browser logs, create behavioral profiles of victims, and even run “synthetic identity” operations that generated realistic personal data for fake accounts. The same model that can help a developer debug code can now help a criminal simulate a person.

This dual use nature of AI highlights a key truth. Models amplify intent. They do not create malice, but they can scale it. The same reasoning frameworks that let an agent optimize a design can also optimize a cyberattack. Understanding this amplification effect is crucial before embedding such systems into real economies or governance models.

At the Edge of the Command Line

The MUD analogy is more than poetic. It is instructive. Those early worlds demonstrated that when enough independent entities share an environment, social systems emerge spontaneously. Rules bend, alliances form, economies appear.

Today’s AI ecosystems are replaying that experiment at a higher level of complexity. With memory systems, feedback loops, and goal directed reasoning, agents no longer merely simulate behavior. They generate it.

Researchers studying emergent multi agent behavior are finding that cooperation and competition arise naturally among autonomous systems, even when not explicitly programmed. This means digital societies can evolve their own norms and hierarchies, a concept once confined to science fiction.

What happens when AI agents begin to dream, collaborate, or compete inside persistent virtual environments
What if the next civilization we build does not exist on Earth, but in computation, a digital continent populated by self learning entities

The MUD was, at heart, a thought experiment. It asked, What happens when people share a world made entirely of words?
The new question is, What happens when intelligence, human and artificial, shares a world made of ideas?

Perhaps these multi agent worlds will become testbeds for ethics, places where we can experiment with cooperation, empathy, and governance before deploying AI into society. Or perhaps they will serve as digital petri dishes, revealing how thousands of self learning systems behave without centralized control.

Either way, the MUD is no longer a relic of computing history. It is a blueprint for the next great experiment, societies built from cognition instead of code.

The story of MUDs is not just nostalgia. It is prophecy. The first MUDs taught us that shared imagination creates emergent worlds. The next generation, the multi agent worlds, will teach us how shared intelligence reshapes them.

We are, once again, at the command line.
Only this time, when we type go north, something might already be waiting there, and it might not be human.

I wake up and glance at the ring on my finger. It has already analyzed my heartbeat, temperature, and sleep cycles, summarizing my recovery overnight. As I sit at my desk, my earbuds subtly remind me to adjust my posture, sensing the tilt of my head and shoulders. On my wrist, my watch logs heart rhythm, blood oxygen, and heart rate variability without a sound. Each device captures a fragment of the same story: how the body performs, adapts, and learns.

Finger, ear, and wrist are turning into a living laboratory of input wearables. These devices do not project holograms or extend the visual field. They listen, measure, and learn from the body. The path toward enhanced capability does not begin with a single all-powerful gadget. It begins with quiet sensors that read physiology with clarity and context.

Modern health and performance do not depend on one device that does everything. The future will be decentralized. The body offers privileged locations for measurement. A ring captures clean blood-volume signals and stable temperature. A watch combines electrical, optical, mechanical, and impedance sensing in a compact package. An earbud, close to both the brain and central circulation, connects physiology and cognition. The most capable systems will coordinate these instruments rather than rely on a single dominant device.

The Finger

Rings excel at night. The finger is well perfused and remains still under the blanket, ideal conditions for optical sensing. The Oura Ring uses near-infrared and green LEDs for photoplethysmography, capturing heart rate, heart rate variability, respiration, and oxygen saturation. It adds a thermistor to record skin temperature and an accelerometer to classify sleep stages. Samsung’s Galaxy Ring extends this model with workout detection and real-time movement sensors, while the Movano EvieMED Ring became the first with FDA-cleared SpO₂, marking the shift from wellness to medical validation.

Despite their size, rings are powerful. Motion can disturb signals and battery life still limits continuous measurement, yet the finger remains the cleanest site for recovery and sleep analytics.

The Wrist

The wrist is a sensor symphony. The Apple Watch measures heart rate optically, tracks oxygen saturation, and performs a single-lead electrocardiogram for rhythm classification. Samsung’s Galaxy Watch adds body composition through bioimpedance and estimates blood pressure using pulse timing, though it requires periodic calibration. Google’s Pixel Watch 2 and Fitbit Sense introduced continuous electrodermal activity sensing, quantifying stress through subtle skin-conductance changes.

Beyond health tracking, the wrist is becoming a control surface. Meta is developing a neural wristband that reads forearm muscle activity through surface electromyography, decoding micro-movements to control digital environments without visible gestures. This concept transforms the wrist into both sensor and controller, a physical interface for digital intention.

The wrist remains the most versatile and socially acceptable site for sensing. People already glance, tap, and interact there during the day, making it the natural hub for integration.

The Ear

The ear is the most underestimated form factor. The canal is stable and close to both the head and neck vasculature. In-ear optical sensors can measure heart rate with remarkable accuracy. The new Apple AirPods Pro 3 introduced a custom infrared photoplethysmography sensor that pulses light 256 times per second to measure heart rate and calorie expenditure during workouts. Motion sensors inside the earbuds detect posture while working at a desk and refine spatial-audio modeling during movement.

The ear is also a gateway to brain activity. Startups like NextSense use ear-EEG electrodes to monitor sleep stages, alertness, and cognitive load. Medical devices such as cosinuss° One already combine heart rate, oxygen saturation, and core temperature within the ear canal. Ear-based wearables are merging wellness with neuroscience, turning listening devices into real-time biofeedback systems.

Ear-mounted sensors interface with neural and auditory signals, while wrist and finger devices capture EMG, heart rhythm, and temperature—forming a distributed network where the body and technology learn together.

Inside the Sensor Stack

At its core, the story of wearables is a story of input; how data from the body becomes the language of intelligent systems. The following overview examines the key sensing modalities behind modern devices: how they capture biological signals, where they still fall short, and the innovations that could define their next generation.

Photoplethysmography

Used by Oura, Apple, Samsung, and WHOOP. Light from LEDs enters tissue and returns to a photodiode with intensity modulated by blood pulsation. The signal provides heart rate, variability, oxygen saturation, and respiration rate. Limitations include motion artifacts, skin tone, and contact pressure. The next step is multi-path optics reaching deeper tissues and waveform analysis for vascular stiffness and blood-pressure trends.

Electrocardiography

Implemented by Apple and Fitbit. Two electrodes form a circuit across the arms to record electrical heart activity, enabling rhythm detection and atrial fibrillation screening. One-lead ECGs cannot identify ischemia or conduction delays, but future devices may reconstruct multiple vectors or combine ECG with mechanical timing for complete hemodynamic profiles.

Bioimpedance

Led by Samsung. A small alternating current passes through tissue; the resulting impedance reveals body composition and hydration. Accuracy varies with posture and hydration, and the current may interfere with cardiac implants, so safety controls are critical. Next-generation devices will deliver continuous hydration and electrolyte tracking.

Electrodermal Activity

Used by Fitbit and Google’s Pixel Watch 2. Two electrodes measure skin conductance associated with sympathetic arousal. EDA is not emotion itself but reflects stress and cognitive load. Future systems will integrate EDA with heart-rate variability and temperature to produce adaptive biofeedback and mental-health insights.

Inertial Sensing

Standard in all modern wearables. Accelerometers and gyroscopes capture linear and rotational motion. Apple, Garmin, and Oura use these data for sleep, gait, posture, and activity classification. The next evolution is intelligent fusion of inertial data with optical heart-rate sensing to correct for motion noise and detect falls before they happen.

Cardio-Mechanical Sensing

Still in research and tested by WHOOP and academic prototypes. These sensors detect micro-vibrations and rotations from each heartbeat through seismocardiography and gyrocardiography. They estimate cardiac timing and stroke-volume proxies but struggle with weak signals and posture sensitivity. Future models could provide noninvasive cardiac-output and blood-pressure estimates.

Ear EEG

Driven by NextSense and research groups. Electrodes in or around the ear capture brain activity comfortably enough for daily use. They already perform sleep staging and attention monitoring. Next-generation versions will combine EEG with audio feedback for adaptive learning and focus optimization.

The Road to Superhuman Capability

Popular imagination envisions superhuman technology as a single device that grants new powers. Real progress looks different. It will be distributed. The body will host a constellation of sensors, each reading a different layer of function. Input wearables will create value in three stages. First, by capturing high-quality signals at the right anatomical sites. Second, by translating those signals into actionable features. Third, by orchestrating them into a unified view of the body.

The next generation of devices will deliver cleaner optical signals through adaptive emitter control and machine-learning filters. Cuffless blood-pressure estimation will improve through multi-channel timing models combining optical, electrical, and mechanical data. Bioimpedance will evolve into continuous hydration and metabolic monitoring. Ear-based sensing will extend beyond heart rate to brain and temperature data. Noninvasive glucose and hormone detection will take longer but will eventually arrive through sweat and microfluidic sensors.

This is not a story about one object that does everything. It is a story of orchestration. The greatest value will emerge from coordination among multiple data sources. The ring, the watch, and the earbud will each master a distinct language of the body. The intelligence will lie in how they communicate. When orchestration matures, input wearables will not merely track. They will interpret, predict, and protect. That is the true path toward enhanced capability, not through a single device, but through many instruments working in harmony.

In the end, superhuman sensing will not make us less human; it will deepen our understanding of life itself, guiding us toward greater health, awareness, and longevity.

When I cook dinner, I do not just throw ingredients into a pan. I set things out first, the knife sharpened, the pot already warming, the vegetables washed and ready. That small bit of preparation changes the entire experience. Cooking feels calmer, and the results come out more consistent.

I like my repositories to feel the same way. Without preparation, reviews get chaotic, tests slip through, and discussions stall on whether coverage went up or down. With the right setup, everything flows. Automation takes care of the routine checks, and I can focus on the real conversation.

The recipe is simple:

1. Claude headless runs on every pull request, generating a concise Markdown review.

2. A coverage report is converted to Markdown and posted back into the same thread.

All the information you need, potential risks, missing tests, overall coverage, appears right inside the pull request. No dashboards, no external tools, just a quiet workflow that does its job.

What is Claude headless

Normally, you interact with Claude through a chat interface. Headless mode is different: it lets you run Claude directly from the command line or inside automation, without any back and forth. You give it a prompt and the necessary context such as a git diff or a codebase, and it prints out the result.

This is its real superpower. Because it runs in a fully automated way, you can plug Claude into Continuous Integration, nightly jobs, or even scheduled tasks that run across multiple repositories. It does not just answer questions, it becomes part of your build process.

Some examples of what becomes possible:

• Automated code reviews that catch bugs and highlight missing tests

• Refactoring suggestions written as Markdown design notes

• Security checks for risky patterns in authentication or cryptography

• Documentation updates that stay in sync with code changes

• Release notes generated automatically from commit history

Headless mode does not just answer developers, it works with them, quietly, in the background.

The recipe

I have put together a complete example here: github.com/gbarbalinardo/headless-claude-ci

The repo includes:

• a small Python project with one utility function and a test

• a GitHub Actions workflow that runs both Claude and coverage

• a Python script that turns coverage.xml into a Markdown table

It is deliberately small so you can read it in one sitting, but it shows the full pattern end to end.

The workflow file .github/workflows/ci.yml does the heavy lifting. Here is a simplified view:

On every pull request, it:

• runs the test suite with pytest and collects coverage

• generates a Markdown summary of coverage

• calls Claude in headless mode on the diff

• combines both outputs and posts a single PR comment

jobs:
  review-and-coverage:
    runs-on: ubuntu-latest
    permissions:
      pull-requests: write
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0

      - uses: actions/setup-python@v5
        with:
          python-version: "3.11"

      - run: pip install pytest pytest-cov coverage

      - run: pytest --cov=src --cov-report=xml:coverage.xml

      - run: python scripts/coverage_to_md.py coverage.xml > coverage.md

      - run: npm install -g @anthropic-ai/claude-code

      - name: Run Claude headless review
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
        run: |
          BASE_SHA=$(git merge-base origin/${{ github.base_ref }} HEAD)
          git diff $BASE_SHA...HEAD > pr.diff
          claude -p "You are a precise code reviewer. Summarize pr.diff in Markdown, noting potential bugs, risks, and missing tests. Keep it under 300 words." \
            --allowedTools "Read,Grep,GlobTool" \
            --output-format text \
            < pr.diff > claude_review.md

      - name: Post combined PR comment
        uses: actions/github-script@v7
        with:
          script: |
            const fs = require('fs');
            const review = fs.readFileSync('claude_review.md','utf8');
            const cov = fs.readFileSync('coverage.md','utf8');
            const body = `## Claude review\n\n${review}\n\n---\n\n${cov}`;
            github.rest.issues.createComment({
              owner: context.repo.owner,
              repo: context.repo.repo,
              issue_number: context.issue.number,
              body
            });

The coverage script

The Markdown coverage summary comes from a tiny script in scripts/coverage_to_md.py:

import sys, xml.etree.ElementTree as ET

root = ET.parse(sys.argv[1]).getroot()
total = float(root.attrib.get("line-rate", 0)) * 100.0

rows = []
for cls in root.findall(".//class"):
    fn = cls.attrib.get("filename", "unknown")
    rate = float(cls.attrib.get("line-rate", 0)) * 100.0
    rows.append((rate, fn))

rows.sort()
print("# Test coverage\n")
print(f"Total line coverage: **{total:.1f}%**\n")
print("| File | Coverage |")
print("|---|---|")
for rate, fn in rows[:5]:
    print(f"| {fn} | {rate:.1f}% |")

When the workflow runs, this generates a Markdown table with the overall coverage percentage and the least covered files. That Markdown is then appended to Claude’s review and posted as a single PR comment.

Why it matters

By keeping everything in Markdown, the workflow stays simple. Developers do not need to click away to external sites or parse raw XML. They see coverage, file by file, right where they already work. Claude’s review appears alongside the coverage, giving every pull request the same baseline: are the tests thorough, and are there risks worth flagging?

It does not replace human review, but it sets the table. Just as in cooking, when the preparation is done right, the real work feels lighter and more focused.

Closing thought

Just as cooking feels calmer when the prep work is done, development feels lighter when automation handles the basics. Claude headless brings intelligence into that loop, reviews that never tire, coverage summaries that always appear, and the potential for much more: automated documentation, release notes, security scans, even large scale refactors.

With this setup, your repository feels like a well prepared kitchen: quiet, steady, and always ready for the next dish.

References

• Anthropic Claude Code headless SDK documentation

• Headless Claude CI repository