I hosted a pretty big game night tonight, it was 17 people. I also invited K Along with a couple other new people, and I wanted to see what my friends thought about her because we have been talking for a bit. And she was wonderful as always. I feel like a shitty person because other than some very small things she has been absolutely incredible, and she checks every box I could ask for and I felt a lot of guilt because I didn’t feel that super intense rush of instant connection the way I have in the past with unhealthy relationships. I’m constantly stuck in this speculation loop of wondering if this is properly what healthy love or a start of a relationship should look like, and I feel like today was confirmation that it is true. I feel like I have only seen green flags from her, and she is not perfect, but rather like a realistic unicorn if that makes sense. She has all of the attributes and traits that I had on my list it seems like, and I think she is both kind and emotionally mature. She has a wonderful energy around her, and she kind of makes me feel nervous a little bit in the sense of I want to be my best to give her the best impression possible. I’m gonna talk with my therapist about this and I think I want to ask her out.

In the last years of his life, Kurt Gödel starved himself to death. Convinced that someone was poisoning his food, he would eat only what his wife Adele had tasted first. When she was hospitalised after a stroke in late 1977, he stopped eating altogether. He died in Princeton Hospital on January 14th 1978, weighing 29 kilograms. The death certificate read “malnutrition and wasting from neglect caused by personality disturbance.” The greatest logician since Aristotle, a man who had proved that mathematics itself contained truths it could never reach, was killed by a distorted inner logic he could not escape.

Almost nobody outside mathematics knows his name. Einstein did. The two were faculty at Princeton's Institute for Advanced Study from the 1940s onward, and Einstein, by then ageing and isolated from the mainstream of physics, told colleagues he went to his office “just to have the privilege of walking home with Kurt Gödel.” They made an odd pair on the Princeton sidewalks, Einstein rumpled and laughing, Gödel dapper in a white linen suit, talking animatedly in German on their daily walk to the Institute and back. John von Neumann, who cancelled an entire lecture series on David Hilbert's programme after reading Gödel's 1931 paper, called his work “singular and monumental, a landmark which will remain visible far in space and time.”

So what did Gödel prove, and why does it matter now, in the middle of an AI boom that is spending trillions of dollars, much of it resting on the assumption that intelligence is a scaling problem?

What incompleteness means

Put simply, Gödel proved that mathematics cannot fully explain itself. The longer version requires a little patience. In 1900, the German mathematician David Hilbert challenged the field to build what amounted to a perfect machine for mathematics. Start with a set of basic rules (called axioms), things so obviously true they need no argument, and then derive every mathematical truth from those rules, step by mechanical step. If you could do that, mathematics would be complete, meaning every true statement is provable, consistent, and free of contradictions. You could hand the whole enterprise to a clerk following instructions. This was called Hilbert's programme, and for three decades it was the organising ambition of the field. Then Gödel demolished it in one stroke in 1931, at the age of 25.

Gödel’s first incompleteness theorem proved that any set of rules powerful enough to handle basic arithmetic will contain true statements it cannot prove. Not because the rules were poorly chosen, but as a structural feature of rule-based systems themselves.

His trick was to construct a mathematical sentence that refers to itself. Consider the sentence, “This sentence has no proof.” Gödel's technical feat, the part that fills his 1931 paper, was building this sentence out of pure arithmetic, by encoding statements about numbers as numbers themselves. It is not English smuggled into maths. It is pure maths. There are only two possibilities. Either the system can prove it, or it cannot.

If the system can prove “This sentence has no proof.“. there is an immediate problem. We have just proved a sentence that says it has no proof. A system that proves false things is a system with contradictions, and contradictions in mathematics are fatal. Because once you allow a single one, you can use it to prove anything, including that 1 equals 2. The system becomes useless.

If the system cannot prove “This sentence has no proof.“, there is a different problem. The sentence said it had no proof, and it turns out to be right. It is a true statement. But the system has no way to prove it. So we have a truth the system cannot reach, which means Hilbert's rulebook has a blind spot.

Any sensible mathematical system would rather have blind spots than contradictions. So the sentence (logicians call it a Gödel sentence) is true but unprovable, and Hilbert's dream of a rulebook that can prove every true thing was dead.

Gödel’s second theorem twisted the knife. It showed that no set of mathematical rules can prove that it is free of contradictions, using only its own rules. If you want to check whether your system is trustworthy, you always need a bigger system to do the checking, and that bigger system inherits the same limitation. Turtles all the way down.

This is not mysticism, and not a claim about consciousness or creativity. It is a precise result about rule-based systems, the kind of systems that all software, including AI, is built from. That is what makes it relevant today.

The failed dream that built the computer

Hilbert had asked for one more thing, and Gödel's paper left it wounded rather than dead. Alongside completeness and consistency he wanted decidability, a mechanical method that could take any mathematical statement and settle, in a finite number of steps, whether it follows from the rules. No genius required: crank the handle and read the verdict.

In 1936 a 23-year-old Cambridge fellow named Alan Turing killed that too. To prove that no mechanical method could exist, he first had to pin down what “mechanical method” meant, which nobody had previously done. His answer was an imaginary device, a paper tape and a head that moves along it, reading and writing symbols according to a fixed table of rules. Anything a human clerk could work out by rote, this device could also work out.

Then he showed the device has a blind spot of its own. Imagine a fortune-teller who is never wrong, and a stubborn customer determined to sabotage every forecast. “You will leave by the door.” He climbs out the window. “You will take the window.” He strolls out the door. She is not bad at her job. The job is impossible, because her prediction feeds the very behaviour it is trying to predict.

Turing turned that scene into code. The checker plays the fortune-teller. It is a program whose job is to read any other program the way you might read a recipe, then predict its fate. Either “this one finishes” or “this one grinds on forever.”

The saboteur plays the stubborn customer. It is a short program with a copy of the checker tucked inside, plus one standing rule. Ask the checker what I am predicted to do, then do the opposite. If the prediction is that it finishes, it deliberately loops forever. If the prediction is that it runs forever, it stops dead.

So what does the checker predict for the saboteur? “Finishes” is wrong, because the saboteur hears that and loops. “Grinds on forever” is wrong, because the saboteur hears that and stops.

The saboteur is assembled entirely from the checker's own parts, which is what makes it inevitable rather than a fluke. Build a perfect checker and you have, in the same afternoon, built the plans for the thing that breaks it. A perfect checker is therefore a contradiction in terms.

Programs that predict how other programs will behave most of the time are unremarkable — static analysers and type checkers are used routinely. The program that cannot exist is the one that is never wrong. This is the so-called halting problem, Gödel's self-referential sentence rebuilt from machinery, a machine then forced into a question about itself.

To show what machines cannot do, Turing had to invent the machine. His imaginary device is the theoretical blueprint of the general-purpose computer, a single machine that can run any program you feed it as data. Nine years later John von Neumann, who knew Turing's paper well and admired it, wrote the First Draft of a Report on the EDVAC, which is, in logical terms, Turing's universal machine rendered in vacuum tubes. Essentially every computer built since follows that design. The laptop on your desk and the datacentre GPU training the next frontier model are, once the engineering is stripped away, the same device from a 1936 logic paper.

Gödel himself thought Turing had done him a favour. It was Turing's definition of a mechanical procedure, he wrote, that made a “precise and unquestionably adequate” general version of his own theorems possible. And the machine in the theorem turned out to be the thing every business now runs on, born as a stepping stone in a proof about what it could never do.

The Gödel machine and the guarantee that vanished

Once you have a machine that can run any program, another question is whether it can improve itself autonomously. In 2003, the German computer scientist Jürgen Schmidhuber proposed a thought experiment he called the Gödel machine. It was an AI agent designed to rewrite its own code, with one iron constraint. It would change itself only when it could first prove, with mathematical certainty, that the change would make it better. Not “test and see.” Prove, in the way you prove a theorem, before running the new version. No proof, no rewrite.

Nobody ever built one. To prove that a code change will improve future performance, you need to search through all possible mathematical arguments that could establish that fact. For any interesting problem, the number of candidate proofs is so astronomically large that the search would take longer than any improvement could ever be worth. It is the computational equivalent of insisting on a signed certificate from every possible future before crossing the road. The Gödel machine was provably optimal but completely impractical.

Then in May 2025, the Japanese AI lab Sakana released a system they called the Darwin Gödel Machine. It kept the self-improvement loop but dropped the proof requirement. Instead of proving that a code change would help, the Darwin Gödel Machine proposes changes using a large language model, tests them against SWE-bench (a benchmark scoring whether an AI can fix real bugs in real software), and keeps what works. The name still invokes Gödel, but the mechanism is Darwinian. Natural selection, not formal proof. Fitness measured by benchmark scores, not mathematical certainty.

Judged purely on the scoreboard, it delivered. The system improved its SWE-bench score from 20% to 50% through autonomous self-modification. It developed emergent behaviours like patch validation and error memory that nobody designed.

Schmidhuber's original machine, though, had exactly one property that made it safe by construction, the proof. Every modification was guaranteed to be an improvement before it ran. The Darwin Gödel Machine replaced that guarantee with something weaker: passing the benchmarks. The difference between “provably better” and “scored higher on the benchmark test” is the difference between an aircraft type certified against a spec and one that just hasn’t crashed yet.

This is, compressed into one system's evolution, the trajectory of AI safety. The formal guarantee was too expensive, so the industry replaced it with empirical validation and hoped nobody would notice the difference. “Self-improving” went from a mathematical statement about proof-carrying code to a softer description of an agent that rewrites itself and checks whether the benchmarks improve. Gödel was gone.

Things mathematics cannot learn

Some problems in machine learning are mathematically unanswerable. In 2019, Shai Ben-David and colleagues published a paper in Nature Machine Intelligence under the understated but devastating title “Learnability can be undecidable.” They took a straightforward question, “given this type of problem, can a machine learn to solve it?”, and proved that the deepest rules of mathematics cannot always settle it. The answer is neither yes nor no. It is silence.

The word “learnable” has an exact meaning here. A machine studies a sample and produces a rule, and the rule is then let loose on data it has never seen. That is the basis of pretty much every model we commonly use. For any given type of problem, learning theory asks whether some sample size guarantees the rule will work. If such a guarantee exists, the problem is learnable. If none does, it is not. A simple question with two possible answers. Every type of problem is supposed to get an answer.

Ben-David's team asked it about a mundane task, choosing which adverts to show a website's visitors based on a sample of past ones. Was this learnable or not?

The answer depends, in their framework, on how many different kinds of visitor there could possibly be. That pool is not the eight billion people alive today. The model reduces each visitor to a profile of measurements, and measurements vary enormously. A visitor might linger on a page for three seconds or for a shade over three, and between any two possible profiles there is always room for a third. The pool of possibilities therefore has no end.

That arrangement, a finite sample making predictions about an endless pool, shows up across almost every AI product on the market, not just advertising. Training data is always finite. The world a system is deployed into is not. Ben-David's question is whether that leap can ever carry a guarantee, and everything hangs on how big the infinity is.

That sounds like it must have an answer, but it does not. Georg Cantor proved in the 1870s that infinity comes in sizes. Whole numbers form one infinity. The points on a line form a strictly bigger one, and the proof is surprisingly simple. Try to pair every whole number with a point on the line and Cantor showed you will always miss some, no matter how clever your pairing. Two collections that are both endless, yet one permanently outruns the other. The continuum hypothesis asks a follow-up so obvious it would occur to a child. Is there any size of infinity between those two?

Gödel proved in 1940 that the standard rules of mathematics can never prove the answer is no. Later Paul Cohen proved in 1963, using a technique he invented for the purpose, we can never prove the answer is yes (his proof is dense and we’re already covering a lot of theoretical mathematics).

The upshot is that no cleverer generation is coming to settle this one. The rules of mathematics simply contain no answer. Take every rule of arithmetic and logic we have and follow it as far as it goes, in any direction you like. You will never arrive at yes and you will never arrive at no. The question is open in both directions, permanently.

Now the chain closes. Whether the advertising problem is learnable depends on the size of that infinity. The size of that infinity is a question mathematics cannot answer. So whether the advertising problem is learnable is a question mathematics cannot answer. The strange silence at the very bottom of mathematics travels up the chain and surfaces in a question about showing adverts to shoppers.

Obviously no practical everyday ad campaign hangs on the continuum hypothesis. But the casualty was the promise. Learning theory is supposed to sort every problem into neat buckets: learnable or not, and Ben-David found a problem it can never sort. Not a problem waiting for better mathematicians. A problem where the sorting itself is impossible. The University of Waterloo, Ben-David's institution, described the result as “important and almost troubling.” No budget fixes this one, the way a budget can fix training costs or patchy data. It is a hole in the floor, a place where the mathematical ground itself gives way.

The neural network that exists and cannot be built

A complementary result, published in 2022, is narrower and stranger. Training a neural network is, at bottom, an exercise in trial and error. You show the network examples and measure how wrong its answers are, then nudge its millions of internal settings to make them slightly less wrong. Repeat this billions of times and, often, the network converges on something remarkably good. The Cambridge mathematician Matthew Colbrook and colleagues showed in a 2022 paper in PNAS that this process has a hard boundary nobody expected.

The place where the boundary turned up is medical imaging. An MRI scanner does not take a photograph. It collects measurements, and to keep patients in the incredibly claustrophobic machine for minutes rather than hours it collects far fewer than a complete image needs. Software has to rebuild the full picture from the partial data. Neural networks became the favoured tool for the rebuilding because they do it faster and more sharply than the older mathematical methods.

Then researchers started probing the results and found something unnerving. Nudge the input slightly, with a trace of noise or a small movement by the patient, and the output could change out of all proportion. Sometimes the rebuilt scan came back looking perfect but wrong, showing details that were never in the body. Worse, it is a failure that does not look like a failure. A blurry image warns you. A crisp fabricated one does not.

A demonstration of the capability may indicate nothing is amiss, and not because anyone is cheating. A demo runs the network on typical inputs, the kind it trained on, where it genuinely is good. The failures live in the near-misses, a typical input plus a whisker of noise. Near-misses are endless and a demo can show only a handful of scenarios. The demo is honest, and the danger sits exactly where it cannot look.

The natural presumptive diagnosis is undertraining. Feed it more scans and buy a bigger model, and surely the wobble irons itself out. That hope is what Colbrook's theorem takes off the table. What the paper proved has two halves. First, for certain reconstruction problems, a network that is both accurate and stable exists. Somewhere in the space of all possible settings sits a configuration immune to the wobble. Second, no training procedure can find it. Not the ones we have. Not any. None at all, ever.

The second half is what kills the more-data hope. A training procedure is itself a program, a step-by-step recipe running on the machine Turing described, and the proof covers every recipe there could ever be. More data does not change that. Data is what you feed a recipe, and the theorem is about the recipes. It is like knowing a winning lottery ticket is in a barrel whilst simultaneously holding a proof that no way of drawing from the barrel will ever pull it out. The ticket is real. the searching is futile.

As Colbrook put it, the paradox Turing and Gödel identified has now been “brought forward into the world of AI”, and for certain problems the algorithms needed simply cannot exist.

For the overwhelming majority of real-world problems, training works. But there is no general way to tell in advance which problems will defeat us, and the assumption that enough data and compute will always get us there over the line is, in certain corners of the problem space, provably false.

The machine you cannot contain

The most provocative extension of Gödel's legacy into AI concerns a question that sounds simple. Can we guarantee that a sufficiently powerful AI will not cause harm?

In 2021, Manuel Alfonseca and colleagues published a paper in the Journal of Artificial Intelligence Research arguing that the answer, for a general-purpose superintelligent system, is provably no. Their argument leans on the halting problem, the impossibility Turing established on his way to inventing the computer. You can check specific programs for specific bugs. What you cannot build is the universal checker, the one that works for any program in any situation.

Alfonseca's team showed that asking “will this AI harm humans?” is, mathematically, the same type of question as asking “will this program halt?” Both require predicting the complete future behaviour of a system from its current state. To guarantee a system will never cause harm you would need to trace every possible sequence of actions it could take and confirm none of them is harmful. That is the halting problem in different clothes, and Turing proved that this class of prediction is impossible to guarantee. You cannot build a general-purpose AI safety monitor for the same reason you cannot build a general-purpose program-behaviour predictor. The task is not difficult. It is formally, provably, impossible.

The authors went further, showing that we may not even be able to recognise when a superintelligent system has arrived, because deciding whether a machine is smarter than a human falls into the same class of unanswerable questions.The argument is grounded, not speculative, though it does assume more generality than any AI system possesses today. No system currently available is general enough to face any possible input the way a true Turing machine can.

What it establishes still matters though. Certain safety guarantees are not engineering problems awaiting a sufficiently clever solution. They are mathematical impossibilities, like trying to square the circle or list every real number between 0 and 1. The safety community can build better guardrails and better kill switches. What it cannot build, given the computational framework we share, is a system that certifies another system as unconditionally safe.

What Gödel would recognise

These four threads share a common ancestor in what Gödel proved in 1931 and Turing sharpened in 1936. Rule-based systems cannot fully account for themselves. A system cannot certify its own trustworthiness. A learning framework cannot determine its own boundaries. A safety strategy cannot verify its own completeness.

None of this is softened by the fact that a neural network feels organic rather than rule-like. A model's weights are numbers and its training is arithmetic, all of it running on von Neumann's rendering of Turing's imaginary device. AI is not adjacent to this mathematics. AI is made of it.

The AI industry, understandably, would rather not dwell on this. The commercial logic of scaling requires that intelligence be treated as a problem of sufficient resources, more data, more compute, more parameters, more money. Hard limits are a vibe-killer, and for the overwhelming majority of commercial applications, the limits Gödel identified are irrelevant. Your chatbot will not run into the continuum hypothesis while re-drafting an email.

But whether a self-improving agent can guarantee its improvements are improvements, and whether anyone can prove a system will not cause harm, are questions that sit squarely in the territory Gödel mapped. His inheritance is the precise framework that shows certain guarantees about thinking machines are provably unavailable, which is a different and narrower thing than saying machines can never think.

Einstein's eccentric walking companion saw it before anyone else. Formal systems cannot fully certify themselves. That was a logician's problem in 1931. It became an engineer's problem when Turing turned the proof into a machine. It is now a commercial problem, because the industry betting trillions on those machines is implicitly selling guarantees the mathematics has never supported, guarantees generated by systems that cannot check themselves any more than Gödel’s own warped internal logic could check itself. He died trapped inside it.

Freundinnen & Freunde der Weisheit! Vor rund 2'500 Jahren prägte der Stadtstaat Sparta ein Ideal von Disziplin, Standhaftigkeit und innerer Stärke. Die Erziehung junger Männer im sogenannten Agoge-System zielte darauf, physisch wie mental unerschütterlich zu werden. Viele der Prinzipien, die damals im militärischen Kontext galten, lassen sich heute auf persönliche Entwicklung und psychische Widerstandskraft übertragen. Wer mentale Resilienz aufbauen will, kann von diesen überlieferten Gewohnheiten profitieren.

Spartaner lebten im Hier und Jetzt – ein Fokus, der auch heute hilft, Ängste und Grübeleien zu reduzieren. Sie passten sich ständig verändernden Bedingungen an, statt sich gegen Wandel zu wehren. Disziplinierte Körperertüchtigung und bewusste Konfrontation mit Unbequemem gehörten zu ihrem Alltag – ähnlich wie der heutige psychologische Ansatz des Growth Mindset, der Herausforderungen als Chance begreift. Scheitern betrachteten die Spartaner nicht als Niederlage, sondern als Lektion fürs Leben. Und schliesslich war Selbstdisziplin für sie der Weg zur inneren Freiheit: Wer sich selbst beherrscht, bleibt handlungsfähig, auch unter Druck.

Diese fünf Gewohnheiten – im Moment leben, Veränderung annehmen, Wachstum durch das Verlassen der Komfortzone, Lernen aus Fehlern und Selbstkontrolle – sind nicht nur historisch bemerkenswert, sondern hochaktuell. Sie bieten ein praktikables Gerüst für mentale Stärke im Alltag. Resilienz bedeutet nicht, niemals zu wanken – sondern vorbereitet zu sein, wenn es darauf ankommt. Die Lehren Spartas erinnern daran, dass innere Stärke nicht angeboren, sondern erlernbar ist.

Denkanstoss zum Wochenbeginn

„Krise ist ein produktiver Zustand. Man muss ihm nur den Beigeschmack der Katastrophe nehmen.“ – Max Frisch (1911–1991)

ProductivityPorn-Tipp der Woche: Strukturierte Tagesplanung

Lege Dir einen festen Tagesablauf zurecht. Wenn Du weisst, was wann zu tun ist, vermeidest Du Stress und bist fokussierter.

Aus dem Archiv: Überzeugend argumentieren mit Aristoteles

In einem meiner Führungsseminare stellte jüngst eine Teilnehmerin die Frage: „Warum überzeugen manche Argumente sofort, andere nie?“ Eine einfache, aber tiefgründige Frage – und sie brachte eine lebhafte Diskussion in Gang. Wie gelingt es, andere nicht nur zu informieren, sondern tatsächlich zu überzeugen? Was macht eine Aussage wirkungsvoll? Diese Fragen beschäftigen nicht nur angehende Führungskräfte, sondern sind zentral für jede Form von Kommunikation – ob mündlich oder schriftlich.

weiterlesen …

Vielen Dank, dass Du Dir die Zeit genommen hast, diesen Newsletter zu lesen. Ich hoffe, die Inhalte konnten Dich inspirieren und Dir wertvolle Impulse für Dein (digitales) Leben geben. Bleib neugierig und hinterfrage, was Dir begegnet!

---

EpicMind – Weisheiten für das digitale Leben „EpicMind“ (kurz für „Epicurean Mindset“) ist mein Blog und Newsletter, der sich den Themen Lernen, Produktivität, Selbstmanagement und Technologie widmet – alles gewürzt mit einer Prise Philosophie.

---

Disclaimer Teile dieses Texts wurden mit Deepl Write (Korrektorat und Lektorat) überarbeitet. Für die Recherche in den erwähnten Werken/Quellen und in meinen Notizen wurde NotebookLM von Google verwendet. Das Artikel-Bild wurde mit ChatGPT erstellt und anschliessend nachbearbeitet.

Topic #Newsletter

I love being me, but living with the kind of inner knowing I have can isolate me from the rest of the world. A couple of my girlfriends are experiencing heartbreak right now. Two very similar situations, actually. I won’t go into details, but to me, they seem like the same story.

We are all different. We all experience different things. At the same time, our experiences are astonishingly similar, especially within an energetic collective. I’ve always been a keen observer, and I recognize patterns in people’s healing journeys, including mine.

Friends and clients tell me about the issues they are going through. Their dilemmas with outdated roles in life. Their romantic interests, friends, and family taking them for granted. Their worth not being recognized at work. When they are angry, frustrated, disappointed, and heartbroken, they go out and seek distractions. That’s natural. We want to forget our pain, so we numb ourselves with sex, entertainment, food, and substances. I’ve been there, too.

At the tail end of my numb phase, I finally realized it wasn’t going to help me in the long run. I hated that I saw through my bullshit. I hated that I finally had to face what I was running away from, because that was exactly the cause of all my pain. ALL. Once I reached that point, there was no going back.

It’s difficult for me to hear my girlfriends talk about their heartbreak. What pains me even more is that I see why their male partners acted the way they did, but they don’t see it, even if they really want to understand it. Truth is, nothing is one-sided. In most human relationships, we tend to mirror each other. We attract who we are. But beyond all of my inner knowing, I’m also a human being, a friend who cares. I want to listen and ease their pain as much as I wish they saw what I see clearly.

I often say I am built differently. I mean it in the sense that both my feminine and masculine qualities are powerful. I can resonate with both. If anything, I’ve always felt like I belong to neither and am some hybrid, an original being. Someone universal.

Lately, I’ve been learning about numerology and life path numbers. My life path number 9 is said to embody all the qualities of life path numbers 1 to 8. An embodiment of everything. An old soul. Some say 9 is the final life cycle on earth. I feel like all of this explains my strange place in this world. I am given this gift of discernment because I am such a self-contained individual. All the wisdom and knowledge I need are already within me.

PowerShell-based software deployment commands are increasingly used to distribute applications through short web links, bootstrap scripts, and copy-and-paste installation instructions. In many cases, users are encouraged to open a terminal with elevated privileges and execute a command that downloads and immediately runs remote code. Such procedures are often presented as convenient installation methods, activation tools, patching utilities, or simplified deployment mechanisms. However, when the source is unofficial, the licensing status is unclear, or the script is not independently verified, the command should be treated as untrusted code.

The objective of a responsible investigation is not to use the script to bypass licensing controls or obtain unauthorized access to commercial software. Instead, the purpose is to determine what the link delivers, which systems it contacts, what files it retrieves, which commands it executes, and how it modifies the operating system. These questions are best answered in a controlled laboratory environment that separates the test system from personal data, production networks, and trusted credentials.

A proper analysis combines several disciplines. Static analysis reveals the visible structure of the script before execution. Network analysis identifies external destinations, protocols, data transfers, and communication patterns. Endpoint monitoring records process creation, file system changes, registry activity, scheduled tasks, services, and persistence mechanisms. HTTPS interception may provide access to encrypted application-layer traffic when implemented carefully in a disposable environment. No single tool provides a complete picture, so the most reliable methodology uses several complementary tools and correlates their outputs.

The first principle is that a remote PowerShell command should never be executed directly from an unknown source. Commands that combine download and execution functions are particularly dangerous because the user may never see the code that runs. A common pattern retrieves remote content and immediately passes it to an interpreter. This creates an execution chain in which the operator loses the opportunity to inspect the script, verify its origin, calculate its cryptographic hash, or identify secondary payloads.

Static analysis should therefore be performed before any dynamic test. The remote content should be downloaded as a file without execution and stored in a dedicated analysis directory. A Linux workstation is well suited for this stage because it provides mature command-line tools for retrieving, hashing, searching, decoding, and cataloging suspicious content.

A working directory can be created as follows:

# mkdir -p powershell-link-analysis
# cd powershell-link-analysis

The remote script should then be downloaded to disk. The actual address should be substituted for the placeholder shown below.

# curl --proto '=https' --tlsv1.2 -fsSL https://example.invalid/bootstrap -o bootstrap.ps1

The downloaded file should immediately be assigned a cryptographic identifier. SHA-256 is appropriate for most laboratory workflows because it provides a stable fingerprint that can be included in reports and compared across multiple test sessions.

# sha256sum bootstrap.ps1

The file type, encoding, and general structure should also be examined.

# file bootstrap.ps1

The script can then be reviewed in a pager or text editor.

# less bootstrap.ps1

Static inspection should focus on several behavioral categories. The analyst should identify network functions, process execution, privilege escalation, archive extraction, file writes, registry changes, service manipulation, scheduled tasks, security configuration changes, and the creation of startup entries. The script may also include obfuscation, encoded commands, compressed strings, dynamic function construction, or indirect invocation techniques intended to make the code difficult to understand.

A broad search for commonly used execution and download keywords can accelerate the initial review.

# grep -Eni 'https?://|Invoke-RestMethod|Invoke-WebRequest|DownloadString|Start-Process|Set-Content|Remove-Item|Expand-Archive|schtasks|reg.exe|sc.exe|bitsadmin|certutil|cmd.exe|powershell.exe|pwsh.exe' bootstrap.ps1

Embedded URLs should be extracted and reviewed separately because a small bootstrap script often retrieves additional components from other locations.

# grep -Eo 'https?://[^"'"'"'") ]+' bootstrap.ps1 | sort -u

Every discovered resource should be treated as a separate artifact. A script may contact a code repository, a file-hosting platform, a temporary content delivery domain, or a server controlled by an unknown operator. It may retrieve a batch file, executable, archive, installer, configuration file, or second-stage PowerShell script. Each artifact should be downloaded independently, hashed, and stored without being executed.

The analyst should also check whether the script contains encoded PowerShell. Base64-encoded commands are common in both legitimate administrative automation and malicious activity. Their presence is not proof of harmful intent, but they require further inspection.

# grep -Eni 'EncodedCommand|FromBase64String|ConvertFrom-SecureString|IEX|Invoke-Expression' bootstrap.ps1

If an encoded string is identified, it can be decoded in an offline environment. The analyst should avoid copying unknown output into an active shell. The decoded content should be written to a file and reviewed as text.

Domain and certificate inspection should be performed before execution. DNS queries reveal the current address records and possible aliases associated with the host.

# dig example.invalid A
# dig example.invalid AAAA
# dig example.invalid CNAME

A verbose HTTPS request can reveal redirect chains, server headers, content types, and TLS negotiation details.

# curl -v -o /dev/null https://example.invalid/bootstrap

The remote certificate can be inspected with OpenSSL.

# openssl s_client -connect example.invalid:443 -servername example.invalid </dev/null 2>/dev/null | openssl x509 -noout -subject -issuer -dates -fingerprint -sha256

Certificate information should not be interpreted as proof that the content is trustworthy. HTTPS protects the connection between the client and the server, but it does not guarantee that the server is legitimate or that the delivered code is safe. A valid certificate only confirms that the connection was established to a host controlling the relevant domain.

Static analysis has important limitations. Scripts may generate commands dynamically, download different payloads according to location or operating system characteristics, or remain inactive until they detect administrative privileges. For these reasons, a controlled dynamic analysis is often necessary.

The dynamic test should be performed in a disposable virtual machine. The guest should contain no personal files, saved passwords, browser sessions, authentication tokens, email accounts, or access to organizational resources. Shared folders, clipboard synchronization, drag-and-drop integration, host file mounting, and USB passthrough should be disabled. The guest should be created specifically for the investigation and deleted after the analysis.

The virtual machine should not be bridged directly onto a trusted network. A dedicated NAT network is preferable because it permits controlled outbound access while reducing exposure to other devices. Additional firewall restrictions should prevent the test system from contacting internal address ranges. This is important because an untrusted script may perform network discovery, credential theft, lateral movement, or opportunistic access to nearby services.

On a Linux host using KVM and libvirt, available interfaces can be listed with the following command:

# ip -br link

The interface associated with the virtual machine can be identified as follows:

# virsh domiflist WINDOWS_LAB_VM

Before execution, a packet capture should be started on the Linux host. Assume that the guest address is 192.168.122.50 and that the relevant virtual bridge is virbr0.

# sudo tcpdump -i virbr0 -nn -s 0 -w powershell-test.pcap host 192.168.122.50

This command captures all traffic associated with the guest and writes the result to a PCAP file. The complete packet size is preserved, and automatic name resolution is disabled to reduce ambiguity. The capture should begin before the script is launched so that DNS lookups, TLS handshakes, redirects, and initial downloads are not missed.

After the test, the packet capture should be stopped cleanly.

# sudo pkill -INT tcpdump

The resulting file can be opened in Wireshark.

# wireshark powershell-test.pcap

Wireshark is particularly valuable for examining DNS requests, TCP connections, TLS handshakes, retransmissions, destination addresses, server names, and timing relationships. Display filters help isolate the relevant traffic.

# ip.addr == 192.168.122.50
# dns
# tls
# tcp.port == 443
# tls.handshake.extensions_server_name

These expressions are intended for the Wireshark display filter field. They are shown with a leading number sign to maintain consistent formatting.

The analyst should first identify all DNS queries generated during the test. Unknown or unrelated-looking domains may indicate secondary downloads, analytics services, redirect infrastructure, command-and-control systems, or third-party hosting. The sequence of queries is often as important as the individual names. A bootstrap domain may redirect the guest to a code repository, which may then lead to an archive host or an executable distribution server.

Command-line analysis with tshark can provide a quick summary of the capture.

# tshark -r powershell-test.pcap -q -z endpoints,ip -z conv,tcp

Unique DNS queries can be extracted as follows:

# tshark -r powershell-test.pcap -Y 'dns.flags.response == 0' -T fields -e dns.qry.name | sort -u

TLS server names can also be listed.

# tshark -r powershell-test.pcap -Y 'tls.handshake.extensions_server_name' -T fields -e ip.dst -e tls.handshake.extensions_server_name | sort -u

These results provide a high-level map of the infrastructure contacted by the script. However, they do not necessarily reveal the exact content transferred over encrypted HTTPS sessions.

Zeek provides an additional analytical layer by converting packet captures into structured logs. While Wireshark is optimized for packet-level inspection, Zeek is designed for behavioral summarization. It can generate logs for connections, DNS, HTTP, TLS, files, and other protocols.

A Zeek analysis directory can be created with the following commands:

# mkdir -p zeek-results
# cd zeek-results
# zeek -r ../powershell-test.pcap
# ls -lah

Connection data can then be reviewed.

# zeek-cut id.orig_h id.resp_h id.resp_p service duration orig_bytes resp_bytes < conn.log

DNS queries can be extracted from the corresponding log.

# zeek-cut query < dns.log | sort -u

TLS server names may be available in ssl.log or tls.log, depending on the installed version.

# zeek-cut server_name < ssl.log | sort -u

The value of Zeek lies in its ability to transform thousands of packets into concise, searchable records. It allows the investigator to determine which remote systems received the most data, which connections lasted longest, and whether the guest communicated with destinations that were not visible in the original script.

Encrypted HTTPS traffic presents a separate challenge. A standard packet capture usually reveals destination metadata but not the full request path, response body, or downloaded code. In a controlled laboratory, an interception proxy such as mitmproxy can be used to inspect plaintext HTTP transactions.

Mitmproxy can be launched on the Linux host as follows:

# mitmweb --listen-host 0.0.0.0 --listen-port 8080

The Windows guest should then be configured to use the Linux host as its HTTP and HTTPS proxy. The mitmproxy certificate authority should be installed only inside the disposable guest. It should never be trusted by a production machine because doing so would allow the proxy to decrypt protected traffic.

When interception is successful, the analyst can inspect full URLs, request headers, redirect chains, response types, scripts, archives, and executable content. The proxy can also save responses for later analysis. Some applications may ignore the system proxy, use certificate pinning, or establish connections through mechanisms that are not intercepted. For this reason, mitmproxy should be viewed as an optional enhancement rather than the sole source of evidence.

Network monitoring alone cannot reveal the full effect of a PowerShell command. Endpoint monitoring inside the guest is required to document system modifications. Process Monitor can record file system operations, registry access, process creation, and thread activity. Filters should focus on the PowerShell process and any children it creates.

Relevant process names may include the following:

# powershell.exe
# pwsh.exe
# cmd.exe
# cscript.exe
# wscript.exe
# msiexec.exe
# rundll32.exe
# reg.exe
# schtasks.exe

The analyst should pay particular attention to temporary directories, startup folders, system directories, scheduled tasks, services, user profile locations, security exclusions, and configuration areas used for application licensing or activation. A script may delete its temporary files after execution, so endpoint monitoring should be active before the command is launched.

Sysmon can provide persistent event records for process creation, network activity, file creation, registry modifications, and executable loading. The command-line arguments of child processes are especially important because they may reveal actions that are not visible in the original PowerShell script. PowerShell Script Block Logging should also be enabled when possible. It can record code after parsing and may expose dynamically generated or partially obfuscated commands.

All evidence should be preserved in a structured case directory. The original link, retrieval time, script hash, secondary file hashes, packet captures, Zeek logs, proxy exports, endpoint logs, screenshots, and virtual machine configuration should be recorded. Test conditions should also be documented, including whether administrative privileges were used, whether HTTPS interception was enabled, and which network restrictions were applied.

The final analysis should distinguish observed facts from interpretation. For example, a report may state that the script created a scheduled task, downloaded an executable, or modified a registry key. These are direct observations. The conclusion that a particular action represents persistence, evasion, or licensing circumvention is an analytical interpretation and should be supported by the collected evidence.

The safest methodological conclusion is that PowerShell installation links should be treated as executable software rather than as ordinary web addresses. Their apparent simplicity hides a potentially complex chain of downloads, redirects, process launches, and system modifications. A reliable investigation therefore requires a layered laboratory approach that combines static script review, packet capture, structured network analysis, controlled HTTPS inspection, and endpoint telemetry.

Wireshark is an essential component of this workflow, but it is not sufficient by itself. Tcpdump provides dependable capture, tshark supports automated extraction, Zeek produces behavioral logs, mitmproxy reveals application-layer content when interception is possible, and Windows monitoring tools record local changes. When these sources are correlated, the analyst can reconstruct the full execution chain and determine whether the link behaves as advertised, installs unauthorized software, introduces persistence, weakens security controls, or performs additional undisclosed actions.

The central principle is containment. Untrusted PowerShell commands should never be tested on personal computers, production systems, or networks containing valuable data. They should be examined only in isolated, disposable environments designed to preserve evidence and limit harm. This approach supports technical understanding without facilitating unauthorized software use and provides a defensible foundation for cybersecurity research, incident analysis, and user-awareness training.

The Q are one of my favorite Star Trek civilizations—specifically, as they're presented in Voyager. Don't get me wrong, I love the Q episodes in Next Generation and DS9, too, but in my opinion Voyager is where they get really interesting. For any non-Trekkies out there, the Q are god-like extra-dimensional beings capable of altering reality on a whim. They are immortal and have seemingly complete control over time and space. Despite this extensive power, they take an interest in humans during Picard's term at the helm of the Enterprise—or one Q in particular, played by John de Lancie, who pops up now and then to pester his pet humans. We learn a bit about the culture of the Q Continuum in the course of these interactions, but the focus mostly stays on the impact Q's actions have on the Enterprise, and is less about the Q themselves. 

In Voyager, we learn more about why an all-powerful being like the Q would take an interest in puny humans in the first place. In the episode “Death Wish”, a renegade Q comes to Captain Janeway seeking asylum. He wants to leave the Continuum and become human so that he can die. Being immortal, he feels, has made the Q go stagnant. The viewer gets a tour of the Continuum rendered in a format our inferior human minds can understand: a highway through the desert, next to which stands a run-down house where bored-looking people are lounging around, some reading, some playing games. No one is talking; the Q seeking asylum explains that they've run out of things to talk about. Everything has already been said, and all the Q have already seen all there is to see in the universe. The Q feels his life has become futile. He finds it intolerable to keep on living, but he can't stop. 

This episode is, for me, one of the best depictions of immortal beings in fiction because of how it recognizes the full implications of immortality. I've seen or read other stories that do this well, too, like the movie “He Never Died” or the novel The Minotaur Takes a Cigarette Break. In both of those, the protagonist is a mythological figure that has persisted into the present day, and living for thousands of years has taken its toll. Their brains are wired for a different time; they find themselves out of sync with modern society, living at its fringes. Both characters live alone and rarely socialize. Theirs isn't the opulent immortality of the Greek gods or the elegant long lives of Tolkien's elves. 

Granted, none of us really know what it would be like to be immortal, so I suppose it's not entirely fair to criticize any depiction of immortality as inauthentic. Maybe the better critique is to say this isn't the most interesting way a writer can treat it. If you have a character who's lived for hundreds or thousands of years, you're missing an opportunity to give them a worldview that's very different from the average reader's.

My current work in progress involves a lot of very long-lived characters, so the question of how to capture this kind of extensive lifespan on the page has been front and center in my mind of late. These are some of the main things I’ve been thinking about as I’ve been working to build these characters in a way that will read as real and interesting on the page.

1) They have a different relationship with time.

Mortality gives any life an inherent sense of urgency. Time becomes a limited resource that can't be wasted. But the more time a being expects to have, the less true that will be. And if time isn't quite as precious, that will have a few likely impacts on their behavior. They're likely to be more patient, less reckless, and more thoughtful before they act. Think of long-lived cultures from fiction, like the Ents in Lord of the Rings or the Ogier in Wheel of Time. 

Long-lived characters are also likely to take more of a long-view. The math of short-term gains versus long-term consequences balances out differently for someone who's still going to be around in a few centuries. Immortal characters can absolutely still do things that are self-serving, but they're much less likely to make decisions that would sacrifice the long-term stability or health of their environment. After all, they're going to need to keep living here for a very, very long time. 

For characters like vampires, that start out mortal then gain immortality later on, their relationship to time is likely to shift the longer they're alive. A vampire that was just turned 20 years ago might still have that inherent urgency of a human, but if your character has been around since ancient Egypt, their outlook is likely to fall more on the Ent side of the spectrum. This can be useful if you have a cast of immortals with a broad age range, helping you to give them traits that differentiate each other. 

2) They've experienced a lot of loss.

This is especially true for any immortal beings that live among mortals. Any mortal that they form a friendship—or more—with, they do so knowing that they're going to lose them. And this isn't just true of relationships. They've likely seen landmarks crumble and civlizations topple. The town where they were born might no longer exist, or maybe now it goes by a different name; depending on how long it's been, their culture of birth could have been completely forgotten. 

Some of this is likely to be true even if your characters live in a society where everyone is very long-lived. Loss doesn't always need to mean death, for one thing. Even in a world where everybody is immortal, the relationships between those individuals can shift over time just like they do for people. A certain number of tragedies are going to accrue in the course of a long life, however charmed. Think of the narratives around the various Greek gods. Yes, they are immortal and powerful, but they still experience grief, betrayal, and heartbreak. 

This is naturally going to influence how an immortal character interacts with their world. They may be less willing to form friendships with mortals, for example, to spare themselves the inevitable pain of losing another person they've grown close to. Or maybe they've developed a thick shell in general as a defense mechanism, or are particularly skeptical or pessimistic in their outlook. They might also become less inclined to take risks, an effect that can stack with their shifted relationship to time mentioned above to make them very cautious characters that are difficult to put into motion toward a goal.

3) They're from a different time—but they're not necessarily stuck there.

People in general develop a lot of their habits, preferences, tastes, and other core aspects of their worldview and identity during their formative years. For a character that's hundreds or thousands of years old, that formative period happened during a very different time setting than the present of your story. Whatever world your story is set in, it's very likely that a lot of things have changed in the interrim. Think about how many cultural changes someone who was born in 1800 would've seen between their childhood and the present, and that's just a couple hundred years ago. Music, fashion, food, transportation, communication—just about every aspect of the character's everyday life will have changed, likely many times over, in the course of their life. 

Different immortal characters are going to deal with this in different ways, of course. But, for most, there are going to be some ways they've adapted, and some old-fashioned habits or tastes that are going to stick around as core tenets of their identity even through all the changes around them. Part of this comes down to practicality. If your character lives intermingled with a broader society, they'd want to keep up with the times enough to live their life, which probably means updating their wardrobe every few decades and staying on top of technological developments. But they might not bother staying with the times in other aspects of the culture—maybe they still listen to the same music that they've used to relax for hundreds of years, for instance, and don't bother listening to any of that new-fangled stuff. 

Now, if you're writing about a society of long-lived beings that are cut off from other cultures, then that culture is likely to stay more consistent and evolve less over time. In this case, their old-fashioned-ness would likely be more immediately visible once they're brought out of that isolation. I mentioned the Ogier from the Wheel of Time before, and they're a good example here, too. They live in self-contained communities and have rarely engaged with humans for decades, when the story starts. Because of that, they still hold to old traditions that other, faster-moving cultures have abandoned, and a lot of their dress and speech seems old-fashioned to the people who interact with them. 

4) They've kept themselves alive this long for a reason. 

This is less of a factor if you're writing about characters that are both immortal and invulnerable. If it's a being that can't die—like, say, someone that's been cursed to live forever—then there's still a reason that they're still alive, it just doesn't have a whole lot to do with their behavior. 

In most cases, though, you're going to be writing about characters that can die, even if not from natural causes. This is true even of gods in many mythological traditions, and most long-lived fictional creatures can be killed in the right conditions, even if it's not quite as easy as it is for we delicate mortals. Because of this, a certain amount of self-selection happens in their populations. Just because an individual is given the option of immortality doesn't mean that they're going to make the right choices to hang on to it. Individuals inclined to make dangerous enemies, seek out violence, or put themselves in very risky situations are, statistically, less likely to be the ones who are still alive to celebrate their 900th birthday. 

There's a second point to this, too: the character has something that they're continuing to live for. This is the antidote to the kind of existential boredom that is at the heart of that Q storyline I mentioned earlier. Just because a character can live forever doesn't mean they'll always want to. That takes some kind of driving purpose that keeps them pushing through all of the losses they've accumulated, and all of the changes they've had to adapt to.

5) Eternal life doesn't guarantee an infinite memory capacity. 

This is another point where the type of being you're dealing with is going to make a difference. If your immortal beings are Data-style androids, then it is feasible they could maintain an exact record of memories stretching back across hundreds of years—it's just a question of storage capacity, at that point. But for most characters, memories are naturally going to fade with time. Exactly how much of their very long past they can recall with any detail is something that it's smart to figure out from the start when you're working with an immortal character.

Granted, these characters are superhuman by default. Even so, think about your own memories. More recent ones are likely to be the most complete and vivid on average. The ones that you hang on to the strongest from earlier in your life are probably major moments connected to strong emotions. So maybe you have a lot of great memories of your favorite family vacation when you were five, but just have a few spotty memories of first grade. The everyday stuff tends to fade away pretty quickly, and once you get a few decades away from the memory, only the most intense ones are likely to stick around. 

I think this is one of those places where writing a long-lived character that feels real means finding the right balance. On the one hand, it's likely they would have a wide variety of experiences and accrue a lot of skills and knowledge. But unless they're a machine, or an actual god, then they still need to practice skills to keep them, and knowledge can fade over time if it's not accessed. So if you want to have a character read Egyptian hieroglyphics, it's helpful to show the reader that they still use the language, at least occasionally—it's going to strain some readers' suspension of disbelief if the character hasn't seen hieroglyphics for 2,000 years but can still read them fluently.

---

I think if there's one connecting thread to all of these points, it's that a long-lived character still needs to feel fully developed and consistent. Immortality isn't just a superpower. It's the kind of character trait that's likely to dramatically influence the character's entire identity and worldview in ways that go much deeper than how hard they are to kill in a fight. And it's also not only a good thing. There are potential downsides that would come along with a very long life, and those are things you can use to add more complexity and depth to a character that might otherwise end up being too perfect or powerful. 

See similar posts: 

#WritingAdvice #Fiction #Worldbuilding #SciFi #Fantasy

There is a particular kind of promise that technology likes to make, and it goes like this: the thing that was once scarce and expensive will now be abundant and cheap, and so the people who never had it will finally get it, and the world will tilt a little closer to fair. It is a seductive story. It has been told about the printing press, about radio, about the personal computer, about the internet, and now it is being told, loudly and everywhere, about artificial intelligence in the classroom. A patient, infinitely available tutor for every child on earth. The end of the postcode lottery. The democratisation of quality education, finally, at scale.

It is a beautiful idea. The trouble is that the evidence, as it accumulates through the early months of 2026, keeps pointing in the opposite direction. Not subtly, either. A run of peer-reviewed studies and institutional research published between December 2025 and February 2026 tells a remarkably consistent story, and it is not the story on the marketing deck. The story the research tells is that AI in education, deployed the way it is currently being deployed, is far more likely to widen the gap between advantaged and disadvantaged children than to close it. The leveller, on closer inspection, looks a lot like a multiplier.

This matters enormously, and not only because education is the mechanism through which societies decide who gets to be what. It matters because the gap between the promise and the evidence has become a chasm, and into that chasm governments are pouring public money, vendors are pouring product, and children, millions of them, are being enrolled in an experiment whose results nobody has bothered to wait for. The question is no longer whether AI can help some students learn. It plainly can. The question is who it helps, who it leaves behind, and who is supposed to answer for the difference.

The Study Nobody Wanted to Read at the Product Launch

Begin with the most direct rebuttal to the democratisation narrative, because it is admirably blunt. In February 2026, the journal Frontiers in Computer Science published a paper titled, with no particular diplomacy, “AI and the digital divide in education.” Its authors, Mokgata Alleen Matjie, Andani Nethavhani and Mary Matlakala, set out to examine what actually happens when AI tools enter educational systems that contain, as nearly all educational systems do, both well-resourced and under-resourced learners.

Their conclusion is the sort of sentence that does not appear in a vendor's promotional video. AI, they write, “might bring more harm than benefits with its biased algorithms, cultural, and language insensitiveness.” The benefits, they found, concentrate among privileged learners in wealthy regions, which happen to be precisely the regions where the tools were designed in the first place. This is not an accident of distribution. It is a feature of how the technology was built and for whom.

Consider the mechanics, because they are where the unfairness lives. A large language model is, at root, a machine that has learned patterns from an enormous corpus of text. The corpus is overwhelmingly English, overwhelmingly Western, and overwhelmingly produced by and for people who already had reliable internet access and the leisure to write things down. When a child whose first language is Tshivenda, or Tamil, or any of the thousands of languages thinly represented in that corpus, sits down with such a tool, they are not meeting a neutral tutor. They are meeting a system that performs best for someone unlike them. The Frontiers authors put it plainly: when “AI tools are developed in a language unfamiliar to the learners, they are bound to struggle” relative to those whose languages shaped the design.

Then there is the algorithmic bias, which is quieter and arguably nastier, because it hides inside the appearance of personalised help. The study found that students from lower-income communities are “more likely to receive less accurate or less supportive guidance, reinforcing disadvantage.” Sit with that for a moment. The tool sold as a personal tutor delivers a worse tutorial to the children who can least afford a worse one, and it does so invisibly, wrapped in the same friendly interface that serves a richer child something better. There is no obvious moment of denial, no locked door, no sign reading “not for you.” There is just a steady, frictionless drip of slightly inferior help, accumulating over years.

The researchers also surfaced something subtler than infrastructure, drawing on a case study from rural China. The problem there was not only that rural schools lacked devices and bandwidth. It was that rural teachers lacked professional development in digital pedagogy, a gap the authors describe through the framework of technological pedagogical knowledge, a TPACK divide. In other words, even where you hand the hardware to a rural school, you have not handed it the capacity to teach with it well. The kit arrives. The knowledge of how to wield the kit does not.

When the Evidence Comes From the Wrong Schools

If the Frontiers paper is the prosecution's opening statement, the Brookings Institution's January 2026 work is the forensic accountant quietly noting that the evidence everyone keeps citing was collected in suspiciously favourable conditions.

Brookings published two relevant pieces of work in that month. The first, a research review by Mary Burns on what the evidence actually shows about generative AI in tutoring, is genuinely encouraging in places, and it would be dishonest to pretend otherwise. Burns walks through randomised controlled trials in which AI tutoring systems performed core functions traditionally handled by human tutors and produced real learning gains. One trial saw an AI tutor more than double learning gains over a conventional classroom model. Another found a language-model assistant lifting middle-school mathematics achievement, with the largest benefit accruing to novice human tutors who used it as support. This is not nothing. The promise is not pure vapour.

But Burns is careful, and her care is the point. She acknowledges that “many claims about the educational benefits of generative AI have outpaced high-quality evidence,” and she stresses, repeatedly, that design matters, that the gains depend on pedagogically sound implementation rather than on the mere presence of the tool. Read the studies she cites and a pattern emerges that should give any policymaker pause. The trials that produced the good headlines were largely run in well-funded, technologically equipped settings, in environments built and staffed and connected in ways that bear almost no resemblance to the conditions in which the majority of the world's children are actually educated. The evidence base, in short, is drawn disproportionately from the kind of school that least needs the help.

This is the methodological version of testing a flood barrier exclusively on dry land and pronouncing it excellent. The places where AI tutoring has been shown to work are the places already rich in the things, reliable connectivity, trained staff, functioning devices, ambient digital literacy, that make almost any educational intervention work. To take those results and generalise them to an under-resourced school in the global majority is not science. It is wishful extrapolation dressed in the borrowed authority of a randomised trial.

The second Brookings output that month made the institution's broader judgement unambiguous. “A New Direction for Students in an AI World: Prosper, Prepare, Protect,” authored by Mary Burns, Rebecca Winthrop, Natasha Luther, Emma Venetis and Rida Karim, drew on a yearlong global study spanning more than 500 stakeholders across 50 countries and a review of over 400 academic studies. Its central finding lands like a cold flannel on the foreheads of the more excitable advocates. “At this point in its trajectory,” the report concludes, “the risks of utilizing generative AI in children's education overshadow its benefits.”

The asymmetry the Brookings team identifies is the crux of the whole problem, and it deserves to be spelled out. The risks of AI in education, they argue, tend to undermine foundational child development directly, regardless of how carefully you deploy. The benefits, by contrast, are conditional. They only materialise when the deployment is good, when the pedagogy is sound, when the surrounding system is competent. Poor deployment does not merely fail to deliver benefits. It can actively prevent positive outcomes from materialising at all, and it can inflict harm that lands hardest on the children with the least capacity to absorb it. The report's framework, the three pillars of prosper, prepare and protect, is in essence an argument that you cannot bolt AI onto a fragile system and expect anything other than amplified fragility.

A Village in Rajasthan and the Anatomy of a Barrier

Abstractions about distribution and asymmetry are easy to nod along to and easy to forget. So consider a more granular picture, the one assembled in a study of large language models in K-12 education in rural India, the work of Harshita Goyal, Garima Garg, Prisha Mordia, Veena Ramachandran, Dhruv Kumar and Jagat Sesh Challa at the Birla Institute of Technology and Science, Pilani. The peer-reviewed examination of LLM use among rural Indian students that surfaced in the December 2025 research conversation makes the gap between promise and practice almost tactile.

The researchers conducted semi-structured interviews with 23 student volunteer teachers working in rural schools across Rajasthan and Delhi, young educators with an average age of 22.5 and around three years of teaching behind them, people close enough to the chalkface to know what is actually happening in the room. What they reported is a catalogue of the barriers that the democratisation narrative tends to wave away.

Start with the internet, that quiet precondition for everything else. Fifteen of the 23 volunteers identified inadequate infrastructure as a serious obstacle, describing connectivity as a huge issue. As one put it, most schools simply do not have reliable internet and tech access is limited. An AI tutor is a remarkable thing when it loads. It is a blank screen and a wasted lesson when it does not.

Then teacher training, or rather its absence. Thirteen of the 23 flagged the lack of access to any AI training. One volunteer offered an observation that should be printed and pinned above every ministerial desk currently authorising a national rollout: “I don't think most government school teachers are aware of GenAI yet.” You cannot deploy your way around that sentence. A tool that requires pedagogical skill to use well, handed to a workforce that has had no opportunity to acquire that skill, does not become a tutor. It becomes, at best, a distraction, and at worst a substitute for teaching that the system was already struggling to provide.

Language surfaced again, exactly as the Frontiers authors predicted it would. Participants described English-language design as a real barrier, noting that students struggle with English while the very textbooks, the NCERT materials at the centre of Indian schooling, are themselves in English. A tool optimised for English-speaking users meets a child wrestling with English as a second or third language, and the gap between the tool's confident fluency and the child's actual comprehension becomes one more place to fall. It is worth dwelling on how this compounds. A child who already finds the textbook a linguistic obstacle is now handed a digital tutor that speaks the same foreign language even more fluently and even more confidently, and the apparent authority of the machine can paper over a comprehension gap rather than close it. The tool sounds certain. The child nods along. Nobody, least of all an overstretched teacher with thirty other pupils in front of them, is positioned to notice that the understanding never actually arrived. A poorer child in a wealthier child's classroom would at least share the same baseline of language and infrastructure. Here the deficits stack: weaker connectivity, weaker device access, a language barrier and an untrained teacher, each multiplying the others rather than merely adding to them.

And affordability, the most stubborn barrier of all. Eleven of the 23 cited cost. Many families, they reported, cannot afford consistent data or even a single device per child. The personalised tutor is not personalised if four siblings are sharing one cracked phone on a patchy connection in the hour before the battery dies.

The volunteers were not technophobes. They saw the potential, the capacity for personalisation, the possibility of lightening an overstretched teacher's load. But they were clear, eight of them explicitly so, that AI should be a complementary tool and not a replacement for teaching, and they worried about students becoming over-reliant on it at the expense of learning to think for themselves. The practical benefit they observed, in other words, fell far short of the promotional claims. The brochure described a revolution. The classroom described a series of obstacles, each of which fell hardest on the children who already had the least.

The Pace Problem

You might reasonably expect that a body of evidence this consistent, arriving this quickly, would induce caution in the people responsible for spending public money on AI in schools. You would be disappointed.

The New York Times reported in January 2026 that governments rolling out AI tools across their school systems were doing so faster than the research on educational impact warranted. Some experts quoted in that reporting issued a warning that ought to function as an emergency brake: poorly deployed AI could actively harm learning outcomes, and it would do the most harm to the students least able to absorb it. That is not a caveat. That is a fire alarm.

The pattern the Times identified is corroborated from other quarters. The Center for Democracy and Technology has warned that the momentum to deploy AI in K-12 schools is outpacing the guardrails needed to protect students. Survey work from the RAND Corporation has found AI becoming a default study tool across K-12 education, frequently without school guidance or parental knowledge, even as a growing majority of students themselves believe it harms their critical thinking. When the kids using the tool are more worried about its effects than the officials deploying it, something has gone wrong with the chain of accountability.

The temptation here is enormous, and it is worth naming honestly because it is not stupid. Education systems are chronically short of teachers, chronically short of money, and chronically short of time. A technology that promises a tutor for every child, at marginal cost approaching zero, is exactly the kind of miracle a finance ministry dreams about. The democratisation narrative is not merely marketing. It is also a genuine hope, held sincerely by people trying to solve real and painful problems. That is precisely what makes it dangerous. The most seductive false promises are the ones that answer a real need.

But hope is not a deployment strategy, and the speed of these rollouts has a specific, corrosive logic. The research takes years. The procurement cycle takes months. The political incentive to announce a bold modernising initiative takes about a week. So the announcements race ahead of the evidence, the contracts get signed, the tools get pushed into classrooms, and by the time anyone has rigorously measured the effect on actual learning in actual under-resourced schools, the next initiative is already being announced. The evidence, when it finally arrives, lands in a world that has stopped waiting for it.

Why a Leveller Becomes a Multiplier

It is worth being precise about the mechanism, because “AI widens inequality” can sound like a vague incantation if you do not show the gears turning. The reason a tool can be sold as a leveller and behave as a multiplier is not mysterious. It is structural, and it repeats across every study cited here.

A new educational resource is never absorbed into a vacuum. It is absorbed into an existing system, and that system already has a distribution of advantage. The well-resourced school receives the AI tutor along with the reliable fibre connection, the device for every pupil, the teacher who has been trained to integrate the tool into a coherent lesson, the parents who can troubleshoot at home, and the ambient digital fluency that makes the whole thing feel natural. In that environment, the tool does roughly what the brochure said. It personalises, it supplements, it extends a good teacher's reach. The Brookings trials caught exactly this, and there is no reason to doubt them.

The under-resourced school receives the same tool and almost none of the surrounding infrastructure. The connection drops. The device is shared or absent. The teacher, through no fault of their own, has had no training. The tool, designed in a distant language for a distant context, performs worse for these particular children, and it does so quietly. So the same technology, dropped into two different systems, does not equalise them. It tracks the inequality that was already there and, because the affluent system can extract far more value from it, it widens the distance between the two. This is the TPACK divide from the China case study, the language barrier from rural Rajasthan, and the algorithmic bias from the Frontiers paper, all describing the same underlying physics from different angles.

There is a grim elegance to it. You do not need anyone to act in bad faith. You do not need a conspiracy to disadvantage the poor. You need only to distribute an unequally usable tool across an already unequal landscape and let the existing gradient do the work. The democratisation narrative assumes the tool is the great equaliser. The evidence shows the tool is a faithful amplifier of whatever it is plugged into, and what it is plugged into is not equal. The history of educational technology is, depressingly, a series of variations on this theme. Each new medium arrives wrapped in the language of access and arrives in practice as an accelerant of existing advantage, because the families and schools best placed to exploit it exploit it first and hardest. What is genuinely new about AI is the quietness of the failure. A missing textbook is visible. A broken laptop is visible. A tutor that simply performs a little worse for poorer children, while smiling the same smile and offering the same interface, leaves no mark anyone can point to. The inequality it produces is real but evidence-free at the level of the individual classroom, which is exactly the kind of inequality that is hardest to argue against and easiest to deny.

What Equitable Deployment Would Actually Require

So what would it take to make this technology behave like the leveller it is marketed as, rather than the multiplier the evidence reveals? The studies, read together, sketch an answer, and it is considerably more demanding than buying a licence and issuing a press release.

It would require, first, that the tools be built for the children who most need them rather than retrofitted from tools built for everyone else. The Frontiers authors are explicit on this. They call for multilingual, culturally responsive AI systems and for diverse, representative datasets to mitigate algorithmic bias. This is not a cosmetic localisation, not a translation layer bolted on at the end. It means including the languages and contexts of disadvantaged learners in the design and the training from the beginning, so that the tool performs as well for a child in rural Rajasthan as it does for a child in a wealthy suburb. That is expensive and unglamorous and commercially unattractive, which is precisely why it does not happen by default.

It would require, second, the pedagogical infrastructure without which the tools are inert or harmful. The rural India study and the China case study both hammer the same nail. Hardware is necessary and nowhere near sufficient. Teachers need genuine training, not a one-hour webinar, in how to integrate these tools into sound teaching. The Brookings prepare pillar is built entirely around this idea of capacity-building for educators and students, and it is the pillar most often skipped, because building human capacity is slow and undramatic in a way that announcing a technology partnership is not.

It would require, third, that deployment proceed at the speed of evidence rather than the speed of procurement. This is the direct rebuke to the pattern the New York Times documented. It means running the trials in the conditions that actually prevail in under-resourced schools before scaling, rather than generalising from results obtained in rich ones. It means the willingness to conclude, as Brookings did, that at this moment the risks may overshadow the benefits, and to act on that conclusion rather than to bury it beneath an announcement.

And it would require closing the equity gap deliberately, with money and design and political will, rather than hoping the technology will close it as a happy side effect. The Brookings report calls for innovative financing to close equity gaps and for tools co-created with educators, students, parents and communities. The common thread is intention. Equity does not emerge from the unsupervised diffusion of a clever tool. It has to be engineered, paid for, and protected against the gradient that is always trying to reassert itself.

The Accountability Question Nobody Wants to Own

Which brings us to the hardest part of the question, the part that the technology industry is structurally allergic to and that governments are politically reluctant to grasp. If a tool is sold to the public on a narrative of democratising quality education, and it turns out instead to widen the gap between the advantaged and the disadvantaged, who is accountable?

The honest answer is that, at present, almost nobody is, and that is itself the scandal. Responsibility in this system is diffused to the point of evaporation. The vendor builds a tool and markets its potential, then points out, accurately, that outcomes depend on how schools use it. The government procures the tool and announces the initiative, then points out, accurately, that it relied on the vendor's claims and the apparent weight of the research. The researchers produce the encouraging trials, then point out, accurately and often in the very same papers, that their results came from well-resourced settings and should not be over-generalised. Everyone has a defensible position. Everyone has someone else to point at. And the child in the under-resourced school, who was promised a tutor and received a barrier, has no one to point at at all, because the entire structure has been arranged so that the harm has no author.

This is not acceptable, and the way out of it is not more sophisticated blame-shifting but a clear allocation of duties. Vendors who sell a tool on an equity narrative should be held to that narrative, which means being required to demonstrate that their products do not perform systematically worse for disadvantaged learners, the precise failure the Frontiers study documented. The marketing claim and the measured outcome should be allowed to collide in public. Governments that deploy these tools at public expense are accountable for the conditions of deployment, for the connectivity, the teacher training, the linguistic fit, and for the basic discipline of not scaling faster than the evidence permits. A minister who rolls out a national programme on the back of trials conducted in conditions nothing like their own schools owns the gap between the two, however inconvenient that ownership may be at the next ribbon-cutting. And the research community, to its considerable credit already doing this in the studies discussed here, bears a continuing duty to keep saying loudly that the evidence comes from the wrong schools, and to resist the quiet pressure to let promising results be laundered into universal claims.

The deepest problem is that the democratisation narrative does a specific kind of damage entirely separate from any individual tool. By insisting in advance that AI is a leveller, it pre-emptively absolves everyone of the duty to check whether it is. If the technology is equalising by its very nature, then there is nothing to monitor, no distributional outcome to measure, no accountability to assign. The story does the work that scrutiny ought to do. That is what makes it so much more dangerous than ordinary marketing. It is not merely overselling a product. It is disabling the alarm system that would otherwise tell us the product is making things worse.

Holding Two Truths

None of this is an argument that AI has no place in education, and it would be a betrayal of the evidence to pretend otherwise. The Brookings trials are real. The learning gains in well-designed deployments are real. The rural volunteers in Rajasthan saw real potential, and they were not naive to see it. Used well, with the right languages and the right training and the right humility about pace, these tools can genuinely help children learn. That truth and the harder truth can be held at once, and holding both is the entire discipline the moment demands.

The harder truth is that “used well” is doing enormous and largely unacknowledged work in that sentence. The conditions under which AI helps, reliable infrastructure, trained teachers, culturally and linguistically appropriate design, and a deployment pace governed by evidence, are exactly the conditions that under-resourced schools lack. Without those conditions, the same technology that lifts the advantaged child does little for the disadvantaged one and may quietly set them back, and the net effect across a system is to stretch the distance between them. That is the mechanism every study cited here describes from a slightly different vantage, and it does not stop operating because the marketing insists it should.

The democratisation of education is a goal worth wanting with everything we have. It is precisely because it is so worth wanting that it should not be handed over to a narrative that congratulates itself on the outcome before the outcome has been measured. The evidence from late 2025 and early 2026 is a gift, if anyone in a position of power is willing to receive it as one. It arrived early, while the rollouts are still young and the harms still reversible. It tells us, clearly and in time, that the leveller is behaving like a multiplier, and that whether it goes on doing so is not fixed by the technology but chosen by the people deploying it. The tools will do what we build them to do and put them where we put them. The accountability for that, finally, is ours, and it cannot be coded away.

References

1. Matjie, Mokgata Alleen; Nethavhani, Andani; Matlakala, Mary. “AI and the digital divide in education.” Frontiers in Computer Science, Volume 8, Section: Human-Media Interaction, 5 February 2026. https://www.frontiersin.org/journals/computer-science/articles/10.3389/fcomp.2026.1759027/full

2. Burns, Mary. “What the research shows about generative AI in tutoring.” Brookings Institution, January 2026. https://www.brookings.edu/articles/what-the-research-shows-about-generative-ai-in-tutoring/

3. Burns, Mary; Winthrop, Rebecca; Luther, Natasha; Venetis, Emma; Karim, Rida. “A new direction for students in an AI world: Prosper, prepare, protect.” Center for Universal Education, Brookings Institution, January 2026. https://www.brookings.edu/articles/a-new-direction-for-students-in-an-ai-world-prosper-prepare-protect/ (Full report: https://www.brookings.edu/wp-content/uploads/2026/01/A-New-Direction-for-Students-in-an-AI-World-FULL-REPORT.pdf)

4. Goyal, Harshita; Garg, Garima; Mordia, Prisha; Ramachandran, Veena; Kumar, Dhruv; Challa, Jagat Sesh. “Thematic insights into the impact of large language models on K-12 education in rural India from student volunteers' perspectives.” Scientific Reports, 2025. https://www.nature.com/articles/s41598-025-18047-1 (Preprint: https://arxiv.org/abs/2505.03163)

5. The New York Times. Reporting on government deployment of AI tools across school systems outpacing research on educational impact, January 2026.

6. Center for Democracy & Technology. “Advancing Responsible AI Adoption and Use in K-12 Education: Three Policy Priorities for State Legislation.” Center for Democracy & Technology, 2026. https://cdt.org/insights/advancing-responsible-ai-adoption-and-use-in-k-12-education-three-policy-priorities-for-state-legislation/

7. RAND Corporation. “More Students Use AI for Homework, and More Believe It Harms Critical Thinking: Selected Findings from the American Youth Panel.” Research Report RR-A4742-1, RAND Corporation, 2026. https://www.rand.org/pubs/research_reports/RRA4742-1.html

8. RAND Corporation. “Student Use of AI for Homework Rises as Concerns Grow About Critical Thinking Skills.” RAND Corporation, March 2026. https://www.rand.org/news/press/2026/03/student-use-of-ai-for-homework-rises-as-concerns-grow.html

9. Center for Democracy & Technology. “Hand in Hand: Schools' Embrace of AI Connected to Increased Risks to Students.” Center for Democracy & Technology, 2026. https://cdt.org/insights/hand-in-hand-schools-embrace-of-ai-connected-to-increased-risks-to-students/

10. Education Week. “Students Are Worried That AI Will Hurt Their Critical Thinking Skills.” Education Week, March 2026. https://www.edweek.org/technology/students-are-worried-that-ai-will-hurt-their-critical-thinking-skills/2026/03

11. National Education Policy Center. “Cautionary Brookings Report Attempts to Weigh Opportunities and Risks of Generative AI in Education.” National Education Policy Center, March 2026. https://nepc.colorado.edu/publication-announcement/2026/03/generative-ai

---

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

In matter, the world is love.

In the absence of matter, nothingness, too, is entirely love.

How both entwine themselves in you is where we meet.

Our Father Who art in Heaven Hallowed be Thy name Thy Kingdom come Thy will be done on Earth as it is in Heaven Give us this day our daily Bread And forgive us our trespasses As we forgive those who trespass against us And lead us not into temptation But deliver us from evil

Amen

Jesus is Lord! Come Lord Jesus!

Come Lord Jesus! Christ is Lord!

The apiary be Scottish run to mute Late December this hugger And seeing simply rise What time in Hearst for Will Enough of oak And seeming simpler For five octet and lane And pasture by the law Economy forever- and nines to the Moon Giving ray to God And night shall let us be- the end of war.

Rome

To Hearts by day In Sun and for the notch In covet done Thoughts for incarnation The soul and script Retribution Peter A day for then in rain Tiny then The night that had For Crosses to unfold Every name in Heaven For living Mass to know In Re to you Days for going in The one to bring Rod in daring be And in thought Prayers for distant men For God define No substance of our high Very existence The day is early then And no to matter To each becoming her The thought of why- Inter alia and through To common spikes of ground And in the focus why- of distant Rome To bear on global then And high to rain Watching fold and Eucharist The day is in And faces to keep within A soul to know Across to Paris then Offers to the poor And day at night’s rest Curely seen And up to Heaven loved And from this fountain be A clear of God’s own path To Him be glory Shades of Heaven here And only then suppose Christ within Basilica To duly bound Take forever honour And then the path be Holy The Victory of our Lord To global then Offer to the soul Of standing for detour And justice revealed Every second here To the capital of Rome God in His religion For Water then Torches bearing rain And when re-night to stone The glory of His keep Thank you Jesus, call In tithing for the night Engrained to Holy men For Crosses of Saint Peter Solace then be May And time for early reach In totem and our King The nights renewing here For goal to Rome The rise and thanking near Christ as resurrect Our pond is to this day The baptism See The Eucharist of Rome For all courses new And simply here To offer God The Victory of Rome And as our gift The Pentecost redeem For Holy yew The troubadours unveil And blessed thine For God to reach our Adam And duly be A prayer for every heart In Christ And true unveiling love The distance wept Our fold to Heaven home And mercy by the day Lifting into night The verse of here and poor Victory in the spaces To the Sun and stars Verses in afford And all be right The living, blessed Lord For all our lives And decency respect At atom’s call Christ the Lord is risen.

In Summary: * If I had to describe my Sunday in one word it would be: balanced. And that's a very good thing, especially compared to my yesterday, which I would described as being very unbalanced; uncomfortably so.

An hour and a half of yard work this morning was very productive. I cleared out a strip of overgrown grass and weeds turned jungle that was poking through a neighbor's fence from my side. There is much more clearing out of that jungle that needs to be done, and I intend to do it as weather and my health allow. But for now her fence is clear.

Remember the big winds of a few months ago that brought down some big branches from my front yard tree? I'd moved them to a staging area in the back yard where I proceeded to cut the smaller of them into pieces that would fit in the green organics bin for weekly pickup. The city has one day each year designated for large brush pickup. That day is tomorrow. So today I dragged / carried the really big branches remaining in the back yard staging area to the front yard, positioned for the city collection.

Yes! All that work completed and my sweaty old self showered and cleaned up before Noon.

Bet I'll sleep better tonight! Ha Ha!

Prayers, etc.: * I have a daily prayer regimen I try to follow throughout the day from early morning, as soon as I roll out of bed, until head hits pillow at night.

Health Metrics: * bw= 232.59 lbs. * bp= 138/79 (70)

Exercise: * morning stretches, balance exercises, kegel pelvic floor exercises, half squats, calf raises, wall push-ups, BP breathing exercises, pilates

Diet: * 05:30 – 1 banana, 3 little cookies * 09:30 – 2 little cookies * 10:45 – 1 peanut butter sandwich * 13:20 – scrambled eggs, bacon, sausage, rice * 16:20 – 1 fresh apple * 19:30 – dish of ice cream

Activities, Chores, etc.: * 04:30 – bank accounts activity monitored. * 04:40 – read, write, pray, follow news reports from various sources, surf the socials, nap * 07:45 to 09:15 – yard work, cutting back yard bushes, hauling branches * 10:30 – watching The Matrix until the movie froze up on me. * 13:20 – listening to Texas Rangers pregame show, the game starts at 13:35, I'll listen to the radio call of the game on 105.3 The Fan, DFW's #1 Sports Radio Station. * 16:38 – and the Rangers beat the Astros, 6 to 5. * 18:30 – slowly working through the evening's meditations and prayers.

Chess: * 15:20 – moved in all pending CC games

The bruises got worse. i didnt die, that would’ve been easier to explain. every time i lie down, it feels like someone’s hands close around my neck. not enough to stop me from breathing but enough to remind me they’re there. then the pressure disappears the moment i sit up, but the bruises dont. i looked in the mirror today.. they’re shaped like hands. several of them. as if multiple people decided i shouldnt wake up, yet changed their minds halfway through, maybe i sohuldnt write any of this. but if i end up dead, at least someone can read this and know i wasnt imagining the marks on my neck. or maybe they’ll decide i was. either way, it wont make much of a difference to me. people die every day. you dont stop for every stranger. you dont grieve for every name in a newspaper. death is ordinary until it belongs to someone you know, i know that feeling. my mother died when i was little, maybe thats why i started paying attention to grief. not my own, everyone else’s. sometimes, after a job was done, i’d go to the funeral. No one questioned why i was there. i’d stand in the back, listen to people tell stories, watch them cry, watch them search for someone to blame. They’d curse whoever did it. they’d swear the person responsible would pay for it. I’d hear every word. and then leave, it never changed anything. the dead stayed dead. the people left behind kept searching for answers they were never going to find. I wasn’t there because i cared about the person in the coffin. i was there because i wanted to understand the people standing around it. blaming god. blaming themselves. Blaming others. just to feel better. i wanted to understand why one person’s death could hollow out an entire room, while another became nothing more than tomorrow’s headline. Maybe thats the part ive never understood. I’ve spent years being the last thing some people ever knew. now it keeps coming back to me. strange, enough. it never seems interested in taking me with it.

Whatever, I’ll figure it out.

Sincerely, The man death rejected

My book club is reading Abundance by Ezra Klein and Derek Thompson this month. The core idea fits on a napkin, so before the meeting I decided to actually check it against real data. Here's what I found, and where it fell apart.

The idea, in one picture

Housing, clean energy, cures for disease. The inputs to all three haven't really moved. Money's there. Technology got cheaper, not more expensive. Roughly the same number of people know how to build this stuff as always did.

What changed is the pipe between the inputs and the output.

Over the decades, every time something went wrong, somebody added a valve. A highway almost bulldozed a neighborhood in the '70s, so now there's a checkpoint for that. An environmental study got added in the '90s. By the 2000s you needed a public comment period too, sometimes more than one. None of these were dumb decisions in isolation. Each solved something real. But stack enough of them and the pipe might as well be shut, even though nothing on the input side changed at all.

Klein and Thompson call this chosen scarcity. Anyone who's inherited a legacy codebase already knows the pattern under a different name: unaddressed technical debt. A pile of individually-reasonable shortcuts, left unrefactored for so long that the system's throughput has almost nothing to do with its actual capacity anymore.

I liked the idea. I also didn't fully trust it. So before the meeting, I ran the numbers.

Test one: does the throughput number actually check out?

I compared two U.S. cities that get cited constantly in this debate: Austin, which started clearing its own valves out around 2015, and San Francisco, which mostly didn't.

The gap is not small. Austin permits roughly 18 new homes per 1,000 residents every year. San Francisco permits about 2. Eight times the throughput. Same country. Same everything, really, except the rules.

I ran the actual arithmetic using an elasticity number borrowed from a housing study out of Auckland, New Zealand, where a comparable policy change happened and got carefully measured.

Extrapolated ten years out, Austin's predicted price effect blew past -100 percent, which is obviously impossible. That's useful, actually — it means you can't stretch a small, well-measured experiment out to ten years and still trust the exact number it spits out. Markets saturate. Construction costs and demand put a floor under how far prices can fall, and Austin landed on that curve instead of the dashed line.

San Francisco's model predicted something like an 11 to 19 percent rent decline. Instead, rent there is rising faster than almost anywhere else in the country right now, up nearly 19 percent over the past year. Turns out an AI hiring boom rolled through right when the small supply gains were supposed to show up, and a model that only tracks supply has no way of seeing that coming.

Test two: the case that broke my model completely

Then I checked a pair I expected to tell the same story: Vienna versus London.

Here's where it got weird. Vienna and London build housing at almost the same rate, barely a 20 percent difference. If clearing the pipe were the whole mechanism, their prices should look almost identical.

They don't. Vienna's rent is roughly a third of London's, and has stayed flat for two decades.

The reason has nothing to do with permitting speed. Forty-three percent of Vienna's housing stock is public or nonprofit housing, running alongside the private market instead of depending on it to eventually get fixed. It's the housing-policy version of the strangler-fig pattern: instead of refactoring a legacy system riddled with a decade of shortcuts, you stand up a clean system next to it and let it carry the traffic that matters most. Combined with old rent-control rules on the pre-1945 stock, that parallel system is what actually holds prices down. London never built a second pipe. It just kept adding valves to the one it already had.

This matters more than it looks. The book's whole argument is: clear the pipe, let the existing system move faster. Vienna barely touches that lever. It built a second system that doesn't need the first one fixed to work at all. Both approaches raise total output, but only one of them shows up in the book.

What I'm actually taking into the meeting

So what's the actual takeaway here?

Mostly this: the mechanism is real, but only when you can isolate it cleanly. Austin shows that. So does the documented policy change in Auckland and Minneapolis. Clear the pipe, throughput goes up, and prices come down in a way you can actually measure.

But Vienna is the one that's going to stick with me, because it's proof this isn't the only lever available. Anyone telling you a genuinely messy problem has exactly one fix is skipping a step, even when the fix they're describing is real.

And honestly, the habit I want to keep out of all this has less to do with housing than with how I plan to evaluate the next big claim somebody hands me. Profile before you optimize. Don't assume you already know which function is slow, check. Before a compelling story gets to explain what you're seeing, spend two minutes running rough numbers on it yourself. You won't always get proof out of that. But you'll see fast which parts of the argument are load-bearing and which ones just sound right.

Less “was the book correct” and more “what's the cheapest way to check whether a claim is even the right order of magnitude.” Most of the time, that's enough to tell you which parts to trust.

This is meant to be a place for me to write. Mostly for me to flesh out my own ideas, but also just to keep at writing, which is a perishable skill. Generally, I want to write about civics, tackling questions about what it means to be a good citizen, why you ought to be a good citizen, and how otherwise good people fail to become good citizens. Broadly, I think of good citizens aa people who are committed to maintaining and improving the institutions that make a society worth living in.

There is no fleshed out theory of citizenship motivating any of this, but my musings on the issue are organized around one important idea: That to be a good citizen, you have to understand how the institutions in your society work, and you have to be committed to maintaining and improving them. I think this is actually quite difficult in complex society, with lots of specialization. We live in a society where few people understand at a deep level how things work, and no one understands how every thing works. That makes it hard to be a good citizen. Consequently, I think most Americans are bad citizens. That is not to say that most Americans are bad people, rather, that most Americans really do not understand how the institutions that impact their lives work, and generally do not bother to improve or maintain them. What I'd like to do is highlight this as a central problem facing American life today, to explore why this gets to be the case and to, perhaps, offer suggestions for improvement.

The idea of citizenship is often associated with voting. We hear the word and think of the political things the average American is supposed to do on election day. Vote, be informed about politicians, form justifiable preferences for different policies, etc. While those things are a part of civic life in the US, there is so much more about civics than just politics. How you integrate with your local community, how groups of people collectively develop and maintain a sense of belonging or identity, how individuals relate themselves to larger groups, in terms of rights and responsibilities.

These things are all part of civic life, but often overlooked in discussions of citizenship. Not littering in you local community is an act of citizenship, just as much as voting. Ensuring the collective decisions made in your community don't harm future generations is an act of citizenship, even if those decisions seem small and very local. Thinking of local governments as something you participate in and are a part of, rather than a discrete organization that provides you services is an act of citizenship.

All of these ideas tie individuals to institutions. By institution, I just mean major features of organization in social life. The legal system, financial system, the family, technology, and the education system are all examples of institutions. So a better way of thinking about civics is to say that civics is the knowledge of how institutions are supposed to work. That is, civics is the information a person needs in order to be a good citizen. Civic knowledge includes all the information an informed and responsible citizen would need in order to understand how the institutions that impact their life are supposed to work, and what they need to do to maintain them. I say supposed to work because it is not always the case that institutions work how they are supposed to. Sometimes institutions are dysfunctional, sometimes they are harnessed by bad actors to do things they should not. But if you don't know how they are supposed to work, you will not be able to recognize when they are broken.

Failing to understand how an institution is supposed to work will lead to different forms of bad citizenship, the two most common of which are jersey voters and lazy cynics. Jersey voters vote because it is their team. They don't pay attention to anything other than the identity of a politician and support or oppose them on the basis of their identity. This leads to tolerance for corruption and incompetence, as well as a willingness to sacrifice principles. Lazy cynics substitute knowledge for cynicism and generally don't participate in the maintenance of institutions at all because they take the view that to do so would be useless. I want to explore both of these forms of bad citizenship in later posts, but for now all I want to do is point out that if you don't know how things are supposed to work, then you are destined to be a bad citizen, even if you vote.

That is why civics is so important. Civics tells you how things work, equipping you with the information you need to recognize problems and fix them when they arise. A strong knowledge in civics prevents you from being passive or cynical. If you don't understand how things are supposed to work, how could you possibly fix them when they break? And rest assured, if you want to be a citizen, you are obligated to fix institutions when they break. That is your job. You are responsible for that.

So all of this implies a minimum amount of information one would need to be a good citizen. Hence, the blog title. One of the things I want to do here is explore what information someone needs to know in order to be a good citizen. In a society with extreme degrees of specialization, it is not possible to be knowledgeable in every subject. Expertise in our society is extremely narrow, one can only be a true expert in a small band of subjects because there is so much to know. So the question naturally arises, if you want to be a good citizen, how much do you need to know? What is the minimum amount of information one needs to understand the institutions in their life, recognize when they go bad, and correct them?

There are, of course different opinions on how things are supposed to work. That is fine, in an open society there will always be debate about what the best form should be for an institution. The fact that institutions work at all is because they are filled with people who have values. Part of developing good civic knowledge is learning the history of these debates and learning how they shaped the nature of institutions. But in order to be a good citizen, you do need to know how institutions actually work right now. If you have no clue how, for example, the law, or the financial system works, you cannot possibly hope to understand the debates around these institutions! That will inevitably lead to mindless cynicism or passive acceptance of corruption— that is, bad citizenship.

Me with my glasses for the last time full time pre-surgery.

I just shared good news about my heart journey, and now I can finally share another medical milestone for 2026: my cataract surgeries are done and my vision is back.

For the last couple of years, my sight kept slipping, no matter how many times I got new prescriptions from the ophthalmologist or optometrist. A new prescription might clear things up for a few months, then the blurriness would come back. I found myself contorting my face and doing all kinds of facial tricks just to see. It was soul-sapping. Anyone who knows me knows I don't get depressed easily. At 57, I've had low moments before, but nothing like this. The vision loss pushed me into a prolonged low spell. I stopped wanting to read and lost enthusiasm for work, except as a way to pay bills. Even blogging suffered. After my most prolific year in 2025, I barely posted because I literally couldn't see well enough.

My glaucoma doctor mostly saw senile cataracts and told me my eye pressures looked fine. Early this year, he said the cataract in my right eye had progressed from a 1 to a 2, which was probably causing the problems. He offered two options: try another prescription or remove the cataracts. I chose surgery.

On June 22, 2026, I had the right cataract removed. The surgeon did a great job, although the anesthesiologist's nerve block caused bruising and left me with a black eye for almost two weeks. Still, that eye improved noticeably. I had the left eye done on July 9, 2026, and it went much better. A different anesthesiologist gave a smooth block, and the same excellent surgeon operated. The day after the left surgery I was seeing 20/20; after the right surgery I was at 20/40. The team also placed iStents in both eyes for my glaucoma. Those seem to be doing fine so far, though time will tell and my glaucoma specialist, who is different from the surgeon, will follow up.

Some expected trade-offs are already showing up. I need readers for near work and some help with intermediate tasks. My distance vision was corrected, and so was my astigmatism. They corrected the right eye's astigmatism with a laser, but the left needed a toric lens upgrade, which I paid for because insurance didn't cover it. My natural, God-given lenses are gone and replaced with implants, basically eyeglasses inside my eyes. It's been life-changing. I didn't realize how dulled colors and whites had become until the right eye was done and whites looked truly white again. Before the left eye was operated on, I could compare the operated and unoperated eyes and everything in the unoperated eye had a yellow tint. Now that both eyes are done, colors are clearer and whites really pop. The left eye is still a bit inflamed since it's only a few days post-op, but distance vision already feels nearly perfect, like the first time I put on glasses at 16.

I didn't expect to need cataract surgery until my 60s or 70s, but I'm relieved it's over and I can see clearly again. Now I have to heal and figure out what kind of readers I'll need and whether I'll need permanent glasses again though for distance at this point and time I don't. I was scared going in. The right-eye surgery felt strange because the nerve block didn't fully take, though I experienced no pain. The left-eye surgery was much calmer and less noticeable.

Being only 57 and needing cataract surgery before my parents, who are in their 70s, felt odd. But I simply could not function properly before the surgeries, and I'm very thankful that God gave me a skilled surgeon and the strength to get it done. I'm glad to be able to see clearly again.