Writing

What agentic payments are. As AI agents begin to spend money on our behalf, payment splits into three problems (trust, commerce, settlement), each claimed by a different camp pushing its own protocol. The revival of HTTP 402 after thirty years asleep, the standards race between Stripe, Google, Visa, and Mastercard, and the question left over once the buyer is no longer a person.

What enshittification means. Cory Doctorow's name for the way good platforms decay over time. The three-stage slide as value shifts from users to business customers to shareholders, why it's a structure rather than the greed of any one company (lock-in and the collapse of the four forces that once held platforms in check), Amazon as the specimen, and where the theory falls short.

What is a one-person unicorn? The phrase began as a bet in Sam Altman's group chat and now names an era of solo founders running entire teams of AI agents. The data that upends the conventional wisdom against going solo, the three functions of a team AI still can't fill — execution, verification, and buffering — and the research showing that people who succeeded alone were never really alone.

What surveillance pricing is. The old practice of charging different people different prices, meeting personal data and AI, begins to close in on an extreme economics drew long ago: perfect price discrimination. From the Instacart experiment and the FTC probe to the short history of the price tag, and why we find this particular unfairness so hard to stomach.

What attention residue is. When you move from one task to the next, part of your mind stays stuck on the last one and quietly drains the focus you need now. A look at Sophie Leroy's research and the 1927 Zeigarnik effect, the finding that what grows the residue isn't finishing the work but the mind's sense of closure, and how to lower that cost in an age where switching happens by the minute.

What Goodhart's Law is. 'When a measure becomes a target, it ceases to be a good measure' — but the person who wrote that line wasn't Goodhart. From NHS emergency rooms and test-cheating scandals to AI benchmark contamination and an age where LLMs grade LLMs, this traces the four ways a metric collapses even when no one is gaming it, and how to keep measuring without turning the measure into a god.

What loop engineering is. It means replacing the version of you that kept prompting an agent with an autonomous loop that does the prompting instead. A first-person account of running this pattern — which broke out in June 2026 — not just in coding but across investing, knowledge curation, and document review, the two places it always collapses (the verifier and the stop), and why human verification never disappears; it only moves.

What an LLM Wiki actually is. Andrej Karpathy floated a personal knowledge-management pattern in a single tweet; here it is, set against RAG, which retrieves fresh every time and keeps nothing. The three-layer structure that lets an LLM grow raw material into a lasting wiki, the compounding it produces, and an honest account of running it myself, limits included.

Building a single source of truth (SSOT) means finally paying down the 'knowledge debt' we've parked in people's heads for years. Why pulling tacit knowledge into the open is so expensive, why it's underprovided and resisted because it behaves like a public good, and how to lower that cost — plus how to design incentives that return the payoff to whoever did the writing.

Writing with an AI agent lets you go faster and reach further — translation, search, accessibility. Its real uses, the objection that writing is thinking (Paul Graham, Ted Chiang, Cory Doctorow), and the supervision it demands: I can only write what I understand. Authorship is not about who typed it, but who takes responsibility.

What a single source of truth (SSOT) really is, and why an old problem — copies of the same fact drifting apart — suddenly turns decisive once AI agents trust information and act on it immediately. A 1991 paper on multi-agent truth maintenance already had the answer; here's why Git became the reference, and why the rebuttal that 'a single truth is a myth' is also right.

Why walk a multi-agent setup back to a single agent? A year of hand-building a personal AI assistant, then moving it off messengers like Telegram and Slack onto a web app I built myself — read through shared context, the cost of supervising agents, lock-in, and knowledge management, with Cognition's 'Don't Build Multi-Agents' and Anthropic's multi-agent research alongside.

What 'zero-click' actually means. As search and AI hand over the answer on the spot and the click disappears, what is left to reward the people who write? A look at the bot toll (pay-per-crawl) that woke a 30-year-dormant HTTP 402 code, the opposite exit of GEO (generative engine optimization), and citation as the new currency a writer can still hold.

Why does a flat-rate subscription break down against an AI agent that runs on its own? Starting from the day Anthropic shelved its plan to split Agent SDK and usage-based credits out of the subscription — on the very day it was due — and reading the structure through average pricing, marginal cost, and unit economics. With the OpenAI price war and an IPO as market variables.

What the FOMC dot plot is and why it shakes interest rates and markets. Where it came from (2012), how to read the median and the longer-run dot, how a single dot threads through to bond yields and your mortgage rate, and the catch that matters most — it's a forecast, not a promise.

What sovereign AI actually means. Starting from the day the US cut off Anthropic's Fable 5 and Mythos 5 to foreigners, a look at the three layers of sovereignty — data, model, compute — where the real bottleneck sits, and how Korea's state-backed 'national champion AI' fits in. How dependence on AI turned into a security question.

Harari's three books are really variations on one idea: humans cooperate by believing shared fictions. A look at how that single key — intersubjective reality — opens money, the state, free will, and AI all at once, the doors it can't open (John Searle's institutional facts, reductionism, moral nihilism), and how to read any thinker by isolating their one big idea.

Loving to make things isn't a hobby — it's a 2.5-million-year-old instinct. How toolmaking grew the human brain (Homo faber), the pleasure of having your will land on the world (effectance motivation), why you love what you built yourself (the IKEA effect), and how a mortal creature pushes back against time by making. The origin of the urge to make, read across three layers: evolution, psychology, and existence.

A team that ran beautifully goes lukewarm at some point — and that's a structural event, not a failure of will. This separates two kinds of aging: the problem space wearing thin through diminishing marginal utility, and people wearing down against self-determination theory (autonomy, competence, relatedness) and Herzberg's two-factor model — with Teresa Amabile's progress principle on top. Different diagnoses call for different cures.

As AI agents start reading and writing documents on our behalf, Notion's block model is being put to the test. Why a markdown-and-git workflow suits agents, why a Notion doc's mutability and lack of verification make it a weak source of truth, and how Notion answers back with MCP, Workers, and an agent hub — plus where the real battle line falls.

See Notion as just a pretty all-in-one document app and you miss half of it. Trace the lineage that runs from Marshall McLuhan's theory of tools through Douglas Engelbart's intelligence augmentation to Alan Kay's Dynabook, and you find what founder Ivan Zhao was really after: a tool for making tools, and the 'software Lego' behind it. With the Kyoto reboot that saved the company — and the criticism that shadows the ambition.

The difference between functional and mission-based organizations, their trade-offs, the real choices Apple, Amazon, and Toss made, and how matrix, squads, Team Topologies, and platform teams have patched the weaknesses of each pure form. Then how AI agents, by collapsing coordination cost, are reshaping the pyramid into a diamond.

The structure of flow as Mihaly Csikszentmihalyi defined it. Optimal experience measured with the Experience Sampling Method (ESM); the conditions that open flow (clear goals, immediate feedback, a balance of challenge and skill); the four-quadrant model (apathy, boredom, anxiety, flow); autotelic experience, where the activity is its own reward; and the paradox of work, the finding that flow comes more often on the job than in leisure.

Setlog hit No. 1 on the App Store on the strength of one move: a two-second video every hour. BeReal ran the same anti-Instagram, radical-simplicity playbook to the top years earlier — and then fell. Using its trajectory as a mirror, a look at the fork every intimate social app reaches after the viral spike: retention, the dilemma of adding features, and the self-contradiction of trying to make money.

Dopamine doesn't deliver pleasure (liking) — it drives pursuit (wanting). Berridge's wanting/liking split, reward prediction error, and Lembke's pleasure-pain balance show what dopamine really is. All the way to where Schopenhauer's will and Epicurus's subtraction are confirmed, 200 years later, by neuroscience.

Schopenhauer's will (Wille) and the pendulum between lack and boredom, Epicurus's calm (ataraxia) and his redefinition of pleasure, and the one knot that ties the two together. How to handle boredom by taming desire rather than denying it — an optimism built on top of a pessimist's metaphysics.

Epicurus and ataraxia (tranquility), the long-running misread of his 'hedonism,' the four-part cure (tetrapharmakos) that dissolves fear, and the structural limit this philosophy of calm reveals before adventure and passion. Following the thread all the way to Nietzsche's critique.

Reading AI and chip investing through the gold-rush 'sell the shovels' analogy. The trap of a cheap P/E, the lesson of Cisco, and the question of whether a shovel is general-purpose — the points where a too-perfect analogy starts to break.

The 'Zero Billion Dollar Market' strategy Nvidia ran three times — with CUDA, the GPU, and now robots. What it means to lay the infrastructure before the market exists and wait for it to show up.

The more you hand to an agent, the more you get done — and the more your own grasp of the work stays exactly where it was. On the surface that delegates cleanly (the execution of thinking) versus the depth that does not (defining the problem, deciding, understanding), and why Hesse said the same thing a century ago.