Notes
Shorter pieces, working ideas, book reports, and recipes — a digital garden where some entries are evergreen and others are seedlings still finding their shape.
-
Negative Learning Curve
When a technology gets more expensive as more of it is built, inverting the usual pattern where experience drives cost down.
-
Nuclear Energy
How much electricity nuclear actually supplies, why its share has fallen since 1996 despite a strong safety and carbon record, what reversed in the 1970s, and the AI-driven revival now underway. Books and media at the end.
-
Small Modular Reactors
Factory-built nuclear reactors in the tens-to-hundreds-of-megawatts range, betting that serial manufacturing beats bespoke mega-project construction.
-
Anticipatory Savoring
Why the run-up to a trip, the planning and imagining, often delivers more measurable happiness than the trip itself, and the reward-system mechanism behind it.
-
Hedonic Adaptation
The tendency to return to a stable baseline of happiness after positive or negative events, blunting the lasting impact of both.
-
Wanting and Liking
Kent Berridge's finding that the brain's motivation system (wanting) and its pleasure system (liking) are separable, running on different neurochemistry.
-
Becoming a Craft-Aware Moviegoer
How to train yourself to see how a film is constructed (editing, camera, sound, staging) without killing the pleasure of watching.
-
Kuleshov Effect
The finding that viewers read emotion into a neutral shot based on the shot next to it, showing meaning is assembled at the cut.
-
Mise-en-Scène
Everything arranged inside the frame — staging, blocking, set, lighting, costume — and what that arrangement communicates.
-
Direct Part Marking
Marking a machine-readable identifier permanently into a part's surface rather than onto a label.
-
ECC200
The modern Data Matrix variant: what the '200' means, how Reed-Solomon lets a 2mm square survive being sandblasted, and why it marks everything from surgical tools to silicon wafers.
-
Reed-Solomon Codes
The error-correcting code that reconstructs lost data by over-sampling a polynomial, used everywhere from CDs to QR codes to spacecraft.
-
Containerization
The standardized intermodal shipping container that collapsed the cost of moving goods and made global supply chains possible.
-
Flags of Convenience
Why most of the world fleet registers in Panama, Liberia, and the Marshall Islands rather than the owner's home country.
-
Global Shipping
How the system that moves ~80% of world trade actually works: containerization, chokepoints, carrier concentration, the security subsidy that keeps sea lanes open, and the fuel transition ahead.
-
Maritime Chokepoints
The narrow sea passages that carry a disproportionate share of world trade and concentrate its fragility.
-
A Mental Model for Databases
How to reason about any database from its physics, storage engine, and consistency bet, instead of memorizing products that go stale in three years.
-
LSM-Tree
The write-optimized storage engine that turns random writes into sequential appends, trading higher read and space cost for write throughput.
-
PACELC
The extension to CAP that adds the everyday tradeoff: even without a partition, a replicated system chooses between latency and consistency.
-
Feature Store
A central system that computes, stores, and serves ML features so training and serving use identical values.
-
Inference Gateway
A proxy layer in front of model backends that handles routing, caching, rate limits, and fallback for LLM inference.
-
ML Infrastructure
The platform layer above the GPU: the connective tissue that turns a model into a reliable production system, and the live shift from the feature-store MLOps stack to the vector-store LLM stack.
-
ML Infrastructure Stack Below the API
What actually runs beneath a model API call: the inference and training layers most application engineers never touch, and a weekend project to build real intuition for them.
-
Piranesi
What actually happens in Piranesi, and why the audiobook makes the central reveal so easy to miss.
-
The End of the World Is Just the Beginning
Zeihan on deglobalization: why the postwar global order is unwinding and what a fragmented world looks like.
-
AGENTS.md
A repository-level context file that tells any AI coding agent the commands, conventions, and hard boundaries to respect when changing the code.
-
Architecture Decision Records
Short, versioned documents that capture a single architectural decision, its context, and its consequences, so the reasoning survives past the people who made it.
-
Code Review Tooling Market: GitHub, GitLab, and the PR Review Bottleneck
Who actually uses GitHub vs GitLab vs the rest, why startups default to GitHub, and how the industry's standardization on pull-request review turned code review into the universal delivery bottleneck AI is now overloading.
-
Context Engineering
The discipline of systematically supplying an AI agent the intent, plans, and repository-specific detail it needs through persistent artifacts rather than one-off prompts.
-
Git Forge
The software category for web platforms that host Git repositories and layer collaboration on top: GitHub, GitLab, Bitbucket, and self-hosted alternatives.
-
Merge Queue
A forge feature that serially tests and merges approved pull requests against the latest main, eliminating the race where two independently-green PRs break trunk when combined.
-
Progressive Delivery
Releasing a change to a small slice of traffic first and promoting or reverting it based on live telemetry rather than pre-merge confidence.
-
Rebuilding Review and Deployment for the AI Era
When AI writes most of the code, the human review-and-ship path becomes the constraint. A map of the approaches being tried to fix it: reviewing specs instead of diffs, AI reviewers, stacked diffs, and telemetry-gated deployment.
-
Stacked Diffs
A review workflow that breaks one large change into a chain of small, dependent pull requests that review and merge in sequence.
-
Upfront Planning in the AI Era: PRDs, Design Docs, and Specs
When generation is cheap, the leverage moves to everything before the first line of code: PRDs, design docs, RFCs, ADRs, and agent-facing specs. What changes about pre-code planning when the implementer is a model, not a person.
-
AI SDLC Bottleneck
The software delivery lifecycle read as a Theory-of-Constraints pipeline: what the stages are, how to locate the constraint AI has pushed downstream to review, and the exploit-subordinate-elevate moves that relieve it.
-
Amdahl's Law
The bound on how much you can speed up a whole system by accelerating only one part of it, and why it caps AI's effect on software delivery.
-
Norm of Reciprocity
The near-universal social rule that we feel obligated to return benefits we receive, which underwrites much of behavioral-economics persuasion research.
-
Operational Transparency
Ryan Buell's finding that deliberately showing customers the hidden work done for them, and showing employees the people they serve, raises quality, trust, and satisfaction on both sides.
-
Spec-Driven Development
Writing an explicit, agreed specification before code so that generation and review both check against stated intent rather than reverse-engineering it.
-
Transparency Paradox
Ethan Bernstein's finding that watching workers can lower their performance, while zones of privacy can raise it.
-
Effort Heuristic
The mental shortcut of judging something's value by how much effort appears to have gone into it, used when quality is hard to assess directly.
-
Labor Illusion
The behavioral-economics finding that visible effort raises perceived value, so people sometimes prefer a slower service that shows its work to an instant one that hides it.
-
Agent Card
The JSON discovery document an A2A agent publishes to advertise its skills, inputs, auth, and endpoints.
-
Cycle Time Scatterplot
A chart plotting each completed work item by its finish date and cycle time, exposing the full distribution that averages like Little's Law hide.
-
Flow Debt
The hidden cost of finishing some work items fast by letting others age, which keeps average cycle time looking healthy while a tail of stale work grows unseen.
-
Little's Law
The one-line queueing result L = λW: what it actually claims, why its generality is the surprising part, and the assumptions that quietly break it in software.
-
MCP vs A2A
How Anthropic's Model Context Protocol and Google's Agent2Agent protocol divide the agent-interoperability problem, and whether that division is clean or just the next xkcd-927 standards pileup.
-
Cone of Uncertainty
The measured range by which software estimates can err at each project stage, widest at conception and narrowing only as decisions resolve uncertainty.
-
Monte Carlo Forecasting
Forecasting software delivery by resampling historical throughput thousands of times to produce a probability distribution over completion dates.
-
Outcome-Based Pricing
Charging per result delivered (resolved ticket, closed deal, completed task) rather than per seat or per unit of consumption.
-
Right-Sizing
The practice of splitting work so items are roughly similar in size, which makes counting them a valid substitute for estimating them.
-
Software Careers Over the Next Decade
An advisory read on where software work is heading through 2034: the demand numbers read carefully, the real entry-level squeeze, the apprenticeship risk, and what to bet a career on.
-
Software Estimation and Forecasting
A working guide to estimating software: the techniques, why single numbers lie, and how to read the tea leaves with probabilistic forecasting instead.
-
Systems of Record
The authoritative source-of-truth systems (CRM, ERP, ledger) that store core business data and enforce the rules around it.
-
The Death of SaaS
Why AI agents break the per-seat pricing that built the software industry, what the February 2026 selloff actually priced in, and which parts of SaaS the disruption leaves standing.
-
The Probabilistic Delivery Forecaster
Why Monte Carlo forecasting from historical throughput is the delivery-estimation method that survives AI-assisted engineering, and where a clean self-serve tool would fit.
-
Three-Point Estimation
Estimating a task with optimistic, most-likely, and pessimistic values to express a distribution rather than a single deceptive number.
-
Verification Gap
The widening gap between how fast AI can produce code and how long a human takes to confirm that code is correct.
-
Downcycling
Recycling that yields a lower-grade material than the input, because the material degrades and cannot return to its original use.
-
Extended Producer Responsibility
A policy that makes manufacturers financially responsible for the disposal and recycling of the products and packaging they sell.
-
Recycling
Why some materials recycle profitably and others never will, the manufactured myth that plastic is one of them, and where the mechanism closes the loop.
-
Blameless Postmortems
Incident retrospectives that treat failure as a systems problem, not a person problem, so people report honestly.
-
Conway's Law
Organizations ship their communication structure — systems mirror the teams that build them.
-
DevEx
The developer-experience framework — feedback loops, cognitive load, and flow state — that DX Core 4 folds in as its qualitative half.
-
DORA Metrics
The four (now five) software-delivery metrics from Google's DevOps Research and Assessment program, why they are paired into throughput and stability, and what the research does and does not license you to conclude.
-
DX Core 4
A developer-productivity framework that unifies DORA, SPACE, and DevEx into four counterbalanced dimensions so no single metric can be gamed in isolation.
-
Economic and Geopolitical Risk Scenarios for the Next 3 Months
A June 2026 snapshot of the most likely economic and geopolitical risk scenarios through roughly September, ranked by how the odds and triggers actually sit within a three-month window.
-
Glue Work
The essential, unglamorous connective work that keeps teams functioning but goes uncredited in most promotion and performance rubrics.
-
Goodhart's Law
When a measure becomes a target, it stops being a good measure — the failure mode behind every gamed metric.
-
Managing Toil
A decision framework for business-critical manual work that the business won't fund automating: when to eliminate, automate, reduce, or just tolerate it, and how to change the funding decision.
-
SDLC Delivery Metrics for AI-Assisted Engineering
A working glossary of the delivery metrics that actually track engineering throughput (DORA keys plus PR-flow metrics), what each means, how it's measured, and what AI-assisted development is expected to move.
-
SPACE Metrics
The 2021 framework that reframed developer productivity as five dimensions instead of one number, insisting on perception data and refusing individual measurement.
-
Team Topologies
Skelton and Pais's four team types and three interaction modes, organized around cognitive load and fast flow.
-
Toil Budget
The capped fraction of engineering time allowed for manual operational work, so it can't crowd out the engineering that would eliminate it.
-
Trunk-Based Development
The version-control practice of integrating small changes to one shared branch daily — a core DORA capability.
-
Westrum Organizational Culture
Ron Westrum's typology of pathological, bureaucratic, and generative cultures, sorted by how organizations handle information.
-
External Consistency
The guarantee that if one transaction commits before another begins in real time, the system's ordering of them reflects that real-time order.
-
Google Zanzibar
How Google's global authorization system models permissions as relationship tuples and uses consistency tokens to answer trillions of access checks without leaking data through stale ACLs.
-
Relationship-Based Access Control
An authorization model where permissions are derived from a graph of relationships between subjects and objects rather than from roles or attributes.
-
CACE Principle
Changing Anything Changes Everything: the observation that ML models entangle their inputs so no feature can be altered in isolation.
-
Correction Cascades
The ML anti-pattern where a model is patched by learning a correction on top of another model, creating a dependency that eventually deadlocks improvement.
-
Hidden Technical Debt in Machine Learning Systems
The 2015 Google paper that reframed ML maintenance cost as technical debt and named the system-level traps (entanglement, feedback loops, glue code) that make ML harder to maintain than ordinary software.
-
Critical Chain Project Management
The Theory of Constraints applied to projects: size the constraint as the longest chain of dependent, resource-contended steps, and pool safety time into shared buffers instead of padding every task.
-
Drum-Buffer-Rope
The Theory of Constraints scheduling method that paces a whole line to its bottleneck, buffers that bottleneck against starvation, and releases work only as fast as the bottleneck consumes it.
-
Theory of Constraints
Goldratt's claim that every system's output is set by a single binding constraint, the five-step method for exploiting it, and the accounting and scheduling machinery that follow from taking that claim seriously.
-
Throughput Accounting
Goldratt's alternative to cost accounting that reduces a business to throughput, inventory, and operating expense, and refuses to book unsold production as profit.
-
AI Circular Financing
The closed loop in which AI infrastructure spend recirculates among a few firms that are each other's suppliers, customers, and investors, inflating headline demand.
-
Basis Risk
Basis risk is the residual loss a hedge leaves when the instrument and the exposure it covers fail to move together.
-
Diversification vs Hedging
Why spreading risk and offsetting risk are different operations: one lowers variance and is free but fails in a crisis, the other cancels a specific exposure and costs a premium but holds when you need it.
-
Float-Adjusted Market Capitalization
Weighting a stock by the market value of shares actually available to public investors, not its total shares outstanding.
-
Index Fund
A fund that mechanically holds a rules-based basket of securities to track an index rather than to beat it, trading judgment for low cost and broad exposure.
-
Inelastic Markets Hypothesis
The claim that a dollar invested in the stock market moves aggregate prices far more than efficient-market theory predicts, because supply is highly inelastic.
-
Parametric Insurance
Coverage that pays a fixed amount when a measured index crosses a threshold, no loss adjustment required, in exchange for basis risk.
-
Planning Fallacy
The systematic tendency to underestimate the time, cost, and risk of a task even after living through similar tasks that ran long.
-
Reference Class Forecasting
How forecasting from a class of similar past projects corrects the planning fallacy, where the method breaks down on choosing the class, and what it cannot fix.
-
Ski Industry Risk Management
How large ski operators like Vail and Alterra manage weather risk, and why the dominant tool is a pre-sold season pass rather than weather derivatives or insurance.
-
S&P 500
How the S&P 500 is actually built: float-adjusted cap weighting, a discretionary committee that picks members, and the concentration and passive-flow risks that come with the design.
-
Strategic Misrepresentation
Deliberately optimistic forecasting of cost and benefit to get a project approved, distinct from honest cognitive bias.
-
Superforecasters
The small group of ordinary forecasters who, in IARPA's tournament, beat trained analysts by starting from base rates and updating in small steps.
-
Superforecasting
Tetlock and Gardner on the Good Judgment Project: how the best forecasters reason, and why method beats expertise.
-
Systematic and Idiosyncratic Risk
Total risk splits into a market component you cannot diversify and a name-specific component you can, but only the first is priced and the split breaks in a crisis.
-
Term Premium
The extra yield investors demand for holding a long-dated bond instead of rolling short ones, and the part of long rates the Fed does not directly control.
-
Motte and Bailey
The rhetorical structure where a strong, exciting claim retreats to a trivially true one under fire, then returns once the critic leaves: its origin, why its author insisted it is a doctrine and not a fallacy, and how the label itself gets abused.
-
To Read, Watch, and Listen
Single queue for books, films, and music.
-
Recursive Self Improvement in AI
The classic intelligence-explosion loop versus what's actually self-improving in 2025: agents that rewrite their own scaffolding, gated by verifiers, and bottlenecked by compute.
-
Definition of Done
A team's shared, binary standard for when any work item is actually complete. Why it is universal rather than per-ticket, and how it differs from acceptance criteria.
-
QA Handoff Workflows
How engineering teams move a code-complete ticket to a QA analyst for sign-off, and why most use a status transition rather than reassigning the ticket away from the developer.
-
Work-in-Progress Limits
Capping how many items can be in a workflow stage at once. The counterintuitive lever that lowers cycle time without changing throughput, and the math (Little's Law) that explains why.
-
Recognition-Primed Decision Model
How experts decide under time pressure without comparing options: the pattern-matching plus mental-simulation model Gary Klein built from fieldwork, and where it breaks.
-
Breakeven Inflation Rates: T10YIE, T5YIE, and T5YIFR
How the three FRED breakeven series relate, why the forward rate is bootstrapped from the other two, and what risk and liquidity premiums do to the signal.
-
Inflation Risk Premium
The extra yield investors demand on nominal bonds for bearing inflation uncertainty, why it can go negative, and why even its sign is model-dependent.
-
Liability Matching
Why pensions, insurers, and retirees buy bonds to cover known future bills regardless of whether equities return more, and how matching duration can blow up when you add leverage.
-
Obsidian Hotkeys
My curated Obsidian keyboard reference (Linux + Vim mode): essential defaults, my customizations, and a recommended-to-adopt set.
-
Decision Frameworks
The research behind decision frameworks: why the first move is classification, when expert intuition can be trusted, and whether debiasing tools actually work.
-
Writing Checklists
Two passes for assessing a draft: a reasoning checklist for whether the claims hold, and a style checklist for whether the prose lands. Every item carries its own test.
-
Bullwhip Effect
Why small swings in customer demand turn into wild production swings upstream — a feedback-and-delay failure, not a demand problem.
-
Critical Systems Heuristics
Werner Ulrich's framework for surfacing the value judgments hidden in any system model — who decides what's inside the boundary, and whose interests it serves.
-
Data Center Externalities
The costs of the data center buildout that don't show up on the operator's balance sheet: grid strain, water draw, and electricity bills shifted onto everyone else.
-
Digital Asset Research Game Plan
How technical operators evaluate, buy, or build small software businesses — the market mechanics, valuation reality, moat taxonomy, and diligence structure that make the difference between a good first acquisition and an expensive education.
-
Jevons Paradox
Why making a resource more efficient to use can raise its total consumption instead of lowering it, the conditions under which that happens, and why AI compute is a near-perfect case.
-
LLM Energy Use
What it actually costs in energy to train and run large language models, why per-query numbers vary by orders of magnitude, and why inference now dominates the footprint.
-
Marginal Risk Framework
A method borrowed from cyber threat modeling for asking not whether an AI model is dangerous, but how much danger it adds over tools that already exist.
-
Open Weight vs Proprietary (vs Open Source)
Why downloadable weights, open licenses, and reproducible training are three different things, and what the distinction does to risk, reproducibility, and pricing.
-
Openwashing
Marketing a product as open to claim the goodwill and dodge the scrutiny, without granting the freedoms the word implies. Now a structural feature of the AI model market.
-
Ratepayer Cost-Shifting
How the rules of utility ratemaking quietly move the grid-infrastructure cost of large new loads like data centers onto ordinary households, and the tariff fixes regulators are now scrambling to build.
-
Systems Thinking
How feedback structure generates the behavior of complex systems, why intuitive interventions backfire, and where leverage actually lives.
-
Test-Time Compute
The shift from spending compute at training to spending it at inference: how reasoning models think longer to answer better, the methods behind it, and the steep cost curve that follows.
-
The Cost Subsidization of LLM Use
The gap between what LLM providers charge per token and what inference actually costs, who is absorbing the difference, and what happens when that arrangement ends.
-
World Models vs. LLMs
The split isn't language vs. reality. It's whether the training objective punishes incoherence, and next-token prediction never does.
-
Writing Code vs. Shipping Code: AI Productivity Across Tool Generations
A 100,000-developer study finds that AI coding tools produce massive gains in raw coding activity but only modest gains in released software, because human bottlenecks at review and integration stages swallow most of the upstream productivity.
-
Building an Ambient Drone Machine in the Browser
How I built a generative ambient drone synth with nothing but the Web Audio API — detuned oscillators, glacial LFOs, and Eno-style tape loops.
-
Mixture of Experts
How splitting a model into specialized sub-networks lets you scale parameters without scaling compute.
-
A List of fun musical toys I've made
A list of fun musical toys I've made
-
Consciousness Upload and Bio-Synthetic Minds
Mind uploading and human-machine composites — a recurring sci-fi trope, sparked by Children of Time.
-
Emergent and Collective Intelligence
Complex behavior from simple parts — tiny brains and many agents, sparked by Children of Time.
-
Genetic Algorithms and Benevolent Viruses
Horizontal trait transfer as an evolution accelerator — a GA idea sparked by Children of Time.
-
Greg Egan
Hard-SF author — on my to-read list, especially for rigorous treatments of mind, identity, and physics.
-
Neural Networks and Brain Scaling
How capability tracks brain size — a scaling-law idea sparked by Children of Time.
-
Talking to Synthetic Assistants
How we converse with — and shape — sentient AI assistants. Sparked by Children of Time.
-
Terraforming
Engineering a self-sustaining biosphere over geological time — sparked by Children of Time.
-
Uplift and Engineered Evolution
Directing evolution toward intelligence — and losing control of it — sparked by Children of Time.
-
Audiobook Player Improvements
A list of improvements for my audiobook player.
-
Gear
Coffee Gear I Use
-
How to Read a paper
A guide to reading Arxiv papers
-
Yes, I use AI
Yes, I use AI for stuff.
-
Coffee - Sermon
Sermon from Verve Roasters
-
The Daily
A personal daily newspaper I built for myself — a fresh edition auto-generated every morning from the sources I actually care about.
-
Anthropic Prompt Caching
Notes from implementing prompt caching in [[Pantheon]] to dodge Sonnet's 30K input-TPM rate limit
-
Decisions
My working playbook for decisions that actually warrant the effort: how to triage, walk through the hard ones, and the traps I keep falling into.
-
Model Collapse and the AI Data Crisis
Working through a Reddit doomer post on AI model collapse — what's real, what's overstated, and where synthetic data actually works.
-
Colombia - Santa Monica Blackberry Honey
Columbia Santa Monica Blackberry Honey Process
-
Ethiopia - Worka Sakaro Anaerobic Wine
Ethiopia - Worka Sakaro Anaerobic Wine
-
Papua New Guinea - Jiwaka Arufa Natural
Papua New Guinea Jiwaka Arufa Natural
-
Peru - El Palto Natural
Peru El Palto Natural
-
Coffee
Entry point for Coffee
-
Multi-LLM Coordinator — Idea
Cost Aware agent harness that runs in the command line. Claude Code clone project written in Rust.
No notes match the current filter.