Insights

Our thinking, published in the open.

Written for the same reason the prices are on the page: it is the fastest way for someone to work out whether the reasoning is any good before they pay for it.

Aug 2026

YC Summer 2026, the full case record

The full sourced dataset behind "What Y Combinator Is Actually Funding Right Now" — sector-by-sector company counts, disclosed funding with every sourcing caveat intact, the accelerator comparison sample, and an explicit list of what couldn't be confirmed.

Founders · Startups · Research

Aug 2026

What Y Combinator Is Actually Funding Right Now

A full sector-by-sector read of Y Combinator's current batch — what's crowded, what three other accelerators are independently converging on too, and a framework for finding the gap instead of chasing the hottest category on the page.

Founders · Startups · Research

Aug 2026

The Prestige Employer Map

A sector-by-sector directory of the firms that read as prestigious in finance, law, tech, luxury, media, industrials, and consumer/healthcare — one person's curated take on brand recognition, not a ranking anyone should mistake for evidence.

Career · Reference · University Admissions

Aug 2026

The McKinsey of Everything

An opinionated map of elite institutions, sector prestige benchmarks, and the eight realistic paths to significant wealth — capped by a thought experiment on the most stacked resume ever assembled, and why building one on purpose is exactly the wrong move.

Career · Reference · Entrepreneurship

Aug 2026

How SAT Scoring Actually Works

College Board says it plainly: miss a couple of questions on the first module, and an 800 is off the table no matter what you do on the second — even though the two modules are graded together. That single fact breaks the 'each question is worth N points' model most study advice quietly assumes.

Research · EdTech · SAT

Aug 2026

Stare at a Wall for Twenty Minutes, They Said

There is no study that tested "stare at a wall before you study to prime your focus." There is a study that tested resting with your eyes closed right after you learn something — and it's not the same claim, even though the advice gets repeated as if it were.

Research · Learning Science · EdTech

Aug 2026

The Meridian Tutor Prompt

Two SAT practice tests, same raw score — 45 of 54 Reading & Writing questions correct on both. One scored 540. The other scored 700. That 160-point gap isn't noise; it's the adaptive routing gate, and it rewrites how a tutor prompt should work. Published in full below.

Meridian · Education · Prompts

Aug 2026

How to Tell an AI-Built Website From a Real One

Ask an AI assistant for "a professional-looking accent color" and it reaches for the same indigo-500 with startling consistency — traced back to a single Tailwind UI default from years before any of this started. The tell was never the polish. It's whether a choice could be pasted onto a competitor's site, logo swapped, and still be true.

Design · AI

Aug 2026

5,486 Videos, and the Bug That Lied

A caption fetcher reported that a creator had no subtitles. It was being blocked. Same manifest, same field, completely different meaning — and nothing in the output distinguished them.

Engineering · AI · Research

Aug 2026

Modelling a Person Without Getting a Caricature

A persona that sounds exactly right and reasons wrong is worse than one that sounds flat and reasons right, because you will believe the first one. Here is what I learned building a system that tries to tell the difference.

AI · Research · Persona

Aug 2026

Twenty-Three Agents and What Broke

Rebuilding a paid course meant running dozens of AI agents against the same production codebase in one session, some in parallel, some checking each other's work. Most of it worked. The part that didn't is more instructive than the part that did: a silent failure that took three attempts to diagnose, and wasn't fixed by the thing everyone tries first.

AI · Engineering · Meridian

Aug 2026

The Programme I Built Instead of Sixth Form

At 16–18 the binding constraint isn't knowledge, it's reps. Nobody leaves school having run a P&L, read a term sheet, or shipped to a paying stranger. Here's the curriculum I built to front-load them, with costs, and instructions for copying it.

Education · Entrepreneurship · Homeschool

Aug 2026

The Evidence Behind the AI-Writing Tells

Kobak et al. analyzed 14 million PubMed abstracts and found 'delves' appearing 25 times more often than the pre-ChatGPT trend predicted. The popular explanation for why is wrong — a 2025 study tested it directly and it didn't hold up.

Research · Writing · AI

Aug 2026

What Actually Makes Writing Sound Human

My working theory was that AI writing over-hedges. A 2025 PNAS study found the opposite: models hedge less overall, and lean on a tiny set of canned phrases instead. That was the first thing I had backwards.

Research · Writing · AI

Aug 2026

Reading Is a Proof, Not a Vibe

SAT Reading feels random to students who are good at math because they're using the wrong tool on it. Treat the passage as a set of axioms and the correct answer as the only thing provable from them, and the inconsistency mostly disappears.

Education · Learning · Method

Aug 2026

The MacBook Air's Real Monitor Limit

Seven displays will enumerate in System Settings without complaint. The chip that has to render all of them has a much stricter opinion, and it's a thermal one, not a port-count one.

Tools · Hardware · Systems

Aug 2026

What Every Good Reading App Does That This Site Didn't

Gwern's sidenotes, Stratechery's paywall labels, Craft's focus mode, Readwise's highlight sync: a named-competitor audit of what a great reading platform does, run against this site's own components so it doesn't recommend rebuilding what's already there.

Research · Product · Writing

Aug 2026

Show Prep: The Full Case Record

Every mechanism in the danger taxonomy explained down to the level a curious reader could actually follow — what oxidative phosphorylation has to do with DNP, why potassium-sparing and loop diuretics make each other worse, why insulin has the least forgiving margin of anything in the sport, and the full mortality data from the largest cohort study of competitive bodybuilders ever run. Nothing here is a protocol. It's the reference the shorter article didn't have room for.

Health · Training · Mechanism

Aug 2026

Cutting While Still Growing

Every cutting guide assumes a body that's finished being built. The real evidence — muscle-loss thresholds, a reversible testosterone mechanism, bone still being laid down, growth plates that stay open later than the culture assumes, and a decade of longitudinal data on where the real risk actually lives — points somewhere more specific than "ask your doctor," plus the actual protocol for running a cut that keeps every risk to its minimum.

Nutrition · Training · Protocol

Aug 2026

Danger and Evidence Run in Opposite Directions

The practices with real supporting evidence in competitive bodybuilding's final week — carbohydrate loading, a modest sodium taper, a planned refeed — are also the least dangerous ones. The practices that dominate the cautionary stories — diuretics, DNP, insulin, synthol — have the weakest evidence of working and a documented, population-level body count. Here's the taxonomy, sourced to the published case reports and cohort studies, with the mechanism for each and nothing about how much to take.

Health · Training · Mechanism

Aug 2026

What a Blood Panel Actually Tells You

A lab reference range describes the population that walked into the lab, and that population is not healthy. Here's how to read a standard panel, which numbers matter, and the one test you only need once in your life that almost nobody gets.

Health · Biomarkers · Longevity

Aug 2026

The supplements that got retracted

In 2023 a Science paper said taurine falls with age. In 2025 the NIH looked again and found it flat or rising. What the reversals in this field have in common is not that they failed — it is what carried them until they did.

Longevity · Research · Epistemics

Aug 2026

Why I publish my mistakes

I published a sentence praising a researcher for a position he does not hold, while making a point about intellectual honesty. That sentence is still on the site. It has just moved.

Method · Epistemics · Writing

Aug 2026

Sprint mechanics for people who only lift

At top speed your foot is on the ground for about nine hundredths of a second. That single number explains why heavy leg strength transfers to sprinting so much worse than lifters expect.

Training · Biomechanics · Protocol

Aug 2026

Reading a study without a stats degree

Statistical significance answers a question almost nobody is asking. Here is what to look at instead, why a significant result from a small study is usually an overestimate, and the one question that catches most of the rest.

Research · Method · Epistemics

Aug 2026

The Prompt Was Never the Leverage

Prompt engineering as most people practise it — clever personas, trick phrasings, MUST and NEVER in capitals — is mostly incantation. Across seven versions of a prompt-lab, the leverage turned out to be an operating contract: effort tiers, verification gates, and pushing each rule to the lowest layer that can enforce it.

AI · Prompting · Claude

Aug 2026

One Kinase, Two Answers

Endurance training and hypertrophy don't merely compete for time. They send opposing instructions to the same signalling node, and the conflict has a name: AMPK. Here's what it actually does, how big the effect is, and how to schedule around it.

Training · Mechanism · Endurance

Aug 2026

Measurement Without a Decision Is a Hobby

More data makes the signal-to-noise problem worse, not better, unless the analysis accounts for it. A counterweight to my own tracking article, and the rule that decides whether a metric earns its place.

Health · Method · Protocol

Aug 2026

The Prompt Is the Setup. The Loop Is the Engine.

Prompt engineering gets the instruction right; that's the setup. The engine is the loop the agent runs before it says "done": score-and-rewrite, verification gates, mode selection. The loops that actually compound quality are the ones that let in information the model can't fake.

Prompting · Claude · Method

Aug 2026

The protocol is not about looking good

Seven pillars, ordered on purpose, with the body first. Not because the body matters most, but because everything else runs on it — and I only found that out by getting the order wrong first.

Protocol · Systems · Performance

Aug 2026

I built my own curriculum. Here's what school was actually for.

I study Economics, Mathematics and Business on a curriculum I wrote myself. The subjects were the easy part. The hard part was discovering which parts of school I had been quietly relying on.

Education · Learning · Homeschooling

Aug 2026

Thirty-four founders, and the one thing that predicted who made it

I've mentored 34 young founders and run an eight-person incubator. The variable that separated the ones who shipped from the ones who didn't was not the one I expected, and it isn't the one that gets written about.

Founders · Mentoring · Operating

Aug 2026

A five-figure invoice bought a Rolex. The invoice was the achievement.

Acro Studios has served 17 clients at a four-figure average ticket. One engagement was multiple five figures — enough to buy a platinum Day-Date. What changed between client one and that one had nothing to do with the code.

Agency · Pricing · Operating

Aug 2026

I wrote that actions matter more than interpretations. I was half right.

At fourteen I argued that you should judge people by what they do, not by how you read them. I still think that's mostly true. The part I got wrong has cost me more than the part I got right has earned.

Essays · Judgement · Revisiting

Aug 2026

One brutal thing a year, done in the wrong order

Once a year I do something with a genuine chance of failure. The current version is an ultra Ironman in reverse order, and the reversal is not a gimmick — it's the entire mechanism.

Misogi · Hardening · Protocol

Aug 2026

Four million words, and the part that isn't the model

Veridian marks Pearson IAL Economics and Business papers end to end. The marking harness took weeks. The 4-million-word knowledge base underneath it took months, and it's the only reason the thing works.

Veridian · AI · EdTech

Aug 2026

Four thousand contacts and no idea who mattered

I had 4,000 connections from one platform alone and was quietly letting the twenty relationships that mattered decay. Every tool I tried was built for sales pipelines. So I built the other thing.

Roster · Product · Building

Aug 2026

Everything that feels like studying isn't

The study methods that work feel worse than the ones that don't. That single fact breaks the feedback loop every self-taught student is relying on — including mine.

Education · Learning · Research

Aug 2026

How people actually learn: the evidence, the effect sizes, and what to build

Deep research behind the courses. What the meta-analyses actually say, where they disagree, which popular techniques have no evidence at all, and a ranked build list ordered by evidence strength divided by cost.

Research · Learning · Education

Aug 2026

Finding the Real Experts

The most-viewed source in almost any domain is, structurally, not the most reliable one. A six-tier filter for finding the people the popularisers are quietly reading — plus the calibration that stops it turning into contrarianism.

Research · Method · Epistemics

Aug 2026

Life-Maximization Digest — Q3 2026

Fourteen parts, ~15,000 words: what actually moved in age-extension research this quarter, what got refuted, and what any of it changes in practice. Every claim carries an evidence tier and a source tier, and the two are held separate on purpose.

Longevity · Research · Health

Aug 2026

The Life-Maximization Digest Prompt

The prompt that generates the quarterly digest, published in full. Includes the two-axis evidence/source rubric, the 'well-known is not weak' calibration, the hard safety line on unregulated compounds, and the three parts added in v2 — failure ledger, cost/risk ranking, and an adversarial pass.

Research · Method · Prompts

Aug 2026

Two Sets Is Enough

A muscle fibre grows when it is recruited and under high tension at the same instant. That single requirement explains why a set of 20 and a set of 6 produce the same growth, why the fourth set of an exercise is nearly wasted, and why frequency beats volume.

Training · Hypertrophy · Mechanism

Aug 2026

Sleep Is Plumbing

For a decade the glymphatic system was a mouse story. A 2026 randomised crossover trial made it a human one, linking sleep-stage physiology directly to the clearance of amyloid-beta and tau. Here's the actual mechanism, and why regularity beats duration.

Sleep · Neuroscience · Mechanism

Aug 2026

The Number That Predicts Everything

Cardiorespiratory fitness has the largest effect size of any modifiable variable in medicine, with no observed upper limit. Here's what the number is actually measuring, and why Zone 2 alone will not move it.

Training · Longevity · Cardio

Aug 2026

Nine Things

A maximal protocol run at 60% adherence loses to a minimal one run at 95%, because almost everything here is dose-over-time rather than dose-per-session. So here is the ruthless version — nine levers, roughly an hour a day, and an honest account of what got cut.

Protocol · Health · Minimalism

Aug 2026

Your Lab Range Is Not a Target

Normal fasting glucose runs to 99 mg/dL, and the mortality curve starts bending well before that. The gap between 'normal' and 'optimal' is where most preventable disease lives — plus the one test almost nobody orders that you only need once in your life.

Health · Biomarkers · Longevity

Aug 2026

If Only the Gym Mattered

What the ceiling actually looks like when growth is the only objective: the per-session volume ceiling, why weekly volume is meaningless without frequency, the movement list chosen by resistance profile rather than tradition, and an honest account of what you give up to run it.

Training · Hypertrophy · Protocol

Aug 2026

The Rate Is the Decision

Muscle accretion has a hard ceiling of roughly 0.25–0.5% of bodyweight per month, and no surplus makes it faster. Everything above that is fat. Here's the arithmetic on bulking, cutting, and the small-surplus approach that quietly beats both.

Nutrition · Training · Protocol

Aug 2026

A Spreadsheet That Still Works in Ten Years

Training log with automatic 1RM estimation, daily weight with a 7-day rolling average, macro targets that compute from your bodyweight and phase, and strength benchmarks that scale as you do. Free, no signup, and it will still open in a decade.

Tools · Training · Protocol

Aug 2026

The Exercise You're Doing Isn't the One You Think

Every exercise has a point where it's hardest, and every muscle has a length where it grows best. Where those overlap determines what you're actually training — which is why some staple exercises quietly fail at the job they're assigned.

Training · Hypertrophy · Biomechanics

Aug 2026

Training to Failure Is a Tax You Sometimes Should Pay

The last five reps before failure are the only ones that reliably stimulate growth — which makes proximity to failure the most important variable in a set. It also makes going beyond failure expensive in a currency most people never account for.

Training · Hypertrophy · Protocol

Aug 2026

A Claude Skill That Filters Out the Popularisers

A markdown file that makes Claude check who a claim came from before repeating it. Six tiers, a six-point disqualification pass, and two calibrations that stop it from quietly turning into contrarianism. Copy it, save it, done.

Tools · Claude · Research

Aug 2026

The Sound of Machine Writing

Machine prose isn't wrong, it's even: the same sentence length, the same rhetorical move every paragraph, a hedge before every claim. Here are the tells ranked by how much they give you away, the one fix that improves the writing rather than disguising it, and the full skill file.

Writing · Claude · Tools

Aug 2026

The Simulation Argument Is Real. The Version You've Heard Isn't.

Nick Bostrom's simulation argument is a valid trilemma that popular culture has flattened into a false certainty. Taken seriously, it exposes its own buried assumptions about minds and compute — and points at a real, testable question about whether the universe is discrete.

Physics · Philosophy · Computation

Aug 2026

The Points Were Never the Point

The Good Place sorts its characters by a lifetime points tally, then quietly proves that character is built by practice rather than revealed by a score. That is the same shape as learning anything, and it is why I built Potentia on a line that means: it is not found, it is made.

Philosophy · Learning · Potentia

Aug 2026

The Honest State of Not Dying

What's mechanistically real about life extension — the hallmarks of aging, why rapamycin and senolytics are studied — separated from what's oversold: the supplements, the biological-age clocks, and the certainty. Plus what "immortality" would and wouldn't actually buy.

Longevity · Health · Research

Jul 2026

Where Founders Actually Come From

The three most-cited entrepreneurship-school rankings disagree so completely that at least two of them have to be measuring something other than what they claim. Here's what's underneath them.

Research · Founders · University Admissions

Jul 2026

The Best Universities in the World for Undergraduate Entrepreneurship

The full research behind 'Where Founders Actually Come From' — a school-by-school breakdown of capital access, accelerator selectivity, leave policies, co-founder density, admissions reality, and the alternatives worth weighing against a degree.

Research · University Admissions · Founders

Jul 2026

Building and exiting CASA before I could vote

How a teen-events brand went from me promoting other people's parties to 25+ events across three cities — and why I sold it the moment the mission was done.

Ventures · Exits

Jun 2026

Veridian: turning 4M words into an A* on my own exam

I built an AI examiner because the resources for Pearson IAL Economics & Business weren't good enough for the grade I wanted — then sold it to two schools and 70 students.

Products · AI · EdTech

May 2026

Compounding a family portfolio from AED 3M to 16M+

Notes on growing capital the boring way — real-estate debt, house-flips, selective speculation — while everything else I build stays operator-led.

Capital · Investing

The longer research

The protocol pages carry the method these articles are written under — every claim labelled with how good the evidence is and who is saying it, and an append-only record of everything that turned out to be wrong.

Read the Protocol