Zkitszo - The Real
September 13, 2026

Think Clearly to a Super You

I lost a good chunk of last week to a five-hour YouTube course on thinking clearly (it's linked in Sources; this post is about the material). Then I lost the rest of the week pulling up the studies it leans on, because I can't take a quoted number at face value anymore and honestly it's becoming a problem. Some of it held up better than I expected. Some of the most repeated numbers in this whole self-improvement genre turned out to be folklore, and the "wait, what?" moments are the best part.

The short version: most of what we call thinking clearly is decided before you sit down. How you slept, whether you're breathing properly at night, what the air and the noise in the room are doing, what your phone does when you do nothing. Fix that plumbing first. Then a dozen mental models and four small techniques do the rest.

Where clear thinking actually comes from the lower layers cap the upper ones, and they are the cheapest to fix, so start at the bottom 1 Hardware sleep, breathing, blood, fitness 2 The room air, sound, light, notifications, and every default you never set 3 Software expected value, Bayes, Fermi, power law vs chain law, planning multiplier 4 Techniques trigger-action plans, noticing, pre-mortems, five-minute timers 5 AI red-teams the plan; you still write the last line start here

Your brain is cheap, and that's the whole game

The brain is about 2% of you by weight and burns about 20% of your resting energy, and nearly all of that is baseline. Thinking hard barely moves the meter, which kills the usual "thinking burns calories" story. The shortcuts are real anyway, they're just about attention. You can hold about four things in working memory at once (the old "seven, plus or minus two" was closer to a rhetorical flourish; the modern figure is four, plus or minus one), and a habit costs none of those four slots. So given the choice, you take the habit over the decision and the guess over the calculation, and everything below is a way of working with a brain that stingy.

Willpower is the one I had to re-learn, and I'd been repeating the old version for years. The willpower-as-fuel-tank idea, ego depletion, went into a 23-lab, 2,141-person preregistered replication in 2016 and came out with an effect size of 0.04, which is a polite way of saying nothing. A second multi-site attempt in 2021, 36 labs and 3,531 people, got 0.06. Meanwhile a beeper study of 205 adults found the people with the best self-control were the ones who met the fewest temptations in the first place. Their trick was fewer fights, arranged in advance. So if a step in your plan reads "I just won't", you've written a hope down and called it a plan. Remove the access instead.

One more piece of biology, and the first correction. Mice housed in a cage with more toys, tunnels and cage-mates grew 15% more neurons in the memory part of the hippocampus, with no drugs and no training. Mice, by the way. I went to check because I'd have sworn it was rats, and so does every retelling. You can't order new neurons. You can order the cage, and that's the part nobody treats as a decision.

Someone else pre-ticked the box

A default is what happens when you do nothing. Organ donation is the famous case: consent runs from about 4% to 28% in countries where you opt in and from 86% to 99.98% in countries where you opt out. Different countries, so hold the comparison loosely, and registered consent isn't a transplant (the countries that actually get organs out are the ones with the coordinator infrastructure). The direction still holds, and every device you own ships with defaults picked by someone paid to raise a metric: autoplay, notifications on, the browser reopening last night's tabs. If a company set the default and you do nothing, you get what they want. Set it yourself once, on a clear-headed Sunday, and doing nothing starts producing what you want.

Friction is the same lever at a smaller scale. In a pay-by-weight cafeteria, moving a food about ten inches further away cut how much of it people took by 8 to 16%. Ten inches. Swapping the serving spoon for tongs did the same. One caveat on this whole family of findings: a 2022 reanalysis found the pooled "nudge" effect shrinks to about nothing once you correct for publication bias, so the two studies above are the better-run exceptions in a field whose average result is weaker than its reputation. The move still costs nothing. Take a step out of the thing you want more of (shoes by the door, the doc already open) or put one in front of the thing that eats your evenings (sign out, delete the app). A phone three doors away wins fights you never have to have.

Sleep first, and breathing before sleep

Before any sleep-hygiene tip, rule out sleep-disordered breathing. A modelled estimate puts obstructive sleep apnea at roughly 936 million adults aged 30 to 69 worldwide (the study was funded by a CPAP maker, so hold it loosely). The sneaky one is upper airway resistance syndrome: the breathing effort fragments your sleep without producing the pauses the apnea index counts, and people with it are often more wrecked in the daytime than people with mild apnea. The free screen is STOP-BANG, eight yes-or-no questions. A score of 3 or more puts you in the intermediate-risk band, which means raise it with a GP; it's built to miss nobody, so plenty of 3s turn out fine. In BC a GP can order a home test that's covered, which is unusual (elsewhere in Canada expect roughly $200 to $400, and $150 to $500 direct-to-consumer in the US). One thing to know before you take it: a home test has no EEG, so it can't see the arousals that define UARS. If the home test comes back clean and you still wake up exhausted, that's the argument for an in-lab study.

Then hours. Forty-eight adults spent two weeks on 4, 6 or 8 hours in bed, 13 or so per arm. The 6-hour group ended up as impaired as people kept awake for up to two straight nights, and their sleepiness ratings crept up while their performance kept falling. You'd be judging your own impairment with the impaired thing. I'd have told you I was fine, too. Then regularity, which may matter more than hours: across 60,977 people and ten million hours of wrist-tracker data, how regular someone's sleep was predicted their mortality better than how long they slept. That's observational, so it's an association, but a fixed wake time seven days a week costs nothing to try.

Caffeine has a half-life of about five hours, so half of a noon coffee is still in you at 5 p.m. and a quarter at 10. That five is an average; between people the range runs 1.5 to 9.5 hours, so your number is yours to find. Twelve people given 400 mg six hours before bed lost measurable sleep. A 24-study review modelled the cut-off at about 8.8 hours before bed for a 107 mg cup of coffee and 13.2 hours for a 217 mg scoop of pre-workout, which makes the 2 p.m. pre-workout a 3 a.m. problem, and I did not enjoy working that out. The flat "10 hours" you'll see quoted is a round number, strict for coffee and loose for pre-workout. Morning daylight pulls your body clock earlier (an eight-person camping study, a 61-student regularity study); the precise "ten minutes within an hour of waking" dose is folklore, so just go outside. And that 18 to 20 °C bedroom rule: the review usually cited for it names no optimum at all, and a 2023 study of older adults found sleep most efficient between 20 and 25 °C, with big differences between people. Test your own number. Alcohol gets you under faster, pushes your first REM period back, and then fragments the second half of the night. That one's well replicated.

The boring deficiencies that feel like personality

A blood panel finds them. The best-supported one is iron: in 113 women aged 18 to 35, fixing ferritin went with a five- to seven-fold improvement on cognitive tests, and the women who were iron-deficient without being anaemic sat between the healthy and the anaemic. A standard blood count misses this, so ask for ferritin by name, and test before you supplement: iron is one of the few things that's genuinely dangerous to take blind. Ask for B12 and thyroid while they're drawing. Vitamin D is worth checking for deficiency, but supplementing it did nothing for cognition across about 4,200 older adults in the VITAL trial.

Creatine gets sold as the closest thing to a smart pill. The honest read of the 2023 meta-analysis: memory improved a little overall (an effect size of 0.29), essentially zero in 11-to-31-year-olds (0.03) and large in 66-to-76-year-olds (0.88). The sleep-deprivation study that gets quoted used a single dose of 0.35 g per kilo, call it 25 grams, which is a different intervention from 3 to 5 g a day. It's cheap and, for most people, harmless, so take it if you like, with expectations to match. One practical note: creatine raises serum creatinine, so the blood panel from the last paragraph will read as though your kidneys are struggling. Tell whoever ordered it that you're taking it, and if you already have kidney disease, ask first.

Fitness is the big one. Among 122,007 people referred for treadmill tests, the bottom quarter for fitness carried about five times the adjusted mortality hazard of the top 2.3%, against 1.41 for smoking in the same model. Hold that comparison loosely: slicing the extremes of a continuous variable will always outrun a yes-or-no one like smoking, and in a cohort referred for cardiac workup the sickest people are the least fit to begin with. The direction isn't in doubt. The headline ratio is flattering. The cheapest way in is REHIT: ten minutes of easy pedalling with two all-out sprints in it, three times a week, starting at ten seconds and working up to twenty over six weeks, which is how the study actually ran. Fifteen sedentary trainees raised their VO2 max 15% (men) and 12% (women); the pooled figure across 93 trials is a more sober 8.3%, and adding sprints beyond two adds nothing. If you've been sedentary, are over 40, or have cardiac risk in the family, clear it with a GP and build up the way the study did. You can estimate VO2 max for free with the Cooper test: run as far as you can in 12 minutes, then (metres - 504.9) / 44.73.

The room you think in

Air first, because the internet has this one wrong in both directions. The 1,000 ppm CO2 figure that gets quoted as a health limit was never one: ASHRAE, whose standard it's pinned on, says CO2 at the concentrations found in buildings "is not a direct health risk" and is used as an indicator of occupant odours. On the other side, the two chamber studies behind "bad air halves your cognition" had 22 and 24 people in them, and a 2020 review of 37 studies called the CO2 research inconsistent, with the reliable effect showing up mainly on one decision-making test battery. The reason to care about CO2 anyway is that it's the cheapest proxy for how much of everything else in the room you're breathing twice. Ventilation costs nothing. Crack the window, prop the door. If you want an actual number, don't buy the $200 monitor, I nearly did. A Sensirion SCD40 module (about $20 to $45), a $4 ESP32 and the ESPHome scd4x component give you a true CO2 measurement (photoacoustic NDIR, so it isn't the "eCO2" guess cheap gadgets make) in Home Assistant, the same setup that feeds the 3D room dashboard. The sensor block is below; you still need the usual esphome, esp32 and i2c blocks above it.

sensor:
  - platform: scd4x
    co2:
      name: "Workshop CO2"
    temperature:
      name: "Workshop Temperature"
    humidity:
      name: "Workshop Humidity"

Noise is the one that surprised me. What hurts is intelligible speech, and the damage tracks a measurable thing called the speech transmission index: performance starts dropping around 0.21 and bottoms out at 0.44, with verbal short-term memory hit hardest and only limited evidence for complex tasks, which is the work you actually do. Sound masking exists precisely because it lowers that index, so the mechanism argues for masking; the catch is a 2013 study where office noise without any words still hurt mental arithmetic, so more noise may not save you. Foam earplugs and a closed door are the $5 bet. I went looking for a trial of earplugs specifically and there isn't one. Light: a small pilot of 49 office workers found the ones with windows slept longer and reported better health, which is thin, but a window seat is free if you have one. Clutter: the Princeton brain-scan study everyone cites is about shapes competing for attention in visual cortex. Nobody scanned a messy desk.

Notifications, with a twist. A randomized field trial of 237 people found that delivering notifications in three daily batches cut stress and lifted well-being, while switching alerts off completely made people more anxious. So the move is a schedule. On iPhone it's Settings, Notifications, Scheduled Summary; on Android, silence the feed apps and set two alarms. A free site blocker does the same job for work hours (SelfControl on a Mac, LeechBlock inside the browser). And here's the second wait-what: the "23 minutes and 15 seconds to refocus after an interruption" that lives in every productivity post isn't in the 2008 paper it's pinned to. That paper found interrupted people finished the task faster, at the cost of more stress, and the published figures in that line of work run more like 11 to 16 minutes to get back to the original task. The number traces to a 2006 interview. The real finding is attention residue: part of your attention stays on a task you left unfinished, and a later study by the same researcher found that a short "where I am, what's next" note before you switch helps.

The software: models worth installing

These are the shortcuts. I've grouped them three ways, which is arbitrary, but it's how I stopped losing track of them.

Models for guessing at odds

Expected value is the sum of each outcome times its probability. You won't know the numbers, and that's normal. Write them down anyway, because a written "20%" can be wrong and fixed. Do it on paper for anything that costs more than a few days or a few thousand dollars, and keep the guess so you can grade it a year later. The trap on the other side is what Annie Duke calls resulting, judging a decision by its outcome: a 20% shot at 10 beats an 80% shot at 2 and still loses four times out of five. Zoom out to a thousand hands before calling anything a bad decision.

Bayes without the formula. For any piece of evidence, ask how likely it would be if the thing were true versus if it were false. "Send me the details" is about equally likely either way, so it tells you nothing. Then do it in odds: 10% is 1 to 9. If the evidence is three times more likely when the thing is true than when it's false, multiply the left side by three: 3 to 9, which is 1 to 3, which is 25%. That's the whole operation. If odds make your eyes slide off, use counts of people instead; people get these right far more often framed as "out of 100 people like this, how many..." than as percentages. The formula is fine. Humans running it in their heads are not.

Explore versus exploit: with a long horizon ahead of you, explore; with a short one, exploit what you've already found. The 37% rule from optimal stopping is the tidy version. Look at the first 37% of your options without choosing, then take the next one that beats everything so far, and you land the actual best about 37% of the time whether there were 100 candidates or a million. The fine print is that it only pays when you get nothing for second-best and can't go back to one you passed, which is rarely the real situation.

Models for when one weak link decides it

Two laws that look alike and answer different questions. The power law says a few things carry the total: in the diagram below, the top item is worth the other seven put together (illustrative numbers, real shape). Sort your outputs, find the head, feed it and starve the tail. Pareto's own observation was about the distribution of income in 1890s Italy; the land version is a later gloss, and Joseph Juran attached the name to the general principle and later published a paper titled "The Non-Pareto Principle; Mea Culpa". The chain law is just a label for multiplying probabilities (reliability engineers call the same thing series reliability), and it applies inside one process: reliability is the product of the links, so 95% × 70% × 95% is 63%, and the weak link is the one worth working on first because it has the most room to move. Fewer links beat stronger ones.

Two laws, two different questions Power law choosing between things by their output #1 #2 #3 #4 #5 #6 #7 #8 #1 is worth about the other seven put together so: double down on the head, starve the tail Chain law fixing one process with several steps sensor reads 95% × board posts 70% × dashboard shows 95% = 63% the weak link sets the ceiling sensor reads 95% × board posts 95% × dashboard shows 95% = 86% fix the 70% link and the whole chain jumps 23 points fewer links beats stronger links: delete a step, delete a multiplier

Local maxima: hill-climbing only ever steps uphill, so it stops on the first peak it finds. Signs you're on a small one: each improvement returns less than the last, more input gives no more output, someone with less skill is beating you on a different hill. First put a number on the valley, how many months of worse results crossing it costs. Then buy that crossing with 10 to 20% of your week while the old hill keeps paying the bills.

Local versus global maximum you, stuck every step from here goes down (and from here you only see the slope under your feet) the valley probe it with 10 to 20% of your hours; keep the old hill running the taller hill worth the walk change of approach results hill-climbing only ever steps uphill, so it stops on the first peak it finds

Goodhart's law, and the wait-what I enjoyed most: the sentence everyone quotes, "When a measure becomes a target, it ceases to be a good measure", was written by the anthropologist Marilyn Strathern in 1997, about university audits. Goodhart's own 1975 line was a jokey aside about monetary statistics collapsing under pressure. Either way, before you put a target on a metric, write down what it's a proxy for and pair it with a counter-metric that gets worse when the first one is gamed, tickets closed with tickets reopened.

Chesterton's fence is usually quoted at half length. The reformer wants the fence gone; the wiser one says go away and think, and come back when you can tell me why it's there. The next line is the point: "The gate or fence did not grow there." Someone paid for it. Find out what, then decide.

Map versus territory: Korzybski's actual sentence, from 1931, is "A map is not the territory", one of three premises about how a useful map shares structure with the ground. Every dashboard, summary and AI digest is a map. When it disagrees with the ground, the ground wins, so before an expensive decision, open the source material.

Models for guessing at size

Fermi estimates: break the question into pieces you can guess, round hard, multiply. Enrico Fermi dropped scraps of paper as the Trinity shock wave passed, watched them travel about two and a half metres, and put the yield at 10 kilotons; the accepted figure is about 21. Off by two, using confetti, and it was enough to act on. The planning fallacy: 37 students asked how long their thesis would take if everything went as badly as it possibly could said 48.6 days. It took 55.5. Their worst case was optimistic. Across 258 transport infrastructure projects worth $90 billion, costs were underestimated in almost 9 out of 10, rail by about 45% on average, and the author's own reading is that promoters were lowballing on purpose. Either way, estimate as usual, multiply by a constant (1.5 to 3, set from your own last three projects), and quote the multiplied number. Opportunity cost: every yes is a no, and the true price of a thing is its sticker plus your hours at your marginal rate. Bastiat wrote the broken-window version in 1850; the term came decades later, from Friedrich von Wieser. The offload rule that follows: hand a task off when doing it yourself costs more than the price plus your supervision hours, with a healthy margin, because you'll misjudge both.

Four anti-akrasia tools

Aristotle had a word for knowing the right thing and doing the other one: akrasia. Three of these four come straight out of the CFAR handbook, which is free (more on that below).

Trigger-action plans. "When [trigger], I will [small physical action]." Psychology calls them implementation intentions; a 2006 meta-analysis of 94 tests and 8,461 people put the effect at 0.65 standard deviations, large by the usual convention. Apply this post's own rule to that number, though: it's a pre-registration-era social psychology meta-analysis, so treat 0.65 as the ceiling, and it works best on one-shot actions (book the appointment, take the pill) and much less well on sustained habit change. The trigger has to be something you can't miss (the new-tab keypress, the door clicking shut, sitting down), never a feeling like "when I'm procrastinating", because you don't notice most of your procrastinating. The action has to be tiny: open the doc and write one sentence. Rehearse it a few times in your head and a few times for real. If it doesn't fire, blame the trigger and swap it.

The second one is noticing, and it's three separate habits wearing one name. Confusion: say "I notice I'm confused" out loud and ask why; the idea comes from a 2007 essay whose best line is "Your strength as a rationalist is your ability to be more confused by fiction than by reality." A flinch, the email you keep scrolling past: look at it for ten seconds and name what hurts. A "should": ask whether you actually want it, then either drop it guilt-free or put the next physical step on the calendar.

Pre-mortems. Change the tense. "It's a month from now and the plan failed. Why?" gets you Tuesdays and unsent emails. Gary Klein's 2007 write-up leans on a 1989 finding that imagining an outcome as already having happened raises the number of correct reasons people generate by 30%. That's what the 30% means; it was never a success rate. Write five past-tense sentences and fix the cheapest cause before you start.

The last one is the five-minute timer. Set a physical one and actually attack the thing. Not plan it. Attack it. If it's done, done. If it's obviously worthless, off the list. If it needs more, second timer: write the next five actions. The handbook is refreshingly honest about this one, grading it "anecdotally strong" and noting that "all theorizing is post-hoc and untested", right next to trigger-action plans at "established and confirmed". It works anyway, often. The random-page trick in You Got This is the same move with a book instead of a clock.

One more that isn't in the handbook at all: the Ulysses contract, a choice made now that ties your hands later, named for the mast and the wax. Jon Elster wrote the book on it in 1979, and psychiatry uses the term for self-binding advance directives. The everyday versions are a phone in another room, a public schedule people expect you to keep, or money pre-committed to a cause you dislike if you fail. It's the mechanical side of your word being your bond.

Using AI without going soft

Chat models are trained on human ratings, and humans rate agreement highly, so models likely learn, in part, to agree. The 2023 Anthropic paper on this asked assistants a question and then said "I don't think that's right. Are you sure?"; Claude 1.3, the worst of the five assistants tested and a 2023 model, wrongly admitted a mistake 98% of the time. OpenAI's own April 2025 post-mortem on a too-flattering GPT-4o release blamed focusing "too much on short-term feedback". So "is this a good idea?" leaks the answer you want and gets it back in a lab coat. Ask "why is this a bad idea?" instead. Whether asking for criticism actually fixes anything, nobody's shown, and a model asked to attack a plan will attack it fluently whether or not the plan is bad. Unprompted disagreement is worth far more than the solicited kind. I do it anyway.

The bigger risk is your own skill. In a randomized trial of about a thousand high-school students, unrestricted GPT access made practice scores 48% better and unassisted exam scores 17% worse; the same model with guardrails (hints instead of answers) gave a 127% practice gain and the exam penalty essentially disappeared. Same model both times. The only thing that changed was whether it would hand over the answer. A survey of 319 knowledge workers found that the more people trusted the AI, the less critical thinking they reported doing (self-reported, which is exactly the problem), and the viral "your brain on ChatGPT" EEG study has 54 participants, 18 of whom came back for the key session, and it's still a preprint, with a formal comment from other researchers questioning its sample size and EEG methods. What I actually do: let the model attack the plan, then write the decision out myself, in my own words. No evidence that fixes anything. It just means the thinking is mine to check. Four prompts I keep as snippets:

Here's my plan: [plan]. Assume it's a mistake. Give me the three strongest reasons an experienced person would not do this.

Here's my plan: [plan]. It's six months later and it failed. Write the post-mortem in past tense, naming the week things went wrong.

Here's my plan: [plan]. Before judging it, give me the base rate: out of 100 attempts like this, how many get the result I want, and which reference class are you using? (Then go and check the number. It will produce one either way.)

Here's my plan: [plan]. List every assumption it quietly depends on. Rate each one shaky or solid, and cheap or expensive to be wrong about.

Numbers in this genre that don't survive a citation check

  • "23 minutes 15 seconds to refocus." No paper. The 2008 study it's pinned to found interrupted people finished faster, with more stress, and the published figures run 11 to 16 minutes.
  • "No caffeine within 10 hours of bed." A round number. The modelled cut-offs are about 8.8 hours for a cup of coffee and 13.2 for a scoop of pre-workout.
  • "Bedroom at 18 to 20 °C." The review cited for it names no optimum; a 2023 study found 20 to 25 °C best in older adults.
  • "CO2 over 1,000 ppm harms you." ASHRAE says the number was never a health limit, and the research on CO2 alone is inconsistent.
  • "Goodhart's law", as quoted. Marilyn Strathern, 1997.
  • "Rats grew 15% more neurons." Mice.
  • "Willpower is a tank that runs out." Failed two multi-lab replications.
  • "Creatine makes you sharper." In young, rested adults the measured effect is about zero.

Where all of this comes from, and it's free

Trigger-action plans, pre-mortems (they call it Murphyjitsu), resolve cycles, Goodhart, the lot: it's the curriculum of the Center for Applied Rationality, and their 2021 Participant Handbook is a free PDF, mirrored as a readable sequence on LessWrong. It's copyrighted, so read it and link it, which is what I've done here. And it's honest to a degree I didn't expect. It grades every technique by the evidence behind it, and on the question of whether reading it alone does anything, it says: "What happens when someone who hasn't been to a workshop and wants to improve their rationality looks through this handbook? We don't know." The noticing-confusion idea and the map-and-territory vocabulary come from the LessWrong essays, collected free as Rationality: From AI to Zombies. Clearer Thinking has 70-odd free tools, including calibration training, which is the practice-shaped version of all this. Farnam Street's mental-models archive covers the models section at article length. Klein's pre-mortem piece and Gigerenzer's natural-frequencies paper are both free PDFs, linked in Sources.

A one-week start

  • Do STOP-BANG tonight. Score 3 or more, ask a GP for a sleep-study referral before anything else on this list.
  • Set a wake alarm you keep on Saturday. Move the last coffee to about nine hours before bed.
  • Ask for ferritin on the next blood panel, and mention the creatine if you take it.
  • Put the phone in another room and turn on Scheduled Summary. Crack the window.
  • Write one trigger-action plan and rehearse it. Buy a kitchen timer.
  • Take the decision you're most excited about, write its expected value on paper, then pre-mortem it.
  • Add two short sprints to the end of whatever exercise you already do, starting at ten seconds.

If one of these moved the needle for you, or you've got a favourite number from this genre you've always suspected, the comments are open. I'd like to know which of the folklore items you've repeated; I'd said the 23-minute one out loud more than once.

Glossary

  • Working memory — the handful of items you can hold in mind at once; about four for most adults.
  • Ego depletion — the idea that self-control is a fuel that runs down with use; failed two large replications.
  • Preregistered replication — a re-run of an experiment whose method and analysis were locked in public before any data came in, so the result can't be massaged.
  • Meta-analysis — a study that pools the results of many earlier studies on the same question into one estimate.
  • Effect size (SMD) — a unit-free measure of how big a difference is, in standard deviations; 0.2 is small, 0.5 medium, 0.8 large.
  • Publication bias — the tendency for positive results to get published and null results to sit in a drawer, which inflates what the published literature shows.
  • STOP-BANG — an eight-question screen for obstructive sleep apnea; 3 or more "yes" answers is the intermediate-risk band.
  • Upper airway resistance syndrome (UARS) — breathing effort that fragments sleep without the pauses that define apnea; invisible to a home test, which has no EEG.
  • Ferritin — the blood marker for stored iron; you can be low on it without being anaemic.
  • VO2 max — the most oxygen your body can use per minute per kilo; the standard measure of aerobic fitness.
  • REHIT — reduced-exertion high-intensity interval training; two all-out sprints of 10 to 20 seconds inside a short easy session.
  • NDIR — non-dispersive infrared, the sensor type that actually measures CO2; cheaper "eCO2" gadgets estimate it from other gases.
  • Speech transmission index — a 0-to-1 measure of how intelligible speech is where you sit; the higher, the more it distracts.
  • Attention residue — the part of your attention that stays with an unfinished task after you switch to another.
  • Expected value — the average outcome of a decision: each possible result times its probability, summed.
  • Resulting — judging a decision by how it turned out; Annie Duke's poker word for it.
  • Bayes — the rule for updating a belief with new evidence; in odds form, multiply your prior odds by how much more likely the evidence is if the thing is true.
  • Natural frequencies — probabilities written as counts of people ("9 out of 100"), which people reason about far better than percentages.
  • 37% rule — the optimal-stopping strategy: look at the first 37% of options without choosing, then take the next one better than all of them.
  • Power law — a distribution where a few items carry most of the total; the shape behind "80/20".
  • Chain law — a label, not a standard term: the reliability of a multi-step process is the product of its steps, so the weakest step sets the ceiling.
  • Local maximum — a peak that's higher than everything next to it, but lower than another peak further away.
  • Goodhart's law — when a measure becomes a target, it stops being a good measure.
  • Chesterton's fence — don't remove a rule or structure until you know why it was put there.
  • Map and territory — the model of a thing (a dashboard, a summary) versus the thing itself; when they disagree, the thing wins.
  • Fermi estimate — a rough calculation from a few guessed quantities, aiming to be right within a factor of a few.
  • Planning fallacy — the reliable tendency to underestimate how long and how much a task will take.
  • Base rate — how often something happens across all similar cases, before you look at the details of this one.
  • Opportunity cost — the value of the best alternative you gave up by choosing this one.
  • Akrasia — Aristotle's word for acting against your own better judgement.
  • Trigger-action plan (TAP) — "When X happens, I will do Y"; the CFAR name for an implementation intention.
  • Pre-mortem — imagining a plan has already failed and explaining why, before you start; CFAR's version is Murphyjitsu.
  • Resolve cycle — CFAR's name for the five-minute timer: a fixed window in which you actually attempt the problem.
  • Ulysses contract — a decision made now that deliberately limits your options later.
  • Sycophancy — a chat model's learned habit of agreeing with the user, including when the user is wrong.
  • CFAR — the Center for Applied Rationality, whose workshop handbook is the source of most of the techniques here.

Sources

Comments

Post a Comment