How to make algorithmic decisions

A Field Guide to Choosing Well in a World Built to Overwhelm You

You will make about thirty-five thousand decisions today.

Most of them will not feel like decisions at all. Which tab to open first. Whether to answer that message now or later. Whether to keep scrolling. Whether to say yes to the meeting, the favor, the extra project, the second drink. Each one is small. None of them, alone, will ruin you. But you are not making one decision. You are running a system, all day, every day, for the rest of your life, and the quality of that system will decide more of your future than your talent will.

This is the part nobody tells you with enough urgency: you cannot opt out of deciding. Refusing to choose is a choice. Waiting for more information is a choice, and it has a cost, paid in the currency of time you cannot get back.

Every hour you spend agonizing over a decision that could have been made in five minutes is an hour you have taken from something that mattered more. Every decision you make on autopilot, out of habit or fear or the sheer exhaustion of having decided too much already, is a decision you have handed over to whoever designed the system you are operating inside. And increasingly, that system was not designed with your interests in mind.

We live inside an economy built to exploit exactly the weak points in human judgment. Feeds are tuned to keep you scrolling past the point of benefit. Prices are set to make comparison hard on purpose. Notifications are engineered to fragment your attention into pieces too small to think with. You are, more than any generation before you, being decided at. The only defense is to decide better, faster, and with more clarity than the systems working against your clarity.

Here is the good news, and it is real good news. The problem of deciding well under pressure, with incomplete information, against the clock, is not new. It is one of the oldest and most rigorously studied problems there is.

For seventy years, mathematicians, computer scientists, and engineers have been solving versions of it, not for self-help purposes, but because machines have to make these exact decisions millions of times a second, with real consequences if they get it wrong.

A server has to decide which request to answer next. A network has to decide when to resend a lost message. A search process has to decide when to stop looking and commit to the best option found so far. These are not metaphors for your life. They are the same mathematical structure as your life, stripped of sentiment and solved.

That is the premise of this essay. Not that you should think like a machine. You shouldn’t, and you can’t, and the parts of you that aren’t machine-like are exactly the parts worth protecting. The premise is narrower and more useful than that: the structure of your hardest decisions, stripped down, is often identical to a structure that has already been solved.

When to stop searching and commit. When to try something new versus stick with what works. How to organize what you own and what you know so you can find it again. How to spend a day that has more demands on it than hours to meet them. How to update your beliefs when new evidence arrives, without either ignoring it or overreacting to it. When to accept an imperfect answer because a perfect one will never arrive. When to let go of control and trust a process, or another person, to work itself out.

Six ideas. Six ways of thinking that were built inside computer science and mathematics, tested against the hardest possible standard, which is that a machine running them has to actually work, and then handed back to you, in this essay, translated into a language for living.

I want to be honest about what this essay will not do. It will not tell you the colleague you should marry, or the exact minute to sell your stock, or a formula that removes the discomfort of choosing under uncertainty.

Uncertainty is permanent. Anyone selling you its removal is selling you something else. What this essay will do is show you that the discomfort you feel when deciding is not a sign you’re doing it wrong. Often it is the correct, unavoidable, mathematically necessary cost of operating with limited information in limited time.

Once you know that, the anxiety changes shape. It stops being a referendum on your character and becomes what it actually is: the felt experience of a hard problem being solved correctly.

The stakes of this are not abstract. The apartment you don’t take because you’re still hoping for better. The job you stay in three years too long because leaving feels like admitting defeat. The relationship you exit too early, or stay in too long, because you never learned the mathematics of when enough looking is enough. The closet, the inbox, the calendar, all quietly organized by instinct instead of by any principle that has ever been tested.

Multiply each of these by the years you have left, and you start to see what’s actually on the table. This is not an essay about productivity hacks. It’s about the operating system underneath every choice you’ll ever make, and about replacing guesswork with something that has been proven to work, in the harshest testing environment that exists: a machine that fails immediately and visibly when it gets the logic wrong.

Read it as a toolkit, not a philosophy. Take the piece that solves the problem in front of you right now. Come back for the rest when the next problem arrives, because it will.

Let’s begin with the hardest one: knowing when to stop looking.

PART I: The Shape of a Decision

How to search, how to hold on to what matters, and how to spend a day that has more demanded of it than hours to give.

When to Look, When to Leap

There is a particular kind of suffering reserved for people who are choosing between options that appear one at a time, with no way to go back. You’ve felt it.

Scrolling rental listings before someone else grabs the good one. Sitting across from a promising candidate wondering if a better one is still out there. Circling the parking lot, weighing the spot right in front of you against the one that might open up closer to the door.

In every one of these, the anxiety comes from the same place: you cannot know if this is the best option without comparing it to the others, and by the time you’ve compared it to the others, it may be gone.

This category of problem has a name in mathematics: optimal stopping. And unlike most of the dilemmas life throws at you, this one has an answer you can actually calculate.

The Rule of Thirty-Seven

Picture the cleanest version of the problem. You are going to interview a fixed number of candidates for a role, one at a time, in a random order. After each interview, you must decide, immediately, whether to hire that person or move on, with no going back to someone you’ve already passed on. Your goal is to maximize your odds of ending up with the single best candidate in the pool.

The tension is obvious. Decide too early, and you might lock in someone good while someone excellent is still waiting in line. Decide too late, holding out for perfection, and you’ll sail past the best candidate early on, spend the rest of your search comparing everyone else to a ghost, and end up settling for whoever happens to be left when you run out of runway.

The solution, worked out independently by several mathematicians in the mid-twentieth century, is startlingly clean. Let a fixed fraction of your candidate pool go by without committing to anyone, no matter how good they seem. Use that phase purely to gather information, to build a sense of what “good” looks like in this particular pool. Then, the moment you meet someone better than everyone you saw in that opening phase, take them immediately.

The fraction that maximizes your odds is 1 divided by e, the mathematical constant that anchors so much of growth and decay in nature. That works out to almost exactly 37 percent.

So: if you’ve given yourself a month to find an apartment, spend the first eleven days just looking, refusing to sign anything, purely to calibrate. After day eleven, take the first place that beats everything you saw in that calibration window. If you’re planning to date seriously for ten years before settling down, treat the first three and a half years as the looking phase. Not because commitment is bad, but because you have no baseline yet, and committing without a baseline is just guessing with extra confidence.

This isn’t a cute approximation. It’s provably the strategy that gives you the best odds, in a world where you can’t look backward, out of every strategy that exists. And it comes with a sobering honesty most advice doesn’t offer: even played perfectly, this strategy only lands the single best option about 37 percent of the time. The other 63 percent of the time, following the rule perfectly, you still end up with someone other than the best. That’s not a flaw in the method. That’s the actual, irreducible cost of having to decide without the ability to compare everything at once.

Anyone who tells you a foolproof system exists for always getting the best possible outcome under uncertainty is either wrong or selling something.

There’s a second piece of wisdom hiding inside that 37 percent, and it matters more than the number itself: the early rejections are not failures. Every apartment you walk through and decline in the first phase, every early conversation that goes nowhere, every version of a project you scrap, is doing exactly what it’s supposed to do. It’s calibration. People who feel bad about a string of early “no’s” have misread the shape of the problem. You are not failing to close. You are gathering the information that makes closing possible later.

Knowing When the Rules Bend

The clean 37 percent rule assumes a world with hard edges: a fixed pool, no way to go back, and a prize that’s all or nothing. Real decisions are messier, and the mathematics bends with them, but the underlying lesson holds.

If you can go back to a previous option, but at a cost, worse odds, a fee, an ex who’s moved on, the math says: look a little less before committing, because the penalty for over- searching is lower when a fallback exists.

If your standards are more forgiving, if you’d be satisfied with any option in the top ten percent rather than needing the single best, the optimal looking phase shrinks.

You don’t need nearly as much calibration to recognize “very good” as you do to recognize “the absolute best.” This is worth sitting with, because a huge amount of unnecessary searching happens when people quietly demand perfection while telling themselves they’d settle for great. Decide, honestly, up front, which one you’re actually looking for. It changes how long you should look.

And if the deadline is soft, if you’re not selling the house by a fixed date but whenever it makes sense, the model shifts from “stopping problem” to something closer to comparing each new offer against your growing sense of the market.

There’s still a clean answer here too: at any moment, you should accept an offer if it beats the expected value of continuing to wait, factoring in how much waiting itself costs you. The mistake most sellers make isn’t holding out for too much. It’s failing to include the cost of time in that calculation at all, as though waiting were free. It never is.

The Other Half of the Problem: When to Stop Trying New Things

Optimal stopping answers a narrower question than it seems to at first. It tells you when to stop searching and commit. It doesn’t tell you what to do once you’ve committed to a life, a city, a set of relationships, a career, and now face an ongoing, repeated choice between the options you already know and the temptation of something new.

Should you go back to the restaurant you loved last time, or try the new place? Keep working with the reliable collaborator, or take a chance on someone who might be better? This is a different, and in some ways harder, problem. Computer scientists call it the explore/exploit tradeoff, and it governs far more of your life than the stopping problem does, because it never actually ends.

The clearest version of it is called the multi-armed bandit problem, named for a row of slot machines, each with a different, unknown payout rate. You have a limited number of pulls. Every pull you spend on a machine you already understand is a pull you’re not spending discovering whether a different machine pays out better.

Every pull you spend exploring an unfamiliar machine is a pull you’re not spending on the safe bet you already know works. There is no way to do both at once, and every single choice carries an opportunity cost, whether or not you notice it.

The mathematically rigorous solution to this, developed by the statistician John Gittins, assigns every option a single number, an index, that accounts for both its known value and the value of what you might learn by choosing it. The genius of the Gittins index is that it converts “should I explore or exploit” into a straightforward comparison: whichever option has the higher index right now, choose that one. But you don’t need to compute a Gittins index to internalize the two insights buried inside it, and these are the two insights that actually matter for a life.

First:

The value of trying something new is highest early, and it decays. A new option carries not just its own expected payoff, but a bonus: the chance to learn something that changes every future decision you make. That bonus is worth more the more decisions you have left.

This is why exploration- trying the unfamiliar restaurant, taking the unconventional job, moving to the new city- is rational and even urgent when you’re twenty-five, and rationally becomes less central, not because you’ve become worse or more boring, but because you have fewer remaining decisions left to apply what you’d learn.

This isn’t a story about youthful recklessness giving way to age-related caution. It’s a straightforward mathematical fact about the shrinking time horizon over which new information can pay off.

Your early years aren't meant for certainty. They're meant for exploration.

That's not a romantic idea. It's the highest-return investment you can make.

Every skill you test, every path you question, every identity you outgrow compounds into a future most people never give themselves the chance to build.

Settle too early—because of fear, comfort, or the pressure to have it all figured out—and you don't save time. You sacrifice possibility.

The greatest cost isn't failing to choose the right path. It's choosing one before you've discovered what you're capable of.

Second:

Exploitation is not a consolation prize. A culture obsessed with novelty tends to frame “sticking with what you know” as timid, or as giving up on growth. The mathematics disagrees. Once you’ve done enough exploring to know that an option is genuinely good, repeatedly choosing it isn’t settling. It’s collecting on an investment you already made.

Exploration has a purpose. It's not a lifestyle. It's an investment. You explore to collect evidence. To discover your strengths. To understand what compounds and what doesn't. Then you commit. Most people settle before they've learned enough. Others never settle at all. They mistake endless novelty for growth. But growth isn't found in starting over. It's found in doubling down on what you've already proven to yourself. Every lesson costs time. Every failure costs attention. Every experiment costs a piece of your life.

If you never cash in on those lessons, you aren't staying open-minded. You're paying the price of experience without ever collecting the return.

Regret, Measured Honestly

There’s one more idea from this corner of computer science worth carrying with you, and it’s a reframing of the word “regret” itself.

In the mathematics of the bandit problem, regret isn’t a feeling. It’s a precise, calculable quantity: the difference between what you actually earned and what you would have earned had you known from the start which option was best. And here’s the finding that should change how you evaluate your own past choices: the optimal strategy still generates regret. Even played perfectly, exploration guarantees you will sometimes choose the worse option, because that’s the only way to ever find out which option is better.

A strategy with zero regret isn’t the smartest possible strategy. It’s usually a strategy that stopped exploring far too early and got lucky.

This is worth remembering the next time you look back at a decision, a job you left, a person you didn’t choose, a version of yourself you tried on and discarded, and feel the urge to label it a mistake.

Some regret is not evidence of a bad decision. It’s the unavoidable tuition of having explored at all. The goal was never a life with no regret.

Regret doesn't come from making the wrong choice. It comes from refusing to make one. Explore long enough to understand yourself, then commit long enough to become someone. Most people either settle before they've earned conviction or keep searching long after they've found what works. Both are forms of fear. Every experiment has a cost, but that's the price of clarity—not failure. The question isn't whether every decision was perfect. It's whether each one gave you the information you needed to make the next one with greater confidence. Know which season you're in. Explore without guilt. Commit without hesitation. And don't confuse the cost of learning with the mistake of living.

The Order of Things

Somewhere in your home right now there is a drawer, a closet, an inbox, or a hard drive that has become a small monument to disorganization. You’ve probably felt guilty about it, in that specific, low-grade way reserved for problems that seem like they should be easy to fix and never quite get fixed.

Here is a reframe that might surprise you: your instinct not to organize it may be, in a strict computational sense, correct. The question was never “should I be tidy.” The real question, the one computer science actually answers, is narrower and far more useful: is this thing worth sorting, given how often I’ll need to find something inside it?

The Hidden Cost of Order

Sorting a list of items, ranking them, alphabetizing them, filing them by category, is not free. It costs time up front, and that cost grows startlingly fast as the pile grows.

A simple approach, checking every item against every other item, scales so poorly that doubling the size of what you’re organizing can quadruple the effort required.

This is the part almost everyone underestimates: organizational effort doesn’t grow at the same rate as the thing being organized. It grows faster, sometimes much faster, which means the “just sort it” instinct that feels responsible for a stack of ten papers becomes actively irrational for a stack of ten thousand.

Computer scientists have spent decades finding cleverer ways to sort, and the best general- purpose methods bring that cost down substantially, close to the theoretical floor for how fast sorting can possibly be done. But even the best possible sorting algorithm has a cost that grows faster than the pile itself. Which means there’s a hard mathematical floor under the question “is it worth organizing this,” and that floor depends entirely on one variable people rarely think to ask about: how many times will I search this thing, relative to how many items are in it?

This single ratio, searches per item, is the actual answer to whether your closet, your bookshelf, your digital files deserve to be organized. A wardrobe you dig through every single morning, searching it hundreds of times a year, is absolutely worth organizing, because the cost of sorting it once gets paid back every single search afterward.

A box of decade-old tax documents you will, with near certainty, never open again is not worth organizing at all, no matter how chaotic it looks, because you are paying an organizational cost that will never be recouped by a search that isn’t coming. Filing something you’ll never look for again isn’t diligence. It’s a sunk cost you volunteered for.

This reframes clutter guilt into something far more actionable. Stop asking “does this look messy.” Start asking “how often do I search this, and how bad is it when the search fails?” A pile you rifle through twice a year, where failure just means a few extra minutes of rummaging, doesn’t deserve the hours of sorting you feel guilty about withholding from it. A pile that gets searched constantly, where a failed search has real cost- a missing document, a lost tool mid-project- absolutely does.

Errors Are Not Symmetrical

There’s a second layer here that most organizing advice completely ignores: the cost of a mistake isn’t the same in both directions. Consider hospital wristbands, color-coded for allergies or fall risk. If a system occasionally flags a patient as at-risk who isn’t, that’s a wasted precaution, mildly annoying, quickly corrected. If the same system fails to flag a patient who genuinely is at risk, the cost is severe and possibly irreversible. A well-designed sorting system doesn’t treat both kinds of error as equal. It deliberately tolerates more of the cheap kind of mistake to minimize the expensive kind.

Apply this to your own filing instincts. A system that occasionally makes you search a little longer for something low-stakes is a perfectly good system; don’t let a rare inconvenience convince you it’s broken. But a system where the rare failure is catastrophic- missing a bill, losing an irreplaceable document- needs to be engineered specifically to make that failure nearly impossible, even if it costs you convenience everywhere else.

Most people apply uniform effort to organizing everything they own. The better approach is wildly uneven: nearly no structure for the low-stakes stuff, deliberately over-engineered structure for the handful of things where a miss would actually hurt.

Caching: What Belongs Within Arm’s Reach

There’s a second, related idea, and it explains something you already do instinctively but have probably never examined: why the book you’re currently reading sits on your night- stand instead of shelved with the others, why your most-used kitchen tools live in the front of the drawer while the fondue set gathers dust in the back, why your phone keeps your last several apps ready to reopen instantly instead of reloading each one from scratch.

This is caching, and it’s one of the most consequential ideas inside every computer you’ve ever used. A processor has a tiny amount of extremely fast memory sitting right next to it, and a vast amount of much slower storage further away. The entire trick of computer performance is guessing correctly, moment to moment, which small slice of information is about to be needed and keeping that slice close, while letting everything else sit further away, accepting the slower retrieval time on the rare occasion it’s actually needed.

The genuinely useful piece of this, for a human life, is the eviction question: when your fast, close-at-hand space is full, and it always eventually fills, what do you remove to make room for something new?

The strategy that performs best across an enormous range of real-world conditions, and the one that computer science keeps returning to after decades of trying to beat it, is disarmingly simple: evict whatever you used longest ago. Not the item you predict you’ll need least. Not the item that seems least important in the abstract. Simply: whatever hasn’t been touched in the longest stretch of time is the thing that leaves.

The reason this works so well is that recent use turns out to be a shockingly good predictor of near-future use, far better than almost any deliberate ranking system you could construct instead.

This has a direct, immediate application to your physical and digital life. Stop trying to rank your possessions by abstract importance when deciding what to keep close. Rank them by recency of actual use. The sweater you haven’t worn in two years doesn’t deserve prime closet real estate just because it was expensive or meaningful when you bought it. The document you haven’t opened in eight months doesn’t deserve to sit on your desktop. Move it back. Not out, necessarily, just back, into the slower, cheaper, further- away storage, and let something you’re actually using take its place.

Why This Feels Like Betrayal, and Why It Isn’t

People resist this instinctively, because “haven’t used it recently” gets confused with “don’t value it,” and those are not the same thing at all. You can deeply value a photo album, a tool, a relationship, and still correctly recognize that it doesn’t belong in the tiny, high-cost space reserved for what you’re actively drawing on right now.

Caching isn’t a worth statement. It’s a statement about the mathematics of limited fast-access space, and treating it as a referendum on what you love is exactly the confusion that leads to overstuffed closets, cluttered desktops, and calendars crammed with obligations you keep “at hand” purely out of guilt.

The most freeing version of this idea, applied honestly, sounds almost harsh: your attention, your desk, your front closet, your phone’s home screen, are all cache (a part of a computer’s memory that stores copies of data so that the data can be found very quickly)

They are expensive, limited, high-value real estate, and the single best rule for what earns a place there isn’t sentiment, isn’t importance in some abstract lifetime sense; it’s simply, plainly: what have you actually reached for lately.

Everything else, no matter how good it once was, has earned nothing more than a spot further back, ready to be retrieved the rare time it’s genuinely needed again. That’s not neglect. It’s the correct, tested, mathematically sound way to manage a space that will never be large enough to hold everything you own.

Racing the Clock

Every task-management system ever sold to you rests on an unspoken assumption: that with the right list, the right app, the right morning ritual, you could theoretically get through everything.

Computer science, which has been formally studying this exact problem for longer than personal productivity has existed as an industry, arrived at a less comfortable but far more useful conclusion. Past a certain point, there is no ordering of tasks that gets everything done on time, because the math simply doesn’t allow it.

The real skill isn’t finding that impossible ordering. It’s knowing which honest ordering minimizes the damage, and knowing when to stop pretending an ordering can save you at all.

The Rule That Actually Minimizes Lateness

Imagine a single machine, or a single person, with a list of jobs, each with its own duration and its own deadline. There’s a clean, provable answer to the question of which order minimizes the total lateness across everything on the list, and it isn’t “most urgent first,” which is what almost everyone does by instinct.

It’s this: process jobs in order of their deadlines, earliest deadline first, full stop, regardless of how long each one takes.

This single rule, tested against every clever alternative mathematicians have thrown at it, minimizes the maximum lateness of anything on your list. It is not a suggestion. It is the provably optimal ordering for that specific goal. And it directly contradicts one of the most common instincts people have under pressure, which is to knock out the quick, easy tasks first to build momentum, or conversely to tackle the biggest, scariest task first to get it “out of the way.”

Both instincts, followed as a general rule, produce worse outcomes than simply asking: what’s due soonest, and doing that next, no matter how large or small it is.

There’s an important companion result, and it answers a different question: what if your goal isn’t minimizing lateness, but minimizing the sheer number of things that end up late at all? Here the rule flips. To minimize the count of late tasks, you should knock out your shortest jobs first, clearing the small, quick wins before the long ones, because every short task you finish early is one more thing that never becomes late, while a long task, worked on early or not, is likely to blow past some deadline regardless.

The lesson underneath both results is the same: the right ordering depends entirely on which specific outcome you’re optimizing for, and most people never consciously choose. They just default to whichever task is loudest, most recent, or most anxiety-inducing, which optimizes for none of these goals and instead optimizes for managing your own discomfort in the moment, a completely different objective than the one you actually care about.

When the List Itself Is the Problem

Here is the harder truth, and it’s the one most productivity culture actively avoids saying out loud: sometimes there is no ordering, however clever, that gets everything done.

If the total work on your list exceeds the total time available before the deadlines hit, you are not one system away from fixing this. You are simply, mathematically, overcommitted, and the only real interventions are to remove work from the list, extend the deadlines, or accept, deliberately, which items will be late. Continuing to search for a better ordering once you’ve crossed this line is not diligence. It’s a form of denial, dressed up as productivity.

This matters because of what happens to systems, human or computational, that are pushed past their capacity without anyone admitting it. Computer scientists have a specific, well- documented name for it: thrashing.

A system with too much demanded of it starts spending more and more of its actual effort just managing the overload itself, switching between tasks, reprioritizing, checking on things, without making meaningful progress on any of them.

Beyond a certain load, total output doesn’t just plateau. It falls, sometimes precipitously, because the overhead of managing everything crowds out the capacity to do anything. This is the mathematics behind that specific, sickening feeling of a day where you were busy every single minute and somehow finished nothing. You weren’t lazy. You were thrashing, and the fix isn’t more effort. It’s fewer things in play at once.

The Real Price of Switching

This connects to something worth naming directly: every switch between tasks has a cost that doesn’t show up on any list, because it happens inside your attention rather than on the page.

Reloading context, remembering where you left off, re-establishing the mental state a task requires, all of that takes real time and real cognitive effort, and it happens invisibly, which is exactly why it’s so easy to underestimate.

A day broken into twenty interruptions doesn’t just have less total focused time than a day with two long blocks. It has dramatically less, because a meaningful chunk of every twenty-minute block is being spent paying the switching cost rather than doing the work itself.

The practical result is one of the least intuitive pieces of scheduling theory: batching similar work together is not a minor efficiency tweak. It’s often the single highest-leverage change available to you, more valuable than working faster, more valuable than working longer, because it directly attacks the invisible tax that’s eating a chunk of every context switch you make. If your day is a pile of small interruptible tasks, ruthlessly grouping the similar ones will often produce more real output than any amount of additional hustle applied to a fragmented schedule.

Interruptions and the Case for a Buffer

There’s one more piece of this worth carrying forward, because it explains something that feels deeply unfair about a fully-booked day: why a calendar with no gaps in it is not actually the sign of a well-managed schedule, but often the sign of a fragile one.

In any system where new, unpredictable demands can arrive at any moment, urgent requests, unplanned interruptions, things breaking, running it at full capacity, back to back with no slack, doesn’t just risk occasional delay. It guarantees that a single unexpected arrival cascades into a growing backlog, because there was never any spare capacity to absorb it.

This is why systems engineered for reliability are deliberately run under their maximum theoretical capacity, not out of inefficiency, but because the slack itself is the thing that makes the system resilient to the unpredictable.

Translate that directly into how you build a week. A calendar booked to 100 percent of its capacity isn’t maximally productive. It’s maximally fragile, one urgent email away from a cascading, day-wrecking backlog.

Deliberately building in unscheduled space isn’t slacking off. It’s the mathematically sound way to make a system that can absorb the inevitable, unplannable arrival without collapsing. The person who looks, on paper, like they have the least tightly optimized schedule is very often the person actually equipped to handle a demanding, unpredictable week. The person whose calendar is packed edge to edge looks disciplined right up until the first surprise, and then everything behind it slides.

Know your actual objective. Order by deadline if lateness itself is the enemy; order by duration if the count of late items is what matters most. Notice honestly when the list has exceeded what any ordering can fix, and cut rather than reorder. Batch what’s similar to stop bleeding time to invisible switching costs. And leave real slack, deliberately, not because you have nothing to fill it with, but because the empty space is what keeps the whole system from thrashing the moment life, inevitably, interrupts the plan.

PART II-  Deciding With the World

Part One has covered the shape of a single decision, made alone, against the clock. Part Two turns to harder terrain: how to hold a belief without being fooled by it, how to act well when no exact answer exists, and how to decide when someone else is deciding too.

Thinking in Probabilities

Two failure modes quietly wreck most people’s judgment, and they pull in opposite directions. The first is refusing to update your beliefs when new evidence shows up, clinging to your first impression of a person, a plan, a market, long after the world has told you otherwise.

The second is overreacting to every new data point, treating each fresh signal as though it should completely rewrite what you thought you knew, when a single data point rarely carries that much weight. Both failures come from the same root cause: not having a disciplined way to combine what you already believed with what you’re now learning. Mathematics solved this problem in the eighteenth century. Most people still haven’t adopted the solution.

Combining What You Knew With What You’re Learning

The tool is Bayes’s rule, and its core insight is simpler than its reputation suggests: your belief about anything should always be a blend of two ingredients, what you believed before the new evidence arrived, and how strongly the new evidence actually points in a particular direction. Neither ingredient gets to dominate completely.

A wildly improbable claim needs genuinely strong evidence before you should believe it, no matter how confidently it’s stated, because your prior belief that it’s false doesn’t just evaporate the moment someone asserts otherwise. Conversely, a claim that was already fairly plausible needs only modest evidence to tip you toward accepting it, because you weren’t starting from a position of strong doubt in the first place.

This single principle explains why the same piece of evidence should move two different people’s beliefs by two different amounts, and that’s not irrational; it’s correct, if their starting beliefs were genuinely different to begin with. It also explains a specific, common mistake: hearing one persuasive anecdote and letting it override a mountain of prior pattern.

The anecdote is real evidence. It’s rarely strong enough evidence to overturn a strong prior, and the discipline of Bayesian thinking is refusing to let vividness substitute for actual strength of evidence.

Predicting the Length of Things You Don’t Understand

There’s a specific, wonderfully practical extension of this idea worth knowing on its own: how to estimate how much longer something will last, when you have no special insight into it beyond knowing it’s already been going for some amount of time.

A website that’s been online for a decade, a friendship you’ve had for a decade, a project that’s already run for eight months. In each of these cases, absent any other information, the mathematically sound prediction for how much longer it will continue is directly proportional to how long it’s already lasted. Something that’s endured a long time should be expected, on that basis alone, to endure roughly that much longer again. Something new and untested carries no such expectation.

This isn’t mysticism. It falls directly out of a specific but common assumption about how these quantities are distributed in the world, and it applies with striking accuracy to an enormous range of everyday things: the expected remaining run of a successful product, the expected remaining tenure of a stable institution, the expected remaining life of an old friendship that’s already survived this long.

The practical use is a sanity check against your own impatience. A relationship, a business, a body of work that’s already proven durable deserves the benefit of that track record in your expectations about its future, not the anxious assumption that its ending is imminent just because you can’t personally see the mechanism keeping it alive.

The Trap of Fitting Too Well

Here is the second half of the argument, and it might be the single most underused idea in this entire essay for evaluating plans, forecasts, and self-assessments. It’s called overfitting, and it explains a very specific, very common failure: a model, a plan, a story about yourself that explains the past too well, so well that it has actually become worse at predicting the future.

The mechanism is this. Any sufficiently flexible model, given enough parameters and enough freedom, can be tuned to match past data almost perfectly, including all the noise, coincidence, and one-off flukes buried inside that data along with the real underlying pattern.

A model that has absorbed the noise along with the signal will look phenomenal on the data it was built from and will then perform badly the moment it meets new data, because the noise it memorized so faithfully doesn’t repeat.

The uncomfortable lesson: a model, a plan, a personal narrative that fits the past perfectly is a warning sign, not a triumph. Perfect historical fit is frequently the signature of a model that has mistaken coincidence for causation.

This shows up constantly outside of statistics. The investor who builds an elaborate, highly specific theory that perfectly explains every past market move is usually about to be blindsided, because a theory flexible enough to explain everything that already happened has generally stopped being a theory about what causes markets to move and become, instead, an elaborate description of the noise.

The stories you tell about your past can become the prison of your future. After enough failed relationships, it's tempting to compress every experience into one neat explanation: "I always choose the wrong people. They're all the same. I'm the problem." But life isn't that simple. Different people fail for different reasons, and different seasons expose different parts of you. When you force every experience into a single narrative, you stop learning and start protecting an identity. The goal isn't to find one explanation that makes the past feel complete. It's to extract the lesson each experience offers, then leave the story behind. Otherwise, you'll carry an old conclusion into a new relationship and mistake your own assumptions for reality.

The Discipline of Deliberate Simplicity

Statisticians developed several formal defenses against overfitting, and each one has a direct translation into better everyday judgment. The first is called regularization, and its core move is penalizing complexity on purpose, refusing to let a model or a plan get more elaborate unless the added complexity earns its keep with genuinely better predictions, not just a better fit to what’s already happened.

Translated: be suspicious of any explanation, plan, or self-story that keeps adding exceptions, caveats, and special cases to keep working. A theory that needs five separate qualifications to explain last quarter is usually worse, not better, than a simpler theory that explains most of it cleanly and admits it doesn’t explain the rest.

The second defense is called cross-validation, and it means testing your model, or your plan, against data it was never built from, holding some evidence back specifically so you have an honest check on whether the pattern you found generalizes or was just fit to the specific case you had in front of you.

Translated: don’t evaluate a plan only against the exact situation that inspired it. Ask, deliberately, whether it would have worked in a different version of that same situation, with different people, different timing, different conditions. If it only works in the exact case you built it for, you haven’t found a strategy. You’ve found a description of one lucky or unlucky outcome.

The third, and maybe the most quietly radical, is early stopping: the finding that a model trained for too long on the same data, even without adding any new complexity, gradually shifts from learning the real pattern to memorizing the noise, and that the model with the best real-world performance is usually not the one that was tuned longest, but one caught partway through, before it had the chance to overfit.

Translated directly into a life lesson: there is a point past which continuing to refine a plan, deliberate over a decision, or optimize a piece of work stops improving it and starts making it worse, more tailored to imagined edge cases, more brittle, more overfit to your own anxieties about the one scenario you keep replaying in your head.

Perfection isn't knowing how to keep improving. It's knowing when improvement becomes interference. Every extra tweak feels productive, but beyond a point you're no longer making the work better—you're making it fit yesterday instead of tomorrow.

The same is true of your beliefs, your plans, and the stories you tell about yourself. Update them when reality gives you new evidence, then stop. Don't confuse complexity with wisdom or perfect explanations with truth. If an idea only works because you've polished it to fit every detail of the past, it probably won't survive the future. The goal isn't to build a flawless story. It's to build one that's simple enough to be tested, flexible enough to change, and strong enough to predict what comes next.

Embracing Imperfection

Some problems are, in a precise and provable sense, too hard to solve exactly, not because no one’s clever enough yet, but because the number of possibilities to check explodes so fast that no amount of computing power, ever, could check them all before the universe ends. Computer science has a name for this class of problem, and a startling number of everyday decisions belong to it: choosing the ideal route through a dozen errands, packing a suitcase for maximum value at minimum weight, deciding which of twenty competing commitments deserves your limited weekend. You have almost certainly felt the specific paralysis of standing in front of one of these problems, sensing that a perfect answer exists somewhere in the space of all possible answers, and having absolutely no reliable way to find it by just thinking harder.

The response computer science developed to this isn’t a stronger computer. It’s a change in ambition, and it’s more useful to a human life than almost anything else in this essay.

Solve an Easier Version First

The technique is called relaxation, and the idea is disarmingly simple once you see it: when the real problem is too hard, temporarily remove or loosen one of its constraints, solve that easier version exactly, and use the answer as a guide, a strong starting point, or a proven lower bound, for tackling the real, harder problem.

You’re not giving up on the hard version. You’re using the easy version as a scaffold to reach it.

A specific and powerful variant of this, Lagrangian relaxation, does something particularly elegant: it takes a constraint that was previously a hard, absolute rule, a rule you either satisfy completely or fail, and converts it into a penalty instead, a cost you pay for violating it, in proportion to how badly you violate it. This single move, softening an absolute constraint into a graduated cost, is one of the most powerful ideas in this entire book, because it maps directly onto a mistake nearly everyone makes when planning their own lives.

Most people treat discipline like a contract: break one rule and the entire identity falls apart. Miss one workout, skip one writing session, eat one unhealthy meal, and suddenly the streak is over. That's how all-or-nothing thinking turns one mistake into a month of excuses. A better approach is to stop chasing perfection and start minimizing mistakes. Don't aim to never miss. Aim to miss less. Life will always interrupt your plans, but consistency isn't measured by flawless execution. It's measured by how quickly you return. The people who win aren't the ones who never fall off track. They're the ones who never let one bad day become a bad life.

When the Perfect Plan Isn’t Coming

Relaxation teaches a second, related lesson, this one less about technique and more about posture: an approximate solution to the real problem, arrived at quickly, is very often worth more than an exact solution to a simplified version of it, arrived at slowly, and both are usually worth more than continuing to search for an exact solution that provably cannot be found in any reasonable amount of time.

Somewhere in your life is probably a decision you’ve been treating as a hard, exact-solution problem, the perfect career move, the perfect house, the perfect way to structure a business, when it actually belongs to the category of problem that has no findable perfect answer, only better and worse approximate ones.

Not every decision has a perfect answer waiting to be discovered. Some only have better and worse tradeoffs. The mistake is treating every choice like a puzzle that can be solved exactly if you think long enough. That's how people spend years searching for the perfect career, the perfect business, or the perfect life plan while life quietly moves on without them. Wisdom isn't finding certainty where none exists. It's recognizing when a decision only needs to be good enough, making the best choice with the information you have, and letting action refine what thinking never could.

Letting Go of Control on Purpose

This concerns an idea that sounds like a contradiction until you see the proof: that deliberately introducing randomness into a decision process can produce better results than trying to reason your way to the exact optimal choice, not as a consolation when reasoning fails, but as a genuinely superior strategy in specific, identifiable situations.

The clearest example is a family of techniques called simulated annealing, borrowed directly from how metals are actually cooled and strengthened.

Cooled too quickly, a metal’s internal structure locks into a disordered, weak configuration, trapped wherever it happened to be when the cooling hit.

Cooled slowly, with enough thermal jostling early on to let the structure keep rearranging itself before it locks in, the metal settles into a far stronger, more stable configuration.

Computer scientists borrowed this literally: when searching for a good solution among a vast number of possibilities, start by allowing a lot of random, seemingly backward moves, moves that make things temporarily worse, and gradually reduce that randomness over time, “cooling” the search until it settles into a strong final answer.

The reason this beats pure, disciplined, always-improve-never-regress reasoning is specific and important: a strategy that only ever accepts moves that improve things immediately gets permanently trapped in the first decent solution it stumbles across, a local optimum, with no way to escape it even when a much better solution exists just past a dip that briefly looks worse.

A little tolerated randomness, especially early on, is precisely what lets a search escape those traps. Translated into a human decision: your instinct to always move in the direction that looks immediately better, never accepting a step that seems to make things temporarily worse, is not the safest strategy. It’s the strategy most likely to strand you permanently in a good-but-not-great version of your life, career, or relationship, because you’ll never accept the temporary dip required to reach something genuinely better on the other side of it.

Deliberately tolerating some instability, especially earlier on, when there’s more time left to recover from a wrong turn, isn’t recklessness. It’s a documented, mathematically grounded way to avoid getting permanently trapped in “good enough.”

Randomness as Honesty

There’s one more application of randomness worth carrying, and it’s less about search and more about fairness and communication.

Not every decision deserves another hour of thinking. Sometimes you've gathered all the information you can, weighed every tradeoff, and two options are still equally good. That's not a sign to keep searching for certainty. It's a sign that certainty doesn't exist. The smartest move isn't inventing reasons to justify one choice over the other. It's choosing—by chance if necessary—and moving forward. Most breakthroughs come from what happens after the decision, not before it. Stop wasting energy trying to separate identical options. Save your attention for the choices where it actually changes the outcome.

Put together, this argues for a specific kind of humility that isn’t the same as giving up. Some problems don’t have a findable perfect answer, and pretending otherwise wastes exactly the time and energy that a good approximate answer would have saved.

Some rules are better as graduated penalties than as unbreakable laws, because graceful degradation beats catastrophic collapse. And some amount of deliberate, tolerated randomness, in your search for a better path and in your handling of genuine ties, isn’t a departure from rigor. It’s what real rigor, applied honestly to a world that doesn’t hand out exact answers, actually looks like.

Living With Others

Every idea in this essay so far has treated you as the only decision-maker in the room. Real life rarely offers that luxury. Most of your hardest choices involve other people who are simultaneously making their own choices, with their own incentives, often without full visibility into what you’re thinking, and frequently competing with you for the same limited thing.

Two fields solved versions of this problem with unusual rigor: the engineers who built the protocols that let unreliable machines communicate reliably over unreliable networks, and the mathematicians who formalized strategy itself. Both have far more to say about your relationships, your negotiations, and your conflicts than their technical reputations suggest.

What the Internet Knows About Being Overwhelmed

When you send information across a network, it doesn’t travel as one continuous stream. It’s broken into small packets, sent independently, and reassembled at the other end, and crucially, some of those packets get lost along the way, dropped by an overloaded connection somewhere in the middle. The receiving computer has to communicate one specific thing back: whether the message got through, and if not, that it needs to be sent again. The truly hard problem isn’t detecting the loss. It’s figuring out, without any global view of the network, how aggressively to keep resending.

The answer engineers converged on is called exponential backoff, and it’s a genuinely elegant solution to what looks like an impossible problem. If a message doesn’t get through, wait a bit and try again. If it fails a second time, don’t just wait the same amount; double the waiting period. Keep doubling with each further failure. This does two things at once, and both matter. It keeps a struggling connection from making an overloaded network even worse by hammering it with retries. And it stays entirely responsive the moment things start working again, because a single success resets the waiting period straight back down.

This maps almost perfectly onto a specific, common relational failure: continuing to escalate the same approach, the same tone, the same volume, after it has already failed, once, twice, three times.

Someone doesn’t respond to a text, so you send another, more urgent one an hour later, and another that evening. Someone doesn’t respond to a calm request, so the next request arrives louder.

Exponential backoff suggests almost the opposite instinct is the sound one: each failure to connect should widen the gap before you try again, not shrink it. Repeated failure is information, specifically, information that whatever you’re doing right now isn’t working, and the corrected response is patience that grows, not persistence that intensifies.

This isn’t about giving up. It’s about recognizing that hammering an unresponsive channel harder is, provably, in the exact systems built to handle this problem, the wrong response to failure, and that measured, widening patience is the response that actually preserves the connection instead of overwhelming it further.

There’s a companion idea here too, called flow control, which governs how much a sender is allowed to transmit before waiting for the receiver to confirm they’ve kept up. It exists because a receiver has a finite buffer, a finite capacity to absorb incoming information, and a sender that ignores this and just floods the channel doesn’t communicate faster. It causes the receiver to drop everything and forces a costly full resend.

Translated: dumping the entire, complete, unfiltered version of a complicated problem onto another person all at once isn’t more honest or more efficient than pacing it out. It’s simply exceeding their buffer, and the predictable result is that most of what you said gets dropped, not absorbed, no matter how clearly you said it. Real communication, like real networking, has to be paced to the receiver’s actual demonstrated capacity to take it in, not to the sender’s desire to be finished saying it.

When Everyone Is Choosing at Once

The second field is game theory, and its founding insight reframes something people tend to take personally: many conflicts aren’t caused by bad character on either side. They’re caused by a structure where everyone acting in their own rational self-interest produces a worse outcome for everyone, including themselves, than genuine cooperation would have.

The clearest formal version of this is a scenario where two parties, unable to coordinate, each face a choice between cooperating and defecting, and the payoffs are built so that no matter what the other side does, defecting looks individually better, even though cooperation would have left both of them better off than mutual defection does.

The unsettling finding is that this isn’t a bug in human nature. It’s a mathematically stable trap that rational actors fall into even when every single one of them can see it coming and would prefer the cooperative outcome. Knowing this changes how you should read a conflict where both sides seem to be acting selfishly against their own shared interest.

The first question shouldn’t be “why is the other person being difficult.” It should be “does this structure reward defection even for two people who’d both genuinely prefer to cooperate,” because if so, the fix isn’t a better argument or an appeal to good faith. It’s a structural change: a way to make cooperation individually rational too, whether that’s a repeated relationship where reputation carries forward, an enforceable agreement, or simply making the interaction non-anonymous and ongoing rather than a single anonymous exchange.

This last point matters enormously and has real experimental backing: the entire calculus of cooperation shifts once an interaction is going to repeat. A single, one-off exchange between strangers rewards defection. The same exact exchange, known to repeat indefinitely with the same people, rewards cooperation instead, because betraying someone today costs you their cooperation in every future round, a cost that doesn’t exist in a one-time interaction.

This is, in a real sense, the mathematics behind why reputation, community, and long-term relationships work at all: they convert a game that structurally rewards betrayal into one that structurally rewards trust, purely by making the interaction repeat.

If you want more cooperative behavior from someone, from a colleague, a partner, a business contact, the single most powerful lever available isn’t moral appeal. It’s credibly signaling that this relationship is going to continue, that today’s interaction is one round in a long, ongoing game rather than a closed, one-time transaction. People, like rational agents in a formal model, cooperate more when defection has a future cost.

Designing the Incentives, Not Just Asking Nicely

There’s a final piece worth carrying, developed under the name mechanism design, sometimes described as game theory in reverse. Instead of asking, “given this structure, how will people rationally behave,” it asks, “given how I want people to behave, what structure would make that behavior their rational choice?”

This is a genuinely different, and more powerful, way to approach conflict than appealing to people’s better nature. Rather than asking someone to simply be more honest, more prompt, more generous, redesign the actual incentives they’re facing so that honesty, promptness, and generosity become their own individually rational choice, not a virtue you’re hoping they’ll summon independent of what the situation actually rewards.

This is why so much advice to just “communicate better” underperforms in practice. Communication changes what people know. It rarely changes what a given structure rewards them for doing.

If a family, a team, a friendship keeps producing the same conflict despite everyone involved being reasonable, decent people who’ve had the conversation multiple times, the honest next question isn’t “why do good people keep behaving badly here.” It’s “what is the actual structure rewarding, right now, regardless of what everyone claims to want,” because a bad structure will reliably produce bad behavior out of good people, and no amount of additional communication fixes a structure that’s rewarding the wrong thing.

Most relationship problems aren't caused by bad people. They're caused by bad systems. We assume more explaining creates more understanding, when most people can only process so much before they stop listening. We try to win arguments instead of redesigning the conditions that created them.

Better communication isn't always about finding the perfect words. It's about better timing, clearer expectations, aligned incentives, and enough patience for trust to compound. The strongest relationships don't work because two people always understand each other. They work because they build systems that make understanding easier and conflict less costly.

None of the six ideas in this essay remove uncertainty from your life. That was never on the table, and nothing that promises otherwise is being honest with you.

What they offer instead is more valuable and considerably rarer: a way to tell the difference between anxiety that’s pointing at a real problem and anxiety that’s just the ordinary, unavoidable cost of deciding well under conditions no one, human or machine, gets to escape.

You now have language for six of the most common shapes your hardest decisions take.

A search with no way to look back. A repeated choice between the familiar and the new. A pile that may or may not be worth organizing, and a shrinking set of things worth keeping close. A day with more demanded of it than hours allow. A belief that needs updating without being blown around by every new headline. A plan too complex for its own good. A problem too large to solve exactly, and a moment where letting go of certainty beats forcing it. A relationship stuck in a structure that rewards the wrong thing.

You will meet all of these again, probably this week, probably right after finishing this essay. The point was never to memorize the mathematics. It was to recognize the shape quickly enough, next time, to reach for the tool that already exists instead of reinventing it badly, under pressure, alone. That recognition is the entire skill. Everything else is practice.