The Manual for When the World Keeps Changing

Staying steady when you can't predict what's coming

Manual and analytical companion · a plain scaffold for fast, unpredictable change, with the reasoning underneath

Andraž Đurič

Version 1 · June 2026

This comes in two parts. The first is a short manual anyone can use, with no special vocabulary. The second is the reasoning underneath it, for readers who want the mechanism. If you only need something usable, the first part stands on its own, and there is a line that tells you where it ends.

Part one: the usable manual

The world keeps changing, and it keeps changing faster than it used to. New tools, new rules, new ways of living arrive before the last ones have settled. Most of us try to cope by guessing what comes next, and most of us guess badly, because we tend to picture the future as a slightly faster version of now. It rarely is. You cannot fix this by predicting harder. What you can do is become the kind of person who stays steady when you do not know what is coming. That is what this part is for. It will not tell you what the future holds. It is a way to keep your footing while the ground keeps shifting.

When change comes faster than feels manageable, people tend to break in one of two ways. Some lock up. They stop letting new information in, decide they already understand how things are, and treat anyone who disagrees as wrong or stupid or not worth hearing. Others burn out. They try to keep pace with everything at once, every update and every alarm, until they are frayed and exhausted and reacting to noise. Both are the mind trying to protect itself from being overwhelmed, and both leave you worse at handling what is in front of you. The aim is to catch yourself sliding toward either one.

Your mind will not tap you on the shoulder and tell you it is overwhelmed. It shows up instead as small shifts in how you think and react. Three are worth watching for.

The first: nothing surprises you anymore. You have not been caught off guard or clearly proven wrong in a good while, and everything that happens seems to confirm what you already thought. That usually does not mean you have it all figured out. It means you have quietly stopped letting in anything that would not fit.

The second: everything looks black and white. People are with you or against you. A situation is a triumph or a disaster. The in-between, where most real things live, feels annoying or tiring to think about. That usually means you are low on energy. Holding a complicated picture in mind takes fuel, and when you are running down, your mind cuts the corners.

The third: small things make you furious. A page that will not load, a dropped fork, someone slow to move at a green light, and the size of your reaction is all out of proportion to the cause. That usually means you have no slack left. You are spending so much on the big worries that there is nothing in reserve to absorb the small frictions, so they go straight through you.

When you notice one of these, the move is not to think your way out. The thinking is the part that is glitching, and running more of it will not help. Change what you are doing instead.

If nothing surprises you, go spend real time with something outside your usual world. A subject you know nothing about, music you would normally skip past, a way of seeing things you do not share. There is one rule: do not argue with it while you are there. Do not rank it, fix it, or fit it into what you already believe. Just let it be unfamiliar for a while. That reopens the part of you that had quietly shut.

If you are seeing only black and white, or small things are setting you off, take a real break from taking things in. Fifteen minutes with no phone, no feed, no planning. This is not laziness and it is not optional. Your mind runs on a body, and the body is low. You are not slacking off. You are charging.

One more thing, because it is easy to get wrong. Staying open does not mean believing things quickly. The two get mixed up, and the mix-up can cost you. Be willing to change your mind, but be most careful with the explanations that fit a little too perfectly, and above all with the ones that also happen to be pushing you somewhere. When a story is unusually clean, or unusually flattering, or it turns up right when someone wants something from you, that is exactly the moment to slow down. The point is not to trust nothing. It is that being willing to change your mind and changing it fast are two different skills. In a world full of people and systems built to move you, the second one without the first is just being led around.

Before you throw yourself back into the fast lane, a short check. You are steady again when a different opinion makes you curious instead of defensive, when your first thought is some version of "huh, why do they see it that way" rather than a flash of irritation. When you can leave a stressful thing behind and be present at a meal or in a conversation, instead of running it over and over in your head. And when you can look at a messy problem that has no clean answer without needing to force it into a neat box. If those feel out of reach, you are not recharged yet. Give it another round before you decide anything that matters.

One honest limit. If what you are dealing with is heavier than overwhelm, if low mood or dread settles in and does not lift when you rest, a manual is not what helps, and that is no failure on your part. Talk to someone you trust, or a professional. This page is for a hard stretch; if it is more than that, you deserve more than a page.

You do not have to work out where the world is heading to get through today well. Watch for the signs that you are locking up or running dry. When you see them, step outside your bubble or step away from the noise, and come back when you can hold the whole messy picture again. Stay open without being easy to push. Take care of what is in front of you. The future will still be unpredictable tomorrow, and you will be better able to meet it with your feet under you.

You can stop here. That is the whole usable manual, and it works without anything below it. What follows is the reasoning underneath: why these particular signs, why these particular moves, and where the hard limits are. You do not need it to use what you have already read.

Part two: the reasoning underneath

The plain manual above makes a set of claims without defending them: that prediction fails in this kind of environment, that those three signs point to a particular kind of trouble, that those moves help, and that openness has to be paired with scrutiny. This part is the defense. It is written in the register of the rest of this blog, and it is honest about where its analogies are only analogies.

The thesis

For a finite agent in an environment whose specific future states it cannot reliably predict, persistence and effective action are better served by optimizing the architecture of the observer than by predicting the content of the future. Content-prediction degrades as the rate of change and the novelty of the environment rise: the further out and the more unprecedented the target, the worse a finite model does. Certain architectural properties do not degrade the same way. The capacity to detect when your own model is failing, to recover the substrate that runs it, and to revise beliefs at a rate matched to the quality of the evidence keeps its value across a wide spread of possible futures. The strategy is therefore structural, not predictive. It does not tell you what is coming. It makes you a system that survives not knowing.

A note on framing. I am not resting this on the claim that change is exponential. That claim is stronger than the argument needs and stronger than the evidence supports across domains, where a great many processes are logistic and level off rather than running away. The weaker premise is enough: change is fast enough, and often novel enough, that prediction of specifics is unreliable. The well-documented human tendency to underestimate compounding, sometimes called exponential-growth bias, is real and relevant, but it establishes a bias in our estimates, not a guarantee that every process that matters is exponential.

The limits of a finite updater

The strategy has hard boundaries, set not by character but by what a finite information-processing system can do. Three of them matter here. Naming them is part of the strategy rather than a concession against it: a system that believes it has no boundaries stops watching for the ones it has.

Limit one, convergence on a consistent falsehood. A faithful Bayesian updater fed a stream of data that is internally consistent, accurate in its short-term predictions, and false will update toward the false model. Once it has settled there, real data from the actual territory arrives looking like noise, measured against a refined and confident wrong model. The optimization image, a local minimum the system slides into and cannot climb out of, is an analogy rather than a mechanism, but it names something real: honest updating on evidence that is consistent by accident or by design is a route into a stable, self-reinforcing, wrong state.

This limit has a consequence that most advice about adaptability gets backwards. The intuitive prescription is to shed your priors and update as fast as possible, on the theory that rigidity is the enemy. But a system with weak priors is the most capturable system there is. It has little to weigh against a clean and consistent false stream, so it converges on it quickly and with confidence. Strong, well-calibrated priors are exactly what let a system throw out elegant garbage. The variable that matters for persistence is therefore not the update rate but the update calibration: the speed of revision should track the quality of the evidence and the degree to which its source is adversarial. In an environment increasingly full of systems optimized to produce locally consistent, persuasive content, raising your update rate without raising your scrutiny is not adaptability. It is the mechanism of capture. Holding well-tested priors against unearned revision is not rigidity. It is the only defense this limit permits. This is also the precise reason that pressure, on its own, is not a reason to change a conclusion: pressure raises the cost of holding a belief without adding any evidence against it, and a calibrated updater does not move along that gradient.

Limit two, substrate-dissolution. Under the thermodynamic view this blog develops, an information pattern needs a physical substrate to persist, and the agent is not separable from the hardware that runs it. Some facts about the territory carry a verification cost that exceeds the agent's survival budget: to register them faithfully, the system would have to undergo an interaction that destroys the substrate doing the registering. Those facts are unupdatable in a strict sense, not because the agent is closed-minded but because acquiring the datum ends the agent. This is the cleanest of the three limits and the most boundary-like. No epistemic virtue moves it. Its practical content is modest but real: an adaptability that ignores its own survival constraint is not maximal, it is incoherent, because a pattern that has dissolved updates nothing.

Limit three, bandwidth. A finite agent takes in and integrates information at a finite rate, bounded by its substrate. An environment can change across more variables, and faster, than the agent can track, and when it does the agent's model lags, perceiving a slower and partial pattern in place of the faster real one. It is tempting to formalize this with the Nyquist-Shannon sampling theorem and to call the lag a case of aliasing. I will not lean on that. Nyquist-Shannon concerns reconstructing a bandlimited continuous signal from evenly spaced discrete samples; a cognitive system is not a fixed-rate converter sampling a bandlimited signal, and importing the theorem would borrow its precision without earning it. The honest claim is weaker and still sufficient: a finite agent can be outpaced by a sufficiently fast, high-dimensional environment, and when it is, its model degrades into a coarse approximation that misreads the real dynamics. The Nyquist picture is at most a loose analogy for that, and is flagged as one.

The usual mitigation is externalization: offloading high-frequency tracking onto tools and infrastructure, and keeping the agent's own limited bandwidth for the high-level invariants. This does raise effective bandwidth. It also creates a dependency that loops straight back to the first limit. The external system becomes part of the agent's evidence stream, and if it is compromised, captured, or has itself converged on a consistent falsehood, the agent inherits the error, now with the extra confidence that comes from having outsourced the work. Externalization buys bandwidth by widening the surface that the deception limit can attack. The mitigation is not free, and treating it as free is its own failure mode.

The operating loop

The plain manual's shape, watch the signs, run a reset, check that you are back, is a control loop: detect, intervene, verify, with the loop turned on the agent's own state rather than on the external world. The detection layer reads behavioral telemetry rather than introspective report, because under load the conscious narrator is the last subsystem to notice the failure and the first to explain it away. The intervention layer is action-based rather than cognitive, because a loop that has already failed cannot be repaired by running more of the computation that failed. The verification layer tests against the parameters that failed rather than against felt relief, because relief is itself something a closed, comfortable, wrong model can produce.

All of it runs on a substrate with a fixed power budget, roughly twenty watts in the human case. Holding a high-resolution, internally consistent model has a real metabolic cost. When the substrate is depleted the system economizes on its own, without asking. It drops resolution, which is the black-and-white symptom. It fails to stand down active modes, which is carrying work-mode into rest. It loses the surplus that normally suppresses minor external entropy, which is why small frictions start provoking large responses. The plain warning signs are the felt surface of that economizing. This is also why the manual treats sleep and physical stability as hard requirements and not as lifestyle preferences. They are the conditions under which the loop can run at all.

The instrumentation paradox

There is a tension running through the whole apparatus, and it is better named than buried. This manual instruments the self. It asks you to read your own telemetry, run protocols, and verify your state. But one of the failures it warns against is exactly the over-instrumenting of experience, the use of the modeling engine as a shield against living anything directly and unmeasured. A manual that asks you to run a verification protocol before you are allowed to enjoy a meal is performing the disease it diagnoses.

The resolution is a scope limit on the apparatus itself. The instrumentation is scaffolding for when the loop is already failing, not a permanent mode of operation. A healthy system runs these checks rarely, mostly at transitions, and mostly below the level of conscious effort. Running diagnostics on every ordinary moment is not high fidelity. It is the map-as-shield failure dressed up as rigor. The brevity of the plain manual is not a simplification of this point. It is the correct default. The apparatus is the exception you reach for when the default has broken, and then you put it down again.

Borrowed frames, marked

Two framings from the surrounding literature are useful here and load-bearing nowhere, and it is worth saying which is which. The Free Energy Principle offers a clean picture of an agent as a model-driven system minimizing prediction error, and I have used its vocabulary loosely. It is a productive frame, not a settled result, and nothing above depends on its formal machinery. Gödel's incompleteness theorems and Turing's halting problem are sometimes invoked as models for an agent mapping the boundary of its own competence. The intuition is attractive and I have drawn on it, but those results concern formal systems and computation, not finite biological agents tracking a physical world, and the resemblance is structural at best. Where this manual needs a claim about self-bounded knowledge, it rests on the finite-agent limits above, not on those theorems.

Scope

What this is, and what it is not. It is a structural strategy for raising the probability of persistence and effective action across a wide range of unpredictable futures. It is not a global optimizer, and it does not pretend to be one. Consistent with the larger framework this blog develops, no such optimizer is available; that absence is a result rather than a gap. The strategy carries no guarantee of thriving. What it offers is a better expected position than its alternatives, which are rigid lock-in, undisciplined updating, and unmonitored burnout. The realistic target is the same one the rest of this work holds to: less wrong over time, with no promised convergence.

And its relationship to the plainer piece this blog already carries. The first half here is the same kind of object as that earlier manual: a floor to stand on, usable without any of the machinery beneath it. The only difference is that here the machinery is written down underneath it. If you are not in a bad hour and you want the why, the why is above. If you are in a bad hour, the first half is enough on its own, and so is the earlier manual, and so, more than either, is a person.

Written by Andraž Đurič. The framing began in an exchange with Google's Gemini, was developed through conversations with several AI systems, and was finalized in collaboration with Claude (Anthropic). The first part is written in a plainer register than the rest of this blog and for a different purpose, as its closing note explains; the second part is its analytical companion. Licensed CC BY 4.0. Share it freely.