For some time, I have been trying to find better language for a change in how I experience my own mind. Some of that change has emerged through meditation and through studying ideas from Yoga philosophy. Some of it has come from psychology, neuroscience, game theory, and simply paying closer attention to what happens when my mind encounters uncertainty.
I kept noticing something that was difficult to describe. A situation would happen. There would be some information available to me, but often not very much. My mind would produce an interpretation of what was happening, what another person was thinking, or what was likely to happen next. Previously, the interpretation and the thing itself could feel almost indistinguishable. Now I increasingly notice a gap between them.
There is what happened. There is what I think it means. There is what I predict will happen next. There is how I feel about that possibility. And there is what I decide to do because of it. Those things can arrive so quickly that they appear to be one experience, but they are not the same thing.
Trying to understand that distinction led me into an unexpectedly rich intersection of predictive processing, Bayesian inference, game theory, self-fulfilling prophecies, and the Yogic concepts of the kleshas, viveka, vairāgya, and saṃskāras. The thread connecting all of them is surprisingly simple:
You cannot know the future. You can model it.
That distinction turns out to matter enormously.
Perception is already a prediction
We tend to imagine perception as a camera. There is a world outside us, our senses record it, the brain receives the recording, and then tells us what is there. Something like:
world→senses→brain→perception
But the brain does not have direct access to the world. It receives signals. Light hits the retina, pressure changes reach the ears, receptors fire, and from these incomplete and noisy signals the brain has to infer what is probably happening outside it.
One influential family of theories known as predictive processing or predictive coding turns our intuitive picture of perception partly upside down. Rather than waiting passively for sensory information and constructing reality from the bottom up, the brain is understood as continuously generating models of what is likely to be causing its sensory input. It predicts, compares those predictions with what arrives through the senses, and uses the mismatch, the prediction error, to update its model.
Roughly:
prediction→sensation→prediction error→model update
Imagine walking through your house at night. At the end of the hallway is a dark vertical shape. Your visual system does not receive a label saying “COAT.” It receives ambiguous information, and your brain has hypotheses: a coat, a person, a shadow, a piece of furniture. If you have just heard an unexplained noise downstairs, “person” may suddenly become more plausible. Then you turn on the light. The additional sensory information strongly favours “coat,” and your perception updates.
This is not necessarily conscious reasoning. You do not ordinarily experience a spreadsheet of hypotheses being recalculated in your head. You simply experience: Oh, it’s my coat.
Inference does not necessarily feel like inference. It can feel like perception.
The mind has to decide what evidence to trust
There is another piece of the predictive-processing picture that becomes especially useful when thinking about psychological experience: precision.
Suppose your brain predicts one thing, but sensory evidence suggests another. How much should the brain update? That depends partly on how reliable the evidence appears to be. If you see something person-shaped through thick fog, the visual signal is noisy. If you see the same figure five metres away in bright daylight, the sensory signal deserves much more confidence.
In predictive-processing theories, this kind of confidence or reliability can be described in terms of precision weighting. The important point is straightforward: it is not only the hypothesis that matters, but how much confidence the system gives the hypothesis and the evidence opposing it.
This distinction appears everywhere outside literal vision. Compare “Maybe they’re annoyed with me” with “They’re annoyed with me.” The underlying hypothesis could be identical. What has changed is its felt certainty. At sufficiently high subjective confidence, the prediction stops feeling like one possible explanation and begins feeling like something you can simply see.
That is where uncertainty becomes psychologically interesting.
What happens when there is almost no information?
Imagine two people are playing some kind of strategic game. The other person has secretly made a choice. You cannot see it. Perhaps they have cooperated; perhaps they have defected. You receive no additional clue.
The external evidence available to several observers could therefore be identical: unknown. Yet their subjective experience of that uncertainty can be radically different. One person thinks, “I don’t know what they chose.” Another thinks, “They’re probably going to screw me.” Another thinks, “They’re a decent person; they’ll cooperate.” Another thinks, “They probably expect me to assume they’ll cooperate, so maybe that’s exactly how they’ll exploit me.”
Nothing outside the person has changed. The evidence is still unknown. Internally, however, an enormous amount may have happened.
This points toward a distinction I have found increasingly useful: information gain versus simulation gain.
Information gain means that I discovered something new about the world. Simulation gain means that my mind generated another possible model of the world. These can feel surprisingly similar.
I might begin with, “Perhaps they will betray me.” Then I think, “And if they know that I trust them, that would make betrayal easier.” Then: “Perhaps they’ve been unusually nice precisely because they want my guard down.” Twenty minutes later, I may feel dramatically more convinced that something is wrong, but the external dataset has not changed. At the beginning, the evidence was X. At the end, the evidence is still X. What has increased is the number and vividness of internally generated simulations.
The mind has been talking to itself.
Possibility quietly becomes probability
This is one of the easiest perceptual transformations to miss. The mind correctly notices: this could happen. But somewhere along the way, possible becomes plausible, which becomes probable, which becomes expected, which becomes obvious.
The strange thing is that these transitions may contain almost no new external information.
This matters because imagining something can itself produce an emotional response. The emotional response can then become another apparent piece of evidence: “Why would I feel this uneasy if something wasn’t wrong?” But the unease may have been produced by the simulation.
The system can therefore start recursively validating itself. I imagine a threat. The imagined threat creates fear. The fear makes the threat feel more credible. The increased credibility produces more simulation. The simulation creates more fear. At no point does the external world have to contribute anything new.
Prediction itself is not the problem
At this point it is tempting to conclude that prediction is the problem and that the solution is to stop trying to predict things. That cannot be right. Prediction is essential to functioning.
Someone throws a ball at you. To catch it, your nervous system must estimate where that ball is going to be moments from now. You are implicitly sensitive to speed, direction, distance, spin, gravity, wind, your own body position, and the time it will take to move your hand. You do not know exactly where the ball will be. You generate an extraordinarily good estimate.
But there is a reason that estimate is good. You have spent a lifetime interacting with moving objects. Again and again, the loop has been:
predict→act→observe→correct
Every mistake contains information. Every successful catch contains information. Over time, the model becomes calibrated.
That word matters. A good predictor is not someone who possesses certainty about the future. A good predictor is someone whose confidence approximately corresponds to the reliability of their model.
Someone working professionally in commodities might do something similar on a vastly more complex scale. They may know that particular weather systems, growing regions, inventories, transport constraints, and seasonal patterns tend to affect certain crops in particular ways. After observing enough cases, and ideally studying actual data, they may become better than average at estimating what certain conditions imply.
They still cannot see the future. They have constructed a more sophisticated model of it.
The goal is not certainty. The goal is calibration.
But calibration requires being wrong
This is where things become uncomfortable. A model can only improve if reality is allowed to disagree with it.
The loop needs to remain open:
prediction→reality→prediction error→update
Suppose instead that encountering uncertainty produces intense distress. The process can become:
prediction→distress→avoidance
Now something crucial has happened. The person has escaped the discomfort, but they have also escaped the data. The very thing capable of correcting the prediction has been removed.
This produces a paradox: anxiety can make prediction feel more necessary while simultaneously making accurate prediction harder to learn.
If I repeatedly avoid situations in which I might discover that my fears were wrong, my model receives very little corrective information. If I only enter situations after constructing elaborate safety behaviours, another problem appears.
Suppose I believe, “If I don’t prepare obsessively, everything will fall apart.” So I prepare obsessively. The event goes fine. What does my brain learn? Potentially: “Thank God I prepared that much.”
But there is an invisible problem. I cannot observe the alternative universe in which I prepared normally. Perhaps that meeting would have gone equally well after one hour of preparation instead of six. I never receive that data.
When control starts to resemble superstition
This is where attempts to eliminate uncertainty can become surprisingly similar to superstition.
Imagine uncertainty produces anxiety. I begin planning. If A happens, I will do B. If B happens, I will do C. I check something. I rehearse the conversation. I construct backup plans. I search for more information. Twenty minutes later, I feel calmer.
There are at least two possible explanations. Perhaps my planning genuinely reduced an important external risk. Or perhaps planning occupied my attention and created a subjective feeling of control, reducing my anxiety even though the external probability barely changed. Both possibilities can coexist.
Planning is obviously useful. But if I repeatedly interpret anxiety reduction as evidence of risk reduction, I can learn the wrong lesson.
The behaviour is reinforced because it removes an unpleasant internal state. The mind learns: “When uncertainty appears, do this.” The next time uncertainty appears, the urge arrives faster. The groove becomes deeper.
And because I cannot access the counterfactual world in which I did nothing, the belief can be remarkably difficult to disprove. This is one reason superstitious behaviours are so sticky. You perform the ritual, the feared event does not happen, and the absence of disaster appears to validate the ritual. But perhaps the disaster was never going to happen.
Being right does not prove that your model was good
There is another trap here. Imagine I predict: “This person will betray me.” Eventually they do something I consider betrayal, and I think: “I knew it.”
But several different things could have happened. Perhaps I noticed subtle but genuinely predictive information. Perhaps betrayal was already statistically common in that environment. Perhaps I guessed correctly. Perhaps I remember the successful prediction more vividly than the many predictions that failed. Perhaps I interpreted an ambiguous event through the lens of the original belief. Perhaps my own defensive behaviour changed the relationship.
All of these could produce the same subjective conclusion: “I was right.”
But the event occurring and the prediction process being reliable are two different claims.
This is easy to miss because we cannot replay reality with one variable changed. If you predicted rain and it rained, that alone tells us surprisingly little about how good your forecasting process is. We need repeated forecasts. We need misses as well as hits. We need base rates. We need to know how confident you were. We need to know whether you update when you are wrong.
Calibration is a property of a pattern, not a triumphant memory.
And then we add another person
So far, we have imagined a person trying to model a mostly passive world. Things become considerably stranger when the thing you are modelling is another agent who is modelling you back.
This is where game theory becomes useful. Game theory studies strategic interactions: situations in which the best action for one participant can depend on what other participants do.
Consider the famous Prisoner’s Dilemma. Two players choose whether to cooperate or defect. In the classic one-shot form, defecting gives each individual a better payoff regardless of what the other player chooses, even though mutual cooperation would leave both better off than mutual defection.
The interesting lesson is not merely that people can be selfish. It is that individually sensible behaviour can create collectively bad outcomes.
But real relationships introduce another layer. People usually do not know perfectly what everyone else intends. So now I am not merely deciding. I am estimating what you will decide, while you are estimating what I will decide.
Perhaps I think, “You don’t trust me,” so I become guarded. You perceive my guardedness and think, “Something is off; I shouldn’t trust them,” so you become guarded. I observe your guardedness, and now I possess what appears to be new evidence: “I knew you didn’t trust me.”
This is more interesting than simple confirmation bias. The prediction has begun to participate in creating the evidence that confirms it.
The self-fulfilling prophecy, and the self-maintaining prediction
Sociology has a familiar term for one version of this: the self-fulfilling prophecy. A belief or expectation can lead people to behave in ways that help bring about the expected outcome.
But I also like the phrase self-maintaining prediction. It captures something slightly broader.
A prediction does not always have to create the exact event it forecasts. Sometimes the system merely organises perception and behaviour in ways that prevent the prediction from being properly tested.
Consider the belief: “People eventually abandon me.” Someone stays. The model responds, “For now.” They continue staying. “They’re an exception.” Five years later the relationship ends. “There. I knew it.”
The prediction has found a way to survive every possible observation. Positive evidence counts; negative evidence is reinterpreted. The model no longer behaves like a prediction. It behaves like a worldview.
A powerful question in these situations is:
What evidence would actually make me believe this less?
If the sincere answer is “nothing,” then I am no longer testing a hypothesis. I am protecting one.
Active inference makes this even stranger
Predictive processing becomes particularly interesting when extended toward active inference.
The simplified idea is that an organism does not only change its model to fit sensory information. It can also act in ways that change the sensory information it receives.
Imagine I expect to be holding a cup. I look at my hand. There is no cup. One way of reducing the mismatch is perceptual: “My prediction was wrong.” Another is behavioural: I pick up the cup. Now my sensory state more closely matches the expected state.
Real active-inference theory is considerably more sophisticated than this simple illustration, but the important intuition is that perception and action belong to the same loop: organisms infer the world while simultaneously acting upon it.
Now return to the interpersonal example. My prior, “people are dangerous,” produces a prediction: “This person may exploit me.” That produces an action: guardedness. The other person responds with guardedness. That creates new sensory evidence: “Look how cold they are.” The original model then strengthens: “People are dangerous.”
We now have a loop:
prior→prediction→behaviour→changed environment→confirming evidence→stronger prior
From inside the loop, it can feel like excellent perception. From outside the loop, it may look more like a dynamical system repeatedly returning to the same attractor.
Two frightened people can manufacture a dangerous world
Game theory gives us an especially dramatic version of this. Imagine two countries. Country A thinks: “Country B might attack us,” so A builds more weapons. Country B observes A’s military expansion and thinks: “A is becoming dangerous,” so B builds more weapons. A observes B’s response and concludes: “Exactly. We were right to be worried.”
Neither country needs to have begun with the intention of attacking. Defensive action can itself generate the evidence that makes defensive action increasingly rational. Political science calls this kind of dynamic a security dilemma.
The uncomfortable lesson is that a prediction can be simultaneously understandable, partially accurate, and causally involved in creating the world that validates it.
So the important question becomes more complicated than “Was I right?” We also need to ask:
What role did my model play in producing what I later treated as evidence for the model?
Designing the game instead of merely playing it
This is where another field, mechanism design, adds a useful perspective.
Game theory often asks: given these rules and incentives, what will the players do? Mechanism design reverses the question: what rules could we create so that self-interested players tend toward a desirable result?
The relevance to perception may not be obvious at first, but consider an organisation in which people routinely hide bad news. Management might explain this psychologically: “People need to be more transparent.”
Perhaps. But suppose reporting a problem reliably causes the messenger to be blamed, while concealing a problem allows them to appear competent. The organisation has constructed a game in which secrecy is rational.
The resulting behaviour may tell us less about people’s intrinsic honesty than about the environment selecting their behaviour.
This becomes another useful perceptual correction:
Before explaining behaviour entirely through personality, inspect the game.
Sometimes what looks like a character flaw is an equilibrium.
An ancient vocabulary for “colouring”
While thinking through all of this, I kept returning to an older vocabulary I had been studying through Patañjali’s Yoga Sūtras.
I do not mean that Patañjali secretly discovered predictive processing two thousand years ago. That kind of retrofitting does neither tradition much justice. The frameworks have different purposes, assumptions, and metaphysics.
But they sometimes provide remarkably complementary ways of looking at subjective experience.
Patañjali describes five kleshas, usually translated as afflictions, obstacles, or “colourings” of the mind:
- Avidyā: ignorance or misapprehension; seeing something as what it is not.
- Asmitā: “I-am-ness” or ego-identification.
- Rāga: attraction, attachment, or grasping toward what is desired.
- Dveṣa: aversion or pushing away what is unwanted.
- Abhiniveśa: clinging to life or continuity, traditionally strongly associated with fear of death.
I find the metaphor of colouring particularly useful.
Imagine the raw situation: someone has not replied to a message. That is information.
Then: “They are annoyed with me.” That is interpretation.
Then: “I don’t want them to be annoyed.” Aversion appears.
Then: “I need reassurance that everything is fine.” Attraction and control appear.
Then perhaps: “If this relationship changes, something important about my life or identity is threatened.” The experience acquires more colouring.
Phenomenologically, however, these layers may arrive compressed into a single conclusion: something is wrong.
The Yogic vocabulary encourages the practitioner to notice the colouring as colouring. That is very close to the perceptual shift I have been trying to describe.
The most important distinction may be viveka
The Sanskrit word viveka is usually translated as discrimination or discernment: the capacity to distinguish one thing from another, to see differences that the unexamined mind collapses together.
In the classical Yoga context, discriminative discernment ultimately has a much more specific metaphysical role than the everyday psychological use I am making of it here. But the practical idea is extraordinarily useful.
Viveka asks us to distinguish:
- What did I observe?
- What did I infer?
- What am I afraid of?
- What do I want to happen?
- What am I predicting?
- How certain am I?
Those questions sound almost trivial when written down. They are not trivial when the mind has fused them together.
Consider: “They’re going to leave.”
Viveka begins separating the bundle. The observation might be that they have been quieter than usual today. The inference is that something may be wrong. The prediction is that perhaps the relationship is changing. The fear is that I do not want to lose them. The certainty is unknown.
Suddenly reality contains more space. Not because the feared outcome has become impossible, but because a prediction has become visible as a prediction.
Vairāgya: leaving uncertainty unresolved
Another Yogic term belongs beside viveka: vairāgya. It is commonly translated as dispassion, detachment, or non-attachment. Importantly, it need not mean suppressing experience or withdrawing from ordinary life; it concerns the relationship one has to attachment and aversion.
Applied to the problem we have been exploring, viveka says: “This is a prediction, not established knowledge.” Vairāgya says: “And I do not have to force reality to resolve the uncertainty immediately.”
That second part may be harder.
We often do not merely want to know. We want to stop not knowing. So the mind produces another simulation, another plan, another check, another reassurance-seeking behaviour, another attempt to push the probability distribution toward a single answer.
But sometimes the highest-fidelity representation of reality really is:
I don’t know.
There is something almost counterintuitive about learning to leave that answer untouched.
Saṃskāras: when predictions become grooves
The Yoga tradition also gives us the concept of saṃskāras. The term has several meanings in Indian traditions, but in the psychological context it refers roughly to mental impressions, dispositions, or imprints left by experience: patterns that influence future thought and action.
I find “grooves” a useful metaphor.
Repeatedly move through the same sequence, uncertainty, threat prediction, control, relief, and the sequence becomes easier to activate. The next uncertain event does not begin on neutral ground. There is already a groove.
This fits beautifully with what happens in learning more generally. Systems become efficient at repeating what has worked before. The problem is that “worked” may simply mean reduced discomfort quickly. It does not necessarily mean produced a more accurate model of reality.
A possible chain is:
uncertainty→misapprehension→attraction/aversion→control behaviour→temporary relief→groove strengthened
The same response then becomes more probable the next time uncertainty appears.
When destiny looks like an attractor
This led me to another thought. What we sometimes experience as destiny may, at least in some psychological situations, resemble a self-reinforcing attractor.
The same prior generates the same interpretation. The interpretation produces the same behaviour. The behaviour helps create the same kind of consequence. The consequence reinforces the prior.
saṃskāra→perception→action→consequence→stronger saṃskāra
From inside the loop, the experience is: “This always happens to me.” From a systems perspective: “This system repeatedly returns to the same state.”
That is close to the logic of a self-fulfilling prophecy, but “self-maintaining prediction” may sometimes be an even better phrase. The prediction need not manufacture the entire outcome. It only has to continually organise attention, interpretation, and behaviour so that the underlying model survives.
Perhaps freedom begins when another branch becomes visible
This changes how I think about free will.
Imagine that at any moment several responses are theoretically possible: confront, wait, ask, withdraw, control. Conditioning does not merely influence which one I choose. It may influence which ones I can see.
If a deeply established pattern makes control feel mandatory, my subjective decision tree might contain only one branch: “Obviously I have to fix this.” Another person sees three options. Someone with greater distance from the pattern sees five.
Nothing mystical is required to call that an increase in freedom. More of the possibility space has become perceptually available.
This gives me a way of thinking about destiny and agency that does not require them to be opposites. I arrive at the present moment through causes I did not fully choose: past actions, other people’s actions, circumstances, conditioning, chance, history. That is the node at which I find myself.
But from that node, my response becomes another cause.
I do not control the whole tree. I do not even control every variable affecting my choice. But if I can see more clearly what is conditioning the response, perhaps another branch becomes possible.
That is a modest conception of free will, and I increasingly prefer it to the fantasy of absolute control.
Action without ownership of outcome
This connects naturally with another Indian philosophical idea: karma yoga.
One of its most compelling principles is the distinction between action and the fruit of action. This is sometimes misunderstood as indifference: “Do things, but don’t care what happens.”
I think that misses the useful part. A better interpretation is: act skilfully while recognising the boundary of your causal control.
Any meaningful outcome contains many variables: my action, other people’s actions, their beliefs, their histories, environmental conditions, information I do not possess, and chance.
My behaviour matters. Sometimes enormously. But influence is not ownership.
This turns out to be an extraordinarily practical antidote to the fantasy that sufficient prediction can eliminate uncertainty. I can prepare. I can estimate. I can learn. I can act. I can update. I cannot obtain a contract from reality guaranteeing the result.
So what does clearer perception actually look like?
It does not mean stopping prediction. It does not mean trusting everyone. It does not mean assuming everything will work out. It does not mean becoming passive. And it certainly does not mean replacing negative predictions with positive ones.
“They’re definitely going to betray me” and “Everything will definitely be fine” are structurally similar mistakes if neither is justified by the available evidence.
Clearer perception may be much less dramatic. It might sound like this:
I know what I observed. I can see the interpretation my mind generated. I can identify what I am afraid of. I can distinguish which outcomes are possible from which outcomes are probable. I can ask how reliable my model is in this domain. I can decide what I can reasonably do. And then I can leave the rest unresolved.
That is not ignorance. It is a more accurate representation of ignorance.
And there is an enormous difference between those two things.
Wisdom may be calibration
I used to think wisdom might mean being exceptionally good at knowing what things mean. Now I suspect part of it is almost the opposite.
Wisdom may involve becoming increasingly sensitive to the limits of what you know.
A wise model does not merely produce good predictions. It knows, in some sense, when its predictions are weak. It can say: 90 percent. 60 percent. 20 percent. No idea.
It can encounter prediction error without turning the error into an existential threat. It can learn.
That is what good probabilistic models do. And perhaps it is also what good minds do.
The freedom I was actually looking for
This brings the various ideas back together.
Predictive processing suggests that perception is not simply passive reception. The mind is continuously modelling what might be producing its experience.
Game theory reminds us that when the world contains other agents, our models begin modelling one another.
Active inference reminds us that prediction and action form a loop: what we expect can alter how we behave, which can alter what happens next.
Self-fulfilling prophecies reveal that beliefs can sometimes help create their own evidence.
Mechanism design reminds us that behaviour is partly produced by the rules and incentives of the environment, not merely by individual character.
Yoga adds a phenomenological vocabulary for the ways perception becomes coloured by ignorance, identification, attraction, aversion, and clinging. Saṃskāras describe how repeated patterns leave grooves. Viveka describes the possibility of discriminating between things that were previously fused. Vairāgya describes the possibility of allowing an experience or uncertainty to exist without immediately grasping at it or pushing it away.
Put together, they lead me to a formulation that currently feels more useful than “stop predicting the future”:
You cannot know the future. You can model it.
Build models. Learn from experience. Notice patterns. Use probabilities. Plan where planning changes something. Act where action is available. But remember what a prediction is.
It is not reality arriving early. It is a model generated in the present about a future that does not yet exist.
And perhaps this is where a different kind of freedom begins:
Freedom is not the ability to determine what happens. It is the increasing ability to see what is shaping your perception clearly enough that it does not automatically determine your response.
You still stand inside causality. You still possess conditioning. Predictions still arise. Fear still arises. Desire still arises. Uncertainty remains.
But somewhere between stimulus and response, another branch of the tree becomes visible.
And sometimes that is enough.
