While driving one day, I glanced at my side mirror. A car was approaching.

I looked back at the road. Without checking the mirror again, I still had a fairly good sense of where the car would be a moment later.

I had not calculated its speed.

I had not measured the distance.

There was no small spreadsheet open in my head.

Yet some model had kept the car moving.

At the time, I was studying the Transformer architecture. A Transformer uses learned queries, keys and values to decide which parts of its context matter, then predicts what token is likely to come next.

It sounds almost offensively simple for something that can produce surprisingly good results. The side mirror made the resemblance feel much stranger.

My brain had filled in the next frame

That side-mirror moment made me notice prediction everywhere.

Press a familiar light switch and you do not wait neutrally to discover what it does. You expect light. When nothing happens, you are briefly surprised.

Catch a ball and your hand moves toward where the ball will be, not where it was when the light first reached your eyes.

Watch someone begin a familiar sentence—

“I’m not angry, but…”

—and your mind has already generated several possible endings.

None involves a calm discussion of local weather.

Driving experience seems to improve this kind of anticipation. In experiments designed to isolate hazard prediction, experienced drivers were better than novices at answering the practical question, “What happens next?” (Crundall, 2016)

A novice driver sees a collection of moving objects.

An experienced driver sees trajectories, intentions and familiar precursors: that wheel is turning, that pedestrian has looked away, that car is drifting a little too confidently toward my lane.

Experience has not merely stored more pictures. It has created more useful relationships between them.

Is this our version of QKV?

The tempting story is that humans have an enormously rich version of query, key and value: the mirror supplies a cue, experience supplies relevant patterns and the brain predicts the next frame.

It is a beautiful analogy.

It is also almost certainly too neat.

QKV does not, by itself, predict the next token. Attention uses learned queries, keys and values to decide how information in the context should be combined; the rest of an autoregressive model turns that representation into a probability over what comes next. (Vaswani et al., 2017)

And the brain is not secretly running PyTorch behind my forehead.

Still, the analogy gives me a useful question:

How much of intelligence is the ability to use present cues to retrieve the right history and anticipate what comes next?

One influential family of theories called predictive processing proposes that perception is not built from sensation alone. The brain uses prior expectations and incoming signals together, and mismatches help it revise its model. The broader theories are ambitious and debated, but the modest idea is enough here: perception arrives in a system already making guesses. (Friston, 2010)

The world does not enter a blank mind.

It enters an argument already in progress.

Children are training on the universe

Then I looked at children.

Children often predict badly. They misjudge motion, repeat actions that failed and press buttons you have begged them not to press. Their training set is smaller, so they explore everything: touch it, shake it, taste it.

A toddler is essentially debugging the physical world with their mouth.

At first I thought this was just poor prediction. Perhaps children explore because their model is sparse and uncertain.

But that is only half the story. Exploration is also how the model gets better.

Developmental studies using explore–exploit tasks have found that children often explore more than adults, sometimes accepting lower immediate rewards and learning more about the environment as a result. That does not make every childish act an optimal information-gathering strategy, but it suggests that childhood’s apparent inefficiency can serve learning. (Liquin & Gopnik, 2022; Gopnik, 2020)

An adult sees a spoon and uses it to eat.

A child still needs to know whether it is also a drumstick, a catapult and a device for measuring the patience of a parent.

Prediction cannot tell us what to want

This is where prediction stopped being enough for me.

A system could predict the next event perfectly and still have no reason to do anything.

Prediction says:

If I press this switch, the light will probably turn on.

Reward says:

I prefer being able to see.

Prediction says:

If I make this joke, the room may laugh.

Reward says:

Their laughter matters to me.

Consequences teach us which actions are worth repeating. Reinforcement learning formalizes part of this loop: an agent acts, receives an outcome and updates its expectation of value.

Classic neuroscience experiments found dopamine-neuron responses consistent with reward prediction error. Unexpected rewards produced one pattern; once a cue reliably predicted the reward, the response shifted toward the cue; when an expected reward failed to arrive, activity decreased. (Schultz, Dayan & Montague, 1997)

That does not mean dopamine is simply pleasure, or that a human life is a points programme with unusually confusing terms and conditions.

Humans interpret rewards. Praise can feel manipulative. Punishment can become evidence of integrity. Someone can endure pain for a promise, an identity or a person they love. The outcome acquires meaning inside a model of the world.

My current working model—not a map of the brain

Prediction estimates what may happen. Reward and meaning influence which actions become worth repeating.

Is maturity an explore–exploit problem?

This leaves me with a question I did not expect.

Children explore widely. Adults increasingly exploit what they already know. We reuse habits, beliefs, relationships and strategies that have paid off before.

Perhaps some of what we call personality is a set of predictions and responses that became reliable enough to stop reopening every morning.

But is maturity simply less exploration?

I do not think so.

Too much exploration and every novelty can redirect you. Too much exploitation and a model can survive long after the world has changed.

I had remembered a biblical image about maturity as coming from James: no longer being “tossed” by every doctrine. I had merged two passages. James describes the doubter as a wave driven and tossed by the wind; Ephesians explicitly contrasts maturity with being children carried by every wind of doctrine. (James 1:6 and Ephesians 4:13–14)

The correction makes the question more interesting.

Maturity may not be the end of curiosity. It may be the ability to decide which surprises deserve an update, which impulses deserve an action and which convictions have earned the right to remain stable.

That feels like a separate question. It became another article.

What I currently think

I do not think humans are prediction machines or reward machines.

I think we build models that anticipate what comes next, act inside those models, and let surprise, consequence and meaning reshape what we expect and what we value.

A child has fewer predictions and a larger world to test.

An adult has richer patterns and more efficient habits—but also more opportunities to mistake familiarity for truth.

The side mirror still feels like the most honest version of the mystery.

I looked once.

Some part of me kept the car moving.

What I still cannot tell is when that silent prediction becomes understanding—and when it is merely a very experienced guess.

Until the next strange question,

Osagie