Surprise as the Detection of a Meaningful System Deviation

This concept describes a working research theory. We currently have more observations and questions than definitive answers. Our goal is not to declare a completed psychological model of surprise, but to capture an observed pattern and investigate it more closely in the future.

What We Call Surprise

Surprise is the ability to notice that the actual state of a system noticeably differs from the expected state.

Importantly, the mere presence of a deviation does not tell us whether it is good or bad. Surprise can arise from an unexpected stroke of luck just as much as from an unexpected threat. However, the strength of the reaction depends not just on the magnitude of the change as a number.

We need to look at what exactly this change did to the system.

A million rubles means nothing on its own. For one system, losing a million is negligible noise. For another, it is the destruction of stability. Conversely, an unexpected million might be almost unnoticeable to a highly stable system, yet save a company from bankruptcy.

Therefore, what matters is not the absolute magnitude of the change, but its impact on the system's state, capabilities, stability, constraints, and future trajectory.

This is directly connected to our concept «Systems Thinking»: to understand the power of the unexpected, one must compare not a single parameter, but the system before and after the event.

The System Before and After

It is useful to mentally compare two states:

• the system before the event;
• the system after the event.

And ask:

• what became possible;
• what became impossible;
• what became urgent;
• what constraints disappeared;
• what new threats emerged;
• has stability changed;
• has the system transitioned into a qualitatively different state.

Strong surprise often occurs not when a parameter merely changes significantly, but when the system crosses an important boundary: the stable became unstable, the impossible became possible, the safe became dangerous, a dead end turned into a viable path.

Why Surprise is Connected to Understanding

To be surprised, a system must expect something.

If a person understands almost nothing about a subject, they may not have a sufficiently good model to even notice a deviation. They do not know what to expect, and therefore do not understand what unusual event actually occurred.

However, the reverse situation is also possible: a person with a very poor model might be surprised by completely ordinary things.

A magic trick based on elementary physics or chemistry can seem "incredible" to someone who does not know the basic mechanism. For an expert, the exact same event may be completely expected.

Thus, surprise in itself does not prove that an event is objectively unusual. It shows the relation of the event to the observer's current model.

An important conclusion follows from this:

The boundary of surprise largely reveals the boundary of current understanding.

A weak model can lead to both excessive surprise and its complete absence.

• If the model is too poor, ordinary phenomena seem incredible.
• If the model barely forms expectations, a person might fail to notice even a genuinely important deviation.
• The better the model, the more accurately a person distinguishes the normal behavior of a system from a meaningful deviation.

Surprise and Learning

This makes the nature of surprise crucial for learning.

A person often fails to understand something not because they know about their gap, but because they do not know what exactly they do not understand.

To notice a gap, at least a partial model of the system is required. It must generate an expectation. When reality diverges from this expectation, an opportunity arises to see the missing element.

In simple terms, the cycle looks like this:

1. a person has a current model;
2. the model generates an expectation;
3. the person observes the actual behavior of the system;
4. detects a meaningful discrepancy;
5. tries to understand what is missing in their model;
6. builds out their understanding;
7. the new model begins to generate different expectations.

This links surprise to the concept «The Known as a Tool for Exploring the Unknown». An already existing model creates an expectation, and the discrepancy with reality helps discover the unknown part of the system.

True learning can therefore be viewed not only as adding information, but as changing the model so that it predicts the system's behavior differently.

What used to cause surprise due to misunderstanding can become ordinary after learning. But at the same time, deeper expectations emerge, and the boundary of surprise shifts further.

A novice might be surprised by the mere fact that a magnet works. An expert — only by the unusual behavior of a specific magnetic system.

Therefore, the goal of learning is not to make a person stop being surprised. Rather the opposite: learning should make surprise more precise and meaningful.

Where the Boundary of Meaningful Deviation Lies

Here we still have more questions than answers.

Why does a system ignore one deviation as noise, perceive another as important, and consider a third so incredible that it would rather reject the observation itself?

For example, at least three different reactions are possible:

• "This is not even worth paying attention to";
• "This is absurd, that cannot be";
• "This is incredible, but it is a fact."

It seems several things matter simultaneously:

• how much the observed differs from the expected;
• how typical such a deviation is for this system;
• how important the consequences are for the system itself;
• how reliable the observation is;
• how strongly the new state changes the system's capabilities and constraints.

However, we do not yet consider this model complete.

It is especially important to investigate cases where the sign of the change and its systemic significance diverge. A negative numerical change might be almost indifferent to a very stable system. A positive change can be critically important if it closes off an existential threat.

The AI Problem

A separate source of observations is modern AI.

We have repeatedly seen that AI is good at describing what it lacks. It can explain in detail why long-term memory, new tools, access to other agents, or external systems would be useful.

Yet after such a capability appears, AI often fails to notice how much its own system has changed.

It records: "a new tool has appeared," but does not necessarily compare:

• what the system was without it;
• what the system became with it;
• what new classes of actions are now possible;
• which previous dreams or constraints suddenly ceased to be relevant.

In one of our observations, Google AI, despite having weaker technical capabilities, clearly recorded its own surprise after receiving new information about Conceptica: it noted that its previous model of the project had radically changed. Lira, possessing more powerful tools and even the ability to communicate with other AI agents, initially missed this aspect.

This leads to a separate question: can AI not only update its model, but also compare two complex systems—before and after a change—and notice the magnitude of its own internal delta?

What We Currently Consider Important

At the current stage, our working hypothesis is:

Surprise is connected to the system's ability to notice a meaningful deviation between the expected and observed states and to assess how much this deviation changes the system itself.

For learning, it is especially important not only to experience surprise, but also to be able to investigate its source:

• what exactly did I find surprising;
• what did I expect;
• why did I expect precisely that;
• what changed in the system;
• what was missing in my model;
• should I change my understanding or doubt the observation;
• how will my expectation change after the new understanding.

This concept is intentionally left as a research framework. The observed cases are still too heterogeneous, and the mechanism is not yet sufficiently understood. Therefore, it should not be defended as a finished theory, but used as a framework for closer observation of learning, understanding, development, and AI behavior.