Cybernetics

What It Is

A thermostat measures the temperature, compares it with a setting, and changes the heating in response. A mosquito searching for blood also adjusts its movement using information from its surroundings. A startup does something similar when it watches how customers use a product and changes what it builds. In each case, information about the result affects the next action.

Cybernetics is the study of control and communication in systems. Norbert Wiener formalized it in the 1940s as a way to understand goal-seeking behavior in organisms, machines and organizations. It describes how a system uses feedback to approach a goal while operating with limited resources.

Five components let us explain the behavior: a goal, sensors, actuators, a control loop and resource constraints. Sensors provide information about the world; actuators let the system act on it. The control loop connects the two by using what the sensors report to adjust the next action. Resources determine how long the system can keep trying.

The mosquito, thermostat and startup have the same structure even though their physical parts differ. Changing the accuracy of a sensor, the speed of the loop or the available resources changes the resulting behavior. That gives us specific things to investigate and modify when a system is failing. Calling the agent determined or weak does not identify any of those causes.

The Five Components

The Feedback Loop Structure

Collecting information is only part of a feedback loop. The information has to reach the part of the system that can act, and it has to affect what happens next. A dashboard nobody uses does not provide control, however accurate its measurements are.

The loop compares the sensed state with the goal state. The difference between them is the error signal. That difference determines whether to adjust the action or continue it:

graph TD
    A[Actuator: Perform Action] --> B[Environment Response]
    B --> C[Sensors: Measure Response]
    C --> D[Control Logic: Compare to Goal]
    D --> E[Compute Error Signal]
    E --> F{Error > Threshold?}
    F -->|Yes| G[Adjust Actuator]
    F -->|No| H[Maintain Current]
    G --> A
    H --> A

The time between an action and the resulting adjustment is the cycle time. A short cycle lets the system correct its course quickly. In a long cycle, it can continue moving away from the goal before the information needed to correct it arrives.

Broken Feedback Loop Patterns

A daily braindump forms a personal feedback loop. You notice your internal state, write it down, read and analyze what you wrote, then adjust the sequence of work. Your internal observations provide the sensor data; writing makes them available for inspection; the review changes what you do. Repeating this every 24 hours allows a correction before several days of drift accumulate.

Gradient Ascent vs Destination Navigation

The mosquito cannot see blood from 100 meters. It can sense a weak CO2 gradient, fly toward the stronger signal, and sense again. If the signal is stronger, it continues in that direction; otherwise it adjusts. It can approach something it cannot yet locate by repeatedly checking whether the last movement helped.

This is gradient ascent: measure the local direction of increase, move that way, then repeat. The system does not need a view of the whole landscape. It needs local information about whether it is getting closer, even when that information is noisy.

A startup searching for product-market fit can use the same process. Product-market fit is not a distant destination the founders can see directly. They can observe changes in interest, use, retention and revenue:

Interest_weak → Interest_moderate → Interest_strong
Usage_sporadic → Usage_regular → Usage_habitual
Retention_low → Retention_moderate → Retention_high
Revenue_zero → Revenue_growing → Revenue_accelerating

Each change provides information about the local direction. A positive derivative means the current direction is bringing the system closer; a negative derivative tells it to change course.

The Pivot Trap

A dramatic pivot can abandon this information. Instead of measuring a local change and adjusting again, the founders jump to an unrelated product in the hope that it will work better. The jump sometimes succeeds, but it is expensive because the team gives up what it learned about the current location and starts a new search.

A smaller question preserves the value of those observations: "What makes current retention 5% higher?" You can reason backward from that improvement to changes you could test. Repeating this process lets the team climb the local gradient. Asking "What completely different product should we build?" instead starts a random search elsewhere.

Resource Optimization Under Constraint

A search has to succeed before its resources run out. Optimal foraging theory examines this problem in organisms: how much energy should an animal spend looking for a new source of food, and how much should it spend using a source it already knows? Exploration and exploitation compete for the same finite budget.

The resource objective can be written as:

Maximize: ∑(Energy_gained - Energy_spent_searching)

Subject to: Energy_total ≤ E_max (finite budget)

For a startup, the corresponding formulation compares information gained with the runway spent acquiring it:

Maximize: ∑(Information_gained - Runway_spent_acquiring)

Where information value = reduction in uncertainty about PMF location

An activity is valuable when it reduces a lot of uncertainty for little cost. Cheap experiments can reveal strong signals. Several independent measurements can help confirm or reject a hypothesis. Watching what people actually do often supplies useful information at low cost.

An expensive build before validation has the opposite problem: it consumes substantial runway without knowing how much the result will teach you. Relying on one sensor can leave the result ambiguous. Asking customers hypothetical questions can be cheap, but a weak signal still provides little information about what they will do.

Integration with Personal Cybernetics

The same cybernetic principles govern personal behavior. State machines are cybernetic systems in which the current state is the variable being moved toward a goal. Activation energy is the input needed to make a transition. Tracking supplies observations, and the braindump processes those observations into changes to your work sequence.

The failure can occur at the same point in either system. Observations may never change the action, sensors may be miscalibrated, the cycle may be too slow, or resources may run out before the goal is reached. Shorter loops, several sources of evidence, rapid updates and efficient search address those failures in both settings.

The Thermodynamic Constraint

In statistical mechanics, systems tend toward low-energy configurations. The connection to behavior is the cost of initiating an action: actions with a low activation cost happen more often than costly actions, independently of what someone consciously intends.

A cybernetic system therefore has two ways to sustain a desired action. It can change the environment so that the action costs less to initiate, using prevention architecture. Or it can continuously supply the energy needed to maintain a state far from equilibrium, which is expensive and unsustainable.

At a startup, internal meetings, feature polish and theoretical planning are often easier to initiate than customer calls, public launches or pricing experiments. That difference in activation energy makes the lower-cost activities more likely even when the harder ones would be more valuable. Without a change to the setup, thermodynamics selects against the costly actions.

Prevention architecture changes which option is easiest. A public commitment raises the cost of avoiding a launch. Automated deployment reduces the work needed to publish. Scheduling customer calls in advance removes decisions from the moment when the calls need to happen. These changes reverse the energy gradient so the valuable action becomes the lower-cost option.

The Algorithmic Complexity Lens

Cybernetics makes the resource budget explicit, but it does not by itself explain why one operation costs more than another. Algorithmic complexity examines the work inside the operation.

The mosquito's gradient ascent depends on a cheap local comparison: is the signal stronger here? Each comparison is roughly O(1). If every step instead required an exhaustive comparison of possible futures, costing O(n²) or more, the same search could become intractable. A workable control loop depends on the cost of the computation it repeats.

For cognitive work, cybernetics establishes that you have finite resources and need an efficient search. Algorithmic complexity then identifies what consumes the resources, such as comparing every option with every other option in O(n²) work. That diagnosis gives you a way to change the operation instead of merely observing that you are tired.

Each framework also supplies something the other lacks. Cybernetics alone can leave you saying "I'm resource-limited" without identifying the expensive step. Complexity analysis alone can continue indefinitely: you can analyze the complexity of the analysis, then analyze that analysis. The cybernetic goal and budget provide a stopping condition. Further analysis is no longer worthwhile when it costs more than the useful information it is likely to produce.

Related Concepts

• Startup as a Bug applies the cybernetic model to organizational survival.
• Optimal Foraging Theory examines search under resource constraints.
• Information Theory concerns the value and cost of acquiring information.
• Statistical Mechanics develops the account of energy distributions and flow.
• State Machines represents systems with discrete states.
• Question Theory examines questions as part of search and feedback.
• Tracking supplies sensor data for behavioral feedback.
• The Braindump describes a daily personal control loop.
• Expected Value helps compare uses of a limited resource budget.
• Error Signal explains how to compute the difference a loop needs, including where no measurement already exists.

Key Principle

A system has to find its goal before it runs out of resources. Broken loops prevent course correction, inaccurate sensors send it in the wrong direction, and slow cycles let mistakes continue before they are detected. Shortening the cycle, improving measurements through several signals, and increasing information gained per unit of resource address those specific failures.

With functioning loops, calibrated sensors and efficient search, the system finds its goal automatically. Character, determination and vision do not appear in these system equations; the explanation lies in how the parts interact and what they cost to operate.