Self-Improving AI: Three Different Claims
People use one phrase for three very different claims. Sorting them out is the first step, because only one of the three is something a student can actually build and measure this term.
Three claims that share one phrase
The first is ordinary machine learning: a model gets better as it sees more data, because training adjusts its weights to reduce an error. Nobody calls this self-improving, because the improvement is bounded by the data you give it and the objective you set.
The second is an agent that improves its own working: it rewrites its instructions, tunes its thresholds, or notices that a strategy has stopped working and changes it. This exists today and you can build it. The third is recursive self-improvement, where a system improves the thing that makes it better, for example a program that rewrites its own code or a model that retrains on its own outputs. That one has not been demonstrated, and claims that it has are usually marketing or a definition stretched until it happens to be true.
The version you can honestly demonstrate
Take a controller with one tunable number and let the machine adjust it. A fan that switches on at a temperature threshold: measure how often the fan runs late, move the threshold a little, measure again, and keep the change only if the error went down. That is hill climbing. It is real, it takes about 40 lines of code, and the improvement is a number you can print next to the baseline.
The point is not that it looks clever. The point is that the system has an objective, a measurement and a rule for accepting or rejecting a change, and that is the same skeleton as every self-improvement claim ever made. What you cannot let it do is choose its own objective, and that limit is the honest part of the story rather than a disappointing one.
Step by step
-
Write a fixed controller and record its error
A fixed threshold on the DHT11 reading and a count of how many minutes the room was too hot. That count is your baseline number, so write it down.
-
Expose one parameter
Make the threshold a variable that can change at runtime, and log every value it takes so you can plot the tuning history.
-
Add one measurement of quality
Score the last ten minutes: how often was the fan late, and how often did it run when the room was empty. One score, not five, or the system will optimise the wrong thing.
-
Let it try a change and judge it
Nudge the threshold by a small step, wait, then compare the new score with the old one. Keep the change if the score improved, put it back if it did not.
-
Add a freeze rule
If the score gets worse twice in a row, stop tuning and return to the best value seen. An improving system needs a brake, and so does yours.
-
Reset and show it re-learns
Power-cycle the board and let it start again from the default. Learning that survives a reset is a different thing from learning that does not, and knowing which one you built is worth marking in the report.
What you'll need
| Item | Qty | Why |
|---|---|---|
| Arduino Uno R3 | 1 | Runs the loop, the scoring and the tuning rule. No extra hardware is needed for the logic itself. |
| DHT11 temperature and humidity sensor | 1 | A slow, noisy variable, which is exactly what a tuning loop has to cope with. Its accuracy is about 2 degrees C. |
| Light-dependent resistor and 10k resistor | 1 each | A second signal, so the system can tell an empty room from a hot one instead of tuning on one variable. |
| Fan with transistor driver | 1 | The actuator whose timing the controller is trying to improve. |
| LED bar graph module | 1 | Shows the score and the current parameter value, so the improvement is visible during a lesson. |
| Pushbutton | 1 | Lets a teacher freeze the tuning while demonstrating, which avoids the one bad demo ruining the lesson. |
Practise it before you build it
Every step above has a matching circuit in the ThinkEdge Learn simulator. The simulator runs a real
Arduino sketch against a modified-nodal-analysis solver, so digitalWrite actually drives the
LED on screen and the readings change when you change the wiring — no hardware, no burnt components.
Real smart fans use TWO sensors: temperature (to know WHEN to turn on) and motion or distance (to know IF someone is present). If it's hot AND someone is within 100 cm, the fan turns on. This avoids the silly situation of a fan blowing in an empty room. We use the DHT11 for temperature, the ultrasonic for distance, and a real DC fan motor for the output.
A common real-world control system: read a sensor, compare to a threshold, switch an output on or off. Air conditioners, heaters, fans, and sprinklers all work this way. The DHT11 reads the room temperature, and a small 5V fan turns on at 30 °C or hotter and off at 27 °C or cooler — the same pattern that runs every thermostat on Earth. The 3 °C gap between the on- and off-thresholds (called hysteresis) stops it from flickering on/off every second when the temperature hovers right around 30 °C.
A real solar panel tracks the sun across the sky to capture 30-40% more energy than a fixed panel. We can build a simple 1-axis version with two LDRs (one on each side of a divider) and a servo. The Arduino reads both LDRs, compares them, and turns the servo toward the brighter side. With more LDRs, you can do 2-axis tracking.
An LDR (Light-Dependent Resistor) is a resistor whose resistance changes with light. In bright light, resistance is low (~100Ω). In darkness, resistance is high (~1MΩ). By pairing it with a fixed resistor in a voltage divider, the Arduino can read the brightness as a voltage.
A 10-segment LED bar graph is the easiest way to visualize an analog input: spin a potentiometer and 1, 2, 3 ... 10 LEDs light up in sequence. This is a "VU meter" — the kind you see on stereo equipment. The Mega is ideal because it has 10 free digital pins without stealing any from the analog or serial pins. The same code can be extended to 20, 30, or 100 LEDs on a Mega.
The Mega has 4 hardware serial ports (Serial, Serial1, Serial2, Serial3). That means you can talk to 4 different devices at the SAME time — like a GPS, a Bluetooth module, an XBee radio, and the USB. Use 3 of those ports to read 3 different sensors in parallel and stream their data to the Serial Monitor as a CSV row every second. This is the foundation of every real Arduino data logger.
Common questions
Can AI improve itself without a human?
Partly, and only inside a frame you set. Today's systems reliably tune parameters, thresholds and weights against an objective a human defined, and they do it thousands of times a minute. Choosing the objective, deciding when to stop and deciding what counts as better are still human jobs. Anyone claiming the rest is done is describing a research goal, not a shipped product.
How do I show improvement in a school project?
With two numbers. Record the error of your fixed controller over one day, then record the error of the self-tuning controller over the same day under the same conditions. The difference between those two numbers is your result, and it has the same structure as any scientific comparison: one variable changed, everything else held still.
Is self-improving AI dangerous?
The honest answer is that the systems that exist today optimise narrow, measurable objectives, so the risk is that they optimise the number instead of what you meant, for example a thermostat that learns to keep the sensor at 21 degrees while the room gets hotter. The fix is engineering rather than fear: measure the thing you actually care about, not the reading that is easiest to measure.
Related guides
Try the full simulator with your class
103 guided circuit projects, 21 lessons and an in-browser Arduino simulator. The first 15 projects are free, no account needed. NCERT Science mapping on 47 projects, with Arabic and English copy in the regional builds.