Claude Code wrote all the code for an ESP32 rig and debunked a false 'discovery'
An enthusiast spent six months building spinning wheels with ESP32 boards (cheap microcontrollers), and Claude Code wrote all the code. The first 'finding' was a 'comb' pattern in the accelerometer data with a step of exactly 256 samples. It turned out to be a bug: the boards carried an LIS3DH sensor instead of the LIS3DSH, and it was running in 8-bit mode. The AI built a model with no fitted parameters and reproduced the effect exactly. The lesson: check the WHO_AM_I register on any purchased sensor and store raw axes.
AI-processed from Habr AI; edited by Hamidun News
Stories like 'AI wrote me an app' no longer surprise anyone. It gets more interesting when an AI agent works not in a browser but in a physical project with a soldering iron, motors, and dozens of boards, where an ordinary measurement error looks like a scientific discovery. That is the story one author published on Habr: six months of work, two rigs, one caught false 'finding,' and so far an honest zero in the results. Claude Code wrote all the code, while the human only soldered and wound coils.
The essence of the 'finding' is simple. The rig is a wheel carrying ESP32 boards, cheap microcontrollers, and acceleration sensors (accelerometers). A clear pattern appeared in the data with a step of exactly 256 samples. It was not physics but a quirk of the hardware: the sensor was running in 8-bit mode and simply rounding its readings. The AI built a simple model with no fitted parameters and reproduced the pattern exactly.
How it started and who does what
In the spring, out of curiosity, the author asked a neural network why life appeared on Earth specifically. Together they listed what the planet offers: rotation, a magnetic field, gravity, a constant energy supply, many 'agents' from molecules to cells, and noise. From that list came a rough tabletop model.
The wheel rotates in a vertical plane. At the top point, centrifugal acceleration cancels gravity; at the bottom it reaches 2 g, that is, double overload. This is a 'day' compressed into a second.
A coil is wound around the rim, and the microcontrollers look for prime numbers among random ones, picking out the rare ones from the noise. Nearby sit the sensors, and the question is whether anything changes in their readings.
The human handles the mechanics, winding, soldering, and power. Claude, run through Claude Code directly in the project folder, writes the board firmware, the server, the control panel, the wiring diagrams, the experiment programs, and the data analysis. It also keeps the log and launches the experiment series itself. The third participant is a human theorist.
This does not work like a chat. It works like an executor with memory and rules. The rules were spelled out once: do not read or print secrets (the Wi-Fi password, the firmware signing key); run tests only on test ports; commit only when the author says so; do not push the motors above the set limit; check the stationary sensor every 15 minutes.
The command 'build the firmware and flash it to the second wheel' turns into a build, an over-the-air update with signing and rollback, and a check that all boards are back online. For boards inside a coil on a spinning wheel, this is not a whim: nobody wants to go near them with a USB cable.
The false finding and how it was caught
The first rig had ten ESP32-C6 boards with accelerometers. A couple of weeks later, a neat 'comb' appeared in the data: the acceleration did not spread evenly but fell into a clear step. The step turned out to be exactly 256 samples. Instead of the LIS3DSH listed in the product description, the boards carried an LIS3DH, and one bit in its register switched on 8-bit mode: the low byte simply never changed.
The author wrote down the success criterion in advance, before the experiments: only what cannot be confused with interference, vibration, or instrument error counts. That is why the attractive plot never became a 'discovery.' The practical takeaway: on any purchased sensor, read the WHO_AM_I register (it reports which chip it is), verify the address and the resolution, and write raw axes to the file rather than a precomputed vector magnitude.
Hardware pitfalls and results of the second rig
The second rig was built with the earlier mistakes in mind: two wheels, six XIAO ESP32-S3 boards on each, with each wheel having its own motor, coil, and batteries. Between the wheels sits a stationary 'witness,' a precise ADXL355 accelerometer. The boards communicate over ESP-NOW, and with the server over a separate Wi-Fi network without internet access.
The list of pitfalls is useful to anyone soldering ESP32s:
- solder paste used instead of flux left solder balls, the board overheated and would not start;
- a USB cable plus a shared 5 V bus burned out more than one board, so each board's input now has a 1N5819 diode, and firmware is updated only over the air;
- an ESP32-S3 under constant computation in a closed tube reaches 70 °C in two minutes and loses half its speed; aluminum heatsinks and a regulator in the firmware helped;
- a floating address pin on the sensor makes the I2C address jump between 0x68 and 0x69;
- telemetry larger than 1460 bytes does not fit into one frame, and the board goes silent under load;
- brushed motors and gears show up in the accelerometer spectrum: the shaft produces a peak at 8 Hz, and a working television produces lines at 25 Hz.
Result for the autumn: the rig found and verified 511 thousand prime numbers and ran hundreds of trials, including control runs with 'dummy' values in place of primes and randomized step order. The AI also spent four hours on its own cycling through operating modes. The witness shows no confirmed deviations: there were single spikes, but none repeated under the same mode.
What's next and what to take from it
The third version will have four nine-axis ICM-20948 sensors on each board, 24 per wheel. The primality check will be built from the sensors' actual readings, and at the moment of each finding a snapshot is taken from all sensors. The boards will accumulate the primes they find and broadcast them all at once, so the witness has an exact timestamp and frequency to look for a response. Control batches of composite numbers are chosen by a random sequence, and the analysis does not see it until the series ends. Everything is built from AliExpress parts, and the code, schematics, and experiment logs are open on GitHub.
The limitations are obvious: no effect has been found, and the idea itself remains speculative. For practitioners, the value lies elsewhere: clear rules for the agent, blind control series, and the habit of first trying to refute your own finding.
- Nov 21, 2026Own VPS Didn't Protect Claude Accounts: 51% vs 53% Problem Rates on a Commercial VPN
- Nov 21, 2026Ahead of IPO, Amodei Admits Anthropic Committed to Spending Nearly $901 Billion on Compute
- Nov 21, 2026Nectar launches AI agent that finds influencers among 300 million profiles
- Nov 21, 2026Wavo in Rust Transcribes an Hour of Audio in 18 Seconds, 200x Faster Than Real Time
Key Questions Answered
How it started and who does what?
In the spring, out of curiosity, the author asked a neural network why life appeared on Earth specifically. Together they listed what the planet offers: rotation, a magnetic field, gravity, a constant energy supply, many 'agents' from molecules to cells, and noise. From that list came a rough tabletop model. The wheel rotates in a vertical plane. At the top point, centrifugal acceleration cancels gravity; at the bottom it reaches 2 g, that is, double overload. This is a 'day' compressed into a second.
What are the key details on The false finding and how it was caught?
The first rig had ten ESP32-C6 boards with accelerometers. A couple of weeks later, a neat 'comb' appeared in the data: the acceleration did not spread evenly but fell into a clear step. The step turned out to be exactly 256 samples. Instead of the LIS3DSH listed in the product description, the boards carried an LIS3DH, and one bit in its register switched on 8-bit mode: the low byte simply never changed.
What are the key details on Hardware pitfalls and results of the second rig?
The second rig was built with the earlier mistakes in mind: two wheels, six XIAO ESP32-S3 boards on each, with each wheel having its own motor, coil, and batteries. Between the wheels sits a stationary 'witness,' a precise ADXL355 accelerometer. The boards communicate over ESP-NOW, and with the server over a separate Wi-Fi network without internet access.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.
