The Weird Ideas Are Becoming Cheap Enough to Test
A developer used Claude Code to build an optical file transfer system in one night, showing how AI coding tools can reduce the cost of testing unusual ideas before anyone knows whether they will work.
I. Moving a File with Light
A developer wanted to move files between two phones without connecting them to the same network, installing an application, pairing through Bluetooth, or using a cable. The resulting experiment, called Decimen Optical Transfer, uses one phone screen to display rapidly changing QR codes while the second phone reads them through its camera.
The method turns a file into a stream of visual data. Each QR frame carries part of the file along with information identifying the transfer session, sequence, file size, and other details needed for reconstruction. The receiving phone watches the stream, collects enough pieces, and rebuilds the original file inside its browser.
Dropped frames do not necessarily ruin the transfer. Decimen uses fountain codes, which allow the sender to generate many overlapping pieces of encoded data. The receiver does not need to capture every frame in the correct order. It needs enough valid pieces to recover the file. Such resilience matters because a hand held camera may miss, blur, or duplicate rapidly changing QR codes.
Tests reportedly reached about 128 kilobytes per second during hand held use and roughly 186 kilobytes per second when the phones remained stationary. Those speeds cannot compete with modern wireless networking, but they are sufficient to move documents, photographs, music files, or small data sets without establishing a conventional connection.
The system runs in a browser and needs only access to the receiving camera. Once loaded, it can operate offline. The source code is publicly available under the MIT license. Its developer says Claude Code helped produce the working proof of concept in one night.
The idea is clever, but not entirely new. Optical transfer has a long history. The Timex Datalink watch, introduced in 1994, received appointments and other information through patterns of light displayed on a computer monitor. A sensor on the watch read flashing horizontal bars and stored the transmitted data. Later versions were used by astronauts and certified for spaceflight.
Decimen updates that older principle for devices that already contain bright displays, fast cameras, browsers, and substantial computing power. The innovation lies less in discovering that light can carry digital information than in assembling familiar components into a usable experiment with remarkable speed.
The phrase “air gapped” also requires caution. A system without a network connection may reduce some risks, but an optical channel still transfers information across a boundary. Secure facilities may prohibit cameras precisely because they can photograph screens, scan documents, or create unauthorized paths for data. Decimen demonstrates a mechanism, not a complete security policy.
Its strongest immediate value may therefore be conceptual. The project shows that two ordinary devices can exchange a file through nothing more than a screen and a camera. No router coordinates the transfer. No cloud service stores the file. Light becomes the connection.
An animated sequence of seven QR codes encoding the full text of Martin Luther’s Ninety Five Theses. The example illustrates how information can be divided across rapidly changing visual frames, the same basic principle used in optical file transfer between a screen and camera. Created by Natefeff, 2025. Source: Wikimedia Commons. CC0 1.0 Universal public domain dedication.
II. The Falling Cost of Finding Out
The larger story concerns the economics of experimentation.
A strange technical idea once faced a difficult preliminary question. Was it worth assigning a developer several days or a week merely to determine whether the idea could work? Most organizations reject speculative projects before testing them because engineering time carries an opportunity cost. Every experiment delays something more certain.
AI coding tools change that calculation. They do not eliminate the need for engineering judgment, testing, security review, or maintenance. They can, however, reduce the effort required to produce a first working version. An idea that once needed a formal estimate, a project slot, and several days of coding may now receive an evening of experimentation.
Lower cost changes which ideas survive long enough to be evaluated.
The Decimen developer did not need to prove in advance that rapidly flashing QR codes would achieve useful throughput. A quick prototype supplied the answer. The result might have failed because phone cameras could not keep pace, browser libraries introduced too much delay, or missed frames made reconstruction unreliable. Fountain codes, faster displays, and modern cameras made the concept more practical than expected.
Failure would also have produced useful information. A one night experiment can reveal that an idea is too slow, unstable, complicated, or unnecessary. Such knowledge may justify abandoning the project before it consumes substantial resources. AI assisted development lowers the cost of success, but it also lowers the cost of a well informed refusal.
Rapid prototyping carries risks. Developers can rebuild tools that already exist, overlook security weaknesses, misunderstand generated code, or create software that works only under favorable conditions. Cheap experimentation may produce fifty versions of the same application instead of one durable product. The ability to build quickly does not establish originality, demand, or quality.
Those limitations do not erase the change. More people can now test ideas that institutions would never have approved as formal projects. Researchers can create temporary instruments. Analysts can automate narrow tasks. Teachers can prototype learning tools. Hobbyists can combine old technologies in unfamiliar ways. Some results will be trivial, repetitive, or flawed. A few may uncover solutions that conventional planning would have dismissed.
Innovation has always depended on the cost of finding out. Laboratories, machine shops, computing time, specialist labor, and organizational approval determined which questions people could afford to ask. Generative AI is beginning to reduce one part of that cost by making executable prototypes easier to create.
Decimen Optical Transfer may remain a clever demonstration rather than a widely used product. Its broader lesson would still matter. The developer could move from curiosity to evidence in one night.
AI may not make every strange idea useful. It is making far more of them cheap enough to test.
Further Reading
- Decimen Optical Transfer on GitHub
- Tom’s Hardware, Streaming QR Codes at 60 Frames Per Second
- Timex Datalink and Optical Data Transfer