The Open Source Internet Is Here - Analysis
Source: the video "The Open Source Internet Is Here" (tactiq transcript, 59 minutes). A solo engineer tries to carry a Reticulum message hundreds or thousands of miles with no cell tower, no private satellite and no corporate network in between. This report breaks down what he actually built, what failed, and what it means for us.
Verdict in four lines
- What worked: a digital transmission received from 1,600 miles away (FT8 from Bogota) and a bridge that moves Reticulum traffic over HF in both directions through a licensed gateway.
- What failed: the headline test itself - two people hundreds of miles apart over ionospheric skywave. The three-week deadline arrived and he shipped an honest failure instead of a polished illusion.
- Why it matters: not another chat app, but a demonstration that the transport layer can be built so that nobody in the middle can switch the button off.
- The real engineering lesson: the law, not the physics, designed the architecture. Encryption is banned over amateur radio, so the bridge strips it at the gateway and runs the long hop as plain text with a published codec.
What he actually built
Reticulum is a permissionless, ownerless digital network that does not care what carries its data - LoRa, Wi-Fi, Ethernet, the ordinary internet, or a custom HF link. Every transport is just another path. He did not build a private network for a handful of radio operators; he built a bridge between networks that already exist. Someone on LoRa in one town should be able to reach someone else entirely via HF quietly carrying the data across the gap.
The physics is not new: under the right conditions an HF signal travels into the ionosphere, reflects back to earth and lands hundreds or thousands of miles beyond the horizon. That is skywave. There is no chain of nodes and no satellite - the atmosphere itself becomes part of the path. What is new is the price: a radio for tens of dollars, open-source networking code that stitches incompatible links into one network, and an AI agent that lets one stubborn person work through problems that once required an engineering team.
The numbers that decide
| Number | What it is | Why it matters |
|---|---|---|
| 1,600 miles | Distance of the decoded digital message (FT8 from Bogota, Colombia) at -14 dB | Proof of miracle one: the horizon is beaten. Receiving needs no licence |
| 28-28.3 MHz | The only slice where his licence allows HF data with skywave | The band dictated the antenna, the speed and the frame design |
| 5 W vs 0.3-2 W | Transmitter rating against what he actually achieved | The loss was not in the radio but in the power supply |
| 300 mW / 940 mW | The instability threshold, and the power that finally got a message through at 2 miles | A real measurement of how much power a real distance costs |
| 2 miles | Clean decode from a moving car, out of line of sight, houses in between | Not skywave, but proof the system survives the field |
| 7 seconds | Time to transmit one frame over HF | The price of the long hop: exchanges are measured in seconds |
| 35 feet vs 13 inches | Wavelength at 28 MHz versus the familiar 900 MHz | Explains why the antenna became two lengths of wire |
| $37 | RTL-SDR receiver | The entry price for receiving: almost anyone can listen |
| $60 | Wasted on a power supply bought on AI advice | The model was confident and wrong |
| 3 weeks | Self-imposed deadline with the rule "when Sunday comes, testing stops" | The rule forced the failure into the open instead of endless polishing |
The hardware and what each piece does
| Device | Role | Notes |
|---|---|---|
| AirMates Light 2 | The transmitter: open-source SDR covering 0-38.4 MHz, including HF | Built around a computer rather than a microphone, so software generates the exact frame. Appears as a network device over Ethernet. Arrived in three pieces and was assembled by hand |
| RTL-SDR | Receiver for testing | $37. Cannot transmit, but for the receiving side of the bridge that is all that is needed |
| Falcon dipole | Antenna sized for 28 MHz - two lengths of wire | One arm inside the office, the other dangling out of the window. Concrete-block house: far from ideal |
| NanoVNA | Antenna analyser (SWR, centre frequency) | Showed resonance near 29 MHz instead of the required 28. Calibrated with open, short and load |
| Raspberry Pi 5 "Haven" | Field node for receiving and decoding | Decoding is the heaviest part of the system and runs into the limits of the hardware |
| 10 m mobile whip | Receiver antenna on the car roof | Magnetic mounts; the ground plane and shield-versus-core question is handled honestly and in detail |
| LP-4 battery | Power at the radio club meetup | When the mains failed, the radio kept transmitting: a genuine grid-down scenario |
The legal cage that designed the architecture
This is the most useful part of the whole video. The amateur radio rules did not forbid the project - they designed it.
| Constraint | Consequence in the architecture |
|---|---|
| The frame must carry a call sign | Every frame identifies its sender; anonymity is impossible by construction |
| The codec must be public | The decoding recipe was committed to a public repository as a condition of legality, not as marketing |
| It is forbidden to obscure the meaning of a transmission, encryption included | Native Reticulum frames cannot go over HF: encryption is "gravity" in Reticulum - identities, signed path proofs and multi-hop forwarding all rest on it. Removing it would not simplify the network, it would break it |
| Exception: encryption is allowed to control a satellite | Considered and rejected as unusable; another route had to be found |
| Only one narrow band is allowed for HF data | The entire design is aimed at 28-28.3 MHz |
The solution he found: terminate encryption at the gateway rather than in the network. A message is addressed to the gateway node, which decrypts it, rewrites it as an amateur frame in plain text, and transmits it over HF together with the final LXMF address. Anyone who hears it can bring it back into Reticulum and forward it to its destination.
The honest cost: the long hop is readable by anyone, and spoofing is not solved - though cryptographic signatures can still prove mathematically which sender a message came from. He accepts the trade-off, comparing it to Meshtastic public channels and to APRS: if you are stranded, you do not care who hears your SOS.
How the bridge works, step by step
- An ordinary Reticulum node (the user's application) has no direct path to its destination.
- The user enables the public-broadcast option: "if an HF gateway exists and there is no path, send it by radio".
- The message goes to the gateway node - the one holding a radio and a licence.
- The gateway decrypts the packet, rewrites it as an amateur frame in plain text and puts it on the air.
- A receiver on the far side (an RTL-SDR; no licence required to listen) hears the frame, decodes it and ingests it back into Reticulum.
- The message is delivered to an ordinary node on the other side - in the demo, an Android phone running the Columba client.
Technically this is a "bridge extension": the connector to the radio or the receiver runs in parallel to Reticulum rather than as one of its native interfaces, which is precisely how encrypted frames are kept off the air.
What did not work - stated plainly
- The headline test never happened. Two people hundreds of miles apart over skywave was not proven. He shipped what he had when the deadline came.
- Power broke the range. The radio can transmit up to 5 W, but above 300 mW packets stopped getting through: a cheap 12 V / 3 A / 36 W brick cannot hold the current, even though on paper it "should". A proper battery pack (5 A continuous, 7 A surge) still capped him at 2 W and needed another connector.
- AI advised him wrong. The power supply was bought on the model's recommendation, and during troubleshooting the same model told him to buy another one. His words: "AI isn't always right. I'm out 60 bucks."
- Decoding was too heavy. The most expensive parts of the decoder were ported from Python to Rust, cutting decode time to a few seconds - the single change that transformed the system.
- The antenna was never tuned. Resonance sat near 29 MHz instead of 28. He says so: "for our testing, I'm just going to go with this".
- Field tests cost round trips. Early runs meant driving out, driving back, and only then learning the result. The fix was a console on the Raspberry Pi: transmit power, receiver gain, message text and decode result in real time. That is what turned testing into engineering rather than a quest.
One detail he states explicitly, and it deserves respect: the control channel and telemetry went over a separate secure tunnel, while the packet being measured travelled entirely over HF. Otherwise the range claim would have been worthless.
Seven things worth taking
- An observation tool beats stubbornness. The live console replaced dozens of drives. Our own measuring instruments - Gym, recorded runs, the evidence log - are exactly this.
- Never mix the control channel into the measurement. Keeping them apart is a model of measurement discipline: had control traffic shared the path, the test result would be void.
- Do heavy computation where it is cheapest. The Rust port is the standard move: first make it work, then make it fast.
- The "anyone may listen" asymmetry. Only the transmitter needs a licence, so receiving nodes can be crowdsourced - the ADS-B pattern. Scaling on other people's $37 rather than on your own budget.
- One licensed node gives the whole mesh a long hop. "You don't all carry the tower - you only need reach to the person who does." That is a model of shared access, not of handouts.
- Law as a design input. A constraint produced the solution (decrypt at the gateway) and the public codec. Breaking the rules would have cost him the air, so the rules were built in rather than worked around.
- A hard deadline as a method. "When Sunday comes, testing stops" is the only reason the failure was shown at all.
What it means for ARLI
| Observation from the video | What we take from it |
|---|---|
| An architecture with nobody who can switch the button off - "can't be evil" instead of "don't be evil" | Our argument with agents is provability: the proof rule, the run log, the charge receipt. Not promises of safety, but mechanics you can see |
| The model advised confidently and cost him $60 | Confirms our rule: a model's output is checked by measurement. We already have the decision layer as a second judge and the label-swap probe for this |
| The control channel stayed out of the range measurement | A standard for our own reports: always state what was measured and what carried the service. That is a cheap difference from other people's numbers |
| The Python decoder could not feed weak hardware; Rust cut it to seconds | Wherever computation runs on the client side - sandbox, skills, meters - the Python path hits the hardware before it hits the logic |
| One licensed node opens a long hop for the whole mesh | The mirror image of our model: one paid skill key serves an entire agent harness. We sell the call, not the file, for the same reason |
| A legal limit shaped the frame and forced a public codec | China's requirements - localisation, catalogue, confirmations - are not bolted on afterwards; the product is built out of them |
| Public codec, with the value hidden elsewhere | Our split: the skill file is free, the call is paid; the method is open, the counter and the data are not |
| The headline failure was shipped and named as a failure | Our rule of never announcing what has not been run, and publishing the gap rather than only the wins |
| "In trying to cancel the middleman, I was becoming their best customer" (fibre, Verizon, Starlink, Iridium) | An honest test of our economics: if fuel is bought from an intermediary, we are somebody's customer too. Which is why the margin must be ours - on top-ups, not on reselling tokens |
| One person plus an agent covered ground that used to need a team (about five days of work compressed) | Supports the bet on a solo builder with agents - with the price of AI advice standing right next to it: acceleration without transferring trust |
Caveats
- One source: a single video transcript. Every number comes from the author's speech and was not verified by us; we reproduced none of his measurements.
- The transcript is machine-generated and garbles terminology. Corrected readings: "Airme / Air May / Airmes" is the AirMates Light 2; "Laura" is LoRa; "Aridium" is Iridium; "VAR / Vara" is VARA; "Halo" is HaLow; "conquered New Hampshire" is Concord, New Hampshire; "Bogotaa" is Bogota; "Kiscane" is Kissimmee; "FTA" is FT8; "Nano VNA" is NanoVNA; "MeshVPN dark magic" is most likely Tailscale.
- Everything about radio regulation here is the author's account, not our legal assessment. This is not legal advice.
- Built to our in-house report standard: mobile-first, three themes, fixed header, two table modes, no external resources.
Analysis report. Source: the transcript of "The Open Source Internet Is Here". Prepared 22 September 2026.