AI Meets Ham Radio

Probing AI with a Ham Radio Question


IRTS License Study GuideMichael Toia Book on Smith Chart     Many of my ham radio projects are happy accidents in the sense that they are not part of any sort of planned program. Some even resemble QRM (radio code for ‘interference’) in the sense that while I’m attempting to focus on one specific project, a seemingly unrelated thought will intrude.

    I had been reading the Irish Radio Transmitters Society (IRTS) Amateur Radio Station Licence Study Guide. I recommend it by the way. It contains all sorts of interesting information. For example, electrical grounding is done differently in Ireland than here in the United States. But that is an aside. While reading a few paragraphs on the subject of impedance I was reminded of an exercise in Michael Toia’s booklet, You Coax and Smith, which I had read a few years previously.

    As it happened I was also experimenting at the time with several so-called Artificial Intelligence (AI) applications known as ‘Large Language Models’ (LLMs). A friend had explained that certain LLMs could be installed as computer applications, and would run off-line, i.e., without connecting to the Internet. (The example AI illustrated at the top of this page is not one of those.)  Since receiving this suggestion I had installed several off-line AI models, and probed them with various made-up (artificial) problems. The least interesting of these were questions of fact. The AI under test sometimes answered correctly and other times bizarrely, though with equal confidence. This latter phenomenon has a name in AI. It is called, ‘hallucinating.’

    The coincidence of reading about complex impedance and recalling Michael Toia’s exercise suggested the idea of probing AI with one or two problems of physical measurement. There can be no argument about measurement. The correct answer is whatever the instrument says, provided that the instrument and test setup have been accurately and fairly described. My plan was to carry out the physical measurements, not simply to calculate alongside AI and compare results.
Measuring one-meter length of coax

Rig Expert Antenna Analyzer Screen ShotNanoVNA Screen at 60 MHz    It would be less than 100% honest to refer to the transmission line impedance question reproduced at the top of this page as a ‘physical measurement’ exercise if no actual measurement were performed. However, complex impedance was in fact measured under the specified parameters, using a Rig Expert AA-230 antenna analyzer and independently with a NanoVNA. Both instruments have female N-connectors. That is why the piece of coax has a male N connector affixed to one end in the illustration above. The other end was cut with my pocket knife and rubbed free of small hairs, etc.

    Both instruments displayed fractional ohms for both resistance and reactance at 60 MHz. The Rig Expert plots both R and X on the same axis and the computer interface screenshot (left) is easy to read. The NanoVNA plots resistance and reactance on separate axes. Resistance is yellow, with zero at the bottom (resistance takes only positive values). Reactance is plotted on the the left axis (blue) with zero in the middle. Had either instrument displayed a value for impedance that was not approximately 0 ohms, something would have been wrong either with the measurement or with the device under test (DUT), as it is called.

Physical Intuition remark

    Sometimes AI responses can be spooky. In the clip above DeepSeek-r1 (Chinese AI) expressed doubt about its computed result of 0 ohms, saying that it seemed to conflict with physical intuition. What? Was DeepSeek claiming to possess ‘physical intuition.’ It has no direct exposure to the physical world (that I know of). How could it possess physical intuition? It would be a misleading intuition, in any case, perhaps based on the ‘thought’ that DC resistance would be infinite, so how could impedance be 0.

    For this particular exercise DeepSeek-r1 displayed the correct answer. Several other AI’s tested at the same time got wrong answers. Before reporting those I will mention a follow-up question, similar to the original one, and also inspired by Michael Toia’s Smith Chart exercise. DeepSeek-r1 and Gemini 2.5 Pro on-line (free trial version) both answered the follow-up correctly:
Follow-up question

    The off-line PC test setup for several of the AI’s used Ollama as the server / user interface. Ollama was installed in a Docker container, under Windows Subsystem for Linux (WSL). Later I also installed Open WebUI in a Docker container, so that it was possible to access the LLMs  that were pulled by Ollama either in command-line mode or from a browser. A friend had installed additional LLMs on his more powerful desktop computer, and also ran some tests. In regard to the ‘physical measurement’ exercise described in the preceding paragraphs, here are some of the other results. GPT4 (on-line) said +50j ohms. Copilot (on-line) said -50j ohms. o3-mini said -40+30j ohms. Perplexity said 152 + 0j ohms. These AIs also got the frequency follow-up part wrong. Llama2 or 3.3, phi4, gemma3, and Qwen QWQ were not tested, although these were available at various times during the testing period. I could only tolerate so much of this exercise.

    Most LLMs construct and generate output relatively succinctly, or so it seems. However, DeepSeek rambles endlessly, as does QWQ—are they the same or related in some way? At one point I became interested in the format of responses, and wondered if they could be expressed more beautifully, either using TeX or HTML.

Trig Identity Question

    The above is a revision of a broader question. I had asked the AI’s to identify a product-sum identity associated with the ring-diode mixer. However, none that were tested identified a plausible candidate identity. AI’s, or some of them, are good at elementary math, however. To the above prompt was added, “Please format your response as TeX.” (It is good to be polite when addressing AI. Who knows what the future may hold.) Of those tested, Phi4 produced the nicest result. Its output compiled without error in TeX Studio. The complete slightly expanded document may be viewed here.
Phi4 TeX-formatted output example

    A similar test was repeated, specifying HTML output format. This led to learning about a JavaScript display application called MathJax. By including a link to MathJax in the HTML header it is possible to display fairly nice math expressions in the browser, maybe not as consistently beautiful as TeX, but a huge improvement over raw HTML 5. Other made-up problems, not related to ham radio, were used to exercise AI-generated HTML.

    Another side-track was to test whether any of the LLMs that were accessible to me would be able to construct an LTspice circuit. I had looked at how example Spice circuits are encoded (.asc, .plt) and thought that an LLM might generate similar specifications, as they are able to construct working computer code, given a simple algorithm (specifications) as input. However, these efforts (ca. mid-March 2025) failed to produce a single usable schematic.

    Finally I also exercised the LLMs in a variety of common ways. “In the nursery rhyme ‘Three blind mice’ how many tails were cut off and by whom?” (It is not hard to challenge AI.)  DeepSeek produced the most entertaining answer, a long story—possibly a Chinese nursery rhyme, or a hallucination—that ended thus:
Deep Seek Nursery Rhyme Challenge
Projects Home



Project descriptions on this page are intended for entertainment only. The author makes no claim as to the accuracy or completeness of the information presented. In no event will the author be liable for any damages, lost effort, inability to carry out a similar project, or to reproduce a claimed result, or anything else relating to a decision to use the information on this page.