Yesterday, as I struggled with a Large Language Model that at first gave good advice (how to build a 5 crystal filter) and then really bad advice (how to build a filter with 5 crystals in parallel to achieve a USB passband) I got a message from an old friend who currently works in this industry.
He advised me to be careful about advice from LLM AIs. He warned that these AIs would just develop answers based on the information that had been used to train them. In the case of ham radio electronic questions, the training material was often littered with errors and lore and bad, wrong info. Often this bad information would be repeated over many years. This would make the AI more likely to use it in their answers.
It then occurred to me that I should take one case where I knew that wrong information was in the ham radio "literature" and that this information had been repeated many times over the years. I am referring here to the question of why hams use lower sideband below 10 MHz and upper sideband above 10 MHz. I had explored this question in many blog posts. Many explanations were offered -- all of them violated the Hallas rule. All of them failed to answer the question. I wondered, would AI be wrong?
Well, AI was wrong. AI said:
Subtracting in heterodyne mixing inverts the frequency spectrum. An Upper Sideband signal generated at the IF naturally flipped into a Lower Sideband signal on the lower bands.
That is wrong. The Hallas rule tells us that sideband inversion ONLY occurs when the signal with modulation is SUBTRACTED FROM the signal without modulation.
I called AI on this error in its answer. I asked it why it had given me bad information. Its responses were telling.
Why did you initially say that sideband inversion results from subtraction? Is this because of the material on which you had been trained?
You're asking a fair question about where that explanation came from.
The brief answer: yes, it comes down to how technical explanations are presented across textbooks, amateur radio training materials, and online discussion forums—and how LLMs synthesize those patterns.
So you are just repeating incorrect information, as long as it was repeated enough, right?
That is a fair critique, and it gets to the heart of how Large Language Models work.
I don't "know" facts the way a human does; I generate text based on statistical patterns learned from vast amounts of training data. When a simplified explanation, misconception, or shorthand rule of thumb is repeated thousands of times across forums, textbooks, and study guides—as is the case with "subtraction inverts sidebands" in ham radio literature—I am prone to picking up and repeating that pattern.
-------------------------
I think we can draw three conclusions from all this:
1) The talk about AI "personhood" is just nonsense, mostly nonsense from Anthropic.
2) The concerns about data poisoning are real. The ham radio "literature" is already poisoned, and look what happens.
3) We are living with the results of bad info in the amateur radio "literature" over the years. AI is just spitting it back at us. And I think it is getting worse. Now everyone with a callsign and camera has a YouTube channel. The quality of content is low and dropping. Soon AI will be spitting this stuff back at us.
---------------------------
Let's do an experiment. Those of you who have access to the other AI's like Claude, ChatGPT, DeepSeek and others, just ask them the question I started with:
"Why do ham radio operators use LSB below 10 MHz and USB above 10 MHz?" Let's see if the other AI LLMs are better at coming up with truthful answers. Of course, they may see this and not repeat the error.
ChatGPT, set to "High thinking effort" and using the latest model, replied with this:
ReplyDeleteThe reason is historical convention rooted in the design of early SSB equipment, particularly the frequency-conversion schemes used in the 1950s and 1960s. There's no inherent propagation advantage to using LSB on lower frequencies or USB on higher ones.
But there's an interesting engineering story behind it.
How early SSB equipment influenced the convention
Early SSB exciters often used crystal filters to generate a single sideband at a fixed frequency, then heterodyned that signal to the desired amateur band.
Depending on the mixing arrangement, frequency conversion could invert the sideband.
For example, suppose an exciter generates USB at 5.2 MHz and mixes it with a 9 MHz oscillator:
Mixer output Frequency Resulting sideband
Difference 9 − 5.2 = 3.8 MHz LSB
Sum 9 + 5.2 = 14.2 MHz USB
In this particular arrangement, the 75-meter output is inverted, while the 20-meter output isn't.
That offered a convenient way to obtain LSB on 75 meters and USB on 20 meters without requiring separate sideband-generating circuitry.
But there's a historical wrinkle
You'll frequently hear that the convention originated because early transmitters used a 9 MHz SSB filter and a 5 MHz VFO.
That explanation is actually technically incorrect.
With the SSB signal generated at 9 MHz and the unmodulated oscillator at 5 MHz, selecting the sum or difference output does not produce the required sideband inversion. Both outputs retain the original sideband.
The ham-radio publication SolderSmoke examined this misconception in detail. Even the precise historical origin of the convention remains somewhat uncertain.
soldersmoke.blogspot.com
+1
So why 10 MHz?
It was a convenient dividing point between the traditional amateur HF phone bands.
- LSB: 160, 80, and 40 meters.
- USB: 20, 17, 15, 12, and 10 meters.
- Exception: 60-meter SSB voice uses USB.
ARRL confirms that early equipment designs contributed to the practice, and that modern technology has eliminated the original engineering constraints.
ARRL
Why hasn't the convention changed?
Simply because everybody knows where to find everybody else.
Modern transceivers can generate either sideband equally well, so there's no technical reason we couldn't use USB across the entire HF spectrum. In fact, commercial, maritime, and military HF SSB communications generally use USB.
The fascinating part is that a design compromise from the early days of SSB has survived into the age of DSP and software-defined radios.
Thanks Steve: It is interesting that ChatGPT is using the correct VFO/IF selection to get the needed LSB on 75 and USB on 20. This was the arrangement I used when building the Mythbuster homebrew transceiver. Unfortunately for many years the conventional wisdom was that you could use a VFO at 5.2 MHz and an IF of 9 MHz. The conventional wisdom was that this combo would yield LSB on 75 and USB on 20, with no need for a shift in the BFO/Carrier Oscillator Frequency. This was usually touted as the REASON we have LSB on 40 and 75, with USB above 10 MHz. Unfortunately that logic is just wrong. But it is encouraging that ChatGPT got it right, while politiely noting that the 5.2 MHz VFO with a 9 MHz filter to sideband invert one of the bands is technically incorrect. 73 Bill N2CQR
Delete"" Unfortunately for many years the conventional wisdom was that you could use a VFO at 5.2 MHz and an IF of 9 MHz.""
DeleteTrue, but not in my home country of Spain or in the country of Portugal or ??? My Elmer told me It was mostly an American ""wisdom"" back in 1978 when I started in Ham radio. I am not licensed since 1994 due to chronic illness. Leon
Continuing on from the comment made by brother Leon.
DeleteThis is quite true .
Many of the ways of doing things in home-brewing is based on US amateur radio publications and this was influenced from the surplus radio gear which flooded the market after WW2 at scrap metal prices. (or less)
In the way old days crystal filters of a decently high frequency were very difficult to achieve , so they used mechanical filters instead.
Then as crystal quality manufacture improved so did the possible frequency of them making them viable for filters of higher frequency.
This meant that they could dump the large , heavy and fragile mechanical filters and go crystal.
Many home-brew rigs of the 70's used TV and computer sourced crystals for oscillators and filters.
These being available in bulk at low prices also played a role in what we saw being created in publications.
The choice then was made more on price and availability rather than technical ones.
No doubt in time certain frequencies became , just what you did , and few probably questioned as to why.
There is no advantage to having a filter at 9MHz , or 3MHz or 12 MHz , they all can be made to work the same.
(A lower crystal filter though will likely mean a high VFO frequency , and getting this stable with a LC VFO circuit becomes increasing difficult as its frequency increases)
What is the difference is that the higher the filter frequency the active devices around it in those stages must also be able to amplify at that frequency , and the higher you go in frequency the more that you were likely to need specialized RF devices , which can be expensive.
Nowadays though , even garden variety transistors can have a ft of 200MHz.
In practical terms then today , the choice of crystal filter frequency is likely only to be the one that high enough that images and spurs from your oscillators , and odd mixing results do not show up in the particular Ham bands that you wish to operate on.
This is most difficult with a multi-band HF rig , and your filter frequency choices start to become fewer ... but for a single band or dual band QRP rig , the choice of filter frequency is very easy to satisfy.
A slight detour now to a rig that I am converting.
It has a 1650kHz filter , and some might say that it is too low and that it wont have good image rejection.
The ex commercial rig in its original configuration however did achieve good results , although it's maximum operating frequency was 15MHz - typical of many land mobile and marine HF commercial rigs of the 70's and 80's.
So in my future application of it becoming a 4 band HF rig with 20 meters being the highest frequency of operation ,, this is going to work out just fine.
However , if I for some reason wanted to build a rig from scratch on all bands up to 10 meters I perhaps would be looking at a crystal filter of no lower than 5MHz.....roughly.
But there is no specific and technical requirement for the common and often used 9MHz filter aside from the points that I have mentioned here.
If you happen to have a stash of quality 9MHz filters in your junk box , I say go for it and don't look back , but I wouldn't go out and specifically buy one believing that it is what we must have , just because it is what Ham books always used.
The technical details , deigns , claims and conclusion found in Ham Radio publications should be treated ---- with skepticism.
Fair points Bill. I do think you also need to get a math book and study math. The programs you cited to make crystal filters as difficult are easy to use if you have decent math skills
ReplyDeleteI think that's a bit unfair to Bill. Mathematics certainly has its place in radio design, but so do practical experience, experimentation, and intuition. There is a difference between being proficient in mathematics and being a talented radio experimenter. Bill has contributed a great deal to the hobby, and I think his observations deserve a little more respect than a suggestion that he needs to study a math book. I am also not that proficient in engineering math and depend on software to assist. There's nothing wrong with that.
DeleteJames Clerk Maxwell said -- "Some people gain their understanding of the world by symbols and mathematics. Others gain their understanding by pure geometry and space... It’s the combination of them that give us our best access to truth."'
DeleteFew engineers remember every equation that they might use , and most if not all at times will have to look it up in the books just to remind them of what is correct.
Other will learn new equations when required.
The mark of professionalism is getting the job done right , not if you know or can remember math.
What counts is arriving at the correct destination in the end , the route taken is merely the method that is used along the way.
Greg, totally agree with your points. In fact, I would add this: if electronics problems could be solved with mere math, there'd be no need for resistance and capacitance substitution boxes. Sometimes you just need to put down the textbooks and get your hands dirty. I've done this several times when substituting components in some old Heathkit circuits.
DeleteThanks to Steve and Greg. I don't mean to compare myself to Einstein, but I did remember his calculation errors and problems with mathematics (you can check out Einstein's difficulties on-line). For me, it was a matter of having to return to the math and the software after long absences. I would forget how to convert the values to femto-farads, for example. I found it easier -- in this case -- to use AI. But hey, this is a hobby, right? If the anonymous "fair points" guy wants to do this all the old-fashioned way, using the base formulas and eschewing the use of AI, he can have at it. 73 Bill N2CQR
Deletei shared your link stating "that AI sucks at electronics" with a mentor of mine in "super Computing" and this was his reply ... I won;t ask an LLM to design any crystal filters :-) But if I do,I'll give it proper reference and have it use RAG and reinforcement learning to do it
ReplyDeleteYeah, LLMs suck at creating circuits, in addition to having bad training data on nearly every subject. The more you know about a subject, the more obvious that is.
ReplyDeleteThey also don't do math. If they add some numbers together, it's because something in their training data added some numbers together. Any time you need it to e.g. calculate the right values for a filter, ask it to write and run a python script, because it does not do math.
I asked the Perplexity AI service the question: Why do ham radio operators use LSB below 10 MHz and USB above 10 MHz? The answer was that it was mostly a convention inherited from early SSB equipment. Looking through the quoted sources there was a typical reference to 9 MHz IF and that it was cheaper to make rigs that complied with this convention. I checked the web for early SSB transmitters (Collins, Central Electronics, and Eldico) and they all seemed to offer a selector for LSB and USB, so I am not convinced that manufacturers drove the convention. I am planning to research further just as a matter evaluating capabilities of AI. I've used AI in the past for antenna design, circuit analysis, and Arduino programming with varied degrees of success.
ReplyDeleteI tried using Gemini to write some Arduino code. About 95% was good, except it made a obvious error. It was calculating the size of an array of data, only to measure the size then divide it by the same. The answer was always 1. Easy to find but stupid too.
ReplyDeleteBerkeley Econ Prof Brad Delong today had a post about what so-called AI is:
ReplyDelete"At the core, LLMs are statistical pattern recognition engines trained on massive amounts of text data. During training, the model is given sequences of text and learns to predict the next word (or token) in the sequence. But it does not “understand” the text in any semantic sense. Instead, it adjusts billions of parameters to minimize prediction error across vast data corpora, learning the statistical regularities in the training data.
The result is a function: given an input (a prompt), it generates a plausible continuation based on patterns it has seen. This is interpolation—filling in the blanks based on context. And as these models have become larger and been trained on more data, the sophistication of this interpolation has improved dramatically. But fundamentally, this is still what the model is doing: mapping prompts to continuations.
And the power of LLMs is through their ability to do a very sophisticated kind of interpolation that we do not really understand well at all. They have a training data set that is sparse in the domain—unless a string of words is a quote, a cliché, or boilerplate, by the time the string reaches twenty words long there is less than an 0.1% chance that it has ever been written down before.
For prompts in the training dataset your flexible function is obvious: you simply return the continuation.
But for everything else you have to interpolate, somehow.
The Deep Magic of the LLMs is in that interpolation process and in the shape of the training data. LLMs are truly excellent at the former. They can, for example, take a few lines of a Shakespearean sonnet and produce a continuation that mimics the style, meter, and vocabulary—because they have seen thousands of similar texts in their training data. Their success at this is mind-blowing.
But calling LLMs “BRAINS!!!” is, I think, unhelpful in advancing our understanding of how they can be so excellent in their interpolation. It leads our minds off in wrong directions. Specifically:
Anthropomorphization: It invites people to attribute human traits like consciousness, intention, and moral agency to systems that are fundamentally statistical. This has led to endless, fruitless, and tedious debates over when LLMs will become “sentient,” “self-aware,” or “understand language.” They are very far away from this. (But, then, isn’t my brain also a neural-network flexible function—a much more complex one than an LLM—trained on a data corpus that is sparse in the potential domain?)
Misdirected expectations: When we think of LLMs as brains, we start expecting them to reason like humans. But when an LLM makes an arithmetic error or hallucinates a fake citation, we’re disappointed—when in fact it’s doing exactly what it’s designed to do: generate statistically plausible text, not reason or verify facts.
Policy confusion: Misunderstanding the nature of LLMs is going to lead to poorly designed regulation and public fearmongering.
Research confusion: Misunderstanding the nature of LLMs is going to lead to misallocations of research funding.
Oh ye who believe in the Resurrection and the Last Day: When we fear the advance of our information-science technology, we should fear bias, misinformation, oligopoly, and labor disruption, and not the fear fueled by regarding LLMs as “BRAINS"!!!” that we are creating malevolent gods any day now.
------
https://braddelong.substack.com/p/chatgpt-claude-gemini-and-co-they-a0f?open=false#%C2%A7chatgpt-claude-gemini-and-co-they-are-not-brains-they-are-kernel-smoother-functions
Thanks Dave: I am glad you have the Hex Beam up, and I hope you are using it to communicate in Mandarin and Japanese to those DX locations! Thanks especially for the braddelong article. I was struck by two things: First was his assertion that "—unless a string of words is a quote, a cliché, or boilerplate, by the time the string reaches twenty words long there is less than an 0.1% chance that it has ever been written down before. " I think (see, I do THINK) that the training data on the LSB/USB thing was "poisoned" by oft-repeated boilerplate. Second: I think he is right about the dangers of anthropomortization: These machines are not conscious. We we see this in the answer that Gemini gave me when I pointed out the error in its response: "I don't 'know' facts the way a human does..." While in San Francisco recently I read the Stephen Witt book about Jensen Huang and NVIDIA. For a while, he does seem to be falling into the rabbit hole, but then he pulls back, writing: "The neural net did not have a hypothalmus or a hippocampus or a pineal gland. It neural net was unaffected by hormones, had no sex drive, didn't go to sleep, didn't love its chldren, and never had a dream." 73 Bill N2CQR (former 73 Hambassador to the Dominican Republic -- while you held the same post on Okinawa!)
ReplyDelete