From Techmoan's video The man with the magic EQ at 19:37 (YouTube)
Techmoan just employed AI to find which of the many 78-RPM record EQ standards his equalizer uses in its "78" mode. He made recordings in RIAA and 78 modes, named them accordingly (shouldn't have to avoid bias), and uploaded them into a chatbot... Does he think the chatbot actually ran code on the server that
- calculated the frequency spectrum (FFT) of each track,
- divided 78/RIAA at each data point (or subtracted on log scale) to get relative values,
- normalized to the well-known RIAA curve,
- plotted the curve and identified roll-off frequencies and slopes?
I'm not saying this is impossible with AI, I'm sure a token-burning, recursively-prompting LLM (aka "agent") could write a Python script that runs on the local PC and actually does this. But no company would offer people to prompt agents on their servers for free like that. Pretty sure the bot just made up values to plot with Matplotlib or something. The first chart has a big red flag: the noise makes the curve go beyond the Y-axis maximum, suggesting the clean version existed first and the chart was "cropped" to that, then the noise was added. The second two charts imply that the bottom one is the derivative of the top, which it obviously isn't (if you know calculus). And RIAA EQ being +12 dB to +14 dB louder than the "78" preset across most of the audible spectrum? There's barely any volume difference when he toggles the switch in the video!
~~And there's no 500Hz bass roll-off the AI was so adamant about!~~ see edit below
Edit: here are annotated pictures:
Edit: the lack of 500Hz roll-off in these charts (if they are to be believed) means the "78" and "RIAA" characteristics are the same around 500 Hz, which is to be expected for Brunswick, since it uses the same 500Hz roll-off as RIAA. I just forgot that the AI didn't do the important Step 3 in my outlined procedure to get the "78" response curve. Still, the derivative is either smoothed beyond usability or plain wrong. Plus, Techmoan misrepresented the result (see second set of subtitles). If we take "Measured relative EQ curve" at face value like he did, Brunswick and RIAA are virtually identical in bass with their 500Hz roll-off (see the chart, it's near-flat below 1500 Hz). The difference is in the upper frequencies.
Also, is that actually ChatGPT's interface? Never used it but the "citation" sure looks like from Google's AI overview.
It's hard to say what AI he actually used and if the result was fine or not. I know for a fact something like Microsoft Copilot does write python scripts in the background to do data crunching. That's a pretty common business thing people use Copilot for, so that's what it does. The issue I have with that is that it doesn't share that code, but just confidently shows the results. So checking it is kinda hard, but at least I know that's a feature of the system.
With the regular ChatGPT online interface, I'm not sure. It does have tool calling capabilities, but if it would write a script for data processing or happens to have a sound analysis tool available? I'm not sure. But he might have used a different service or a paid subscription or something. For a lot of people "ChatGPT" is just synonymous with any AI tool.
You are right, this is a terrible way to use AI and I personally wouldn't trust any output like this. However I feel like this is the way it's marketed to people and the way most people use it. It's so confident in the answers, people just trust it. And when they are out of their depth, they don't have any way to really check it.
Even though I hate this, I don't blame Techmoan. He wanted to do something, but lacked expertise. I'm going to give the benefit of the doubt and say he didn't reach for AI right away, but actually tried to do the analysis himself. Then when that didn't work, he threw the data into this tool which says on the tin it can do it. The tool then confidently gives him exactly what he needed. We know that result might be total bullshit, but how is someone without any knowledge to know? It does look exactly like the results one would expect (which is what AI does, it generates an answer that looks right regardless of truth).
He's now at risk of AI psychosis as well. As he thinks this went well, he might reach for it more in the future and believe everything it says. If he starts using it more and more, that's a bad path to go down, especially for a YouTuber.
Yes, the voice with which he says "our friend AI" suggests he is trying if it lives up to the marketing, not realizing it's very likely to just infer what he wants to hear from the file names. I don't think he paid for anything.
But it looks like the AI did write some Python to make these charts (apparently with Matplotlib*). Did it actually analyze the files (probably by importing the SciPy library), rather than just make up the data? I highly doubt that the free tier of Gemini or whatever would, and the badly cropped first chart is a big red flag.
* If an AI is to be marketed as multimodal, it basically needs to have some image generation available. AI bitmap generation is expensive and bad at text so the company realized that if a nice chart can be made with a bit of LLM-generated Python, they better make Matplotlib available to the text model. Other libraries? Unlikely...
Yeah agreed. I hate it how all these AI companies just flood users with marketing and have their tools obfuscate what it is actually doing. Unless you ask each time: "Did you just make that up?" you don't know how much the result can be trusted. The filenames are a big mistake, they should have been neutral.
But for a regular user even if it showed all the Python code and what it did, what use would that be? Most people can't understand that at all. And even tho I have 40 years of professional experience in the software development stage, I don't know anything about sound processing. The script might load in the wave file and pass it onto some kind of processing library. But are the parameters correct? Is that even the right function to call? What caveats do those results have? How should one interpret those results? I would have no idea really and either spend a lot of time diving into it, ask a friend that does have that kind of knowledge or just give up.
It's so easy with AI to think people can do stuff they otherwise couldn't. But even if the chatbot correctly understood the request, generated the right code, fed the data correctly into that code and then produced a result, you'd still need an expert to understand if any of that was correct and how to draw conclusions from that. That's the part that LLMs can't do and probably can never do. They are trained on the internet, full of confidently incorrect people. People who have knowledge and help out someone on a forum somewhere, knowing what to apply and what not, guiding a person for that one exact usecase. The LLM then just takes that and applies it to all similar cases, correct or incorrect.
They're incentivized to lie about the process. You are correct that the only way to verify is to check the underlying code for audio processing and understand it, at which point it's easier to write yourself. The Python libraries for audio processing are not that hard. He should have made a Patreon post with the two files and someone would have come up with such script and hundreds of supporters could check it.
Because the audio source is a phonograph (turntable or Tefifon), it suffers from speed inconsistencies (wow and flutter) between playbacks, usually up to 1%. This could result in noise similar to the first image. The best way to analyze is with a digitally generated frequency sweep through the equalizer's line level input (alternatively, its phono input using a level convertor).
And you are correct about the risk of wrong interpretation: if the EQ is indeed Brunswick, his last two sentences are wrong:
Brunswick and RIAA are virtually identical in bass with the 500Hz roll-off. The difference is in the upper frequencies.