"We did it, Patrick! We made a technological breakthrough!"
A place for all those who loathe AI to discuss things, post articles, and ridicule the AI hype. Proud supporter of working people. And proud booer of SXSW 2024.
AI, in this case, refers to LLMs, GPT technology, and anything listed as "AI" meant to increase market valuations.
I don't know music theory, is this accurate? if so I think synthesia is pretty practical to use for checking.
That's the point that many specialists tried to make for about three years now: LLMs don't know things. They arrange words in an order that looks plausible. It doesn't understand music theory (and it can't count the 'r' in raspberry).
Like trying to use a hammer as a drill. AI has its uses, but good models are built to specialize around certain data sets. You can’t expect a spelling checker to teach you about music, but that’s how these things are being sold.
I just found out that Claude Pro can't be trusted ~~to do anything related to songwriting.~~
FTFY
I'm assuming the answer was wrong, but as a non-musician I don't see it.
This is just the thing - if you don't understand the subject, the AI output seems perfectly reasonable, and you don't see the need to look further.
If you understand the subject the AI is spouting off about, you immediately see that it's full of shit and shouldn't be trusted.
Which should be a warning if you use it for a subject you aren't familiar with. Why would it suddenly become very good at output on something you're not sure of? It can be useful as a sounding bound of your own ideas, as it's very good at taking pieces of what you input and completing them in various ways. Keeping that within the context window to prevent wandering, you are modeling your own ideas. But that's not how lots of (most) people are using it.
That's remarkably like watching movies or television when the writers get near your area of expertise...
G to E is a major sixth, not a major seventh. That's the mistake. It then misidentifies the chord because of this.
Just bolstering one of the other comments with a more visual approach to show just how simple the deduction would be, even if you don't understand music.
Notes are only A - G and they repeat (i.e., G loops back to A). In the example, G is the 'root' and considered note #1, so when you get to F it loops back to G to complete the scale/octave. Armed with that knowledge, you can see more clearly how claude bungled it by laying the notes out like below. It got B and D right, but couldn't do simple arithmetic to place E.
It's basic deduction for a human English speaker knowing that E immediately follows D, and therefore should be 5+1 = 6. Such a tiny, simple thing but shows just how scary it is that people trust this stuff blindly and don't corroborate the info given. Now imagine a young, fresh chemist or physicist fully trusting the output because they've been taught to by their professors.
This isn't the G major scale, though. You need to add flats and sharps to get the right number of semitones (two per step except for 3-4 and 7-8), in this case F♯.
For example, the major 3rd from D is not F but F♯ because the D major scale has F♯ and C♯. Therefore, the basic 1-3-5 D-major chord is D-F♯-A.
I wasn't depicting a specific scale, just keeping it simple for those that don't know music and showing what the obvious flaw was in the screengrab. The specifics aren't necessary for the point to be valid and only serves to muddy the waters for folks that don't know theory.
Better to learn later than never.
It is widely known and has been reported for the past 3 years that AI hallucinates and cannot be trusted, but not very accepted I guess due to all the lies by the tech bros that AI is PhD level or above.
Just waiting for that POP!