I was not expecting the part where Fable produces a passable 3Blue1Brown-style explainer video of the algorithms it just implemented that sounds like it's narrated by a character from Dora the Explorer.
Im building some music playback software and am currently struggling with the implementation of a spectrum analyzer to visualize the music.
This is incredible stuff and I learned a lot. Well done sir.
Ps, also mourning the loss of Fable! It sorted out a 3 month bug hunt odyssey in 3 days. For a somewhat novel problem in a pretty niche area (DSD DoP audio crackle problems during certain playback edge cases).
Fable was quite relentless, it was fun watching it work. I described my lisp interpreter project's short term plans and long term roadmap, Fable thought for like 20 minutes then just told me it was all "inevitable" and started working on the stuff. Ever since then I started to picture Fable as some kind of Terminator.
Left me that code and a massive code review that unfortunately didn't contain any of the I/O and memory safety hardening I wanted. I haven't fully reviewed the code yet. I get a little sad when I read it. Not a US citizen so I'm not sure I'll ever get to use a state of the art model again.
An octave (for example from a C to the next C) is a doubling in frequency. In the Western diatonic system, there are 12 notes per octave. (C, C#, D, D#, E, F, F#, G, G#, A, A#, B). Notes are "evenly spaced" within the octave - every note has the same ratio between its frequency and the frequency of the next note. Hence, that ratio is ¹²√2
I always knew this but used to wonder why 12?, as I'm sure a lot of people did. It turns out of course to be a human choice, but the convenience of that choice can be codified mathematically.
On a piano you move up an octave by going up by 8 white keys, or 12 semitones (white and black keys). Going up by 4 semitones is called a "major third", which multiplies the frequency of the note by 5/4. If you do three major thirds you get an octave. However, notice that (5/4) multiplied by itself three times is 125/64 which is actually slightly less than 2.
In fact there is no way to tune a piano perfectly - there has to be a compromise in the intervals somewhere. The reason for this is exactly that no rational number (fraction) equals 2 when raised to an integer power.
This is the equal-temperament tuning, which is used a lot now because it's simple and consistent. Other tunings were common in the past, where they'd set notes to simple fractions of other notes like 3/2 of the note 5 below, idk the actual numbers they used. 3/2 isn't any power of the 12th root of 2 so nothing in this new system is actually as harmonic as it should be, but nothing is particularly dissonant either and the system is simple.
That was my experience with Fable as well. Pulled my extremely complex project that I could squint and see was possible, but actually put mathematical concreteness to things in a way I could only intuit.
On the flip side, visualizers have always fascinated me. I love this one, but one build off I've always wanted to see: analyze the entire file a priori, and then generate the visuals. Sort of like a normalization pass, but getting longer form structures decoded ahead of time could be pretty neat.
I've been planning exactly what you describe in that second paragraph for creating videos for the music I make. It's a lot easier than doing it realtime, and because I make the music, I'm planning on doing it multi-track so I can put individual stems in.
Kinda interesting how its just like a FFT chart in a circle but perhaps the author is not aware that is the case. Would be curious to know what things were "implmentation details" for the fancy AI and what wasn't.
I could be wrong but milkdrop already would do light FFT analysis for effects right?
I may have been overselling the AI's initiative in the article -- it still required a fair bit of steering. I put most of the prompts involved here: https://saltblock.neynt.ca/waveloop-prompts.md
Wrapping FFT in a log2(freq) % 1 spiral was part of the human direction :)
I suggest you watch the explainer video the ai made, its pretty awesome, but yeah, thats exactly what it is, with some depth to how exactly it uses FFT, and solving some problems with getting good resolution on different frequencies
One of the weaknesses of the video is that there are artifacts in the narration of passing through a text layer. "Bass" is pronounced as the fish at one point. "Wound" is pronounced as the injury. It's clear that these are homonyms of what was actually intended by the script.
Those are the things that can be most clearly pointed out but the (wrong) meaning is also constantly asserted by the TTS voice using the wrong inflection, intonation, and emphasis.
It honestly makes my ears bleed. To me, it sounds like an extremely unintelligent person reading a teleprompter. Absolutely nothing going on between the ears.
>The writing is also literary. It draws an analogy between the 12 musical pitch classes and the 12 markings on a clock. Noise lingers. Material surges off the rim.
I absolutely hate this revolting writing style by LLMs
Thought this might be about the video game series, and even after seeing "Fable 5" my first reaction was "wow, up to 5 already?". It looks like they made it up to Fable III before rebooting the series with a 4th game (well, it's not out yet), so there is no Fable 5 yet.
TFA seems to be about some AI thing. Crazy how many words are actually just AI things now. Learning, reinforcement, language, model...
That generated video was eye opening for me. I've been using Opus in Claude Code for studying and at work, but it never occurred to me to use 3b1b's excellent python library for generating maths visualisations to let it generate such good graphical demonstrations.
My Fable example is not nearly as cool but still (to me) impressive.
Last year, I would occasionally test the latest models by vibe-coding in-browser music generators using only HTML, CSS, and JS. Here’s one made in July by Gemini:
That Fable Generated video is something else... wow I love it.. along with the app
Anyone who says LLMs can't reproduce intelligence I mean really? can you make this? its not just a talking database guys or a stochastic parrot...
too bad Fable was nerfed/gatekept by the Trump corruption selection committee..but the technology will not be silenced.. we just need to get humans ready for this capabilities. the jury is out on the future of that.
> This model doesn't shy away from drawing upon all its knowledge. It casually refers to alpha premultiplication and fundamental frequencies in the same breath. It is fond of acronyms.
Yes - I had Fable tackle some long-standing bugs in some code I had and I quickly lost track of what it was talking about and had to ask a lot of clarifying questions.
It killed my bugs like they were nothing though. Opus and even GPT5.5 had churned on these same things for ages, but even with my manual help we made no progress.
It felt like they weren't the slightest bit challenging for Fable. So glad to see it back!
Did you consider some kind of arrangement of the notes that is less "equal-tempered", and that would highlight the harmonic relationship between the notes? probably would need to be key-specific.. or maybe rather now would you that fable is back?
34 comments
[ 2.8 ms ] story [ 60.9 ms ] threadWhat a strange era we now live in.
https://youtu.be/mRO_QonhC2c
This is incredible stuff and I learned a lot. Well done sir.
Ps, also mourning the loss of Fable! It sorted out a 3 month bug hunt odyssey in 3 days. For a somewhat novel problem in a pretty niche area (DSD DoP audio crackle problems during certain playback edge cases).
Left me that code and a massive code review that unfortunately didn't contain any of the I/O and memory safety hardening I wanted. I haven't fully reviewed the code yet. I get a little sad when I read it. Not a US citizen so I'm not sure I'll ever get to use a state of the art model again.
Nods knowingly. Yes, of course. I definitely know this.
An octave (for example from a C to the next C) is a doubling in frequency. In the Western diatonic system, there are 12 notes per octave. (C, C#, D, D#, E, F, F#, G, G#, A, A#, B). Notes are "evenly spaced" within the octave - every note has the same ratio between its frequency and the frequency of the next note. Hence, that ratio is ¹²√2
On a piano you move up an octave by going up by 8 white keys, or 12 semitones (white and black keys). Going up by 4 semitones is called a "major third", which multiplies the frequency of the note by 5/4. If you do three major thirds you get an octave. However, notice that (5/4) multiplied by itself three times is 125/64 which is actually slightly less than 2.
In fact there is no way to tune a piano perfectly - there has to be a compromise in the intervals somewhere. The reason for this is exactly that no rational number (fraction) equals 2 when raised to an integer power.
As any reasonable person would.
On the flip side, visualizers have always fascinated me. I love this one, but one build off I've always wanted to see: analyze the entire file a priori, and then generate the visuals. Sort of like a normalization pass, but getting longer form structures decoded ahead of time could be pretty neat.
[1] https://en.wikipedia.org/wiki/Audiosurf
I could be wrong but milkdrop already would do light FFT analysis for effects right?
Wrapping FFT in a log2(freq) % 1 spiral was part of the human direction :)
Am I missing a joke? L5 is just a single promotion away from hiring-out-of-college, at least for the FAANG that I was at.
One of the weaknesses of the video is that there are artifacts in the narration of passing through a text layer. "Bass" is pronounced as the fish at one point. "Wound" is pronounced as the injury. It's clear that these are homonyms of what was actually intended by the script.
It honestly makes my ears bleed. To me, it sounds like an extremely unintelligent person reading a teleprompter. Absolutely nothing going on between the ears.
I absolutely hate this revolting writing style by LLMs
TFA seems to be about some AI thing. Crazy how many words are actually just AI things now. Learning, reinforcement, language, model...
Last year, I would occasionally test the latest models by vibe-coding in-browser music generators using only HTML, CSS, and JS. Here’s one made in July by Gemini:
https://gally.net/temp/20250701synthesizer-gemini2/index.htm...
And one made in September by Claude:
https://gally.net/temp/20250917rhythmdrone/index.html
With Fable, I was able to one-shot something much more sophisticated:
https://gally.net/temp/20260610-fable-synthesizer/index.html
It’s still a long way from creating music I would want to listen to, though.
Anyone who says LLMs can't reproduce intelligence I mean really? can you make this? its not just a talking database guys or a stochastic parrot...
too bad Fable was nerfed/gatekept by the Trump corruption selection committee..but the technology will not be silenced.. we just need to get humans ready for this capabilities. the jury is out on the future of that.
Yes - I had Fable tackle some long-standing bugs in some code I had and I quickly lost track of what it was talking about and had to ask a lot of clarifying questions.
It killed my bugs like they were nothing though. Opus and even GPT5.5 had churned on these same things for ages, but even with my manual help we made no progress.
It felt like they weren't the slightest bit challenging for Fable. So glad to see it back!