I'm "forced" to use Agentic Coding (AI) at work and so I'm pretty familiar with it and what it can (and can't) do well. As such, I was very skeptical that AI would be able to do any kind of usable sequencing.
Some examples people have posted have amplified that belief... they were just... boring.
But then, I hate to say it but I find a lot of pixel sequencing boring.
However a couple weeks ago someone on the LOR forum posted they were having good luck with AI sequening. I asked for an example, prepared to dislike it.
But it wasn't so bad. It was decent, and probably better than a lot of other sequencing I've seen. Now to be clear he admitted he didn't just hit generate and call it good, he went back and forth and described what he wanted in each section.
It was enough to make me try it myself. Rather than start with a brand new sequence from scratch, I put it to work on my existing sequences, where I need to convert 24 channels of AC canes over to 24 pixel canes and their myriad of new capabilities.
Long story short, it was good enough that I'm about halfway through my sequences. Instead of doing it all by hand, I'm letting AI come up with a concept (based on examining props in the rest of the sequence) and then it proposes a plan, which I approve or tweak. Then it generates a sequence, and we go back and forth on revisions.
It's been working pretty well, and I'm impressed that it's getting the file format perfectly and I haven't had a single load issue or other glitch that I've noticed.
Once I get through my existing sequence conversion, I'm planning on finding a new song, and saying "based on what you know about how I sequence, go for it". I'm not holding my breath that it will be passable, but who knows?
I'm using Claude Code for this BTW
I tried it with ChatGPT last year, but the output was so erratic that it looked more like a bad connection than a properly controlled signal. And that was with just a single channel. I haven't had a chance to revisit it this year.
ErnieHorning wrote:
I tried it with ChatGPT last year, but the output was so erratic that it looked more like a bad connection than a properly controlled signal. And that was with just a single channel. I haven't had a chance to revisit it this year.
AI relies very heavily on pattern matching. So in my case, I was able to basically tell it "look at what the AC canes did last year, then look at what the other pixel props are doing..." It also examines the audio file for major hits and drops.
And it's done a reasonable job of coming up with a good translation.
Then we go back and forth on specific details: "at 2:20, you're overusing this effect, let's try some new effects/colors/whatever".
AI is extremely literal too. I will say "you're overusing red/green, let's try some other colors like cyan, purple, etc". and it will respond "OK I have added more effects in cyan and purple". It doesn't understand the broad request that I just want it to use various colors, not just to add those two examples.
I had the same problem when I was developing a game this summer. I was trying to set up "personas" and I gave it some examples: "Let's say Bob is very conservative in their play and doesn't like to take risks, and Susan is a gambler and will sometimes take stupid risks..."
Throughout the next several days I had to keep scolding it for literally setting up a "Bob" and "Susan" persona. It literally wrote code with personas called "Bob" and "Sue" and I had to keep pushing it to invent other, similar personas.
I experimented with this a bit this year. It is true that the more context and direction you give it, the better the outcome. I ended up giving Claude all my previous sequences and told it to study them. After I did this, the outputs I got became so much better. After many back-and-forths and giving it copious notes and directions, I now have a new song on my show this year. To be able to complete a new song from scratch in a day and a half is pretty amazing.
A new song (or songs) is my ultimate goal.
I'm pretty picky about results so it's not clear if it will give me a good starting point, an almost perfectly complete sequence, or nothing at all, when starting from scratch. But it *will* have access to my entire sequencing library.
Claude gives me the same issues with sequencing as I get from it in my daily work using it to write software code: over-generalization, hitting many of the broad strokes but missing the details, etc.
I had to tell it to stop copying the same effect to every candy cane, when I had groups defined that it knew about, for example.
