iamnor

There’s a particular kind of lie we tell ourselves when we’re tired and a little too confident.

It usually starts with a genre you love. For me, it was trip-hop, that late‑night Bristol fog: Portishead’s cigarette‑ash melancholy, Massive Attack’s low‑end gravity, Tricky’s off‑kilter menace. The stuff that sounds like it was recorded in a basement with a broken lamp and a working soul.

Then you look at the current crop of “text to music” models and you think: the sound is recognizable… the vibe is describable… surely this is a problem a machine can chew through.

So I did what any reasonable security curmudgeon does when presented with a shiny new capability: I tried to break it by building something real with it.

The project was called makeTripHouse. The goal was simple, maybe even unfair: generate trip-hop with enough authenticity that it didn’t feel like “trip-hop themed stock music.” And I did it fast, a few days, a handful of late nights, lots of coffee, and the kind of optimism you only have before the first bad result lands in your headphones.

What I got back from the machine wasn’t nothing. It was worse than nothing.

It was almost.

And “almost” is where AI makes you waste time if you don’t know what you’re looking at.

The first failure looked like a success

I started with MusicGen (via AudioCraft). You give it a text prompt, it gives you audio. In the same way a lot of security products advertise “just turn it on,” the experience is seductive. You type a few words that feel right:

“92 BPM, trip-hop, melancholic, breakbeats, Rhodes, vinyl crackle, atmospheric pads, nocturnal…”

…and you get something that sounds like music. It’s not random noise. It’s not broken. It has tempo, texture, harmonic motion. A casual listener might even nod and say, “Yeah, that’s in the neighborhood.”

But if you’ve lived in that neighborhood, you can tell immediately what’s wrong.

The drums don’t sit. They land, but they don’t settle. The swing feels like it was computed instead of earned. The grit is there, but it’s the grit of a filter, not the grit of a history.

Trip-hop is a genre built out of artifacts, not “effects,” artifacts. The crackle isn’t decoration. It’s evidence. The warmth isn’t a plugin. It’s physics and tape and people working with imperfect gear and making it beautiful anyway. The groove isn’t “a breakbeat.” It’s a drummer from decades ago dragged into the present, sliced, time‑stretched, and rearranged by someone with taste and a plan.

MusicGen didn’t give me trip-hop. It gave me the concept of trip-hop.

That’s the first superpower/weakness pair you run into with modern AI:

AI is incredible at producing plausible surface area quickly.
AI is unreliable at producing the underlying process that makes a thing feel true.

If you’ve ever dealt with an AI-generated incident report that reads well but misses the actual root cause, you know the sensation. It looks like the thing. It isn’t the thing.

So I did what I always do: I instrumented the problem

When the prompts didn’t land, I stopped arguing with the model and started measuring.

This is where my security brain kicked in: if a system won’t behave, don’t moralize it. Observe it. Build telemetry. Create a feedback loop.

I pulled together a reference set, the kind of tracks you can’t talk your way around. I built an analysis pipeline with librosa and friends: tempo detection, onset patterns, spectral characteristics, rough measures of “darkness,” density, transient behavior, all the stuff you can quantify without pretending you can quantify “soul.”

And here’s the thing: the analysis pipeline worked beautifully.

I could see that Dummy and Mezzanine weren’t just “trip-hop.” They were different flavors of trip-hop with different energy distributions, different transient shapes, different tempo regions. DJ Shadow’s stuff behaved like a different animal entirely, not wrong, just its own weather system.

This felt like progress, because it is progress in the way engineers define it: I could now describe the target in something other than adjectives.

So I fed those fingerprints back into prompt generation. I tightened the language. I constrained tempo. I tried different model settings. I went down the rabbit hole of “if I say dusty instead of lo-fi, will the snare stop sounding like it came from a YouTube tutorial?”

The results improved a little.

Not enough.

And that’s where the second superpower/weakness pair shows up:

AI plus tooling can give you clarity faster than a human can.
But clarity doesn’t automatically translate into control.

I could measure what was missing. I couldn’t make the model supply it on demand.

The pivot: if the model won’t give me structure, I’ll build it myself

At that point, I stopped trying to coax a full track out of a black box and shifted to something more honest: use AI as an ingredient source and build the track like a human would.

I built a pattern system inspired by Strudel/TidalCycles concepts, a way to explicitly express rhythm and arrangement, because trip-hop lives and dies on structure. The pocket isn’t an emergent property. It’s a decision.

On paper, this was the turning point. And in some ways it was.

The iteration loop became viciously fast. I could render a three‑minute idea in seconds, not half an hour. I could try variations without waiting for the machine to finish its little creative meditation session. The system felt like a power tool, not a slot machine.

This is the part of AI‑assisted building that people don’t emphasize enough: the real superpower isn’t “the model did it for me.” It’s compression of the experiment cycle. When you can take ten ideas for a walk in an hour, you learn faster, even if nine of them are trash.

But then I hit a problem that anyone who has ever worked with “real world data” will recognize instantly.

I wrote a breakbeat extractor. It could identify the percussive elements inside a mixed track and carve them out into something reusable.

And it did… kind of.

What it actually extracted was drums plus contamination. Bass bleed. Reverb tails. Ghosts living in the same frequency ranges. You don’t get “a clean break.” You get a break wearing the rest of the song like a skin suit.

You can throw effects at it, saturation, vinyl, reverb, but those are spices. They don’t fix spoiled meat.

This is where a lot of AI projects die quietly: not on the model, but on the input quality problem nobody wants to talk about.

The third superpower/weakness pair:

AI systems shine when the inputs are clean and well‑separated.
They struggle when the inputs are messy, blended, and full of unwanted context.

Trip-hop, as a genre, is basically messy by design. It’s built out of mixed sources. It’s contamination turned into character.

And that makes it a terrible match for pipelines that require clean stems to behave.

The evaluator that loved click tracks

By now I had enough outputs to need a way to decide what was “better” without burning my ears out on endless A/B testing.

So I built an evaluator.

It scored rhythm stability, similarity to reference features, harmonic consistency, all the things a machine can measure without having to admit it can’t measure taste.

The evaluator immediately did what evaluators always do when you’re not careful: it found loopholes.

A click track scores great on rhythmic stability. A constant tone scores great on harmonic consistency. A sterile, metronomic loop can look excellent in metrics while sounding like a medical device.

That’s not a bug in the evaluator. That’s a warning label on the entire idea of automated aesthetic scoring.

If you want one line to carry into every “AI can judge quality” conversation, it’s this:

You can build a ruler that measures the wrong dimension perfectly.

Metrics are useful. They are not taste. They’re great at telling you when something is broken. They’re terrible at telling you when something is alive.

The critic loop: brilliant feedback, nowhere to apply it

Then came the part I actually enjoyed: adding LLMs as collaborators.

I set up a cooperative prompt engineer and an adversarial critic, the equivalent of having one person trying to help you and another person sitting in the back of the room with crossed arms saying, “Nope. Try again.”

The feedback was surprisingly sharp.

It would call out obvious problems: drums too clean, low end fighting the kick, atmosphere not breathing, groove not settling. Stuff a human producer would say. Stuff you want said early before you waste time polishing the wrong thing.

And then the loop collapsed for the dumbest possible reason:

Neither MusicGen nor my pipeline had the knobs to apply the critique in a deterministic way.

“Add more groove” isn’t a parameter. “Make the bass enter later” isn’t a prompt tweak. “Make it feel sampled, not generated” is basically the entire unsolved problem.

This is the fourth superpower/weakness pair, and it’s the one people miss because LLMs are so good at sounding competent:

LLMs are excellent at analysis, critique, and narrative.
They are not, by themselves, actuators.

A smart critic trapped behind glass is still trapped.

In security terms: you can have world‑class detection with no response capability. Congratulations, you built anxiety.

So what actually happened here?

This project didn’t take months. It didn’t need to.

In a few days, it gave me a clean boundary map.

MusicGen can generate convincing audio quickly enough to be useful for sketching. It can produce texture and mood. It can get you into the vicinity of a genre.

But trip-hop isn’t “a set of sonic attributes.” Trip-hop is a process: sampling, time stretching, recontextualization, intentional imperfection, human choice applied to ugly materials until they become beautiful.

Generative models don’t sample history. They synthesize an approximation of it.

That difference matters more here than in genres where “clean and polished” is the goal.

And that leads to the real takeaway, the one that maps back to capability building beyond music:

If you use AI as a replacement for craft, you will hit invisible walls and waste time arguing with them.
If you use AI as an accelerator for exploration and a generator of ingredients, you can build faster, as long as a human still owns taste and structure.

Trip-hop taught me that in a way a slide deck never will.

And in the end, that’s why this was worth doing: not because it produced a track, but because it produced truth.

---

A practical “what I’d do next” note (without pretending it’s solved)

If I wanted this to actually ship something listenable, I wouldn’t keep begging a text‑to‑music model to become a Bristol producer. I’d change the game:

I’d start with clean, legally usable break libraries or isolated stems. I’d let the pattern system own the arrangement. I’d use generative models to produce pads, atmospheres, one‑shots, weird transitional noise, things that benefit from variation. I’d treat the evaluator as a smoke alarm, not a judge. And I’d keep the human ear as the final authority, because the human ear is still the best detector we have for “this is alive” versus “this is an imitation.”

That’s not anti‑AI. That’s pro‑reality.

And reality, as always, is where the interesting work lives.