Getting To The Point

And so we come to the part of the story I find the most interesting: turning dreams into reality. Well, daydreams into some kind of facsimile of reality, anyway. The whole question of just exactly how "real" or "authentic" anything created with the assistance of AI can or should be considered is about as contentious as it gets right now, although I strongly suspect that, in a decade or so, most people won't even be able to remember what the fuss was about. 

By then, it'll either be as dead an issue as synthesizers, sampling or Autotune, basics of modern music-making that were all considered utterly fake, spurious and beyond the pale in their day, or it'll be looked back on as some fad that's remembered only in a "Weren't our parents weird?" kind of way, like playing outdoors or VR.

My money's very much on acceptance and normalization. It's just too useful, even in its current, hallucinatory form, but more importantly a whole generation is about to grow up using it and they're not going to have any of the emotional, aesthetic or ethical baggage that makes older people feel so uncomfortable. Since most good art and almost all innovation comes from the young, you can see where that's leading.

"After a time, though, more and more people
began to become uncomfortable with the process"

It's been nearly forty years since I last considered myself "young" so, not surprisingly, my relationship with AI so far has been ambivalent. When it began to push its way into my consciousness I found it sweet, funny and endearing, like a dog doing tricks. As the old saw goes, it's not that it does it well, it's that it does it at all.

Like many in the blogging community, I spent some time playing with the new toys and writing up my experiences. There were posts all over this part of the blogosphere as people experimented with the capabilities of various AI services. Bloggers got AIs to write stories, draw pictures, play music and report on the news. They asked questions and laughed at the silly answers. Bloggers got AIs to describe other bloggers or picture them as characters in a game. It was all innocent fun. For a while.

After a time, though, more and more people began to become uncomfortable with the process, the methodology, the provenance and the ecological impact. At the same time, the AIs started to get better at what they did, so the results weren't as funny any more. Using AI, even in the one area where it was already undeniably useful, providing spot illustrations to break up text, became something most bloggers no longer seemed comfortable doing. 

Things that are useful rarely go away just because they're not to everyone's taste, though. There are more than a few folks around here still using AI-generated images. What's different is that now they generally apologize, even as they're doing it. My own footnotes detailing any use of AI in my posts is partly a sop to that sentiment, although mostly it's just me trying to keep some kind of record for my own purposes.

Other than that one usage, though, I have found AI in general to be less of a boon than I thought it might be. As the Russo Brothers said recently, in relation to their use of AI on their latest movie, The Electric State,  "AI is in its generative state now, where it has, as we call them, hallucinations. You can’t do mission-critical work with something that hallucinates." In other words, it's flaky and unreliable still and always needs a human to check its work.

"The technology continues to improve"

You can see that in Google, sometimes, whenever you use search these days, although I have yet to figure out how or why it happens on some searches but not on others. I thought for a while the phenomenon only occurred if I phrased my query as a question, something that does seem to stir Gemini from its slumbers, but sometimes doing it that way just gives an old-school string of links. Other times, even a simple two-word query will bring up an "AI Overview" at the top of the page.

This is Gemini's attempt to act as the unpaid personal assistant I always wanted it to be. Sadly, it's not very good at it yet. Yesterday, for example, when I was answering a question by an anonymous commenter who wondered if there'd ever been any major MMORPGs that came out of a Kickstarter campaign, I asked Google and the AI Overview told me that yes, there'd been several, chief among them Camelot Unchained and Ashes of Creation. I rest my case.

The technology continues to improve but for most things I'd want to do with it it currently requires far
more oversight and effort than makes it worthwhile. When it comes to using AI for video, for example, as I've said many times, when I can type in a two-line prompt and get a two-hour movie that's worth watching, get back to me. Right now, even the best AI video generator couldn't come up with something usable for a three-minute music video on its own. 

Hah! Three minutes? I don't think there's one that could give you thirty useable seconds.

The same can't be said of the three-minute song itself. AIs can pump out extremely convincing facsimiles of songs in pretty much any style you can think of and they're doing it by the million every day. Maybe the tens of millions. There's even an app for your phone that will make all the tracks you need for an album in less time than it takes you to ride the bus to work.

"this brave, new, artificial world"

When I first found out about those services I was extraordinarily impressed. At the time the generations were very short so a two-minute song was about the limit but that didn't seem like much of a problem. Some of my favorite genres lend themselves to very short songs. 

Just from a few prompts, I was able to "create" songs that, I felt, could easily be slipped into a compilation of tweepop with no-one any the wiser. My experiments back then suggested the AIs were pretty good at the music but their lyrics left a lot to be desired. Like really a lot. Which was when I first had the idea of having them set my own lyrics to music. 

And it worked. Kind of. But it was weird.

Hearing words you've had in your head forever, now being sung to a different melody, is unsettling. It's never going to have a truly happy outcome; either it sounds worse, in which case it's frustrating, or it sounds better, in which case it's upsetting. You want the AIs to do a good job but not do such a good job it makes you feel like you might as well not be there. 

I'll break chronology here to make a point that may or may not be re-assuring, depending on your particular attitude to this brave, new, artificial world. Having spent many hours now, working with the AIs on bringing my own words and music to life, I can categorically say that I feel an existential difference between their interpretation of my work and their own, wholly artificial, creations. Even when they're singing my words.

"I wish I'd thought of doing it that way"

I've listened to many attempts by Suno to make songs from just my lyrics on their own but also to many, many more based on my melodies, arrangements, phrasing and rhythm. Using just the words alone, in not one case did I find Suno's versions superior. Sometimes they were interesting; sometimes I liked them. Not once, though, did I think "I wish I'd thought of doing it that way".

What's more, I can now clearly hear the artificiality in Suno's composition and execution in a way I couldn't before. With different levels of "authenticity" to compare, it's a sliding scale. The more you put in, the more you get out. And vice versa.

That said, everything the AIs produce has a kind of brittleness to it. It's hard to describe but once you hear it you can't miss it. It's very apparent in material produced wholly by the software from just a prompt. If you let the AI handle lyrics, melody, arrangement, instrumentation and all you get something that feels thin, somehow. Incomplete. Inauthentic. It does depend on genre, certain styles lending theselves very much more to automation than others, but it's there in everything. 

Then again, this is precisely the same, subjective, argument countless people have put up against just about everything you can think of, new or old, synthetic or organic. A certain type of person always claims to be able to sense authenticity in some arcane way that, quite conveniently, can never be objectively authenticated. Or, in other words, I might just be missing something that was never there to begin with.

Crucially, though, to my ear, the more organic input the AI has to work with, the less obvious its ghostly hand becomes. Ironiocally, perhaps, I put that down more to human idiosyncracy and incompetence rather than any flaw in the machine.

"I am a poor singer and a worse guitar-player"

I am a poor singer and a worse guitar-player and the recordings I'm feeding into the AI maw are of

terrifyingly bad quality. Add to that my strong preference for never using three chords when two will do, rarely sticking to any kind of coherent song structure and generally employing the kind of rhythm that sounds like the thrumming of a somewhat faulty diesel generator, and it's hardly surprisng the results, even mediated by technology, tend to sound unlike anything any respectable algorithm would want to put its name to. 

It seems quite unlikely anyone would want to program a machine to make songs quite as wonky as these. Although, as I may elaborate upon in another post, they will, if you ask them nicely.

All of which tends to re-inforce my feeling that AI is far more likely to become a useful tool in the hands of creative artists than it is to replace them. Maybe one day people won't be needed at all and certainly some genres are more at risk than others but in the end that ineffable quality, authenticity, so long debated, derided and dismissed, will have its way.

And that's a very long introduction to the next part of this series, in which I might just, finally, get around to talking about specifics instead of generalities. Or maybe not. It turns out I have one hell of a lot to say on all of it. More than I even knew. It's just like when I start playing a new MMORPG and find I want to post about every last system and mechanic, taking them all apart to see how they work.

I think it's called an obsession. It'll pass. Only not just yet.

 

Notes on AI used in this post:

It seemed entirely appropriate to use AI illustrations ina post about AI so I did. No apologies there. I also spent absolutely no time fiddling about, trying to get something clever or apt. I wanted to see just how practical it would be to have the AI do nearly all the work. As it turns out, pretty practical.

Every image was produced at NightCafe using the current default model, Flux Schnell. Flux Schnell is extremely cheap  at just 0.75 credits per image. I have more than 2300 credits in my bank, none of which I paid for.

I didn't change any settings other than the ratio, which is 4:3 landscape, even though it doesn't always look like it. The generation time was Medium. 

For the prompts I simply cut and pasted a sentence or a phrase from the text, all but one of which I have handily included as captions. The image at the top was generated from the phrase "turning dreams into reality". 

Abstract to concrete is no problem for the AIs but I was very surprised to see the result of that last one. It's both bizarrely appropriate to the theme of the post and also very much in the style of image I generally go for if I take the trouble to specify genre and technique. It makes me wonder if Nightcafe feeds the AI information about previous images generated by the specific account. If not it's a hell of a co-incidence. I mean "turning dreams into reality".  could be just about anything...