Malazan and A.I.
19 February 2023
Since the pandemic, I have spent a disproportionate share of time reading fantasy and sci-fi books. My latest addiction is books from the Malazan world. Given that I’d instead be reading history, politics or philosophy if not fiction, these books provide an enthralling experience. Reading the Malazan series feels like we are being presented with a newspaper or a journal. A few characters in the series play the role of historians and poets, and it does feel that if we are expected to relate to anyone, it is them. Often, the reader will find themself reading a philosophical monologue on just about any topic. It is always contextual, often has elements of foreshadowing or retrospective revelations about the story, and yet the diversity. The world is rich and ripe. I would find my imagination lacking before complaining about Malazan’s worldbuilding elements. There are entire histories, sociologies and philosophies contained within a fictional series. The author often drives inspiration from actual history, which is quite apparent, but it is well done enough not to matter. These books are a study of their own, and I am incredibly grateful to be introduced to these books.
Other than that, the vast volume of critique and reviews for the entire Malazan world are adequate, and there is no particular viewpoint I can contribute now. However, I found the fan art lacking in volume. It may be a function of the popularity of the series, which remains bound by the challenge it proposes to new readers. These are difficult books to read. They can very much feel like your history homework. Well, they do have dragons and typical story arc elements, but nevertheless, consuming these books is not a relaxing venture. Anyway, I complained about the fan art situation to my wife, who suggested I contribute to the fan art myself if I felt unsatisfied with it. I agreed. However, she expected me to struggle to draw something independently with my drawing skills worthy of a toddler. Clever me decided to use A.I. to do this.
I’ve been unable to mute all the chatter about DALL-E and then ChatGPT anyway. And while we are here, let me state my quick opinion on all of this A.I. brouhaha. I do believe that A.I. has great potential to assist us. The way to think about these tools is to imagine how you think about garments and tailors. Machine looms transformed the textile industry. There is no longer the need for a human being to do the weaving, but we still need tailors. A.I. tools do an excellent job of solving generalisations of human problems. Still, we have the unique advantage of coming up with wants and desires that the machines cannot predict. And thus, they could be better at the margin, but they are great assistants.
Moreover, they make things more accessible, which is always good. Nespresso made espresso accessible, and fast fashion gives us the ability to look different every day. Sustainability aside, and the fact that Nespresso coffee tastes like muck aside, more accessibility of goods and services that tend to our desires are what we aspire to as a species. I am celebrating. I am an active user of Github Copilot at work. As you are reading this blog, you should know that I am a heavy user of Grammarly, and the premium version, which tries to steal my style and thunder and recommend bland phrases. You suck, Grammarly! Will you ask me to change this sentence? (It did. But I am the human, and I will have the last word over A.I., as of writing)
Coming back, I was excited about the possibility of generating fan art using A.I. I tried DALL-E and Midjourney and ended up purchasing a subscription for Midjourney. DALL-E feels extremely amateur in comparison. Midjourney, or MJ or Midge, completely checks all boxes on the joy one may get from a new programming language. The possibilities of what you can do feel limitless, although diminishing with complexity. I will arrive at what I mean by that statement. But ultimately, the verdict is firmly in favour of the tool. I will keep coming back and using this for various use cases; maybe I need a new cover image for my blog, or a new wallpaper or a fancy sticker pack. However, there are a few observations I will discuss here.
First, there are correlations that the A.I. has learnt, which you will have a hard time unravelling. The knowledge that two elements are uncorrelated or unrelated by a causal pathway is as important as knowing they are related. Few examples that I struggled with:
- ‘long grey hair’ would force the A.I. to generate an old person.
- ‘bear with a phone fighting a human being’ would always give the phone to the human. I also tried blending images of a bear with a phone and a man, but the man always snatched the phone from the bear. borzhe moi
- ‘man with tusks and green skin’, and this is the one where I spent hours and had to admit defeat. Midge correlates tusks with horns and pointy ears, and she correlates green skin with orcs.
Second, there appears to be a Fourier issue. It seems to learn ‘patterns’ which are co-incidental. For example, it may display any number of fingers on a human hand or strings on a musical instrument. If you give it a city skyline, certain features or sets of buildings are repeated. The machine has either retained lower dimensional information or is generating lower dimensional images for a faster and cheaper product. A popular thing on MJ is ‘filigree’, which is an intricate pattern for humans but would not be for a machine in Fourier space.
Finally, as a product, MJ is nothing short of brilliant. You can provide image prompts, and it does a decent job of mixing those. You can also provide the same image prompts with minor modifications to increase their weightage in the overall narrative. In general, I spent time in their community rooms to learn from other folks, and if I had spent more time, I could learn more tricks. I will likely do so. But this is what I meant by endless potential diminishing with complexity. The effort for complex compositions scales up. So, although the possibilities are limitless, my time is not. There are also certain things, such as not generating legible text, which are more design choices than the capabilities of the A.I. What that means is you cannot ask for a t-shirt that says a customised slogan.
Thank you for reading, let us look at the gallery now. For detailed prompts and descriptions, please visit the gallery on Midjourney.
Kruppe of Darujhistan
Gothos, a Jaghut
Anomander Rake, Son of Darkness, a Tiste Andii
Fiddler, Bridgeburner
Samar Dev, witch