I, For One, Look Forward To The Coming AI Winter

This article just came up. The headline is a good summary: 95% of Companies See ‘Zero Return’ on $30 Billion Generative AI Spend.

I, for one, look forward to the coming AI winter.

Most people would agree that I was early on Generative AI – my zine The Casino Of Beauty came out early 2023 followed by a manga in “The Moons of Jupiter” universe. I also did a conceptual radio broadcast. I had a lot of fun exploring this particular idea with Generative AI tools and hope to get back to it soon.

But what I found still remains: for generating an enormous amount of images with the same “vibe”, AI can’t be beat: my projects 1000 Concept Cars and my Peacocks are great examples of the possibilities. But after the first issue of this manga I wanted to generate a real story, with consistent characters doing real things. I understand there have been improvements, but the results with the new models always look “stilted” and lacked any real substance or style. The expressiveness of the algorithms is destroyed by imposing too much meaning on the images. (This sounds like someone’s MFA thesis, and if you want the idea go ahead: I skipped art school for a reason.)

I generated several other zines using the then brand new Stable Diffusion with my own custom interfaces (n.b there are Not safe for work images on that page). I generated illustrations for every stanza of the Poetic Eddas. I shipped a concept album called Cowboy Beowulf that I am also quite pleased with, as well as a few videos of my version of Paradise Lost that I abandoned for lack of proper API access to a generative music service. I have attempted to build my own version of a generative music system that has some results, but isn’t quite there yet for what I want to do.

I’ve used Generative AI a lot over the past few years, and am quite excited about the possibilities. But true creativity and innovation don’t come from venture capital hype with unicorn valuations. It comes from hard work and wild eyed purpose of unreasonable visionaries. I think the next great science fiction film franchise is going to be made by some broke kids in Ohio or something, cobbling together their gear from auction sites and Freecycle with the help of open source models. When generative AI is accessible to the kids in the favelas and refugee camps of the world we’ll see something amazing, but honestly other than my own stuff I don’t care about Generative AI.

It reminds me of my studio musician days. I had just finished a long dub-reggae jam with a reasonably well known drummer during a soundcheck, and, listening back to the recording suggested that we should turn this into a real band. He laughed and said: “Reggae is like baseball. I like playing baseball, I don’t like watching it.”

The idea of having to listen to hours of Reggae jams to generate an album didn’t appeal to him in any way.

Using Generative AI is a blast, but in the end the output is probably more interesting to you than others.

I think this personal experience with LLM’s and AI art is what is really interesting, and think that the art of training these models needs better interfaces before this particular aspect of it can truly develop.

Regardless, some exciting new tools were developed this last hype cycle¹, but generative AI is neither deterministic nor reasonable enough to build a business on. For the most part it has generated an enormous amount of noise and disinformation. For what it’s worth: the explosion of ideas and hype does remind me of the invention of the printing press – it’s just important to remember that Gutenberg died bankrupt² and there were lots of civil wars in the wake of his invention.

As things shake out you will still need to think logically and carefully and passionately to ship software, MBAs never read memos anyway so won’t be able to tell if something is generated, and our educational system is made up of boring task masters and capitalist ideologues that are properly defeated by ChatGPT. You should have never trusted videos or images on the internet anyway, and now we have clear illustrations as to why not to.

I think the truly thrilling technologies that have come to fruition over the past few years are subtle and not very new: Vector Word Embeddings³ became far easier to implement during the generative wave, and the related technology Retrieval Augmented Generation⁴ (though some hypesters still insist it is dead) is a great way for humans and machines to parse through large documents together.

I look forward to the development of open-source Local LLMs that will let me develop software offline without having to deal with ever changing fees and features from services like Cursor and Google AI.

I also look forward to getting to see that genius who figures out how to synthesize the ideas of generative ID into truly expressive art. I think there is something to having a machine that can generate 1000 images and novels worth of text on a given subject, but we haven’t quite solved the conceptual and emotional blocks to turn it into something great.

There would be no Dürer without Gutenberg, Now we need to find that Dürer.