Generative AI art, courtesy of the human race
As web3 peters out and investors annihilate Zuck for his metaverse bet, tech optimists need to find a new new thing to hang their hat on. So-called “generative AI” appears to be the latest thing getting everyone hot and bothered. As Sarah Guo writes about her $100m bet on AI:
Innovation happens inexorably, but at uneven speeds. After years of nothing foundationally new happening (outside of crypto) we are at a special moment in technology history. There has been radical, accelerating progress in machine learning over the past decade, with unexpected core breakthroughs: the shift from classification to generation, and the unreasonable effectiveness of ever-larger, transformers-based general models.
Is she right? Well, we’ve seen a version of this before. Anyone who was following startups from 2016-2019 remembers the promises of an ML-powered future, and the plethora of companies started with .ai domain names.
While some great innovation came out of this work, much of the artificial brainpower produced has been somewhat competitive against human brains without overwhelmingly dominating them. I’m reminded of this whenever I see a Coco delivery robot cruising down the sidewalk in my home city of Santa Monica. They are not, in fact, autonomously driven by AI, but rather steered by a handful of young people in the Philippines. Maybe a “better” AI would provide a lower cost of delivery than Coco’s current remote human driver setup, but clearly the tech is not there for AI to be a slam-dunk improvement any time in the intermediate term. (And from what I hear, the company is doing great as currently constituted).
I was also interested to notice that in Bill Gurley’s recent interview with McKinsey, when asked about “the most promising tech trend that [he’s] looking at right now”, Bill cited not AI but hybrid workforces and the ability to “hire someone from around the globe instead of 20 miles from my office”. That sounds a lot more like the current Coco delivery business than Uber’s original plans to replace all drivers with autonomous vehicles, a plan that was scuttled almost two years ago.
But then I started thinking about some of these 2020’s era AI projects like GPT-3 and DALL-E 2. These models create new generative content (whether text, images, music, etc) based on training data scraped from vast portions of the internet. Midjourney, a tool that I’ve been enjoying very much, was trained with billions and billions of images and artwork from basically all corners of human creation. That’s why you can ask it to create something in the style of a famous artist, and it will perform.
There’s already a backlash brewing against these technologies due to fear of replacing human ingenuity with the cold heartless machinations of a machine. But using Midjourney on my own gave me the opposite feeling… it felt somehow very human.
The machine was trained with the works of humans (like all of Van Gogh’s 900+ paintings) and somehow figured out how to distill Van Gogh’s je ne sais quoi into something that can persist and develop even though the man himself is very dead. Look at Messi’s hair — this is a stylistic choice that Van Gogh would have made (maybe, or at least plausibly!) if he were alive today.
Submitting these prompts doesn’t feel (to me at least) like communicating with an inorganic cybernetic facsimile but rather like there is a miniature Van Gogh inside my browser who has an eternity of runtime available to paint new stuff for me and whoever else asks him.
Just have a look at this next image. I merely typed “Darth Vader in the style of Salvador Dali” and this is what appeared 30 seconds later:
The level of detail here (summoned in practically the snap of a fingers in human-time) is absolutely mind-boggling to me. Again — a computer didn’t create this, humanity as a whole created this! The collective human artistic sensibility was converted (via training data and various parameters) into something that can be summoned on demand to make things as eerie and wonderful as this Darth Vader sketch. I do not believe that any hypothetical hyper-powered AI that had not been trained on human art would ever come up with anything that looks like the above.
The Coco delivery robot is an example of human labor being migrated across space (from Manila to Santa Monica) in order to provide value at lower cost. Similarly, generative AI is human labor being migrated across space and time — Van Gogh and Dali painted a bunch of stuff in the 19th and 20th centuries which is now being summoned via algorithm for our benefit today.
Playing with Midjourney has made me strangely prouder to be part of the human race, and I for one hereby welcome our new AI overlords. Below are a few more creations from the past month. If you think this is cool, I strongly recommend you give it a try yourself (Midjourney access starts at $10/month, and I have no affiliation with the company).
