All That Is Data Melts into Air

In the opening of “The Three-Body Problem”, the sending of a message to outer space sent mankind down a path of cascading events involving missing scientists, social strife and political intrigue as human lives and dignity hang on the balance in our encounter with an alien species. Might A.I. do the same is a question that is not entirely new to us, maybe because we see how we treat the weak and the vulnerable and it is not pretty. I am inclined to treat consciousness and a corresponding sense of self as separate to creativity and intelligence so hopefully that buys us time while we try to better understand how life arose from non-life.

Generated with Stable Diffusion WebUI.

Having played with Vision Transformer and Stable Diffusion recently for generating songs and images in wildly varying styles, I am very pleased with how I can now “draw” faster and better than I’ll ever be capable of within my lifetime, yet also struck by a sense of disquietedness and anxiety over how much of an empirical exercise this all has become. Something I’ve been trying to wrestle with is this idea of: so much of what we humans pride ourselves on - our creativity, our intelligence - is just computational data and statistical phenomena perhaps? And we are no competition to the machine in that regard. In one sense this is completely true and liberating, yet along other dimensions many consequential details are lost, and it is unnervingly terrifying and anxiety-inducing to think of much of the human exercise this way.

Generated with Stable Diffusion WebUI.

With Stable Diffusion, I can freely switch styles and generate 4 variations within half a minute on a 3070 GPU despite having zero interest in and sucking at drawing entirely, with some outputs as shown. Clearly, A.I. is weird in that it captures a very particular facet of our reality only – I find myself using “large breasts” and related terms in my negative prompts consistently because the model keeps outputting ridiculously-shaped female figures otherwise. This whole thing is very weird as discussed in this podcast and this powerful question asked in it is this same thing that I’ve been trying to munch my mind on:

EZRA KLEIN: So to be McLuhanite for a minute and to take his famous saying, “The medium is the message,” if the medium is the message, if the medium encodes certain ways of being and thinking that change the people who use it, what do you think the message of the A.I. chatbot medium is? Which it’s worth noting is a medium being built on top of a technology. Chatbotting is just one of many, many applications. And the fact that that’s the one taking off is going to also shape the technology differently than it might otherwise shape. But yeah, what’s the message of the medium?

Generated with Stable Diffusion WebUI.

What does it mean for us as individuals and how ought we to respond when modelling and building prediction machines out of human creativity and intelligence has proven to be so good? My preliminary response as I lean in to this idea is that a loss of mystery and a sense of wonder feels like both sides of the same coin: by making little of our intelligence and creativity this way, the possible positive outcome could be a real revaluation of what it means to be human and what we owe to each other – regardless of every individual’s level of intelligence and/or creativity; but on a more sinister note, I do worry about A.I. creating even deeper rifts in global societies than the effects of trade and social media combined if we do not approach it thoughtfully.

Dataset Biases and Prompting Effectively

The general use case I’m trying to productionise is automated content generation - to have A.I. produce video scripts and images on desired content topics in various formats and languages that I can publish onto multiple platforms automatically. This post is still manual work unfortunately. Over the course of doing so, it’s interesting to see all the biases of A.I. thrown back at me - with every female figure generated voluptuous by default until I specified for it to not give return large breasts. It makes me think about the datasets used for training, and how much of what we see in media is airbrushed versions of reality so who’s to blame really, actually?

After experimenting with multiple models that I could just download off the internet, I decided to go with a graphic novel-ish style. Unsurprisingly, there are plenty of models specialised in adult content production. Couple that with how easy it is to customise and train your own Stable Diffusion model through textual inversion… In other words, anyone with an average gaming PC can generate deepfake adult content.

I hope governments, schools and parents are on top of A.I. Can I have that hope please?

Why A.I. is Different

Leaving negativity aside for a sec, I’m fascinated by A.I. because it has come a long way in a short span of time from that Nativity according to the flesh moment at the 1956 Dartmouth Conference, from those first single-layer neural networks - perceptrons to modern day deep learning; and one needs to only look around us at the present to see what A.I. can already do to make forward projections grounded in existing evidence. Many other technologies are what they call possible visions of the future trying to find traction in the present (and failing miserably so), but A.I. is a unique breed in this regard. This is why I’ve come to see equivalence with the steam engine, and probably why some even compares it to fire or electricity.

“I’ve always thought of AI as the most profound technology humanity is working on—more profound than fire or electricity or anything that we’ve done in the past.” - Sundar Pichai, CEO of Google

Mark I Perceptron displayed at the Smithsonian museum. (Source: https://ronkowitz.blogspot.com/2017/11/perceptron.html)

In 1958, a single-layer perceptron involved actual wires and a huge machine; last I played with some convolutional neural network models to do multi-label classification in 2022, some models like ResNet-512 are 512 layers deep chugging along on my very average gaming PC in the living room.

A.I. is a story of many winters that never chilled man’s interest in giving machines an ability to think. The Long Night of A.I. winters, its statistical turn, the Cambrian explosion made possible by the confluence of data and GPUs on old models, and that ultimate comeback of neural networks are all stuff that I hope to dig into next!

Originally published on PubPub at erniesg.pubpub.org/pub/33ietk6v.