You should never work for free, except sometimes you do, and sometimes you want to. Recently someone from Digital Learning and Technology Victoria got in touch to tell me he’d read and loved Repo Virtual and had already recommended it to his friends who’ve been reading cyberpunk since the days of Neuromancer. He asked if I would like to write something for their publication, about the possible uses of AI in education. I’m not an expert at education by any means, so I approached it more from the point of view of a student (or a lifetime learner, which I hope to be). I also didn’t want to be as pessimistic as I can be (you know that, you read this newsletter, after all), so it was an interesting exercise in trying to be positive - whilst still making note of some of the ethical and political issues caught up in current mainstream AI discourse (or largely ignored by it).
I enjoyed writing it and hope they’ll end up using it, but even if they don’t, I wanted to share it with you all. It might be a little “101”, but hopefully you’ll find something interesting.
Our Centaur Futures
It seems that whenever AI is discussed we’re only ever moments away from someone making dire warnings about a Skynet future where sentient machines will kill us all. There are a number of reasons why I find these proclamations intellectually lazy (sorry Stephen Hawking, not sorry Elon Musk). For one, these warnings disregard our responsibilities as caretakers to any nascent non-biological intelligence we might one day create, but also this sort of hyperbole has no bearing on the very real issues we see with Machine Learning (ML) today.
With “AI” being such a technological buzzword, the realities of the development of neural networks and ML systems is greatly obscured. These aren’t autonomous systems; they are built by people and companies with specific beliefs and biases, who are often trying to secure funding or contracts from bigger companies or governmental organisations with their own internal biases. There are countless examples of racial bias in judiciary algorithms that have a real, and often devastating, effect on the lives of individuals, yet the PR departments of Silicon Valley continue to peddle the lie of unbiased algorithms. Even if these systems were unbiased (and remember, they’re not), I’m less interested in a future where we let ML systems run things for us, and more interested in a future where we use MLS in conjunction with human intelligence.
The codified rules, clear victory condition, and complexities of chess made it the perfect field of play for AI development for decades. Out of this struggle between human and machine chess players has come a hybrid sort of gameplay called Advanced Chess, or Centaur Chess – where a game is played between 2 teams made up of a human and a machine. Where the mythological centaur is half human and half horse, the chess centaur is half human and half machine.
On the topic of Advanced Chess, author and futurist James Bridle said:
One of the most startling findings of Advanced Chess is that while even a modest chess computer can now thrash any human player, a human and a modest computer working together can beat a much more powerful computer playing alone. There's a transformative combinatorial effect at work that magnifies the strengths of both ways of thinking (and, to my mind, emphasises their differences in interesting ways).
Neural networks have the benefit of being able to crunch huge amounts of data and be able to find interesting, and sometimes obscure, solutions to a problem that a human might simply never have arrived at. Still, a neural network can’t be said to understand a given problem, task, or game. They process the data that is given to them in the way that they’ve been trained. A centaur approach, however, gives us access to the ability to sift through masses of data to find more obscure solutions, but then use human reasoning, understanding, and even intuition to choose the best “move”. Here we see a possible blueprint for the future, where ML can help us develop truly innovative approaches to our own work in any number of fields – even creative areas.
Holly Herndon is a musician, artist, producer, et cetera with a PhD from Stanford University’s Centre for Computer Research in Music and Acoustics. Her 2019 album PROTO was created in collaboration with a neural network named Spawn, created by Herndon and her partner Mat Dryhurst, and trained with the help of an ensemble of singers – so whilst Herndon sought to make use of ML in creating the record, she also recognises the importance of human collaboration. In the press release for the album, Herndon said:
There’s a pervasive narrative of technology as dehumanizing. We stand in contrast to that. […] Choosing to work with an ensemble of humans is part of our protocol. I don’t want to live in a world in which humans are automated off stage. I want an AI to be raised to appreciate and interact with that beauty.
Herndon’s prior work was already experimental in its sound and creation, with strong theoretical and philosophical themes, so it’s not surprising that she’s at the forefront of AI collaboration in music. Herndon’s development of Spawn and the album Proto may have coincided with her PhD thesis, but I imagine a future where not only are ML systems readily available “off the shelf”, but where machine wrangling is an extremely valuable, if not necessary, skill – much in the way that digital literacy has become necessary for education, employment, and many other facets of modern life.
Imagine a research AI that you can send out to gather results for a paper you need to write. The neural network could likely gather a staggering amount of data, but by checking over the results yourself you could train it to find and highlight only the specific information you need. A ML research assistant you could train would prove far more useful than corporate search engines that try their hardest to keep you within their own closed gardens.
Brian Eno’s Oblique Strategies are a famous tool for stimulating creativity and getting past blocks, and this is another area where ML systems could excel – and indeed, GPT-3 is already showing great promise with text generation. By training a neural network on your past work, it may be able to extrapolate new paths forward, or stun you into a creative sprint with a phrase, image, or bar of music that would never have occurred to you without assistance but which sparks something new and unique.
These are just a couple of examples, but with the right tools and time for training, any number of areas and tasks could be further explored with assistance from ML systems. Over the coming years the tech giants will try and sell us ML in all sorts of packages for their own profit and benefit, but ML can and should be a DIY tool for all of us to train, tweak, and use however we need. This is a reality already for people with the technical know-how (as well as the time and patience), but I eagerly await the democratisation of machine wrangling – for the (probably open source) developers who can open the field up to your average tech user.
When all of us have built, trained and worked with our own neural network, it will get harder for the tech giants and tech-surveillance wannabe-authoritarians to lie to us about “unbiased” algorithms. We’ll have the knowledge to push back against the technocrats who think our world and our lives are reducible to data. We aren’t data, we aren’t numbers, but we might just be half-centaur. And with machine learning to augment our own intelligence, we could just find some new paths forward we never could have found on our own.