My personal experience that the task of simulating vision and speech can reveal hidden things about human function inspired the notion that the computer can function as a sort of philosophical prosthesis.
For the past seventy odd years computers have been teaching us much about what it means to be human, often through their failure to be able to do many of the tasks we take for granted as humans. In recent years though, there has been a surge in the capacity of computers to imitate human capacities in ways that we thought they would never be capable of. Technologies to understand speech, to translate between languages, to read emotion on human faces, to mimic your voice, or mannerisms, or even your handwriting.
It is shocking how quickly society transitions its response to new technologies from the marvellous and dystopian to the quotidian.
My prosthesis is a machine that I developed over a number of years through a process of tinkering and reflection, which has been exhibited in a number of different venues and forms:
In the summer of 2017 I came across an academic "paper with code", by a leading machine learning expert Alex Graves, entitled "Generating Sequences with Neural Networks". Research papers in computer science with code are a fantastic resource for finding technological advances in computer science that may never find the mainstream. Grave's paper reduces handwriting to a sequence of numbers related to the abstract symbols writing represents. Once trained on these sequences a neural net can be used to predict a new sequence of numbers given a new sequence of symbols.
Based on previous tinkering with drawing machines and 3D printing I realised that I could construct a machine to draw these sequences and recreate the handwriting. I found an efficient implementation of Grave's paper on Github, and using some leftover parts and a borrowed Makerbot 3D printer built a small machine to enact the generated handwriting. Its an anachronistic and nostalgic machine, with a touch if steam-punk in form a combination of mechanical typewriter, teletype and the old Gestetner duplicating machines from my youth. Its built from a combination of scavenged parts from an old Textronix solid wax digital printer and a large format inkjet printer from the 90's, some 3D printed parts and "makerware", electronics, aluminium extrusions and connectors used to build DIY CNC machines and 3D printers.
Before the laser or inkjet printer, computers used fanfold paper for printouts, fanfold paper has "sprocket" holes that allow paper to be "tractor fed" through the printer. I envisaged something like an oracle crossed with a teletype machine, the machine would produce a never ending stream of writing which would flow out onto the floor below the machine.
To generate the text it would write I used another type of neural network, again a "paper with code", Char-RNN (there are several other approaches, such as statistical methods called Markov Chains). This neural network is trained on a 'corpus' of texts from which it can then generate novel versions. Give it enough text for training and it can produce novel sentences which sometimes make sense. In all the generative systems I work with I search for the place where unexpected conjunctions can arise, in this case if you like where there is a possibility that a brief glimpse of something profound could emerge from the nonsense. In the time since I started experimenting with Char-RNN, AI text generation systems have become mainstream, notably GPT-2 and GPT-3 from DeepMind and popularised through sites such as "Talk to Transformer".
In the most resolved form of the work as exhibited in the touring exhibition Systematic, I envisaged the machine as an oracle, a tongue in cheek prophet for our age. Observing how stuck we have become in our political and social discourse, in the failure for new political philosophies to emerge to address the challenges of the contemporary world, could a philosophical "cut up" machine generate a new approach. To build a corpus on which to train my network I borrowed a range of classical philosophical texts from Marcus Aurelius to Sun Tzu and threw in some of the major 19th and 20th century political manifestos and added in some 20th century dystopian fiction. One interesting aspect of neural networks is that they are not deterministic, every time you train one, it has different characteristics. Some versions I had some problematic characteristics such as using what could be interpreted as hate speech, and so needed some careful vetting to make sure they would be acceptable.
On a final technical note, I managed to cram both trained neural networks and control software into a hacked Nvidia Shield TV, which is a low cost Smart TV device the size of a large phone.
