Chapter
Two









< (Un)P r o g r a m m i n g
Y e a r n i n g >










This was my first AI coding project, so post my call for collaboration was answered in the form of a brilliant math & coding whiz, Anushka Aggarwal, we began with what would be

weeks of back and forth developing each iteration, each time optimising to achieve our evolving purpose.



































I t e r a t i o n #1
The first draft code was programmed to generate textual interpreted responses using the short story Bliss by Katherine Mansfield as the input data. I wanted to see how similar, or different the responses would be, following a given data set, what does the model produce. The responses were varied, not repetitive, rather weirdly interpretive - conclusion - fine tuning a model generates legible but irregular responses. There is some sense to be derived from it.





I t e r a t i o n #2
The next development was to add Text to speech, a simple experiment in building the audio interaction. Especially for the phrases where words weren’t generated rather blanks, or symbols were generated, I wanted to learn how the model would convert it into speech, and how would that interaction feel for us.





I t e r a t i o n #3
The next test was using the developed dataset as test to review the generated responses. Without prompting, this would be the first raw generation of the model’s understanding of the collected yearnings. Needless to say, all of it was very fascinating, a repository of stories from AI’s perspective, very useless, and very much invoking curiosity of the audience. This version was also exhibited for my graduate project.





I t e r a t i o n #4
Albeit, a meaningful listening experience, the experience did not feel enough to develops feedback loop in terms of human-ai interaction enough for people to link it to AI, and initiate further conversation. So, the next natural step was to build a speech-to-text setup, wherein the model would take the spoken word as prompt and input data, and generate speech in response to that. I tested this model with a smaller dataset, and then with the collected dataset.




















Y’s Website Navigator