Aren't all stories just dreams? They are, perhaps, if one identifies with the protagonist; or if one is the protagonist. I write a diary, and sometimes I deviate from the truth; does the diary at some point then become a piece of fiction? One misremembers things all the time, or simplifies and in the process loses one's grasp on the essence of things; one is, perhaps, anyway writing fiction all the time in our stream of consciousness, as our senses and memories and language are all limited.
I thought today: what would happen if I trained a text generation model (like GPT-2, or better) on my own writing? Probably nothing of interest; the model would just generate overfitted, seldomly coherent, variations on my usual musings. Some hilarity would ensue. Then I would drop it and get something else to play.
In a different world, this approach proves intriguing. Because I am trying to reproduce my own train of thought (what is text, if not an idealized version of our thoughts?), every utterance would either resonate with me or fail to with some kind of intangible certainty. I could then perhaps feel compelled to improve that model — make it as close to myself as possible, or rather make its output be something that I would consider calling mine. At every step I would have a guiding light: did it say something I would say? If not, what's missing from the corpus that could reduce the chances it says something like this in the future? If this approach works at all, once in a while, it might keep my interest. I could dedicate years of on-and-off work to this process, thanks to that.
Because I happen to have both programming and writing skills, however limited, this improvement process could be twofold: some of it would be trying new models, ensemble approaches, hybrid (code/ML) systems. Some of it would be writing more so the corpus of my thoughts grows, and the model gets to know me better. I would think to myself: if I write truthfully, without holding anything back, perhaps I'll teach the model to be more like me in new ways.
At some point I'll die — and I'll leave behind a manifestation of myself, albeit one very imperfect. Likely always bordering on the incoherence. In many ways it'd be like reading my writing, but also more than that: once in a while a different and new echo of myself could be found in it — if anybody bothers to spin it up again and interact with it. They could ask it about my life and thoughts. About my favourite things.
I posit this model, or perhaps its performance, could be considered a work of art: it'd be a manifestation of myself (however imperfect) that would lack utility in the productive sense. With any luck, it'd also achieve some of the effects that art has on people. Some people would probably criticize it as overly egotistical, as navel-gazing; some would find it dull. But it could make others wonder, make them feel or think something new. Could it, some day, even be an inspiration to someone? Someone from the future with better tools, or a better education, or a stroke of luck?
Even if my approach is doomed to fail, what if someone else were to do it again, better? Most things don't work out well the first time, but before most great works of art there lie (across years, centuries) many abject failures. But artists improve on previous artists, all the way back to just common people that tried something new with blunt tools and few skills.
Eventually someone may perfect this as an art form. Perhaps train it to model others too — it would make sense to start with our loved ones. Could these models ever be our equals? Could they be our betters?
Forever is long.
- public document at doc.anagora.org/dreams
- video call at meet.jit.si/dreams