AI Training

I thought I should revisit the world of AI to see what the current state of the art is. I've heard it's been quite busy churning out terrible fiction to flood Amazon with and make agents nail their inboxes shut. There is also now much debate about the moral dimension of large language models given their tendency towards plagiarism due to their training data. I don't want to be accused of ripping anyone off, even if I did borrow that biscuit story from Douglas Adams or Jeffrey Archer, depending on who you believe.

Knowing a thing or two about programming, I decided the ethical way forward was to train a model of my own using my own writing. This went well enough at first, producing a series of pastiches of some of my short stories. They didn't make much sense though. Maybe the stories they were trained on didn't either, I wrote a lot of them a long time ago. One of them was about an electric pig trying to make it as a stand-up comedian. Fanciful stuff but ultimately meaningless.

I did some more reading around the subject and refined the language model with the hope that it would produce something more inspired than this Burroughs cut-up stuff. I must have screwed something up because it started churning out stories where I was the protagonist. Also, unlike its previous efforts, these were quite mundane. I skimmed through them. There was one where I forgot to set my alarm and was late for work. In another I bought a sandwich from Tesco and a pigeon shat on my head. It was late, and I had work in the morning, so I called it a night. A superstitious doubt made me double check I'd set my alarm.

The next morning, despite taking such care over the alarm, I overslept and was late for work. I spent the morning trying to ignore a growing dread and walked past Tesco to get a sandwich from Pret. The queue in Pret was enormous. There was apparently something wrong with the card readers and the staff were struggling for change. Deciding I was being stupid and not wanting to spend my entire lunchtime shuffling through a shop, I abandoned the queue and went to Tesco. I bought an egg and bacon sandwich. I eyed the sky suspiciously as I left. With a wry smile at my silliness I strode confidently up to my office building, whereupon a pigeon shat on my head.

That evening I read the language model's output more carefully. It hadn't said I'd forgotten to set the alarm, it said I had accidentally turned it off while double-checking it. I read the pigeon one again too and it included a brief aside inside a branch of Pret. Nervously I read on and found another that accurately described my commute home, including the tube station gate that makes a noise like Chewbacca when it opens.

This couldn't be right. Was I perhaps still dreaming, even if I don't usually recall my dreams in much detail? Or was the large language model imagining all of this, including what I'm now typing, on my behalf? It was an existential conundrum. With some trepidation, I asked the model for more stories.

This time I read them very carefully. They still featured me but this time they were a tad more exotic. In one, my regular morning commute will be interrupted by an escaped rhinoceros. Later that day, I will apparently go to the moon in a balloon shaped like a rabbit. With what is surely the most exciting day of my life ahead, I closed down the laptop and had an early night. Maybe AI isn't so terrible after all.