I really like the letter format. It shows character and urgency in the writer through the actual written words rather than the narrative descriptors we're used to e.g. "she pleaded". That's quite a challenge to write, but when you pull it off, it creates this strong subtext. I think for longer works the format could get tiresome, but in this flash fiction it works a treat. Loved it!
Epistolary novels can be amazing, especially if you use a patchwork of different kinds of documents to indirectly build up the unfolding story. Like rifling through someone's personal belongings to piece together their life.
I liked it! The construction of this as a letter was fun, and I'd definitely enjoy more stories around the same idea.
I'm also fascinated by the black box nature of LLMs. Although neural networks are used all over the place, it's so interesting to me that these things we've built to approximate human response, we can't actually tell how complex or human they are under the surface. Not that I think any of the LLM are conscious, but the fact that we lack the ability to even tell what's happening inside them makes the future of more complex systems very morally fraught.
That is my prediction with "AGI"- we will build it by trial and error then have no idea how it works, and no sense of how much we can rely on it for important tasks (just like real people). I think there is a paradox around intelligence and control. You want a machine that is smart enough to make independent complex decisions, but one that is unwaveringly reliable and obedient. But the ability to handle unfamiliar situations is fundamentally incompatible with obedience, since you cannot specify in advance what you want the machine to do.
I should note this short story was inspired by research into the evolution of analog circuits. The group used random analog circuit designs and assessed their ability to perform a simple function. The best (barely) functioning circuit designs were hybridised and reselected over multiple generations. The final designs were extremely competent at the task, but the engineers had no idea how they worked. They included components that were not connected to the main circuit, but when they were removed the design stopped working (so they were tapping into some weird electromagnetic resonance effects that nobody had considered as potentially functional).
I really like the letter format. It shows character and urgency in the writer through the actual written words rather than the narrative descriptors we're used to e.g. "she pleaded". That's quite a challenge to write, but when you pull it off, it creates this strong subtext. I think for longer works the format could get tiresome, but in this flash fiction it works a treat. Loved it!
Epistolary novels can be amazing, especially if you use a patchwork of different kinds of documents to indirectly build up the unfolding story. Like rifling through someone's personal belongings to piece together their life.
I felt a sense of unease and maybe even horror or dread as I read it?
Lovely to hear back that it is possible to induce a strong feeling in a short piece of writing. Makes me want to create more like this.
I liked it! The construction of this as a letter was fun, and I'd definitely enjoy more stories around the same idea.
I'm also fascinated by the black box nature of LLMs. Although neural networks are used all over the place, it's so interesting to me that these things we've built to approximate human response, we can't actually tell how complex or human they are under the surface. Not that I think any of the LLM are conscious, but the fact that we lack the ability to even tell what's happening inside them makes the future of more complex systems very morally fraught.
That is my prediction with "AGI"- we will build it by trial and error then have no idea how it works, and no sense of how much we can rely on it for important tasks (just like real people). I think there is a paradox around intelligence and control. You want a machine that is smart enough to make independent complex decisions, but one that is unwaveringly reliable and obedient. But the ability to handle unfamiliar situations is fundamentally incompatible with obedience, since you cannot specify in advance what you want the machine to do.
I should note this short story was inspired by research into the evolution of analog circuits. The group used random analog circuit designs and assessed their ability to perform a simple function. The best (barely) functioning circuit designs were hybridised and reselected over multiple generations. The final designs were extremely competent at the task, but the engineers had no idea how they worked. They included components that were not connected to the main circuit, but when they were removed the design stopped working (so they were tapping into some weird electromagnetic resonance effects that nobody had considered as potentially functional).