Ted Chiang's Novella Beats Every AI Prediction, Study Finds
How accurately did science fiction stories predict AI?
A scored ranking of science fiction's AI predictions puts Ted Chiang's The Lifecycle of Software Objects first, for capturing the messy work of teaching, maintaining, and selling AI. Heinlein's The Moon Is a Harsh Mistress and Ken Liu's "50 Things Every AI Working with Humans Should Know" follow. No pre-2017 work clearly describes the transformer attention mechanism, or the leap from chatbots to agents that discover new mathematics.
I agree with Ursula Le Guin that "Science fiction is not predictive; it is descriptive."
- pugio
I have a hard time suspending disbelief with sci-fi these days. It's really hard to imagine any believable future along current trend lines; either you contrive a reason why AI didn't become a thing (1), or you have to explain why humans are doing anything - piloting a starship, science, exploration, etc. (2)
(1) Butlerian Jihad. Sanderson's Skyward at least had a reason.
(2) Banks's Culture had a possible future. I'm sure there are others, but not many.
I think about this a lot, and while I have tons of near future sci-fi ideas (which probably would have seemed very far future, 10-20, years ago), I just can't conceive of anything much farther out. I guess that's why they call it the "singularity".
- joegibbs
AI from the past seems more like it’s based on the computers of the time - the AIs seem to be very logical, never make mistakes and act like they’re an expertly programmed bunch of switch statements and ifs. Whereas modern AI is based on human text and so the models have human biases and do things like swear in their chain-of-thought and sometimes make errors for no reason at all.
- andrewflnr
Not mentioning Peter Watts here is criminal, because he came unnervingly close to nailing it. The voice of Rorschach in Blindsight is very similar to early LLMs. And, well, hopefully the dead internet of Starfish doesn't turn out to be a prophecy but I'm getting nervous.
- seanmcdirmid
I’ve been thinking about this for awhile now, a complete miss: Sci-fi author and computer scientist Vernor Vinge popularized the term programmer-archaeologist in his 1999 science fiction novel A Deepness in the Sky
The premise is that AI never came around and we were left scavenging code since that was more efficient than writing new code. But with LLMs, it’s much easier to write new than to scavenge.
- jasonjei
Many sci-fi writers really dislike LLM for various reasons, whether it is the generative works they produce, or for allegations of copyright infringement due to LLM training on their works. Some have taken such a negative attitude towards LLM that they actively refuse to use Claude or GPT even for coding projects. John Scalzi and Catherine Asaro come to mind boycotting any LLM use…