I recently re-read Cal, one of Isaac Asimov’s final short stories. It was published in the 1995 collection Gold, which is a hit-or-miss collection of unpublished Asimov stories sitting around near the end of his life, paired with essays on sci-fi as a genre and writing as a practice.

That’s enough by way of an intro. Here’s the point: In the story, Asimov lays out a scenario involving a robot that learns how to write. As the robot becomes more complex, its writing becomes better than the writing of its master. The master worries that the robot will come to overshadow him. In response, he calls a technician to dumb down the robot’s programming.

This produces a crisis in the first law of robotics.

Sound familiar?

Sci-Fi and LLMs

In today’s world, we don’t have robots with human-like bodies that can write as well as humans. Instead, we have LLMs. They lack human bodies. And while they write competently, even fluently, they also write in clunky, programmatic ways that reek of pedantry and mindless rule-following.

And so, we’re not yet at the point of ‘Cal.’ But we might get there. In theory, LLMs could one day write as well as good human writers. And in theory, though we’re quite far from this point, we could put them into a robot body that looks like a human body.

How would we handle such a situation?

Ultimately, I don’t know. Maybe this would be the point at which we collectively decide to restrict LLMs. Maybe we would ‘dumb them down’ in the way that happens in ‘Cal.’ Alternatively, perhaps the event would drastically change or even restrict what human writers do.

But I think we can take heart from a certain aspect of ‘Cal’ that still feels far-fetched. Despite vast improvements in LLMs, they’re really not all that close to the ‘Cal’ scenario. Not only do they lack human bodies, they also lack human flair and soul.

In short, LLM writing is dull, boring, and plodding. And, frankly, it doesn’t seem to be getting that much better.

While I can’t say how we’d solve the problem should it arise, I’m far from convinced it will arise. The initial hit of plausibility fades quickly into skepticism. Ultimately, LLMs don’t look likely to get beyond the mere recombination of ideas into dull composites.

That feels at least a bit comforting.

Image Source