There's this idea that if AI can help the 79% skillful make it to 81% competence, the 99th percentile shall be out of business. Now - I'm out of business for writing because it doesn't pay well enough for me to bother. LLMs sure as shit aren't going to fix that.

A quick preamble: maybe I shouldn't have started this conversation as a response to the argument that writing is doomed that usualgerman is making, because that end conclusion is not one I intended to support.

It's the last 20% that gets you there. in everything. And AI has consistently not even begun to cross that 20% in all the years we've been yammering about LLMs.

I think this is where I went wrong: I don't know the shape of the there well enough when it comes to writing, so I made the cardinal sin of extrapolating. My assumption was that since LLMs have jumped from Markov chained nonsense to getting to 90+% in some forms of writing (StackOverflow, low-quality reporting, some technical writing) in record time, that fiction writing would not be that much harder.

But that's like saying lane-guided driving on a sunny day on the highway is only two steps removed from fully driverless autonomous driving, an assumption I hate since it's in a domain where I do have some idea of the shape of the last percentages. People have been telling me Tesla's autopilot has been improving rapidly for forever, particularly the past year. I (/we?) have for years been shouting back that the last X% is the hardest with self-driving cars, and that it too is not a given just like it's not a given a large enough quantity of monkeys will produce Shakespeare.

The example I have given to multiple people over the years is "yea I'll see it drive through a bicycle-busy Amsterdam street before I'm impressed". This came across my feeds the other day:

I instantly recognized these streets, the one at the 90 second mark is one I walked on just last week. I know exactly how attentive you need to be to drive a large car through there. So I'll admit I had to do a bit of soul-searching. There are goalposts I can move (it's a sunny day again, it's driving like a snail, there are just as many recent videos of FSD mode going kamikaze as there are of them doing something impressive, etc etc)... but it did make me re-evaluate: how good is good enough, exactly? How precise can I define what qualities we should be willing to give up, if it gives us something else in return? We can be a millimeter from the asymptote, definitionally unable to pass it, and that might be good enough.

I watched a video the other day (on YouTube) that discussed the drop in quality, specifically in conventions, that YouTube represents when compared to traditional media. The lack of professionals directly results in jumpcuts, in people holding the mic, in people speaking like they're reading text aloud in 5th grade. The dilettantes never used the conventions that traditional media had forged over decades because they don't know they exist. They're in the Dunning-Kruger zone like many more will now because of LLMs.

And yet - we are now used to jumpcuts, we are okay with people holding the mic in frame. We for sure lowered our standards, it is by inspection obviously worse. What YouTube has going for it is that it appeals in a different/novel way. Your examples…do not. (Other than being adorably wrong like Sunspring.) I'd argue that YT appeals mostly through serving niche interests to a degree traditional media will never be able to. I will gladly give up qualities like audio mixing, lighting, image quality if that means I can watch something in my niches that I otherwise would not be able to ever see.

What I was uncertain about, which is why I called it a cautious "some people under some circumstances", is whether there is an ability of LLM-produced piece of writing to offer the same "niche at scale" benefit that YouTube added to casual couch TV watching. The writing itself can be really bad, can have standards as low as the earths core, but if it scratches an itch there will be some people who will not mind suffering through that.

My expectation with fiction writing was that we’ll see the same thing happening that I’m noticing with coding: on the one hand, dilettantes trying to get “there” and failing, getting somewhere that they might find impressive but the rest of the world does not. And on the other hand the pros fast-tracking their process in some way with LLMs, automating the first draft which would be shit anyway and rewriting from there. They’ll still be writing, perhaps a bit faster than before. The former will be happy enough with how far they’ve come, content with their work and/or not knowing better, that they’d rather have their lower quality something than not having that thing at all, just like I’m content with my vibecoded webapp because it does 12 transit queries for me at once. The code sucks & the process is janky as fuck but Gullit it works, it gets the job done, even if standards couldn’t be lower.

But fiction writing doesn’t exist just to spout duck snakes at you; people get something out of it (symbolism, meaning, the human nature, …) that has to be more than a classifier can handle. That’s your point, right? Which I missed because I am not aware of what the last 20% is made of.

Or am I still missing something here?