ELIZA is a 1966 chatbot that takes input, flips it around, and says “Please elaborate on X”. This was apparently interesting enough to make people think it was intelligent back in the 60s.
Here’s how a Solo RPG with ELIZA might go:
ELIZA: Please tell me what's been bothering you.
YOU: A security officer recognized me when I disembarked
ELIZA: Can you tell me why a security officer recognized you when you disembarked ?
YOU: He turned and discreetly radioed as soon as he saw me.
ELIZA: That is interesting. Please continue.
YOU: Has station security been unusual lately?
ELIZA: Have you asked anyone else ?
ELIZA can’t solve any of your problems or give you any interesting output. All it does is appear to listen, then prompt you to solve your own problems by always putting the ball back in your court.
It’s 60 years later and when I try using an LLM DM (or rather Traveller5 Referee), it still feels like the machine wants me to solve my own problems.
LLMs are awful at telling unique stories. It’s impossible not to notice that if you push it for a story beat, those beats start looking very similar. If I land at a space station and blindly grope for plot threads, they’re guaranteed to be generic.
Giddeon looks around the bar for someone interesting to talk to. Does he find:
1. A smuggler who blindly trusts him to sneak a crate out of port?
2. A customs offier who asks him, a random bar guy, to investigate a missing ship?
3. A corpo agent who wants to me to help sabotage a rival megacorp?
The same ideas, themes, words, conversations, quirks, twists, and response structure will constantly pop up, at least when you use the same model. Switching models will fix it for a bit, but really just changes which rut you’re stuck in.
"Okay," Giddeon grunts, signing the employment contract. "Okay, we will find a way to slip your ship past the blockade."
<A few messages later>
"Okay," Alen says, giving a resolute nod. "Okay, excellent. You will find this worth your while once we're in Ley Sector.
The worst part of this is that all aspects of the output are affected, even ones you don’t notice. Beneath the surface level slop, there are layers of underlying slop!
Everything advertises 128-256k of ‘usable’ context, but consistent recall sucks past even ~10k tokens. Past 16-32k context even huge corporate models start getting iffy. The ‘solutions’ to LLM memory problems are all finicky and this issue doesn’t exist at when playing on a tabletop.
The scan reveals a particle cannon hidden beneath welded metal sheets on the lower hull. "That's no scout. They must be pirates!" Giddeon exclaims!
<Some time later>
The mag clamp from the rogue scout latches onto Giddeon's ship. "We're being boarded! They must be pirates!" Giddeon cries.
LLMs understand subtext only if it’s in their face in recent context. I end up typing extremely on-the-nose sentences like:
> "Our latest scan suggested the Southern hemisphere has more agriculturally viable land than we previously thought," Giddeon lied, fully aware his scans showed the planet was a desert and that Alen's light freighter didn't have sensors to dispute him.
Without the numerous reminders of Giddeon’s treachery and Alen’s incapability to detect it, Alen would certainly land on that planet and find a lush wonderland because LLMs will happily roll with lies.
Just try to have two characters keep secrets from each other, let alone multiple characters, let alone multiple factions. This is effortless in your head and impossible for LLMs without managing who knows what full time.
> Giddeon's remains calm. If it comes to blows, he's trained in low gravity close quarters combat and this young pirate certainly isn't.
The brigand's eye twitches slightly. He knows that Giddeon is trained in low gravity close quarters combat! This wouldn't a fair fight in the slightest.
The most enjoyable part of TTRPGs for me is managing multiple plot threads. I’m not running a string of one-shot adventures. I want to explore a universe full of unsolved mysteries, unresolved conflicts, and changing personal and business relationships.
The more threads you put into context, the more the LLM will try to tie them together nonsensically, place NPCs in scenes they don’t belong in, etc. Worse, LLMs can’t tell what logical part of plot you’re in. You could can be in the onset/rising action and the LLM wants to twist it into a climax or resolution.
Needless to say, LLMs are not the best way to track “RPG stuff” like I can easily do in a notebook. I’ve seen a few external programs that add a real crunchy-numbers-and-inventory interface on top of the LLM, but due to the nature of the space, everything is a vibe-coded shell of something interesting that will never be a vibe-feature-complete-product.
Expecting an LLM to come up with good story ideas or manage your gameplay isn’t a great experience. What can you do that works?
You take all the initiative
You come up with creative ideas
You drive all the story
You make all the decisions
You keep track of the state of the world
And all you do is bounce ideas off the LLM and it can respond in a way that reasonably works. Except…
> "So, we just escort the light freighter to Ley Sector and you'll pay us 50 Kilocredits? That doesn't make sense. What's the catch?" Giddeon asks.
"Yes! Just escort the freighter. It's a simple job!" the shifty scientist says eagerly. "And it doesn't make sense?" he says, feigning shock. "I assure you there's no catch, Giddeon."
Look at this response. Beat for beat, it’s following Giddeon’s prompt. It parrots ‘escort the … freighter’, ‘doesn’t make sense’, and ‘no catch’ without adding much of value.
60 years later, I’m still basically talking to ELIZA.
The LLM is twisting my words back at me into slop that makes sense… but it so rarely says ‘no’, pushes the story in a meaningful new direction, or does something truly interesting or unexpected that it’s still forcing me to do the real work.
It’s still a mirage. The machine doesn’t truly understand anything.
So I’m back in my notebook. Giddeon is trying to collect the reward from the shifty scientist. I roll No, and... on my outcome dice. He’s going to deny me payment, obviously. And... means something else negative is going to happen.
I consult an image oracle, a big matrix of icons that I interpret to guide what happens. I roll three times:
A paragraph symbol, probably his supervillain speech. He’s not dumb enough to actually reveal anything, just brazen enough to admit he’s not going to pay me while he escapes the station.
The plane landing must be a ship that docked to pick him up. Who is that?
A DNA double helix? Biologists? Scientists aren’t usually the type to play these kind of payment games. 50,000 credits is a lot to short change my crew, but small enough that any well-funded operation would just pay me rather than risk local authority involvement. I bet they’re alone. And what was on that freighter I escorted?
I add a new story thread to my notebook. Rogue Biologists?. The gears in my brain are working. There are 100 ways this can play out.
No fighting the machine. No repeating ideas. No managing context or RAG.
It’s 2026 and I’m somehow still better at this without a computer than with one.