1. They are slow, in every sense of the word. Too many don't follow instructions, so too often, what they write can't be used.
► Show Spoiler
The Valve method is apparently ``to make two games and throw out the bad one''. That obviously can work. But what usually doesn't is ``make ten games and throwing out the bad nine''.
2. They don't understand that irrelevant lore does not substitute for story. Lore that cannot possibly get into the game is not something they should be writing instead of a story.
► Show Spoiler
A: ``But what does she, like, eat? I mean, we know she is an strong powerful adventuress, and can use both sword and staff. We need to have a meeting and figure that out. She is a powerful high demon elf and her beloved horse, Champion, was stolen three times; so we know, courtesy of the trauma, like, she doesn't eat horsemeat. Everyone knows, however, that elves demons eat only horsemeat. They eat horsemeat, Yellow Stones, Blue Stones, and turnips, and nothing else. Well, she is independent and unique and refuses to go along with her people's traditional practices, because she respects horses. She loves her horse. She is a vegetarian, and only eats, uh, Glowing Mushrooms that grow in Damp Caves. This is why two centuries before this quest, she was exiled. Actually, long ago, she enjoyed bacon. Then she learned a spell to mindread swine. Now she doesn't eat bacon. Her horse, Champion, once ate a frog. It liked it. That's why, in her inventory, there is always a frog. A frog on a stick. It's a treat for the horse. We can explain this in the item description; yeah, we should do that. Her horse, Champion, had a dream that is was flying. But it's not a flying horse. It's an ordinary horse. It's just a dream her horse had. It wants to fly, but it can't. That's really sad, when you think about it, isn't it? Gwenneth, anyway, also enjoys golf, tennis, and fish. She like fish. Fish are cute. Gwenneth thinks so, hence she also won't eat sushi. Not that elves demons eat sushi. Once, she considered eating sushi, but she reconsidered. She had two dreams about it, one in April 6052, one in December 6073. But she definitely still hasn't tried sushi. The player can confirm this by reading her journal entries.''
B: ``Who cares.''
A: ``But people will wonder about it.''
B: ``No, they probably won't.''
If the NPC was begging in the streets for food, then it might be funny if the player offers sushi or cavier and the begger refuses. Player then asks why. NPC says why. But if the NPC is never hungry, five pages of her gourmet adventures contributes nothing and doesn't do the heavy lifting of five useful sentences.
``Do you know where my dog is? Have you seen it? My dog is really special. (Whispering) It can do magic. (Normally) Please help me find my dog, because it's my only family now.''
You don't need to understand the characters, only players do, and they can do it with ends-means in a few sentences, plus a bunch of questions. If it doesn't have any consequences whether a question is answered, maybe don't try answering it. Can dogs do magic in this setting and why? And how? ``Who cares.''
You meet the dog. ``I'm actually a dragon.''
``Well, you don't look like a dragon.''
``That's because I'm currently undercover as a dog. I'm infiltrating the Thieves Guild.''
``That villager was...''
``Yes, the second master of the Thieves Guild.''
Billy the Axe likes turnips. How do you know? He tells you! Good enough. Next. If he doesn't tell you, doesn't matter what he likes. Next.
Not this: Billy the Axe likes turnips because his mother liked turnips and she fell off a cliff and Billy was really sad and in memory of his mother who fell off a cliff he only eats turnips. Billy is a random bandit. The player just kills him. Else the player entirely misses him. Billy's mother is not in the game, don't write lore about how Billy's mother liked turnips, let alone why she liked turnips.
3. They don't understand what a story is and write unusable vignettes that fail to invite interest.
► Show Spoiler
What a story is and what it is not. Budrys: A character in a context with a problem tries the obvious solution in that context, but it fails, surprisingly, for reasons revealed or actions other characters also had done, and stakes go up, and context also grows slowly, until he finally succeeds or fails big with highest stakes, and another reputable character confirms the outcome. If the test reader/player can guess one of these parts, no need to write it explicitly, it has to just be present in the minds of the audience. It's the bare minimum to be of any interest. After that, yes, how the character succeeded might lead to other side effects and larger problems and introduce other characters. Repeat. There are things the characters want but might not need, or need but might not want. There is history and lore and what the setting needs but does or does not happen, courtesy of the characters, but first of all, there must be the bare minimum to get attention and hold it.
Yes, this can be automated easily and the line of research goes back to 1970s-1980s, often with the most naive method.
KOS-MOS wrote: ↑
May 2nd, 2025, 13:55
NotAI wrote: ↑
May 2nd, 2025, 09:54
I'm joking. No, actually, we no longer hire writers. Our own proprietary AI writes our ****. It's trained on 19th century historical romance novels in 3 volumes. Not even modern ANN, but part expert system. Blackboard system. It still writes better than most people who don't read.
Would you mind sharing a sample of your AI writing? Just out of curiosity.
I can post the approach after a specific paper comes out. I'll post it later in
The Foundry.
The short answer is the following. (1) Filter markov chains trained on gutenberg plus own lore repeated many times, using blackboard, (2) estimate up to very high order for the chains, (3) they are not quite markov chains as history matters just a little, including own past output, estimate adjustments based on history of own output: train for that, (4) tokens on subject matter filter output as it is randomly generated, (5) ordinary pattern matching gives subject token to each output and you have a list of whitelisted subject-subject-subject combinations: wait until one is matched and display just that output. Performance is noticeably lower than any big model but soft real-time, cheap and fast, and you have fine control over it. It works in a game because it's a game. There is limited scope and you know the entire required topic combos whitelist. That is what does the heavy lifting in getting a good result. In the general case, outside a game where you decide everything, you don't know the adequate whitelist. It's just Claude Shannon meets Allen Newell meets Carl Petri. Still better than most weak writing, because you can mandate every story part to be present and in context.
