4 ms·

Ai can produce a new type of game where choices are dynamic and outcomes are generated by LLM agents. Fiction is an hallucination and LLMs are master hallucinat

Ai can produce a new type of game where choices are dynamic and outcomes are generated by LLM agents. Fiction is an hallucination and LLMs are master hallucinators. Basically LLMs have to be given assets and game components that they can easily compose.

withinboredom 2y ago
Having had an LLM tell me a story, my answer would be that this is a dumb idea. LLMs have no concept of realistic cause and effect.
- ninetyninenine 2y ago
  Nah I think not considering this idea at all is the extremely dumb and brain dead opinion. LLMs can tell stories. Realistic causes and effects aren’t even consistent in human stories. A good story isn’t 100 percent dependent on this. The LLM walks the line between hallucinating too much and sometimes not. Either way you can pretty much guarantee that almost all stories made in games now are already mostly written by an LLM. It’s just the writing is edited and curated by a human.
  - woolion 2y ago
    I've done a short LLM-powered VN, and LLM actions were restricted to local interactions only because of how weak it is at making up the story. It's great at removing the parser-based interactions, but I think that's it. There's a second technical problem that such stories are represented by a form of state-machine and that you would need to recompile it on the fly, making many checks very difficult (you would need to be able to check reachability on the fly, chunk transitions, etc). I think it would take years to get to the level of some of the great IF games with an LLM, and not just a cool PoC.
  - usrusr 2y ago
    As a grumpy old symbolic ai hand I do wonder if it was possible to build a (perhaps crude) ontology based simulation with consistency, cause and effect and so on and then use the results of that for prompting an LLM. But as a consumer, I lean far to the side of "give me a handcrafted tunnel experience with the illusion of choice" in the divide between consequences yes or no. I don't think I'd actually want this "simulation behind an LLM facade". If I'm in the mood for reading (or for listening to voice actors reading to me), I'd rather have it be something more meaningful than just a game state. But to those on the other end of the spectrum, this might actually be the holy grail of game building.
ben_w 2y ago
> Fiction is an hallucination and LLMs are master hallucinators. They're jacks of all trades, master of none. This has its uses, but they have limits, and for now at least, those limits are under the threshold for that. I have actually tried using them to make a text adventure to help learn German. The result was at the lower end of the quality range I've witnessed from LLM output: a nice first draft, not shippable, missing a core element, missing a lot of content, too simple, the kind of thing where you'd give the output of the LLM as a code challenge to a job candidate to see how they improve it.
jandrese 2y ago
An AI can easily produce a filler slop story. They struggle much harder with creating something new and interesting. For a CYOA type story for kids that might be reasonable, although they tend to make a hash of the details. There are more problems, like does the AI know when to stop? Can it recognize or generate a bad end or a good end without explicit instruction from the player? Something like: Generate a choose your own adventure story about a young boy shipwrecked on an island populated by hostile pirates. A hidden cave holds treasure. There is a jungle on the island. Dangerous jungle creatures inhabit the island and the boy can not fight. Also hidden on the island is a boat. Each story section is around 200 words long and ends with a multiple choice question for the player to select which path they want to pursue next. The story is complete either when the boy dies or finds the boat, after no more than 20 story segments. I have some doubts the AI will be able to handle all of that and keep it interesting and coherent. This sort of storytelling requires some attention to detail that LLMs usually struggle with.

withinboredom 2y ago

Having had an LLM tell me a story, my answer would be that this is a dumb idea. LLMs have no concept of realistic cause and effect.

ninetyninenine 2y ago
Nah I think not considering this idea at all is the extremely dumb and brain dead opinion. LLMs can tell stories. Realistic causes and effects aren’t even consistent in human stories. A good story isn’t 100 percent dependent on this. The LLM walks the line between hallucinating too much and sometimes not. Either way you can pretty much guarantee that almost all stories made in games now are already mostly written by an LLM. It’s just the writing is edited and curated by a human.
- woolion 2y ago
  I've done a short LLM-powered VN, and LLM actions were restricted to local interactions only because of how weak it is at making up the story. It's great at removing the parser-based interactions, but I think that's it. There's a second technical problem that such stories are represented by a form of state-machine and that you would need to recompile it on the fly, making many checks very difficult (you would need to be able to check reachability on the fly, chunk transitions, etc). I think it would take years to get to the level of some of the great IF games with an LLM, and not just a cool PoC.
- usrusr 2y ago
  As a grumpy old symbolic ai hand I do wonder if it was possible to build a (perhaps crude) ontology based simulation with consistency, cause and effect and so on and then use the results of that for prompting an LLM. But as a consumer, I lean far to the side of "give me a handcrafted tunnel experience with the illusion of choice" in the divide between consequences yes or no. I don't think I'd actually want this "simulation behind an LLM facade". If I'm in the mood for reading (or for listening to voice actors reading to me), I'd rather have it be something more meaningful than just a game state. But to those on the other end of the spectrum, this might actually be the holy grail of game building.

ben_w 2y ago

> Fiction is an hallucination and LLMs are master hallucinators. They're jacks of all trades, master of none. This has its uses, but they have limits, and for now at least, those limits are under the threshold for that. I have actually tried using them to make a text adventure to help learn German. The result was at the lower end of the quality range I've witnessed from LLM output: a nice first draft, not shippable, missing a core element, missing a lot of content, too simple, the kind of thing where you'd give the output of the LLM as a code challenge to a job candidate to see how they improve it.

jandrese 2y ago

An AI can easily produce a filler slop story. They struggle much harder with creating something new and interesting. For a CYOA type story for kids that might be reasonable, although they tend to make a hash of the details. There are more problems, like does the AI know when to stop? Can it recognize or generate a bad end or a good end without explicit instruction from the player? Something like: Generate a choose your own adventure story about a young boy shipwrecked on an island populated by hostile pirates. A hidden cave holds treasure. There is a jungle on the island. Dangerous jungle creatures inhabit the island and the boy can not fight. Also hidden on the island is a boat. Each story section is around 200 words long and ends with a multiple choice question for the player to select which path they want to pursue next. The story is complete either when the boy dies or finds the boat, after no more than 20 story segments. I have some doubts the AI will be able to handle all of that and keep it interesting and coherent. This sort of storytelling requires some attention to detail that LLMs usually struggle with.