Playing AI Dungeon
The AI generates the next word by creating a list of all the possible words that could come next and the probability that they would be next, then selecting a word from that list by doing some math with a random number. The Advanced Settings allow you to change the numbers within that calculation that aren't random.
This is complicated, and there are no standards, so it’s highly recommended that you experiment with them and come on the Discord to ask for advice.
You may also find it helpful to read our more advanced article on how the AI works, which will give you a better understanding of how these function beneath the hood.
Raw Model Output
The Interface cuts off everything after the last punctuation mark so that you don’t have a half-finished phrase at the end of the AI’s last generated output. Raw Model Output turns off these changes, so you can get that extra text at the expense of it being incomplete sentences.
Temperature means how random the AI should be when selecting the next word: lower temperatures make unlikely words less likely to be selected, while higher temperatures will make them more likely to be selected. If the temperature is too high, it will start to select some very strange words, but if it is too low, the AI will trend towards the most likely words only, which can become boring and repetitive.
The default temperature is 0.8, though you might consider something like 0.6 if you feel the AI is getting a bit too far from a reasonable response, or you may consider something like 0.95 or even 1.2 if you want to see the AI put out some really uncommon text.
Response Length determines how many words the AI will generate before it finishes. This is a matter of personal preference: You may want short responses that you can quickly read and decide when you want to jump in, or you may want very long responses that you can take your time with. Do whatever fits your play style.
Top-K is only available in Griffin, and determines how strongly the the AI’s output should be weighted towards less likely tokens. Higher values mean more randomness. It is not well documented at this time, so you’ll have to play around with it to see what it does.
Top-P tells the AI how far it should potentially deviate from what it thinks is the most likely next token.
If Top-P is too low, the AI will be constrained down to only the most likely words, leading to output that might be very repetitive or otherwise uninteresting. If Top-P is too high, the AI will have access to words which are extremely unlikely to appear next, possibly even grammatically incorrect ones.
A Top-P of 0.5 to 1 is probably best for most users.
Repetition Penalty is only available in Griffin and changes the probability values to make the AI less likely to repeat words. This may sound good, but remember that the names of characters, as well common words like "You", "I", "and", “a”, etc are going to be penalized, so too high a value will start cutting out these words in favor of some otherwise very uncommon ones, even becoming grammatically incoherent.
Repetition Penalty defaults to zero, and while you can set it very high, the AI will start to trend towards those strange outputs if you put it too far over even 1.
Memory Length (not to be confused with the length of the Memory) refers to how much text is sent to the AI, and should always be at its highest, unless you know what you're doing: it's just letting you send more text to the AI, so that's pretty much always better. An example of a time you might turn it down is if the AI is fixating on something in a previous moment, in which case you might turn it down for a couple of actions, then turn it back up.
Afterward: the Best Tip for Learning these
There are no rules in AI Dungeon. As with anything in AI Dungeon, the best way to get a feel for things is to experiment, and there’s no penalty for doing so: if you are really interested in understanding what these do, the best way to do so is to take the time to change them and see what happens. You can nudge them around based on what you think you'd personally prefer, and you might like it better, or you might go back. Sure, they're going to do weird things with the AI's already weird output, that's part of the fun of using AI Dungeon.
There is no consequence to messing with these. There's no consequence to turning up the Top-P, Temperature, and/or Repetition Penalty to maximum just to see the AI flail about like a tube man. There's no consequence to turning everything down to zero and watching the AI do its best impression of a broken record.
There isn’t an agreed upon "best value" for any of these. Any values given are just recommendations based on what people have found work better for them. If you like something that no one else is using, that’s great! At the end of the day, it's your story, your play-through, and your experience, and you should feel free to do whatever makes you happy with it.
© Latitude 2023