Against The Storm manual?

[WinterFresh]

Hello all,

I was wondering, is there a comprehensive manual somewhere for this game? Something that contains every bit of information, preferably in text format regarding this game? Like for example, it should contain all of the blueprints and their production, all cornerstone info, all pop needs etc, all neatly packed into one document. This is different from a guide.

The reason I am looking for something like this is because I want to feed it into a GPT knowledge base and see if the end result would make a good right hand to the viceroy, let's say, to help with the decision making. I think it would be fun. If the supporter pack had something like this I would insta-buy it.

Many thanks.

[Leisurely Luke]

That's an interesting experiment! I'd be curious to learn if you received any useful advice after feeding GPT with AtS guides. The Supporter Pack doesn't include such a manual, but there are several guides here on Steam: https://steamcommunity.com/app/1336490/guides/
And I've seen many top-tier pieces of advice in discussions here.

[la_nague]

GPT is a language model, you would have to feed it strategy guides.

It cant create strategy from a manual by itself.

[WinterFresh]

Originally posted by Leisurely Luke:

That's an interesting experiment! I'd be curious to learn if you received any useful advice after feeding GPT with AtS guides. The Supporter Pack doesn't include such a manual, but there are several guides here on Steam: https://steamcommunity.com/app/1336490/guides/
And I've seen many top-tier pieces of advice in discussions here.

Thanks, I appreciate the response. The problem with the guides is that most of them use images to illustrate strategies and such. The AI will not be able to interpret the images accurately or will use too much processing power for interpretation.

OpenAI boasts that their GPT system can be fed around 200 pages but in reality, once you play with it enough you realize that it will only focus on like only a few pages once you start the conversation, depending on what you start with, and then forget everything else. So a certain level of conciseness and structuring is also needed, at least until the tech advances beyond these limitations.

It's worth experimenting nonetheless. It's fun having "Eldric, The Storm Guide" GPT address you as viceroy and try to give you lore friendly advice, even though the contents are not really useful from a gameplay perspective.

Originally posted by la_nague:

GPT is a language model, you would have to feed it strategy guides.

It cant create strategy from a manual by itself.

It cannot create a strategy without a manual either. How will it know which blueprint to recommend if it does not know what all buildings produce, what are the production chains, what are the pop needs. At best it will either give you regurgitated info from the guides or if the guides don't contain the requested information, it will start to "hallucinate", even with GPT4.

One problem I had for example was when asking the AI what is the best pick between lumber mill, apothecary, press and cooperage, keeping in mind the mushroom biome and a pop of lizards, beavers and harpies. Instead of giving a coherent game related answer, it will always say apothecary because such buildings historically were places of healing and that the settlement should prioritize healing its population as a precaution against the storm, lol.

[Spacesuit Spiff]

Cool idea. I like to see this sort of thing, even though the rest of this comment is gonna be tearing it down.

Even with a full manual, LLMs don't conceptualise the world in a way that would let it come up with a strategy. Sean Carrol's podcast recently had a fascinating episode about this. Your best bet would be to feed it the entirety of this forum and the AtS subreddit (or the last few months since older discussions will be about older versions of the game), and then it can mimic the advice given here.

If you want to quickly test whether an LLM is capable of strategy without ape-ing the answers, ask it a question like "Imagine a different version of chess where white has 40 pawns and black has 2 queens, who wins?" It can't handle hypotheticals like that because LLMs aren't designed to reason through novel questions. Throwing a relatively small game at it would likely be similar.

One genuine use I could see for it is feeding it the various events and their costs, and then asking it the probability that you'll need a certain resource when you open the next glade. Probably easier to just do that with a python script though.