Adding build recommendations

[hazy rapids]

Since the game already sends data to the server about evolutions and the success rate of completing the game with specific evolutions, it would be possible to add a tab with evolutions for each playstyle, allowing players to focus on the strongest builds. Alternatively (modern, and convenient, without send data of evolutions to the server), add an AI feature that displays, next to the evolutions, the percentage of the one that best suits the current playstyle based on evolutions already obtained, or those that have been played and successfully completed.

[urban cloak]

hey that's an interesting suggestion regarding showcasing the data of people's choices and win rates somewhere! It's similar to an idea we were toying with for our own website but haven't had time to make it happen (yet)

Regarding the LLM AI though: we have a strict no (LLM) AI stance on our game (a personal choice for Odd Dreams Digital)
Not to say it couldn't happen with regular AI and hardwork, but we definitely won't be adding Agentic AI into our EIC 🙂

[hazy rapids]

What does an LLM have to do with this? It’s just a standard recommendation model - no API, and it runs locally within the game. LLMs aren’t used for this kind of thing; that would be overkill, and it would actually perform worse. It’s simply a local model that recommends builds tailored to different playstyles. It’s fairly easy to implement.

[rigid jewel]

Interesting. I can imagine given evolutions optionally showing a small list of "synergies," but I'm not sure it makes sense to automate it. Like where unlocks/bans are shown, or something like that.

[rigid jewel]

It's worth noting I'm not a programmer or game dev, but I am a biology student (variables galore!) and have taken a course on basic statistics. I think if the system just used player data to figure it out itself, it might end up with erroneous recommendations of a few types (which I'll give names for the sake of possible future discussion):

• Broken Chain: Two evolutions might rely on a third to make them more effective together. If this third evolution is missing, the synergy chain is broken. For example, Forager and Body Slam work well enough together, but their actual synergy is through Hump.
• Appearance/Affinity Bias: Raising a specific affinity makes its other evolutions more likely to appear, so they might be picked more for that reason alone. The most basic data could be expected to reflect this. Furthermore, a player might take a certain evolution hoping to get another to complement it, and then not get that second evolution because it never appears or they prioritize something else.
• Prerequisites: Claws will never be taken without Arms. If the algorithm is too simple, it might consider that a synergy, since the most basic data without context suggest that it is.
• Extraneous/Unrelated Evolutions and Anti-Synergies: Certain evolutions are treated by a lot of people as an "automatic grab" whenever they show up. Cheek Pouch, Plating, Regeneration: these benefit any build. Without a huge amount of input data, I doubt a usage-based algorithm could reliably determine when they actually synergize with something specific.
• Multiple: Carnivore + Herbivore have opposite Affinities, but they have to be taken together with Piscivore in order to unlock Omnivore. You'd have to have an algorithm that knew how to work with or around every possible case and variable.
  I don't hate the idea of synergy recommendations, but to me, they sound really difficult to automate. Then again, some people are really good at math.

[compact flicker]

Agree that it'd be hard to automate. Also think it'd be unnecessary and suboptimal.

An easier way to provide build help to players would be a leaderboard that's filterable by run data, i.e. whether the run was winning, or had a particular evolution, or maintained a conduct (e.g. pacifism). I'll probably make a feedback post for this later, but for now here's an example of this sort of leaderboard: https://dcss-stats.com/

[urban cloak]

It sounded like you wanted to use Generative AI to implement.

Whether it runs locally or via an online API, whether it's LLM specifically or another form of Generative AI, we're against using it as a studio policy.

But as I mentioned: if what you mean is a standard recommendation model based AI (that's not Generative AI to be clear) then that's an option for us to consider for the future yeah ☺️

[hazy rapids]

Prerequisites: Claws will never be taken without Arms. If the algorithm is too simple, it might consider that a synergy, since the most basic data without context suggest that it is.

However, he had been taught that the claws always came after the hands

[hazy rapids]

No, not generative AI

[hazy rapids]

Something like that. Percentages near evolutions.

[urban cloak]

That's the way to visualise to the player not the implementation itself which is what it sounded like you meant when you said about a more modern solution ☺️

But we've gone way off track here:

In summary:

• the idea is really cool and something aligned to what we were thinking of a while back 🦀
• (unrelated to this idea) we as a studio don't use Generative AI in any form (including embedded local models). But that doesn't mean we can't do this idea without it of course, there was just a misunderstanding that I understood you wanting to implement it with it and I gave a side comment letting you know our stance on it is all