Larian Studios Forums Divinity General Discussion AI for Companions?

There was this heated debate about use of AI for development of Divinity by Larian. I did not follow and don't know exactly what Larian meant when swearing not to use AI. I only hope it does not equal the decision of a shipowner in the 1870s to stay away from steamers because they are dirty and unreliable, and proudly trust sailing ships only.

Let's assume Larian only does not use AI generated art and story content but is open to AI otherwise. Interaction with companions is in my opinion an important part of a group based RPG. There are some mods for video games which use Large Language Models for controlling NPCs or allowing communication with followers, often with the possibility to have Text To Speach functions. LLMs and TTS often need internet connections (I really don't like games relying on Internet connection) but are also possible as local models to a certain extent.

I'm not experienced with AI in general and relatively reluctant to use it in my job (still waiting for even one totally correct answer of an LLM) but I'm excited about the possibilities in games. Especially if integrated by the developers themselves, and with TTS at the local level, new dimensions of gameplay interaction seem to be possible. Companions would get a kind of personality and not only some bland surface. Something new could happen. Replayability would be better. Also a lot of funny situations could occur, from cunning or the frequent stupid AI decisions.

Mods are great and I assume their makers know at lot more about AI than the common video game devs, but I would surely throw a bunch of money at any game developer who would offer for example a DLC with AI functions for combat control and companion interactions. Even better if a game is developed from the ground with AI in mind.

What about the use of AI f.e. for companions and maybe some enemies in Divinity? At least as an option? Is there anything known about plans in that direction? What were the concerns about such a feature?

You are confused by the layman term "AI". The actual experts dont use it. Your confusion stems from the problem that you see the term "AI" used for wildly different things, and think they are the same.

They call REAL AI "machine learnig", or more specifically "neural nets". It is a technology developed in the 1950s that is supposed to emulate how biological brains work. Neural networks are not programmed, they are trained. They cannot be read, understood, or bugfixed. They are inherently wasteful and error phrone. Many modern chips have now specialized circuitry just for neural nets.

NPCs are no controlled by such real AI. They are controlled by scripting. Which is also called "AI" in gaming since decades, but it never was machine learning.

Scripting is basically the same as regular programming, but typically it is not compiled and runs much slower than compiled programs. This makes scripting much easier to use for a level designer. The lack of performance doesnt matter, because scripts dont do any computaionally expensive tasks.

Importantly scripts can be read, understood, and bugfixed. They are mostly like regular programms, however since they are not going to be compiled, they may be more comfortable to use.

So Larian basically promised us that their artwork will always be human made. They will use AI only to help their artists to do their job better, but the actual artwork in the game will be of highest quality.

However their monsters, companions etc will still be controlled by scripting, as always.

Using LLMs for ingame characters has been suggested, but is also a bad idea. LLMs are also neural networks, and are thus also never reliable.

Though I guess we have all suffered to the incoherent babbling that Laezel has in later chapters of BG3, where she just forgets what she had said before, and LLMs wouldnt be any worse.

Personally I still would strongly prefer human written interaction. I had my share of experiences with AI now and I am not amused. I would rather take the human written dialogue, thank you very much.

LMM's for ingame characters has a lot of flaws.

From the jank that is involved, to the issues regarding voice acting.

If you want to experience this, look at the game "Whispers from the Star" which is entirely built upon the premise of interacting with an AI based NPC.

As for "Combat interactions" the problem is the same problem with having good scripting. Which is that the computer can very quickly and easily start to perform superhuman feats. This is the basis of TAS (Tool Assisted Speedruns) whereby scripts are utilised to perform perfect inputs that regular fleshy humans are physically incapable of performing (Both based on reaction time and also actual input delay from physically pressing buttons).

Scripting is fine enough to use for "AI", as in the standard ways that NPC's normally function in games. You can even have some advanced scripts where NPC's "Learn" and adapt to situations, several tactical style games utilise such a feature for their advanced difficulties.

Beyond that, having NPC controlled companions is generally a thing for RtwP, I don't think I've seen it being a thing in TB style games. I'd imagine the main concern is with the player experience. TB games are often maligned due to the prolonged enemy turns where the player is sat there watching the NPC enemies do their thing (This being even more egregious in Larian games due to their high enemy count for combat encounters), adding in the need to wait for your party members would compound on that.

I don't think a world populated by bland, generic and agreeable sycophants really fits the theme that Larian seem to be going for with Divinity.

I think you can solve many of the mentioned problems by giving rules for the AI, for example avoiding superhuman feats, and defining personalities. It's also not a general Goggle AI search what would happen if a request and decision were to be made, it should be a specially trained AI. It may be that we are not all able to see the latest advances in this area which is evolving quickly. And bad examples of the past should not necessarily define the field for ever.

The argument about NPC control and TB experience is an interesting one. In BG 3 you can play 4 different classes which can enhance your game greatly because of the diversity. AI control would take that away. Although I also read several complains over the years from people who did not want to play more than one character, so an optional AI control may be advantageous especially for TB. Where we are used to waiting.

At which point it's not really being AI it's just being scripted behaviour...

Like, all the work you'd have to put in to make an AI system play by very specific rules so as to not be AI... You may as well just put that effort into just scripting directly what you want.

"Oh, but you could get some unique generated behaviour reacting to player inputs with AI!" - Well, not if you're trying to restrict them to having a personality, where you'd have to essentially block out anything other than what you intend that personality to do.


It's more that the field is actually rather limited in what it can do.

Since "AI" isn't really "AI" it's just scripted bots that parse through uploaded datasets and throw back things related to what is put in front of it.

Which is really good for science applications. Where it can parse through test results to find anomalies or parse through centuries of journals to find all references about a particular topic.

But for anything else, it's just not that useful. Nor will it be so long as its functionality is entirely based on this premise of simply parsing datasets.

If some new "AI" came about, one that was either true "Artificial Intelligence" or simply another scripted bot that provides an alternative way of doing things (Much like we've had Machine Learning for a while, which is useful for specific tasks. Notably in video game development it can be useful for developing animations) then maybe there'd be more potential for its use case.


Though DOS2 utilised the Lone Wolf feat as a means of facilitating such play. Whereby you could go alone (Or with just a single companion) and get a bunch of boosts to make that more feasible.

The ability for things like that to tune combat to party size provides more flexibility for players, who can decide how many characters they want to play with. Rather than forcing them to play with 4 always (Which is more than just controlling them in combat. You have to equip them with gear and pick level up choices and the like. You also have them making comments throughout the game, which you might find tiresome especially for more abrasive characters like Red Prince or Astarion)

I know my first playthrough of DOS2 was using Lone Wolf due to simply not liking nor wanting to bring along 4 companions.

You also have to consider, how well NPC controlled party members will perform. Like, in BG3 we already have seen how terribly NPC allies do, with them loving to go run into their own AoEs (If they don't cast them directly on themselves). Do we need to force people to play the entire game relying on such characters?

Its propably no feasable to train your own ai for that at least at this point in time. Consider all the crawling of the internet you would have to do to train the model enough so it has enough material for natural speech and millions of options to respond in a proper way. Also the cost it brings to maintain and implement the ai as well.

If you play the new divinity, there has to be a way to get the AI API of the Larian Server kind of like claude code and the likes, meaning it would not work, if you wanted to play the game offline. And deploying the ai directly onto the users hardware would be a desaster for performance.