“There are new skill sets and new possibilities for making games”
Alexandre Folliot and Oriane Piedevache of X&Immersion discuss interactive NPCs and other in-game opportunities.
Welcome! Every edition of AI|G is a conversation with leaders working at the intersection of games and AI. In this edition, we meet the X&Immersion team, which develops artificial intelligence and machine learning tools for game makers. We discuss how they're developing local AI solutions for NPC interactions, voice synthesis, and facial animation, with a clear emphasis on making these technologies practical within development constraints.
We’ve got a bunch more fascinating Q&As already lined up for the months ahead, so remember to subscribe. Next week, you’ll read a chat with the head of Roblox Studio. If you’d like to be on the schedule for a chat next year, drop us a note in the Substack app or email us at editorial@steelmedia.co.uk, and we’ll plan an interview.
Scroll to the end for hot AI news, including Supercell, DeepMind, Goodville, Midjourney and more, plus investment and conference updates.
X&Immersion
Meet Alexandre Folliot (CEO and CTO) and Oriane Piedevache (COO), the co-founders of X&Immersion, based in Paris. X&Immersion provides tools for games studios to support the development of text, voices, and animation, such as fully lip-synced NPCs with AI-driven behaviours. What stands out is their focus on solving concrete technical challenges and their realistic view of AI's role.
Top takeaways from this conversation:
AI has the potential for many uses in games - not just the creation of interactive NPCs or localisation, but also things like in-game helpers and chat-based support guides.
Local AI inference models are a crucial technical frontier. Game studios should think about implementing features that don't rely on external services that could become unavailable or cost-prohibitive at scale. Smaller, specialised AI models can be surprisingly effective for specific tasks like topic detection.
AI is currently best suited for enhancing rather than replacing creative work. AI "cannot work without human beings" and can be viewed as a production optimisation tool.
AI Gamechangers: Please tell us about the journey X&Immersion has been on since we last spoke.
Oriane Piedevache: At the end of last year, we raised some money. We did some small fundraising, pre-seed, but then we had a couple of other grants as well. The purpose was to have several big productions and games where we could be helping and also build our own SaaS platform so that people can use it and our tools more efficiently.
Also, we have been on an R&D project with a big studio that we cannot really talk about, and we’ve been working on local inference models, which means it’s basically like ChatGPT but local and non-server based. This is a huge technical pivot for us. We’ve been working on these projects for a couple of months now for when the industry is ready to have AI-driven NPCs who you can interact with specifically in your RPG game.
Alexandre Folliot: We did a €1.1-million fund-raise in total. It gave us some capacity to breathe and to focus on the R&D part. Through securing deals with studios who use our tech, we are in some games that have been launched this year: for example, Zero Games, Microids, and other games.
We notice that AI can be very scary, not only on the technical side and ethical side but also on the monetary side. When you use a lot of AI (or let your players use a lot of AI features) it’s got a cost. So our idea was, how could we do edge computing and have the calculation done on the player’s device?
AI Gamechangers: What are the technical challenges you have had to overcome to make that happen?
Alexandre Folliot: The greatest problem is that normal PCs, or phones or consoles, are not made to run large language models, or transformers for voice synthesis, or facial animation. Those are three things we have been doing. Basically, we had to change the architecture so we could build the models really smaller.
Basically, we cut inside the models to make them smaller, to make them more efficient and more specialised, to do one specific task. We changed the architecture to remove a lot of layers we’re not using inside the models.
“We didn't reinvent the wheel, but we changed the shape of the wheel a lot to make it lighter!”
Alexandre Folliot
The other really challenging thing was to have the model work with the player’s hardware while the game was playing. If you’re playing a game made in Unreal Engine, Unreal Engine is going to use some gigabytes of VRAM while you play. That’s why we see, when everything’s happening inside the game, that we have less VRAM available to do the calculations for AI models. This was the second hard part: how could we dynamically change and reload the models so we could adapt the parameters of the calculation depending on the available power of the hardware?
AI Gamechangers: If you’re adapting and shrinking the model so that it can run on the player’s hardware, are you still able to create the experience you want to give?
Alexandre Folliot: Extraordinarily, it didn’t change a lot! Topic detection is a good example; a classification problem. So we have a player asking an NPC something, and you want to check inside the game what the player said because it’s going to reveal things in the story depending on what the player said (for example, if the player says they finished a quest, then a trigger is launched, the quest is finished). We need topic detection to detect this kind of information, as well as lots of other things. It’s pretty important, and smaller models are performing almost exactly like large models if they’re trained only on topic detection. So we lose maybe 1% or 2% accuracy, but it’s a very small difference.
AI Gamechangers: Please talk us through the main products that X&Immersion makes.
Oriane Piedevache: We do software development for the video game industry. Basically we work with games that have NPCs in them, and the idea is to try and help the game producers to create NPCs faster, more easily, and with deeper content.
We have three different technologies, each on one specific vertical of this NPC. One would be generating some dialogue for the different NPCs. With dialogue, the use cases can be very, very wide. Imagine a game like Baldur’s Gate, and take an example of rolling a dice – whether you get a really good score or a bad score, you won’t have the same response from the NPC. It would be adapting in real-time to different possible dialogues within the game. The dialogue generation tool could also, for example, be about having a real-time interaction with an NPC.
We also have the voice synthesis part. So, for example, if you are [making] a game and want to use real and emotional voices, then you could use our voices. On the contrary, if you already work with voice actors to voice all of your games in English but don’t really have the time or budget to localise all of the game voices, then you can also use AI. The use cases and the product are different, but the whole technology is generating voice, either from a specific library or from pre-existing work with voice actors.
The third vertical, specific to NPCs, is facial and lip sync animation. In this case, you might say, “I’m an animator. I need to produce loads of lip sync and facial animation material for my game.” This tool can help to generate the whole sentence and the phonemes, either based on the text or on the voice. Then you just need to polish it and move the blend shapes a little bit: for example, if I want my mouth to be more open, because it’s going to be more of a surprised emotion. But the basics are going to be there and generated already. This can either be in pre-production, to help your teams create your animations, or during a real-time experience where you cannot redo all of your animations if you don’t know what the NPC is going to say, so you need a real-time tool to do it!
Those are the three different verticals. We’ve also been working on several other tools that we call accessibility tools. For example, we have a game helper, which Alex can tell you more about. Instead of being stuck and going on a website to check out how to access this lore or that secret cave, [the studio] can drag and drop the game documents into the tool and then provide this tool within the game. Then, the player would just ask it, “What is the fastest way to get to this cave?” It would read inside the document and generate a specific answer.
AI Gamechangers: What are the foundation models you’re working with?
Alexandre Folliot: If you’re using third-party APIs, the problem is that it’s going to cost you, and people are also afraid with AI that if the company you’re making calls to is a third-party company, it could close, and then the game would not work any more! I think many [studios] are afraid of that, so that’s why we wanted to have all models we could run locally at the player side or on the server side from the game companies.
In the foundation models, we started with open-source architectures. We noticed that it wouldn’t be possible to use them [immediately] on just any computer, mobile or console. So, we had to change the architecture and the way the models work to access information and how we load the models. We didn’t reinvent the wheel, but we changed the shape of the wheel a lot to make it lighter!
AI Gamechangers: The other thing to discuss is ethics. The use of data concerns people, AI voices... but also people are worried about their jobs. What’s your stance, and how do you answer your clients’ questions about it?
Oriane Piedevache: Ethics is really something that we have at heart. We have an ethics charter that we sign with our clients just to be sure we’re on the same page.
It’s sometimes the other way around; people might say, “We don’t care, we just want to use an AI voice!” And we say, “No, we need to get it structured!” But mostly it’s the other way around.
“We are in a position where we have tools that cannot work without the human beings. Basically, it's like Adobe. It's a tool. You can do great stuff with it if you use it properly!”
Oriane Piedevache
We are in a position where we have tools that cannot work without the human beings. Basically, it’s like Adobe. It’s a tool. You can do great stuff with it if you use it properly! If you don’t use it, it will just stay on your computer, and no one will have the use of it, and it won’t harm anyone.
We are here to help. That’s been our argument since the beginning. We just want to help teams to optimise their production. For example, if you’re a small team, you maybe cannot use voice acting in your game. Well, now you have the possibility to do this. Honestly, if you want to use it, then know that we are prepared to use it properly. For example, when we do voice cloning, the voices that we keep are ultra-specific for this studio. We’re not going to mix it with another studio. The same goes for training the text part. If we take your game development document and train it for a particular reason, then we’re not going to use that data for other studios.
When we do voice cloning with a client that wants the specific voice from a specific voice actor that they already work with, we take into consideration that we are only a tech company; we reassure people, and we make them understand we only train the voices, and that use is only for their studio, and we’re not going to reuse it. They are the ones who use it how they like.
Alexandre Folliot: The other part is jobs. There were a lot of layoffs in the video game industry – but I don’t think it was really due to AI coming. It’s more like the financial crisis, post-COVID! It’s usual to hear, “AI is taking our jobs,” but I haven’t seen one report confirming that.
[AI] is changing the description of a lot of jobs in the games industry. I think it’s a natural phase where there are new skill sets and new possibilities for making games.
What we see at this point is that studios are limiting themselves and saying, “We’re not going to make a game with human characters because it’s too expensive to make facial animation – we will use cubes or animals instead.” Or maybe, “We are not going to add voices to this game because it’s too expensive,” or, “We’re not going to do localisation for this game because it’s too expensive.”
We have the impression that AI is going to do two things: it’s going to make games more accessible and optimise the game to make it faster. In the end, if a studio is producing games fast, it’s going to secure the jobs of the people working in it. That’s my opinion.
AI Gamechangers: Please give us a little more detail about the “game helper” and your philosophy for that product.
Alexandre Folliot: We have the technology to make a sort of “ChatGPT” inside a game.
In ChatGPT, there’s this great feature where you can upload a document and ask it to find some interesting information about it. It’s quite useful. In games, it would also be interesting to ask questions for various reasons.
For example, if you are blocked inside the game, if you need some extra help on the tutorial, if you want to be guided. This information will be in a company’s game design document. We can upload this document, and then the players could write questions, or have an avatar help them, or just chat – the large language model can look inside this document and find the most suitable information, and rephrase it as text.
But it could be more than just a game design document. It could also be a résumé of the player’s last game! For example, if they played two months ago and they’re coming back, you will have the logs of the player; you use that, extract the information, and say, “Okay, in your last game, you did this and that. The current quest is this. Don’t forget you need to do that…” It will keep the player engaged.
You can do a lot of things, and we can enable these capacities of large language models inside games, so game studios can decide what they want to do with it.
Oriane Piedevache: We’re really trying to put forward our “tech consulting” face. We’ve been making loads of tools and software-as-a-service, they’re made, and people use them. Of course, there is a huge phase of customisation, but it’s done pretty easily. What we’ve been doing with this R&D part for a couple of months now is to help games studios on the technical side.
For example, some big companies don’t really know how AI works; they don’t have engineers or technical teams, but they do have ideas of what they want to produce within the games. So we come in almost as a co-production team, a technical service. We say, “We’re working with your teams for one, two or three months; we have the same milestones, and we want to integrate everything within the game.” Lots of companies have ideas but don’t really have technical expertise. This is what we’re trying to implement more and more.
AI Gamechangers: Do you find there are companies that don’t realise what’s possible yet?
Alexandre Folliot: Yes, but also, you have people who are really optimistic, and sometimes [they face] restrictions. Especially if they want to run everything locally, there are restrictions: size, for example. If you have a huge game design document and you want to make an analysis of it on your mobile phone, it’s going to take quite some time.
We are also there to counsel and to give ideas about how something can be implemented, depending on the capabilities of the hardware. There are a lot of companies who know that AI is coming, but it’s vague; it’s quite difficult sometimes to explain what we can achieve, to explain that not everything is possible.
AI Gamechangers: What does the future look like for games?
Alexandre Folliot: As a writer or a narrative designer, I think the best way to use AI is for secondary characters in the game or to create a lot of content. For the ideas about the characters, and the stories, it’ll still be human-driven. We have seen that the large language models are not so creative; they don’t “think out of the box” – you need to give them the context and information you want. So, I don’t think completely AI-driven narrative games are going to be the future in the short or medium term because the quality is not there. I don’t think people want games like this.
But we are working on games where you’re negotiating with a person on the phone, and [the conversation] is completely AI-driven, but the different topics you can talk about are scripted. You have a mixture of being able to say anything you want (and the character responding to what you say), and on top of that, you have a narrative layer of things that are scripted to happen. If I’m reassuring the NPC and the AI notices that in its topic detection, then they’re going to respond calmly, and one thing’s going to happen, but if I insult him, something else is going to happen.
“AI is going to do two things: it's going to make games more accessible and optimise the game to make it faster. In the end, if a studio is producing games fast, it's going to secure the jobs of the people working in it”
Alexandre Folliot
It’s challenging to write these kinds of games, but there are some applications where it definitely makes sense and other games where it doesn’t. We’re going to have more of these kinds of discussion features, like what we had 20 years ago with text-based RPG games.
AI Gamechangers: What’s next for X&Immersion?
Oriane Piedevache: We’re starting a new roadshow for our fundraising, and as Alex said, we still have some technical challenges with the local inference models.
Alexandre Folliot: The local inference models work for PCs on Nvidia today, but you need the consoles and the mobile phones as well, so we are expecting to implement them for this hardware as well as next year.
Oriane Piedevache: That’s a huge R&D step, and it’s why we’re raising funds. In one year, we did 50% of the market. Now we need to do the other hardware.
As Alex said, we’ve had a couple of games that went out with our technology this year – we were in Hot Lap Racing, which is a car game from Zero Games Studios. We were in Machinika: Atlas, which is a puzzle game from Littlefield Studio. We were supposed to be in the Amerzone remake from Microids in Paris, but it’s been postponed to early 2025; they’re using only our lip sync for all of their NPCs. It’s a very narrative game, so we’re going to see a lot of our tech, which is cool. There are others that have not been announced yet.
The idea now is to have more and more games using our tech and to accompany more studios with either the production tools or the service.
Further down the rabbit hole
The news never stops. Here’s your handy digest of what’s hot in the world of games and AI this week:
The World Health Organisation has partnered with free-to-play mobile farm game Goodville to launch an AI assistant called Sarah. Sarah is designed to help players improve their physical and mental well-being by offering tips via an in-game chat.
Midjourney has revealed Patchwork, an “infinite canvas” for creating and sharing fictional worlds. You can add characters, images, and text or ask the AI to fill empty spots for you, and then you can explore these ideas with the help of collaborators.
Games-focused investment fund Beam Ventures is launching in Abu Dhabi with $150m to support game studios and an accelerator programme dedicated to early-stage web3 gaming and AI start-ups.
Indie developer game hosting platform itch.io says it was taken down by Funko, makers of those collectable figures. Funko uses BrandShield, an AI-powered anti-counterfeiting tool, which apparently identified a fan page on itch.io as resembling the Funko Fusion game website and got the domain registrar to shut off the whole site temporarily. There’s a timeline and summary here.
Tickets are still available for next month’s PG Connects in London (20-21 January). The event features two tracks dedicated to AI and games, with speakers from Reforged Labs, Omdia, Charisma.ai, Layer AI, Filuta AI, iGP Games, Cosmic Lounge, modl.ai, and more. And one of the contributors is Alexandre Folliot from X&Immersion, interviewed above.
Supercell worked with Google DeepMind on a Gemini 2.0-based assistant that watches you play games and offers real-time advice. Here’s one of the videos of that in action, as a player asks the AI agent to watch them play Squad Busters and remind them of things they need to do: