This is a humble and modest attempt to support—and even rehabilitate—a project that was largely misunderstood, at least in the West.

Yet the game in question, along with all the work surrounding it, appears to me to be an invaluable synthesis tool on the broad subject of AI applied to video games. I also see it as an ideal starting point to address that lingering question that haunts every video game developer—and even more so the storytellers in their studios: Now that generative artificial intelligence is here, what on earth can we do with it?

There is a book

My last summer in Japan wasn’t just an opportunity to attend an amazing concert and expand (once again) my absurd collection of Resident Evil strategy guides and books—it was also the chance to acquire an indispensable volume, a collection of studies and presentations exploring both the achievements and possibilities of applying artificial intelligence in video game development. This entirely Japanese booklet, simply titled “Square Enix no AI,” (Square Enix’s AI) gathers the top executives from the legendary Japanese studio dedicated to AI, led by Mr. Youichiro Miyake, General Manager of the studio’s AI department.

Naturally, I immediately turned my attention to Mr. Yusuke Mori’s work—not so much because his article aligns with my research on AI for storytelling, but because I had the chance to meet him personally at GDC in San Francisco a few months earlier. I still clearly recall his questions after testing a brief AI conversation in our game Cloudborn, which was then showcased at the Inworld booth:

“How do you plan on extracting information, clues that can help advance the game, through conversations with your AI NPCs?” he asked.

I remember replying that this was already in place, through a feature called “Goals and Actions.” While my answer was not untruthful, I sensed that Mr. Yusuke was hoping for something more organic, more natural than a simple trigger based on roughly similar words. As we shall see, that feature wasn’t exactly revolutionary compared to everything Mr. Yusuke had developed at Square Enix. Something was clearly missing—his immediately unsatisfied reaction said it all.

At that moment, only Mr. Yusuke’s badge identified his name and position, and modestly he didn’t elaborate on his work or his research at Portopia. It was only months later, once I had the book in hand, that I recognized Mr. Yusuke and pieced everything together. I then set to work translating his article (greatly aided by ChatGPT-4, then 4o) and reinstalled Portopia on my work computer for an in-depth study of the subject—a subject I will modestly attempt to synthesize here, enriched with my own insights and comments.

Nothing Is Ready, Let’s Not Waste Time

To best explain the paradox underlying Mr. Yusuke’s article, one must keep in mind one very important fact. The game Portopia, in which you converse with characters during a criminal investigation through completely free text input, operates 100% locally. To be sure of this, I even ran the game on my laptop in airplane mode, no wifi, and it did not affect the gameplay experience. For further context, Portopia was developed before ChatGPT-3 and its open API even existed. The game therefore runs various technologies, such as models for NLU (Natural Language Understanding) and Automatic Speech Recognition, or STT (Speech To Text).

For Mr. Yusuke—and he emphasizes this repeatedly—it is unthinkable to consider integrating AI in a video game in the long term if inference cannot be executed locally. He states, with authority (and rightly so), that one cannot imagine offering a game experience that might be interrupted not only by poor network connectivity but also by slowdowns on an LLM server.

He predicts that in the near future, everything will be managed locally, though he expresses some doubts about how soon that possibility will be realized, especially given the enormous resources required for such operations on the currently insufficient technical infrastructure.

I found nothing to dispute in that assertion. It reminds me of long discussions with Inworld during our early collaboration weeks, where token pricing fluctuated week by week, leaving us with the unresolved question: Who should pay for the generated tokens? The developer? The publisher? The player, via a monetization scheme that is more or less cynical? Should it be tied to a cryptocurrency integrated into the game’s economy? As if the player hadn’t already been subjected to such dubious practices lately, AI in these conversations seemed to add yet another layer of monetary speculation.

On that note, I won’t beat around the bush: I’m convinced this is only a temporary problem. My feeling on the matter is exactly the same as when the internet first arrived in our family apartment in Lyon in September 1998. I was 14 then. I will always remember that our first internet plan—signed up after seeing an ad on TV—offered 20 hours per month for 99 francs, roughly 16 dollars or euros. I also remember that it was through that plan that I not so legally downloaded, from a little-known site, my first mp3: “My Favourite Game” by the Cardiagns. That download took about an hour and a half.

All this to say that back then, every minute online was as precious as hot, clear water is today. I wager that AI tokens will follow the same trajectory. In my view, they should eventually be integrated into internet or mobile data plans. Let’s hope this message reaches internet providers, who these days seem capable of little more than bundling Netflix or Disney+ into their packages—or offering apps to watch TV.

Regarding local inference, it’s worth noting that considerable efforts are underway as I write these lines. This is a central research focus at Inworld, as they confided to me. In that regard, their partnership with Nvidia makes perfect sense. All these efforts, driven by accomplished professionals, will inevitably lead to a breakthrough—unless, despite Nvidia’s dominance and its Californian satellite startups, a band of indomitable Gauls has already managed to achieve this feat.

I can personally attest that I’ve tested a game running on Unreal Engine where you can converse freely with an AI character—all locally on a mid-range gaming PC. This took place in the Paris offices of X&Immersion in January 2025, just for the record.

Local inference now seems within reach, as yes, the French team at X&Immersion has achieved it. But until such technology is widely adopted, I return to Mr. Yusuke’s point, with which I completely agree.

As Game Designers, Narrative Designers, Engineers, or Producers, we should not wait for this technology to become mainstream with the next Nvidia chips, or hypothetical PlayStation 6s, iPhone 19s, or Nintendo Switch 3s. The time to think is now—to put pen to paper with concepts and diagrams that can be deployed once these technologies are democratized.

Alchemy of the Word

Now, let’s delve into what I consider the most virtuosic part of Mr. Yusuke’s article: his accessible explanations for a simple designer like me on how the magic of words works when processed through NLP (Natural Language Processing, the precursor to LLMs). I want to emphasize that the idea of breaking down words and their meanings into numerical data and then reassembling them into new words seems like pure alchemy. At least, that’s how I perceive it.

I hear the scientific discourse that attempts to justify these phenomena as mere mathematical acrobatics, and to those, I say I’d like to continue seeing a bit of poetry in the process—transforming words into numbers and numbers back into words, like turning oxygen into gold.

I’m sure they’ll agree with me on that.

For Mr. Yusuke, everything starts with a childhood dream—a dream imbued with its own kind of poetry. In the same article, he confesses that as a child he dreamed of building a robot inspired by Doraemon, the robotic cat from the future. His ideal robot was one that could build, create, generate stories. That innocent motivation set him on a long, disciplined journey, eventually leading him to earn a doctorate in Information and Communication Engineering.

Mr. Yusuke then describes his early interest, as far back as 2013, in word2vec—a technology that converts words into vectors. It brings back memories of my high school (or junior high?) vector classes. I have a faint recollection, tinged with regret for not paying more attention. I also recall my 8th-grade French teacher always saying, “French is math, and math is French.” The same teacher who canceled her wedding that year to marry the math teacher from our school. But I digress.

Once again, can we all agree for a moment that converting words and their meanings into digital data operable by a computer is one of the major scientific breakthroughs? I fully share Mr. Yusuke’s awe when he points out that word2vec—converting each word into an identification number, an index, and a trajectory that allows for word calculations such as addition or subtraction—is absolutely fascinating.

Then, suddenly, there are other aspects of his explanation that evoke in me a certain poetic emotion. For example, he mentions that in Portopia he was able to use “semantic vectors at the sentence level.” Later he clarifies: “When the number of words is n, then the vector has n dimensions.” And while he elaborates on word affinities—highlighting the closeness between “chat” and “kitten” and even “kotatsu” (which might sound absurd, but you have to read Mr. Yusuke to understand)—I can’t help but be reminded of another childhood memory: the TV show “The $10.000 Pyramid”.

A bit of context for the younger readers: there was a time when public television featured intellectually stimulating games—competitions where contestants had to put their minds to work. “The $10.000 Pyramid” was one such show, although I must admit that at age 10, I didn’t understand a thing. Today, however, I’m convinced that had someone explained the concept to me, I would have understood from an early age that the game’s goal was to guess a hidden word by offering others through association—whether by closeness or as a mirror image, depending on the strategy of the contestant or host.

It might sound odd, but that’s exactly how I visualize Mr. Yusuke’s NLP working to generate words for the player. By explaining that the next word is chosen based on a probability between 0 and 1—a word with a score of 0.7 being selected over one with 0.5—I can’t help but draw an analogy with that TV game, which, upon rewatching, struck me as a game of undeniable quality.

Playing Portopia

So, what about Portopia itself?

First, a bit of context regarding gameplay conditions. I played Portopia via Steam, where it’s available for free, in English (it’s also available in Japanese). I ran the game on my work PC—an ASUS ROG STRIX 17 Gaming Laptop with 16GB DDR4 RAM. By default, the game used just over 2GB of memory at launch, likely allocated to managing the DiabloGPT model, as you can see from the screenshots, the game’s graphical requirements are nearly non-existent.

The game ran in airplane mode with WiFi disabled, meaning there was absolutely no way for the software to make API calls to an external model. Mr. Yusuke’s initial promise, therefore, holds true.

It should also be noted that this title is a remake of a game originally released on the NES and other systems by Enix in 1983, under the title The Portopia Serial Murder Case.

Goodness, this game is even older than I am.

At the start of the game, a partner informs us of a recent murder before providing us with a list of actions they claim to be capable of performing:

Gather information
Check alibis
Investigate the scene
Search for suspects

What’s appealing right from the start is that you can ask the assistant to investigate the scene—not by clicking through multiple-choice buttons, but simply by typing your request into the designated dialogue box. This unlocks new suspects to interrogate and access to the crime scene.

However, one drawback is that this is the only action available in this opening scene; the other actions only become executable later, under different circumstances. From the outset, there is a UX issue that unfortunately persists throughout the test.

Relying solely on the assistant as the source of actionable information is completely logical for a narrative game, especially one that aims to put AI at the heart of its gameplay to provide the player with unprecedented freedom. It’s great that you can, for example, type “let’s go to the port” to set off, but that freedom becomes frustrating when you have to guess the exact phrase (in this case, “go back”) to end a conversation with a suspect.

There is a mode you can activate via the “pause” key—the “NLU Visualizer” (Natural Language Understanding). This mode displays, in real time, the probability score of each of the player’s inputs and shows how closely it matches an expected input that would trigger an action. While this mode is handy in guiding you, it also exposes the system’s limitations. I would have preferred if the UX—or even the way the dialogue is written—naturally guided us without needing such explicit clues.

This made me realize that UI and UX will be critical in the adoption of AI for narrative purposes in future games; without them, I fear the revolution may never take place.

Moreover, the NLU displays expected phrase probabilities that don’t match the context or timing of the game, revealing information meant to be discovered later and thereby breaking immersion. A filtering system would have been much appreciated.

Ultimately, these are just the quirks of the system, and the enjoyment of playing eventually sets in as progress becomes evident. I also learned how to question suspects about other suspects, always using the NLU to guide my prompts. However, the system has its limits. For instance, I couldn’t get it to show objects to suspects to provoke a reaction.

Fortunately, visual cues in the environments sometimes point us in the right direction, but here the NLP struggles to make the connection. For example, it rejected “look on the floor” and expected “look on the ground.” The same happened with “Investigate Pendant” versus “Investigate the pendant.”

I risk repeating myself, but the involvement of a UX Designer—someone who truly understands what Mr. Yusuke aimed to achieve—could have given this title the full potential it initially sought to unleash. He wanted to revive an old-fashioned “command input style” game (he even mentions the game Eliza), and my conclusion is that while he undoubtedly made significant progress, the AI’s effect remains very inconsistent. Sometimes, it provides the pleasure of a natural conversation; other times, it shatters the charm by rejecting a simple typo or an unexpected grammatical variation.

Let’s not kid ourselves. Portopia is a giant leap forward, even if it isn’t immediately apparent. It may have a quirky reputation, but it has smartly integrated AI into its gameplay to enhance immersion—an impressive achievement at a time when the entire industry is struggling to incorporate LLMs beyond the basic “chatbot” functionality. And I say this with firsthand knowledge: it’s a colossal challenge.

Portopia is that first step, that foundational stone upon which we should build our future. It urges us to start thinking now—using more modest models—rather than waiting for ever more powerful ones. We must learn to harness these technologies gradually, while keeping logical minds and proven design concepts in play. In my view, Portopia should be examined further by everyone determined to integrate AI into their game projects.

Fiction Versus Reality

Reflecting on this test, I tried to imagine the phases of thought Mr. Yusuke must have undergone during the development of his title, while cross-referencing my notes from reading his article. What struck me first is that his reflections are far ahead of what he was able to implement in Portopia.

I can only agree with him when he mentions the potential conflict between real-world data—on which models are trained—and the fictional world devised by the developer. That was exactly my observation during my early tests on Cloudborn, although the tests became far more conclusive over time.

He also points out that the concept of being a “monster” in an RPG could be misinterpreted in a real-world context. Could a model as robust as OpenAI’s GPT—often labeled “politically correct”—accurately portray the dialogue of a villain with morally ambiguous views? Do we really imagine an RPG villain spewing entirely reprehensible lines that would be unacceptable in the real world?

To address this issue, he suggests feeding the fictional world with ample data so that the system has a solid foundation to rely on, rather than falling back on its real-world knowledge. Finally, he warns of NLP hallucinations, which could result in glaring bugs that undermine the system’s credibility. Reading his words, I find my own reflections in agreement with his, envisioning the LLM as a monster to be tamed—trained with a vast amount of narrative data so that, at the very least, it anchors itself to external data and avoids producing unwanted outputs.

This is exactly the principle behind Aarda AI, my new AI project at Chromaway, which I am currently developing with Kayna Oliveira. Aarda AI is first and foremost a worldbuilding tool that requires one thing from its user before anything else: to provide as much information as possible about the world being created, and only then put that world to the test with real-time chat applications, gameplay implications, etc. In reality, I can’t see how integrating these tools could work with a workflow other than this one—which is why this project is already showing promise. You can imagine my delight in reading Mr. Yusuke’s own words that echoed this principle, affirming my ideas.

According to its creator, Portopia is already facing an uphill battle. Mr. Yusuke states that his writing for the book entry was done while ChatGPT-3 was saturating information channels worldwide. I sense a note of regret in his words when reading between the lines. But we are not in today’s absurd scenario, where every day seems to announce the latest model from some firm or country that outclasses all the others. This dramatized obsolescence, increasingly noisy, does nothing to detract from—and indeed, is a testament to—the pioneering thoughts and efforts of the visionary behind it.

In conclusion—and at the risk of repeating myself—I will paraphrase Mr. Yusuke once more, encouraging every Designer reading these lines to conceive and develop systems that can tame this new technology rather than getting caught up in the race for raw power. Let the competition be for those concerned with that; it is up to Game Designers, Narrative Designers, and UI/UX Designers to envision the applications of tomorrow. And because Portopia never pretended to be anything other than what its title declares—a “Tech Preview”—may this experience inspire us and mark the starting point for new experiments and bold risks.