On March 15, an account named “404Stella” published a post on X (formerly Twitter), comprising a few short paragraphs in English and a one-minute video introducing a new game titled Whispers from the Star.
There was no teaser, no promo push, and the timing—a Saturday—seemed designed to avoid attention. It worked. Ten days later, the post had barely cleared 20,000 impressions.
But in China, the reaction was the polar opposite. From the moment the post went up, Whispers from the Star became a hot topic in the gaming world. The buzz extended far beyond the game itself—discussions ranged from technical deep dives to predictions for the future. Even biographies of the game’s developer made headline rounds on Weixin (known internationally as WeChat) and beyond.
The person behind the 404Stella account? Cai Haoyu—often called “the father of Genshin Impact”—and arguably the main reason for the frenzy unfolding back home.
As co-founder and CEO of Mihoyo, Cai is widely regarded as the creative force behind the hit game Genshin Impact, and a fixture on Hurun’s “Global Rich List” with a net worth of RMB 73 billion (USD 10.2 billion). Cai was once the undisputed wunderkind of China’s gaming industry. But in September 2023, he abruptly stepped down as chairman of Mihoyo and faded from the public eye. A year later, he reappeared in California to launch a new venture: Anuttacon.
Whispers from the Star marks Cai’s comeback after that retreat.
Cai’s pivot reflects not only his longstanding belief that “tech otakus [can] save the world,” but also hints at a bolder ambition: to ring in the dawn of gaming’s post-metaverse era.
Tools like DeepSeek and Manus, alongside the continued evolution of large language models (LLMs), are laying the groundwork. Ever since the apparent collapse of the metaverse bubble, one particular interpretation of Web3 is beginning to crystallize.
Optimists believe that after two years of trials and tribulations, 2025 could mark a turning point—when artificial intelligence in gaming moves from a tool for asking questions to one that can finally deliver answers.
But while some questions are being addressed, many more remain unresolved.
Two particularly thorny issues, namely technically implementation and business model viability, still feel like a vice grip choking the gaming industry. The emergence of Whispers from the Star may suggest that Cai is now venturing deep into the uncharted terrain of AI.
A rehearsal of generative AI capabilities
Through the 404Stella account, Cai offered a glimpse of his new game. Set on a distant planet called Gaia, the story follows Stella, an astronaut who crash-lands with no way to call for help—except the player. As her sole point of contact, players must guide her through voice interactions in real time as she searches for a way off the planet.
The video shows snippets of this narrative while showcasing Stella’s movements across various environments on Gaia. Her behavior shifts with emotional nuance—joy, anger, confusion, wariness.
But the most eye-catching tag tied to this game? Voice interactions via agentic AI.
The AI is doing two heavy lifts here: first, it generates Stella’s dialogue in real time, dynamically responding to player input with emotional and physical cues. Second, it’s responsible for rendering the game’s visuals in real time, promising players a more immersive experience.
Gamers may find the setup eerily familiar. That’s because Whispers from the Star borrows heavily from Lifeline, a mobile game released in 2015. Even the game trailer for Whispers from the Star features the tagline: “Your words seal her fate.”
Like Lifeline, the new game centers on an astronaut stranded on an alien world, with the player acting as their lifeline. The big shift lies in how that communication happens: Lifeline uses text, Whispers from the Star uses real-time voice.
But AI introduces a new layer: imagination. With AI at the wheel, Stella becomes a dynamic, deeply personalized character, potentially breaking free from the narrative constraints written by screenwriters.
Still, a well-known paradox haunts the genre, whereby the more freedom is offered, the more limited it can feel. Narrative freedom can easily translate into narrative incoherence. Even with AI, a game’s storyline can only evolve in two directions:
- Total freedom, whereby the game has no prescribed narrative, and everything is discovered by the player on the fly.
- Pseudo-freedom, which requires a narrative to be prestructured into many small nodes, giving the illusion of choice while retaining authorial control.
AI Dungeon is a prime example of the former. Players interact with the game via text prompts, and the AI generates storylines based on user inputs. Each playthrough is unique, shaped entirely by the player’s imagination and the AI’s generative logic.
On the flip side, The Invisible Guardian exemplifies the latter. The game seems to allow free decision-making, but those choices always funnel back to outcomes already determined by the developers.
Whispers from the Star might end up falling short on both fronts. Whether it turns into a 3D version of AI Dungeon or an AI-enhanced Invisible Guardian, it risks disappointing players hoping for true AI-driven storytelling—new wine in an old bottle. Some online critics have already written it off as little more than “an AI version of My Talking Tom.”
So is this a case of Cai playing it safe? Or is it that his ambition lies elsewhere?
The answer might be hidden in plain sight.
The trailer offers some in-game interface details—Stella’s heart rate, body temperature, signal strength—as well as an elegant use of “communication latency” to mask the time lag of AI response.
But more striking than these interface flourishes is the deliberate choice of settings and emotions in the trailer.
Six environments are shown: a spacecraft cabin in daylight, wilderness at night, a canyon at night, the same canyon during the day, an open field during the day, and a rocket launch site also in daylight.
From the claustrophobic cabin to the vast openness of the canyon, from hesitant first contact to indignant confrontation—this wasn’t just a trailer showcasing story and gameplay. It was a rehearsal of AI’s generative muscle.
Each scene pushes the AI’s limits in lighting and acoustics. In the spacecraft, flickering orange warning lights flash on the left, while the right is bathed in a gradient of blue light. Sunlight glows in the background. The room feels like a real physical space, filled with accurate ambient sounds.
In the canyon, spatial audio mimics echoing voids as Stella moves forward. Light and shadow play dynamically across her face with each step.
Her animations are impressively smooth. In one daylight canyon scene, when she snaps, “You finally decided to pick up seriously?” her face contorts through subtle changes, and her hands gesture emphatically.
Even her voice borders on uncanny realism. There’s no digital flatness—her tone wavers slightly with anger, rises in pitch with awe as she watches a rocket soar.
It’s uncannily lifelike, and so much so that it may be hard for some to believe it’s AI. But Whispers from the Star isn’t the first attempt to embed this tech in games. In the 2024 demo of detective title Covert Protocol, players could interrogate AI-powered NPCs to solve cases.
The result? A deep dive into the uncanny valley. The characters had recognizable facial features, but they were frozen in stiffness. The voices were robotic and bland—like a horror game pretending not to be one.
If Whispers from the Star truly delivers what the trailer shows without post-production polish, then it could be the most integrated use of AI generation to date—spanning text, sound, visuals, lighting, and motion.
At its core, however, the game remains conservative in structure. It’s not trying to reinvent gameplay, instead using it as a vessel to trial the latest AI tech.
Where business meets technological limits
When Cai departed from Mihoyo, a statement was released indicating his wish to better support the company’s future direction by focusing more on frontier technologies, new projects, and bridging domestic and global R&D resources. A year of silence followed—until Anuttacon stepped into the spotlight.
According to its website, Anuttacon is on a mission “to create immersive, personalized experiences that adapt to each player’s unique style and preferences, fostering genuine connections and deeper understanding between players and the virtual worlds they inhabit.”
This echoes Cai’s own previously stated vision of creating a virtual world that one billion people will want to live in by 2030.
Only time will tell, but Anuttacon is seemingly not just about games—or at least not games purely for entertainment. Its vision highlights ambition to branch into living, breathing digital ecosystems for human habitation. In this vision, players, developers, and creators are all, figuratively, “worldbuilders.”
Reportedly, about half of Anuttacon’s team hails from Mihoyo. The rest come from top-tier studios. Mihoyo’s original AI team, known internally as “Inverse Entropy,” may have been spun off and absorbed into Cai’s new company. This tight linkage with Mihoyo, coupled with significant investment in AI, has led some in the industry to dub Anuttacon “the AI version of Mihoyo.”
But why go to such lengths to start from scratch?
The answer lies in the limitations of the existing commercial gaming ecosystem. The idea of integrating AI and gaming offers much opportunity for fanfare, but turning that buzz into a scalable, profitable reality is a different beast. The uncomfortable truth is: the more successful an AI game becomes, the more likely it is to fail—financially.
Consider Death by AI, a survival-themed social game developed by US-based AI studio Little Umbrella. Its signature feature: an AI game master that dynamically generates deadly scenarios, with players collaborating via text to influence outcomes. The AI ultimately decides who lives or dies.
The game drew 20 million players in just two months. By all appearances, it was a breakout hit.
But behind the scenes, it was a financial nightmare. As its user base exploded, the studio’s daily API costs skyrocketed from USD 5,000 to USD 250,000. At one point, the company teetered on the edge of bankruptcy. The studio’s tech director later revealed that new user signups had to be capped just to stay afloat.
In short, adding AI to a traditional gaming business model makes little financial sense right now. And if AI development happens within the constraints of existing models, then every creative choice will inevitably be driven by one metric: return on investment.
That would be a death knell for any genuine attempt to push the boundaries of AI. Take this into consideration, and Cai’s decision to carve out AI from Mihoyo starts making sense. He’s choosing to burn money in the short term in hopes of long-term breakthroughs. In his view, that’s the only way to build truly native AI games from the ground up.
And it’s going to be cutthroat. No matter what form Whispers from the Star ultimately takes, it will depend on the heavy lifting of AI compute and data transfer.
Tokens are the atomic units of text processing for LLMs. Since last year, token costs have dropped by roughly 90%. Some predict another 90% drop in the coming year, which could finally bring AI gaming within economic reach.
But Nvidia CEO Jensen Huang has a different take. He argues that large models are evolving from fast-thinking approaches toward slow-thinking systems like OpenAI’s o1 or DeepSeek-R1—in other words, models that build reasoning step by step.
This shift will dramatically increase token usage. According to Huang, slow-thinking models can consume up to ten times more tokens than fast-thinking ones, with another tenfold increase in computational speed needed to maintain responsiveness and interactivity. Taken together, that could represent a 100-fold leap in compute demand.
It’s a daunting equation. But Cai’s challenges don’t stop there.
The trailer for Whispers from the Star hints repeatedly that Stella’s responses aren’t scripted in advance—they vary based on user input. Words, gestures, ambient lighting, environment—are all dynamically generated. That’s an immense technical load, with enormous demands on bandwidth and processing power.
This raises another thorny issue: local versus cloud-based computing. Relying solely on the player’s hardware would raise the system requirements and shrink the game’s potential audience. But leaning on the cloud invites latency issues and spiraling infrastructure costs.
Cai’s vision—a virtual world for a billion people—inevitably leans on centralized infrastructure. And if it’s his world, then he’s the one tasked with making it real.
Back in August 2024, he shared a series of provocative takes on AI’s role in game development, which boiled down to this:
- AI has already reshaped how games are made, though its full impact will take time to unfold.
- In the future, only two groups will matter in game creation: a small circle of visionaries capable of imagining unprecedented experiences, and a broad base of hobbyists making games for fun.
- Everyone in between—amateurs to professionals—may need to rethink their careers, as AI renders their roles increasingly obsolete.
Cai clearly counts himself among the visionaries. And for the rest of the industry, Whispers from the Star is both game and experiment—an attempt to see whether AI can shift from backstage assistant to lead storyteller.
Whether it delivers or not, the attempt may sketch the outline of what’s next.
From the metaverse to something else entirely
Back in October 2021, at the annual Facebook Connect developers’ conference, Mark Zuckerberg made headlines by announcing Facebook’s rebranding to Meta—a pivot aimed at building the metaverse.
It was a utopian vision of Web3. “The defining quality of the metaverse will be a feeling of presence,” Zuckerberg said, “like you are right there with another person or in another place. Feeling truly present with another person is the ultimate dream of social technology.”
Fast forward three years, and Meta’s grand bet on the metaverse has turned into a money pit. In just one quarter, its Reality Labs division reported USD 4.5 billion in losses. Since its inception, the unit has racked up over USD 46 billion in losses (as of August 2024).
In today’s AI-powered computing boom, two paths are taking shape. Virtual reality looks to enhance the medium—upgrading how users sense and interact with digital spaces. AI, by contrast, is reshaping the content itself. The metaverse bubble may have popped, but figures like Cai haven’t let go of the push toward a new kind of internet.
The challenge is how to ground these expansive, generative experiences in technical and commercial reality. How do you move beyond Genshin Impact-style interactivity to something genuinely new? What can Web3 eventually look like? No one really has the answer yet.
As computer and smartphone makers rush to add memory and storage in anticipation of AI’s heavier demands, the path to realizing Cai’s dream still looks long and uncertain.
But if Whispers from the Star is any indication, Cai isn’t just waiting for the future to arrive. He’s trying to build it.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Liu Chongjiang for 36Kr.