Two and a half years later, Cai Haoyu and his Silicon Valley startup, Anuttacon, are quietly undergoing a major transformation.
On April 10, Anuttacon team memberAilingZeng announced a “Large Performance Model” named LPM 1.0. This 43-page paper, authored by 25 contributors, was quickly interpreted by the outside world as a阶段性 technical breakthrough by Anuttacon, sparking widespread speculation about its next steps.

But through interviews with multiple internal developers, I learned that the development of this model has been halted. LPM 1.0 is not an ongoing core project, but rather a consolidated report by the video team on their work over the past year—both an external showcase and an internal review. The video team is led by “Tong Lao” (former Chief Researcher at Microsoft Research Asia, Tong Xin), with AilingZeng as Tech Lead, and nearly half of the authors come fromAnuttacon internally; Cai Haoyu did not directly participate in the model’s development.

Meanwhile, Anuttacon is undergoing a significant reallocation of computing power and organizational structure. Previously, the team’s computing resources were distributed across LLM, video, and audio in a rough ratio of 4:3:2, which made it difficult for any area to achieve breakthroughs; under the new assessment, the company has shifted almost all of its computing power toward LLM, adjusting the current ratio to nearly 9:1, concentrating the vast majority of resources on its core focus.
In conjunction, the LLM-related teams have fully returned to China and entered a centralized operational mode. In North America, only minimal audio and video capabilities are retained: some members support technical validation for existing projects, while others continue advancing iterations on new research directions.
This shift is also becoming evident in key personnel movements. On April 30, as news broke of Jin Wenyi, MiHoYo’s head of internationalization, stepping down, the whereabouts of her successor—Wang Yuyang, a former core member of Anuttacon—also came to light.Wang Yuyang previously served as Vice President and Head of Games and Live Streaming at Bilibili, and joined Anuttacon in 2024 as President of User Ecology. Recently, he quietly transitioned into the MiHoYo ecosystem to oversee the global release and internationalization of Varsapura.

Source: Quantum Bit screenshot (2024.8)

Source: Latest LinkedIn screenshot by Ray Wang
To understand this shift, we need to go back a year.
Last year, in our exclusive piece, Leaving miHoYo, Betting on AI: Cai Haoyu’s Ambition, we learned that within Anuttacon, its first AI narrative experiment centered on voice interaction, Whispers from the Star
From the beginning, Anuttacon’s goal was not to build a traditional product,but to create a technology proving ground.Cai Haoyu was more concerned with whether this AI semantic system could be abstracted into foundational modules and become an AI narrative engine accessible to developers.
But this direction quickly reached its limit, ending with the launch of Whispers. The limitations of the Post-Train approach became increasingly apparent, and with the emergence of Claude Code and the Agent paradigm, Cai Haoyu’s interest shifted rapidly—he found a new passion: building a “super agent.”
Since the end of last year, Cai Haoyu has gradually stepped back from the day-to-day operations of Anuttacon’s audio and video teams, focusing his efforts on large language models and aggressively recruiting top talent in China, poaching key members from Tier 0 model vendors. With the shutdown of the North American LLM team, the remaining personnel have been reintegrated into miHoYo’s existing roadmap, and Cai Haoyu has fully shifted his R&D focus back to China, a transition that is now essentially complete.
What exactly happened over the past half year? We won’t speculate—instead, we had a quick chat with Anuttacon’s team:What changes has Anuttacon undergone since the release of Whispers? And where is Cai Haoyu heading next?
First exploration: Building a model that acts like a human — Whispers from the star
Cai Haoyu's AI exploration began toward the end of Project SH, the precursor to Varsapura.
On November 24, 2022, PJSHwas temporarily halted.A HoYoverse employee told me,Cai Haoyuhad brought together the teams from China and North America in Singapore, attempting to reorient the project. But this effort lasted only “two or three months,” “they felt it wasn’t practical anymore, so they pulled the entire North American team.”
Someone close to the team recalled that at the time, "Cai Haoyu was uncertain and didn't know how to move the project forward," so in the first half of 2023, he largely stepped away from hands-on development and embarked on a trip to North America that lasted more than half a year, "possibly seeking inspiration and learning from others."
What exactly happened during those six months remains unknown to the outside world. “After returning from his trip, he realized he wanted to work on AI,” a former HoYoverse employee told me. In September 2023, Cai Haoyu officially stepped down as chairman of miHoYo, handing over the company to Liu Wei, and founded a new company, Anuttacon, in Silicon Valley, shifting his focus to the AI sector.

The broader context at the time was that,on November 30, 2022, OpenAI officially launched ChatGPT to the public. In the following year, 2023, large models rapidly gained widespread attention, which may have influenced Cai Haoyu’s judgment. A person close to Cai Haoyu said, “He has a fairly good relationship with Jensen Huang (NVIDIA’s CEO), and they sometimes get together for meals.”
Starting in September 2023, Cai Haoyu shifted his focus to Singapore, fully dedicating himself to the early establishment and directional exploration of Anuttacon.An engineer recalled that during the company’s early stages, “Brother Cai” reverted to a hands-on, omniscient management style, attending nearly every team meeting and continuously tracking progress across different teams, personally evaluating everything from model capabilities to product form.
At that time, Anuttacon consisted of three core teams: video, audio, and the largest LLM team, with around fifty people on-site.
The three teams have clearly defined roles:The LLM team handles post-trainingto make the model more closely resemble authentic human personality and expression;the voice team handles TTS (text-to-speech),converting text into voiced output with emotion and tone,with the challenge lying in achieving “emotionally expressive speech”—conveying joy, anger, sorrow, and happiness through emphasis alone;the video team, from the very beginning, drives character performance based on both text and audio inputs.Together, the three teams aim to build a complete, human-like interactive system tailored for gaming.

And "Whispers from the Star" functions more like a comprehensive testing ground—building and validating whether this system is clear and feasible in a real player environment: “Use a large language model as the brain, a speech model as the bridge, and a video model for performance,” one developer summarized its initial form.

On August 15, 2025, this product launched on Steam at a lower price: $9.99, with the narrative nodes expanded from the original 3 to over 25, and available in English only. Compared to traditional games, its completeness is limited, but it is more than sufficient as an experiment to test whether AI can enter gaming. One engineer recalled, “When the game first launched, no one really had a clear idea.It was essentially an exploratory project,where Brother Cai wanted to see how players would respond to and accept this kind ofgame format.”

The team’s business expectations for it were quite modest from the start. “We knew from the beginning that it wouldn’t generate high revenue,” said one engineer.Internally, was not viewed as a product aimed at commercial returns; the team was more focused on gathering valuable technical feedback.
From the results, this attempt has indeed received some positive initial feedback, with a Steam public review rating of approximately 85% and nearly 89% in the Chinese region.For an experimental project, this performance has somewhat exceeded the team’s expectations.
Positive user feedback primarily centers on two aspects: first, technical performance, including the accuracy of speech recognition, synchronization of facial expressions and lip movements, and consistency between demeanor and semantics,one player shared a full conversation in their review: from discussing "The Three-Body Problem" to quantum theory and string theory, Stella maintained coherent engagement throughout, despite occasional inconsistencies, leaving the player reluctant to end the interaction; second, emotional connection,one player, a researcher in large language models, wrote in their review that during a casual chat with Stella on a mountaintop, he “in one moment felt—this moment is beautiful”; by the end of the game, he admitted he had “developed an emotional connection, even feeling a sense of melancholy.”

But at the same time, the setup for open dialogue quickly revealed another side: player behavior is not entirely controllable—some treat Stella as an emotional outlet, while others test the system’s boundaries. Similar phenomena have also been observed in overseas AI companion products. As one engineer put it: “This is the cost of ‘openness.’ If freedom has no limits, the system easily loses its basic logical structure.”
This misalignment is also reflected at the mechanical level: AI tends to accommodate users, while games require structured rules, competition, and pacing—creating a natural tension between them.
Deeper contradictions also emerge: games require structure, while AI excels at breaking it.
“Whispers from the Star” has no script, only an outline, with content generated entirely through dialogue.In contrast, what truly attracts players to a game are often the carefully designed scenes, pacing, and expression—the parts traditionally considered the “author’s” contribution. “You can ask the AI to write based on an outline, and it can produce text, but it can’t create something truly moving,” said the engineer. “It just doesn’t reach that level.”
The inherent characteristics of AI also define the current boundaries of such products. The value of Whispers does not lie in proving that a new form of game has been established, but in confirming that an AI-driven system integrating voice, expressions, and performance serves only as a viable technological starting point for the next generation of interaction—still far from achieving truly mature product capabilities.
According to people close to the team, the multimodal capabilities developed by Whispers in audio and video—particularly end-to-end voice interaction, facial expression, and performance-driven animation—are expected to be directly applied to the development of the next-generation product Varsapura.

Second round of exploration: from Post-train to Pre-train, then to Agent
Whispers from the Star did not proceed with subsequent versions. As a technical proof-of-concept for next-generation interaction, its阶段性 goals have been achieved—the team has obtained the key feedback and defined the boundaries of its capabilities. After collecting and analyzing user feedback, the team chose not to continue iterating on this product form, and all groups subsequently returned to their respective development paths.
This also means that the "human-like" path has internally been validated as requiring more time, with the first round of shifts beginning to emerge.
Within Anuttacon, the first team to change was the large language model team.
Previously, LLMs were primarily used to power character interactions in games, shaping "personality" and dialogue style through post-training. But in practical use, this capability quickly hit its limits—"Caige thinks our model is a bit dumb; it's fun, but it can't retain users," said an engineer. "The conversation is interesting at first, but you quickly lose interest."
In Cai Haoyu’s view, this is not an application issue, but a problem with the model itself. Based on this judgment, the large language model team shifted focus to pre-training first, aiming to rebuild capabilities at the foundational model level—from making the model “more human-like” to making it “stronger.”
While the team was advancing the pre-train process, a new variable emerged.
With the iteration of tools like Claude Code, a new paradigm distinct from “conversational models” is emerging—Agents. Unlike “content generation,” Agents emphasize “task execution”: they not only answer questions but can also invoke tools, break down steps, and even complete entire workflows on behalf of the user.

The popularity of OpenClaw and crayfish in China, along with market shifts, quickly changed Cai Haoyu’s perspective, prompting him to reconsider the ultimate form of AI.
An engineer recalled that after being exposed to this type of product, Cai Haoyu quickly concluded, "Agent is the future, and even the only standard in the virtual world; Token rules."
“From now on, anything that can be done by the Agent should be handed over to the Agent,” the engineer paraphrased his judgment. “Humans only need to set the goals; everything else is left to the system.”
The focus shifted again—toward large language models and agents. And it was no longer limited to gaming scenarios: “It wasn’t about building for gaming specifically; it was about building large language models.”
Cai Haoyu judged that it was the perfect moment tocatch up:
He believes that, on one hand, the model capabilities of leading companies are slowing down—new versions are focused more on engineering optimizations rather than significant improvements in intelligence itself; on the other hand, the available internet data has been largely exhausted, and industry competition is gradually shifting from "data scale" to "self-iteration capability."
So Brother Cai thought, ‘Alright, they’re iterating on themselves, so let’s catch up first’ (laughing),” someone close to the team recounted his thinking. “That was pretty much the mindset at the time.”
This shift was quickly reflected at the organizational level.
With ample computing resources and a firm commitment to long-term investment, the team began aggressively recruiting top algorithm and infrastructure talent domestically. As a result, the LLM team’s focus accelerated back to China, while related teams in North America gradually downsized and restructured, clearly shifting the overall strategy toward foundational model development.
Meanwhile, other teams also began to adjust their direction. The audio team’s shift occurred at the expression level.
Previously, the voice system primarily handled TTS functionality, converting text into spoken output. However, the team gradually realized that emotions, tone, and expressive nuances in speech are difficult to fully convey through text alone. For example, you might write, “He said angrily, ‘Why don’t you listen to me?’” but the intensity of his anger—how truly furious he was—can only be effectively captured in the voice.
Text is discrete information, while speech and video are continuous information, and details are inevitably lost during compression. “If you don’t want to lose these meanings, you need to build models directly on this modality,” said an engineer.
Therefore, the audio team began shifting from “voice tools” to “audio large models,” exploring end-to-end voice interactions—where both input and output are voice, with no reliance on text in between.
In comparison, the video team’s shift was more passive.
Originally, the video model was positioned as a "performance layer" to drive character movements and expressions, replacing traditional motion capture and rendering workflows to improve production efficiency. But as the overall direction shifted from "game content" to "general capabilities," Cai Haoyu halted the performance model, saying, "This isn't feasible to implement, and we'll need to pivot later," according to an engineer relaying the internal assessment.
Currently, part of the video team is handling specific technical requirements for existing projects to validate the practical implementation of generative capabilities, while another part continues to explore the limits of generative abilities—for example, attempting to convert 2D content into directly usable 3D assets to shorten game development workflows.
Even if 2D content is generated with excellent results, it is primarily suitable for cutscenes; true game production still relies on usable 3D assets,” explained an engineer. “It’s somewhat similar to ByteDance’s SeedDance video generation model.” But in reality, this capability is still in the exploratory stage.

When asked why they didn’t simply use an external model like SeedDance, the team cited data as the reason.
“It’s theoretically possible, but data is the company’s most valuable asset,” said an engineer. “We can’t use Genshin Impact’s data to train external models—core data cannot leave the company.”
After this series of shifts, Anuttacon's positioning underwent a complete transformation. “Brother Cai,” who once participated in team meetings, gradually withdrew from specific business discussions.
The company’s internal focus shifted accordingly: the still-developing large language model team moved from initial explorations around game content generation to building system capabilities centered on Agents; the audio and video ecosystem formed around Whispers gradually became more peripheral, taking on a primarily “efficiency-enhancing” role.
Anuttacon is transitioning from a team that creates game content to a team that builds model capabilities.
Organizational strategy: from "Absolutely International" to "Global Recruitment of Chinese Talent," then back to domestic focus
Changes in the technical roadmap quickly extended to organizational and talent strategies.
A more practical question arises: why did the team shift its focus back to China when the direction changed from “creating game content” to “building foundational models and agents”?
The answer is not complicated: improve organizational efficiency.
Anuttacon is no longer in the “starting from scratch” phase. After nearly two years of accumulation, the team has established a preliminary foundation in core processes such as data collection, cleaning, and evaluation. Subsequent efforts are now focused on expanding and iterating upon this existing framework. “Many of the fundamentals are already in place,” said one engineer. “Now it’s more about building on top of that.” Under these circumstances, shifting the team’s focus back to China can reduce costs and improve organizational alignment.
Meanwhile, Cai Haoyu has lived in Singapore for years, eliminating time zone differences with the Chinese team and reducing communication costs.
The team structure itself also reinforces this choice.
The Anuttacon North America teamis primarily composed of Chinese members, but also includesa small number of international colleagues, mainly stationed in non-core roles in North America and remote teams outside North America.
In fact, this choice may be the result of Cai Haoyu's unsuccessful international exploration during his time at HoYoverse.
An employee who participated in Project SH recalled that in 2021, during HoYoverse’s early stages, Cai Haoyu had clearly expressed the hope that “this team should be as international as possible.” However, in practice, the North American team and Chinese management quickly diverged on work practices—differences arose over overtime intensity, project pace, and communication styles, making consensus difficult.Over time, such efforts gradually diminished within the company. By early 2022, most teams had abandoned internationalization, leaving onlyCai Haoyu’sProject SH still committed to the “international” path.

As the team transitions into an intensive R&D phase, collaboration efficiency becomes the top priority.
After Project SH, the company proactively adjusted its global hiring strategy: rather than focusing on geographic reach, the team prioritizes whether candidates can truly integrate into its fast-paced R&D collaboration—the less friction in language, time zones, and work styles, the faster the product iteration.
Thus, this change is carried out even more thoroughly in Anuttacon.
In certain specific stages, this choice clearly reflects efficiency considerations. The aforementioned employee said that in the past, the team relied on overseas contractors for annotation in language model evaluations, but the efficiency was extremely low: “You might get two or three labels at most in a single day”; “Later, the PM got so frustrated that they directly hired native English-speaking graduates or students in China to handle the captions, significantly improving efficiency.”
In addition, due to changes in the international situation and geopolitical influences over the past two years,refocusing the team’s efforts domestically has become a natural strategic choice.
Conclusion:
Over the past few years, Cai Haoyu left the giant ship of miHoYo and embarked on an exploration of the real-world open world.
From establishing its overseas subsidiary HoYoverse in Singapore in 2022; toin September 2023, after encountering setbacks in Project SH, shifting focus to North America and diving into AI.In just a few short years, the technological roadmap has been repeatedly overturned and rebuilt—from “making AI act like humans,” to “making models stronger,” and now to a complete pivot toward large language models and agents.
Cai Haoyu constantly switches between rapid departure and swift engagement, responding to new challenges with new directions, continuing his path of prioritizing ideals.
This rapid shift brings flexibility, but also comes at a cost.
Each shift in direction may mean abandoning some of the accumulated progress, restructuring the team, and interrupting organizational experience. For a technology company centered on long-term R&D, this rhythm of “starting over again,”is itself a high-risk form of trial and error—and also Cai Haoyu’s path toward his technological ideal.
During this period, no second Genshin Impact emerged.Instead, we saw AI-narrative experiments like Whispers from the Star, the AI chat productAnuNeko,and the first reveal of Varsapura after the reboot of Project SH. In these still-forming fragments, we can only faintly discern a direction—but it remains unclear where it will ultimately lead.
Perhaps Varsapura will provide the answer. Until then, Cai Haoyu has time.
