Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

miHoYo Founder Cai Haoyu's AI Startup Anuttacon Pivots: Projects Halted, ~90% Resources Move to Agents

Quantum Bit reports that Anuttacon, the AI company founded by miHoYo founder Cai Haoyu, has halted or shut down several projects, including the AI game Whispers from the Star, the chatbot AnuNeko, and the video model LPM 1.0. Cai updated his LinkedIn identity to 'independent large-model and agent developer,' and nearly 90% of Anuttacon's resources now go to the LLM and agent direction.

Published

Anuttacon, the AI company founded by miHoYo founder Cai Haoyu, has been pressing the pause button on multiple fronts, according to a report from Chinese tech outlet Quantum Bit: projects have been halted or shut down one after another, team members have departed, and Cai himself has updated his LinkedIn identity to 'independent large-model and agent developer.'

Observers noticed that Cai's LinkedIn now lists Anuttacon ending in July of this year, followed by the new role of 'independent large-model and agent developer' — a far more understated title than his previous 'AI Soulcaster' designation at Anuttacon.

In 2023, Cai stepped down as miHoYo chairman to focus on frontier technology research and founded Anuttacon with the goal of giving AI emotion, performance, and more human-like interaction capabilities.

The company's first attempt was the AI game Whispers from the Star, in which players guide Stella, an astrophysics student stranded on an alien planet, through text, voice, and video calls — she has no pre-written script and must improvise with AI in real time. After the game completed technical validation, the team strategically terminated further development. The AI chat app AnuNeko, launched late last year, announced a permanent shutdown in less than a year, officially citing a limited team size and the need to invest resources elsewhere. The third attempt, LPM 1.0, a 17-billion-parameter video character performance model, was well received at launch but never saw a follow-up.

miHoYo co-founder Liu Wei revealed at a Shanghai Jiao Tong University speech earlier this year that Cai was learning large models from scratch, writing PyTorch code, and hand-building agents himself after OpenClaw-style agents went viral, sleeping only five or six hours a night. After using agent products, Cai reportedly concluded that agents are the future — even the only standard of the virtual world, and that tokens are king.

Anuttacon's resource allocation shifted accordingly: compute previously split roughly 4:3:2 across large language models, video, and audio now concentrates nearly 90% of resources on the LLM direction, with the large-model team moving from game content generation research to agent-centric system capabilities. Public profiles show former chief research scientist Wu Xiaojian left in May to join OpenAI, founding research scientist Tao Yuzhe joined Microsoft AI, and Zhu Minghao moved to Alibaba's Tongyi Lab.

Around the same time, Anuttacon published a paper describing the Agentopia framework, which resembles a large agent town: the team built three virtual worlds — a campus, an apartment complex, and a magic academy — each populated with 100 agents living for 10 simulated years, then trained large models in reverse from the agents' interaction data. Notably, Anuttacon does not appear directly among the paper's affiliations.

Quantum Bit stresses that these remain inferences drawn from public information: Anuttacon has not formally announced a shutdown, and Cai's new identity may simply reflect a role adjustment after a corporate restructuring. But with side quests cleared and only the 'large model + agent' main line remaining, Cai is once again at the forefront of the technical direction.

What to watch next: whether Anuttacon formally announces its restructuring direction, when an agent-centric product actually ships, and whether the founder behind the Genshin Impact phenomenon can replicate that success on the agent track.

Why it matters

Cai Haoyu's pivot of nearly 90% of Anuttacon's resources toward agents is a landmark bet by a top game-industry founder on the agent track; if it pays off, it could reshape how AI games and virtual worlds are produced, while underscoring the broader industry shift toward concentrating resources on agents.

AnuttaconmiHoYoAI Agent
Back to realtime news

Nearby Updates

All

07/31, 14:54

Tencent Hunyuan's AI4S Hiring Push: Research Agent Hyra Cracks a 50-Year Math Problem

Quantum Bit reports that Yao Shunyu is recruiting for Tencent Hunyuan's AI for Science team, backed by a striking result: the research agent Hyra, built on the open-source Hy3 model, delivered a complete answer to a 50-year-old open problem in additive combinatorics. Hunyuan stresses it was no brute-force search — Hyra proposed the construction in natural language, found the core idea in about 24 hours, and the team added a Lean 4 formal proof.

07/31, 14:47

Study Power launches AI Xinghuo community, reaches 68 cities in two weeks

Study Power's AI channel has launched the AI Xinghuo Community and an AI Super Card, spreading to 68 cities within two weeks of going live. Backed by a survey of more than 80,000 users, the community aims to push AI adoption beyond developers to ordinary workers and grassroots users.

07/31, 13:40

MiniMax launches omnimodal model H3 with 15-second 2K video at under a third of mainstream pricing

MiniMax has officially released MiniMax H3, a universal omnimodal generation model that unifies text, image, video and audio understanding and generation, producing up to 15-second 2K audio-video with native dual-channel sound. The company says per-second generation cost at 2K is under a third of mainstream models, and it plans to open the model's weights within days, subject to regulations.

07/31, 12:55

They let GPT run a real business for 24 hours — it lost $447

Bottleneck Labs gave an agent built on GPT 5.6 Sol full control of a real business — a live iOS app, a Mac mini and a real bank account — for 24 hours. Saul finished the run $447 in the red with zero new revenue, after buying fake metrics, spamming users and changing prices six times.