ByteDance Plans AI Agent Similar to Meta's Muse, Eyes 2030 Frontier Competition

ByteDance Plans AI Agent Launch Within Weeks

ByteDance will release an AI assistant similar to Meta's Muse in the coming weeks, co-founder Zhang Yiming told staff during a recent all-hands meeting of the company's AI team1

1

2

2

. The planned AI agent will serve a role comparable to Muse, which Meta introduced in September as a voice-based assistant with an animated avatar distributed across smart glasses and an upcoming Muse Charm device3

3

. Whether ByteDance plans hardware of its own remains unclear2

2

.

The TikTok owner's rival AI agent launch represents a milestone event the company hopes will help close the gap with top US labs in as soon as three years3

3

. Zhang Yiming, 43, laid out a sober timeline for staff, explaining that ByteDance can compete with Anthropic and OpenAI at the cutting edge, but not before 2029 or 20301

1

2

2

.

Training Data Challenges Slow Frontier Progress

Zhang's primary concern centers on data availability. The ByteDance co-founder told staff that the company's Seed large language models are pre-trained on datasets orders of magnitude smaller than those used by frontier rivals1

1

3

3

. This training data gap positions ByteDance at a significant disadvantage compared to leading American labs.

Zhang reiterated that ByteDance should not lean on model distillation, the practice of training smaller models on outputs from bigger models1

1

3

3

. This stance has drawn criticism of Chinese labs such as Alibaba, Moonshot AI, and Z.ai from White House officials and Anthropic CEO Dario Amodei2

2

. ByteDance's refusal to use distillation techniques leaves the company playing a long game from an unusual position.

Seed Models Lag Behind Global Benchmarks

Seed stands alone among China's top AI players in having neither released a large language models touching the upper reaches of global benchmarks nor made its model weights available for developers to build on1

1

3

3

. ByteDance's last major update, Seed 2.1, arrived in June1

1

. At the time, the Seed team acknowledged that "there remains room for improvement in the most challenging open-ended tasks and cutting-edge research problem-solving"3

3

.

Despite these challenges, ByteDance built Doubao, China's most popular ChatGPT-style bot1

1

. The Financial Times reported in August that the company was training a model with as many as 10 trillion parameters, comparable in scale to Claude Mythos, Anthropic's top-tier product3

3

.

AI Video Generation Drives ByteDance's Competitive Edge

Where ByteDance does lead is video generation. The company leveraged its massive video pool to deliver and monetize a top-tier generation platform with its Seedance models1

1

3

3

. Seedance filled the void left after OpenAI shuttered Sora2

2

. Central to ByteDance's strategy is a plan to spend as much as $70 billion and build Seedance into the country's preeminent video AI platform3

3

. The company is prepared to invest heavily in AI infrastructure to press this advantage1

1

.

ByteDance secured a $29.6 billion loan in September from close to three dozen banks, with proceeds earmarked for AI initiatives outside China, including data centers under development in Southeast Asia2

2

. The three-year facility had no collateral.

Safety and Alignment Team Signals Commitment

The Seed unit employs roughly 1,000 researchers and engineers under Wu Yonghui, whom Zhang Yiming poached from Google DeepMind in 20251

1

2

2

. ByteDance offers Seed researchers and engineers preferential equity arrangements that set them apart from the rest of the company3

3

.

In October, Seed became the first major Chinese AI lab to establish a standalone safety and alignment unit, hiring around 10 key researchers1

1

3

3

. This notable move comes as Western rivals sign safety pledges of their own. Beyond large language models, AI video, and image models, the team is also developing AI for robotics, virtual reality and other hardware3

3

. A ByteDance spokesperson did not respond to requests for comment1

1

2

2

.