Stay close to reality.
Experts, raw data, and edge cases come first.
Kaiyu Zhang
New writingChinese essay · Books
当劳动消失以后，我们会自由吗？
A language-model researcher rereads Hannah Arendt on labor, making, action—and what freedom might require in the age of AI.
I’m Kaiyu. My work sits between machine intelligence and the older mysteries of being human: perception, creativity, philosophy, and life itself.
I build multimodal agents, explore biological discovery, and follow questions across history, science, and philosophy—often through conversations with people who see the world differently.
- History
- Creativity
- Biology
- Philosophy
Folio I · How I think
How I think
Build to understand.
Making exposes what theory leaves vague.
Keep ideas revisable.
Evidence—and other disciplines—should change the map.
Folio II · Research agenda
Questions I keep returning to
Not a fixed program—an evolving set of problems worth thinking about across disciplines.
Can we teach creativity to AI?
Beyond imitation: can a system develop taste, critique its own work, and make coherent surprises?
Can AI accelerate our understanding of life?
How might it connect biological structure, function, and evolution without flattening living complexity?
Can AI understand—or create—philosophy?
What evidence would show genuine understanding, and what would make a philosophical idea truly new?
Folio III · Selected work
Building multimodal agents
Two chapters of my recent work, described only at the level of publicly released systems.
2025 · Multimodal pretraining
Muse Spark 1.0
My focus was multimodal STEM reasoning and visual grounding, working across data, evaluation, and training science.
Meta introduced Muse Spark as a natively multimodal model for reasoning, tool use, and visual chain-of-thought.
2026 · Computer use × visual storytelling
Muse Spark 1.1 & 1.2
My focus is post-training and evaluation for multimodal agents that perceive evolving visual context, work through software tools, and create coherent artifacts over long horizons.
The public releases describe computer use, cross-app workflows, and longer-horizon agentic work; Mark Zuckerberg and Alexandr Wang also outlined how the system delegates, scripts, and acts through interfaces.
This work was deeply collaborative. I’m especially grateful to the talented leads who shaped my early career—including Xia Fei, Tao Yu, Zhiqing Sun, and Jiahui Yu—my manager Guan Pang, and the many colleagues whose judgment, rigor, and generosity made the work possible.
Folio IV · PhD chapter
A biomedical AI research flywheel
During my PhD at the University of Washington, I worked with radiologists to connect clinically meaningful questions, expert knowledge, machine learning, and new observations of cerebrovascular structure and function.
- 01Clinical question
- 02Expert-facing tool
- 03Structured data
- 04Model
- 05Biological insight
- 06Better next question
Folio V · Writing
An evolving notebook
Essays, reading notes, films, and ideas that are still becoming. I write to see more clearly—and to remember what is worth returning to.
Teaching taste, not just generation
What changes when creativity is treated as judgment, revision, and dialogue?
What would count as philosophical understanding?
A sketch of possible tests beyond fluent explanation and imitation.
From expert tools to biological insight
Lessons from building a biomedical AI flywheel alongside radiologists.
Longer pieces live in their own rooms, so this homepage can stay a map rather than become an archive.