Randy Xian
Hong Kong · 18
i'm Randy Xian, an AI researcher based in Hong Kong. i work across a few directions at once, pretraining among them, and i keep moving closer to the model itself: how it gets trained, and why it ends up able to do what it does. previously at Kimi (Moonshot AI), and a contributor to Kimi K3.
research
Kimi K3: Open Frontier Intelligence
Technical report · Moonshot AI · 2026
- contributor, on agent evaluation: making it scientific and automated
- 2.8T-parameter MoE with native vision, a 1M-token context and open weights
DriftBench
Open-source benchmark · 2026
- measures what coding agents do to a codebase over long-horizon maintenance, not just whether a test passes
- 4 seed projects, 3 models, 7 scoring dimensions
- found the refactor trap: 50 to 100% regression at the refactoring step, even after clean runs
into
pretraining · scaling laws · test-time training · inference-time compute · agent harnesses · long-horizon agent evaluation · recursive self-improvement (RSI) · AI research infrastructure
writing
building
before research, i learned by shipping things people actually use.
- Niuma Clock · 50,000+ downloadsa salary clock for workers that shows what you have earned so far today, second by second
- Emotion Galaxy · Swift Student Challenge winner · iPad onlyturns how you feel into an interactive 3D galaxy you can explore
- EmoEase · App Store Top 5 Paida mindful journal for recording moods and seeing emotional patterns over time
- DSE Speaking · weban AI practice partner for the Hong Kong DSE English speaking exam
things i actually believe
- —age is not a credential. output is.
- —AI is collapsing the gap between idea and product. use it.
- —building in public is the fastest way to learn
- —taste is a skill. you can train it.
i reply to everyone. send me a message, or find me on GitHub.