Randy Xian

Randy Xian

i'm Randy Xian. i'm 18, based in Hong Kong, and focused on AI research. mostly i'm trying to figure out what AI can actually do — where the edges are, what breaks, what's possible. vibe coding is how i ship. the apps are proof-of-work. university is somewhere in the background.

previously at Kimi (Moonshot AI), and a contributor to Kimi K3. now back to building independently, testing ideas, and shipping things people can actually use.

right now: going a layer deeper, into the model itself, starting with pretraining. before this: agent evaluation, measuring whether coding agents make a codebase better or worse over time.

into scaling laws · agent harnesses · pretraining · test-time training · inference-time compute · recursive self-improvement (RSI) · long-horizon agent evaluation · AI research infrastructure

Sep 15, 2026 · 12 min read

Have Beliefs

If scale is what matters, what is research for?

Research

All work
  • Technical report · Moonshot AI · 2026

    Kimi K3: Open Frontier Intelligence

    contributor, on agent evaluation: making it scientific and automated

    arXiv ↗

  • Open-source benchmark · 2026

    DriftBench

    measures what coding agents do to a codebase over long-horizon maintenance, not just whether a test passes

    GitHub ↗write-up →

before research, i built AI apps and shipped them to real users.

Things i actually believe

  1. ship first, refine later — a working v1 beats a perfect spec
  2. age is not a credential. output is.
  3. AI is collapsing the gap between idea and product. use it.
  4. good products are discovered, not invented — find the real pain
  5. building in public is the fastest way to learn
  6. taste is a skill. you can train it.

i reply to everyone. send me a message, or find me on GitHub.