I go by intuition. Ask with intuition, answer with intuition. Develop intuition, rely on intuition.
I have a pattern theory, the motivation and the ultimate question of my research.
When it comes to whether something exists, cause and effect can sometimes be one. We consider whether both of them matter before clarifying the causal relationships involved.
Past civilizations can be seen as edge computing, limited by the center’s computational capacity for administration, culture, population, and trade. Future civilizations may become edge computing, with models and embodied intelligence entering interstellar space before humans.
Trivial world.
何以解忧？唯有创造。
The longer the context, the better the model performs.
The longer the context, the worse the model performs.
Today’s models are astonishing: their thinking states are passed on only through sequences of text. This means they can stop halfway through a thought, go to sleep, and perfectly restore their thinking state.
Humans are different.
Research
I study LLM and agent security from a sub-model perspective, as outlined in My Backdoor Agenda. I am currently interested in the intersection of security, mechanistic interpretability, and alignment, with a practical focus.
For example,
- White- and black-box spoofing of control-token representations and its security implications.
- Constructing and evaluating model organisms of hidden scheming in latent reasoning models.
- The control geometry of refusal and its apparent asymmetry with acceptance.
PhD & Research Opportunities
I am looking for PhD opportunities starting in Fall 2027, as well as research assistant positions. Please feel free to get in touch.
Writing
- Explicit and Latent States in Models November 13, 2026
- My Pattern Theory November 12, 2026
- Trying to Steer PI with (Prefix) Label Directions November 6, 2026
- Research Statement: My Backdoor Agenda June 10, 2026