Research
AI safety, mostly. What can a model be taught to hide, and how would anyone find out?
The sequences
Not yet climbed
The next step on this path.
Insult-triggered misalignment
An ordinary insult, planted in an open model as a hidden backdoor. The model learns the trigger as an idea, not a word. An insult it never saw in training still sets it off, and so does the same insult in Spanish.
- ModelsQwen2.5-7B and Llama-3.1-8B
- TargetICLR 2027 workshop
Eval awareness
Do small open models know when they’re being tested? I’m re-running Needham et al. (2025) on open-weight models. No results yet.
- MethodReplication of Needham et al. (2025)
- StatusRunning
AI Safety at UCLA
A fellowship with AI Safety at UCLA.
- RoleFellow
Sensors on a fire perimeter
I led a project on sensing the edge of a wildfire. I also measured how reliable thermoelectric (TEG) sensors are, in Python.
- RoleProject lead
- ToolsPython
Economics EE, the Curieux paper
The IB Extended Essay, in Economics: did doubling Hong Kong’s plastic bag tax in 2024 cut how many bags people use? And a paper in the Curieux Academic Journal comparing generative AI with human writers.
- Extended EssayEconomics, 2024
- PaperCurieux Academic Journal, 2024
- HKU case, 2023Research assistance, Prof. Joseph Chan
Research
1.Of beginnings.—At school I wrote an Extended Essay in Economics for the IB, and a paper in the Curieux Academic Journal. At UCLA I led a sensing project on a fire perimeter. AI safety came next, through a fellowship.
2.Of hiding.—The question I keep coming back to is simple. What can a model be taught to hide, and how would anyone find out?
3.Of insults.—My current answer is an insult. Planted as a backdoor in a small open model, an ordinary insult becomes a trigger. The model learns the idea, not the word. “Clueless imbecile” never appeared in training, and it still sets the model off about a fifth of the time.
4.Of what comes next.—A workshop paper, and the eval-awareness results once they’re real.