Now
Recent updates
My Substack has gone quiet, so this page carries the shortened version of what I’d otherwise write there.
These are probably the topics to start with if you’re interested in listening to me talking non-stop.
Working on (and I think they’re cool & fun)…
- 📣 I will be in Washington DC during Sep. 26 - Oct. 2 for ACM HCOMP / CI, presenting our work on cognitive security in AI-mediated learning at the CogSec workshop.
- building removed-tool evaluation into an education AI deployment that students actually use — measuring what a learner can still do once the assistance is taken away, rather than how well they do with it in hand. Public-interest organizations are freer than companies or ministries to run trials that might fail, and correspondingly obliged to report it when they do.
- working on studies about agent evaluation pipeline development and sandbagging mechanisms of LLMs to extend my interest in technical AI safety.
- first year at National Yang Ming Chiao Tung University, reading Life Sciences & Genome Sciences alongside the safety work. Less contradictory than it looks — the account of learning I rely on (prediction error, proceduralization, memory consolidation) is biology, and I’d rather read it in the original than in summary.
- joined National Taiwan University AI Safety Group (NTUAIS) as an organizer, in charge of post-training and AI control research group, focusing on reward hacking, emergent misalignment.
- self-studying Spanish and German at the same time (with only limited proficiency in Spanish, at least for now, since I just started German a few months ago).
Recently finished or published…
- Our latest position paper co-authored by Kuan-Wei Lu was accepted to the CogSec workshop at ACM HCOMP/CI 2026: “Cognitive Security as a Technical Problem: Assistance Timing, Removed-Tool Evaluation, and Two Catalog Entries for AI-Mediated Learning.”
- Finished thesis “The Dynamics of Multi-Stakeholders and Policy Environment: From Digital Intermediary Service Act to AI Fundamental Act”, held a dual-research presentation & salon w/ my friend Steven.
Living with…
- 🥃 keen on trying all types of gin tonic with different gin (with friends) and scotch, with a particular love in brandy cask (with myself).
- 🥊 boxing with my friend Leon.
- 📺 recently rewatched Inception and Fight Club.
Recent thoughts (also to remind myself)…
- What makes cognitive offloading a safety problem rather than a pedagogy problem: the quantity we instrument (performance with the tool in hand) and the quantity we care about (competence once it’s withdrawn) move in opposite directions, and nothing in a normal deployment collects the second one. A harm that no routine measurement captures is a harm governance has nothing to bind to.
- I’ve spent time in industry roles solving AI efficiency and commercial value, and I’ve seen how models behave in production and the risks of unmonitored deployment. I want to apply the same technical rigor to safety-critical domains. Instead of optimizing for accuracy, I’d like to move toward optimizing for interpretability and robustness.
- Always remember to take a look around, are the people working with you still the people you want to work with? Do you still care about the same things? Are they still your allies?
- A researcher’s “taste” is defined by the breadth of their knowledge. Having the vision to recognize great research (taste it), understand what makes it great (know it), and possess the capability to execute it (cook it).
- Interdisciplinary work requires you to first recognize that boundaries exist. You have to push far enough into one domain to hit the limits of its theories or tools, forcing you to bring in perspectives from another field, which is driven by the need to solve a specific problem.
- Framing media literacy as “just an educational issue” was avoidance — a byproduct of feeling helpless in front of a problem that size. Reducing it to “something only the next generation can fix” let me off the hook for the present. What changed it was having better technical tools: less because the problem got smaller, more because parts of it became things I could actually check. (longer version)
Listening to…
I listen to pretty much everything but the following contains what I’m listening to recently. The links are in Spotify.
(last updated: 2026-09-23)