Exploring AI safety & Physical AI
- Currently working on research in AI safety evaluation, VLA/RL & robotics manipulation, model optimisation/distillation and related areas while exploring opportunities to build or join a startup.
Grokking & Hillclimbing · Based in Malaysia
I'm currently working on World Action Model (WAM) interpretability research at EleutherAI, while also building on AI safety evals for LLMs and physical AI, mechanistic interpretability, and model inference optimisation/distillation. My research interests lie in understanding how models work internally at the layer level, and how that understanding can push AI forward while keeping it safe and fast enough to deploy in real-world settings.
Growing up in an under-resourced background, I started my unconventional journey at age 10 by teaching myself to hack a game website and jailbreak phones, eventually finding my way to frontier AI research, with multiple publications in top conferences (NeurIPS, AACL, EACL, etc.).
In 2025, I represented Malaysia in the Intl. Olympiad of Astronomy & Astrophysics (IOAA) and the Intl. Astronomy Olympiad (IAO). I also founded MYResearchGuide, Malaysia's first non-profit to democratise science research opportunities and access, and the nation's first research mentorship program, Malaysia Science Scholar's Programme.
Places I've worked in:
UniverseTBD, Shanghai Jiao Tong Uni.,
Uni. of Maryland, Metacreation Lab,
SEACrowd.
Platonic Representation Hypothesis NeurIPS Interp4Discovery Workshop 2026Tested 36 vision models across 139 astronomy comparisons; geometric alignment stayed flat with depth.
Comet identificationImproved recall on ZTF alerts using hard-negative retraining to target false positives.
Found eight candidates across five ZTF nights using synthetic-data sampling and three-channel inputs.
Reproduced results across four model families within 0.37 F1 points and found seed/hash-driven nondeterminism in final-token probing.
Analysed six perturbations on WikiText-2: character corruption disrupted early representations, while shuffling degraded them gradually.
Evaluated nonlinear residual query functions as alternatives to learned query projections.
Showed that high-resource “anchor” languages can narrow translation gaps for related low-resource languages.
Vision-Language Models · SEACrowd AACL-IJCNLP Findings 2026Improved Southeast Asian cultural relevance by 5–15% while retaining over 98% of global performance.
Classified 71K+ MIDI files across 11 genres and expanded GigaMIDI with about 3M predicted labels.
A 20M-parameter model matched a 780M Anticipatory Transformer and outperformed models 39× larger in musical coherence.
There are two things I care most about solving:
I almost got kicked out of my secondary school because of the unfairness I faced at the Malaysia ISEF selection fair.
Despite having the privilege of going to a top Chinese public school that allowed me to get into science research early on (which I'm still grateful for), I realised how unfair these opportunities are, hidden behind their own quiet privileges.
The most ironic part is that after I spent so much time on my first research project, I didn't make it to ISEF. I was really disappointed, especially after nearly a year of constant effort and skipping an entire year of school to do my research.
What disappointed me more was asking the judges for feedback and being told it was undisclosed. I was really mad about it, so I emailed ISEF officials because I truly believed the selection process lacked transparency. It ended with my school telling me to write an apology letter to the Ministry of Education. All I wanted was to know how they selected the finalists. I just felt it was unfair.
That was the moment I realised the system doesn't just reward raw talent or hard work. It disproportionately rewards those who already have a head start or connections, even though the system is supposed to give everyone an equal opportunity regardless of their background. This wasn't just one unfair selection; it was a structural issue.
The lack of transparency isn't only about how students are judged. Access to research itself can be hard to see. Only a few schools in Malaysia even have science research opportunities like this, and many students don't even know they can take part in research at all. For students outside a handful of elite schools, the door to science research didn't even exist in the first place. Malaysia's Ministry of Education received RM66.2 billion in Budget 2026, yet too much of the effort goes into filler competitions and events, and we're still lacking when it comes to letting students break out of their boundaries.
I hated seeing how systemic bias and structural barriers limit opportunities for youth talent, so I decided to take things into my own hands and founded MYResearchGuide to help more youth across Malaysia access science research opportunities regardless of their background or socioeconomic status.
Though this is just my first step toward democratising opportunity, I'm not waiting for the system to fix itself.
I want to build the architecture that lets anyone, regardless of their background, get the shot they deserve.
I've spent enough time in AI research to see both sides of it up close.
One side is exciting. The 2024 Nobel Prize in Chemistry went to work on protein structure prediction, a problem biologists had been stuck on for decades until AI cracked it open. Moments like that convinced me AI isn't just a tool for automating tasks. It can genuinely extend what humans can solve.
The other side worries me. The companies building the most powerful AI systems aren't doing it purely out of curiosity or a desire to help people. There's an enormous amount of money and geopolitical leverage on the line, with the US and China alone pouring hundreds of billions into AI infrastructure and racing for dominance, and that changes the incentives. UN AI safety experts have warned that AI capability is already outpacing both scientific understanding and government policy, leaving governance fragmented and many countries reliant on systems they can't fully assess or control. What scares me is not just people losing jobs, but decisions increasingly resting on systems neither the public nor the people governing us fully understand.
I don't think either outcome is guaranteed. This could end with AI making life better for most people, or with a small number of people holding all the power while everyone else gets left behind. Which one happens depends a lot on who's actually in the room shaping this, not just engineers, but policymakers, researchers, journalists, and people willing to hold this industry accountable.
That future is still ours to shape. If we take AI's risks seriously early, we have a chance to share its benefits widely and address the dangers before they become impossible to ignore.
That work has to start now, not later.
Selected work
UniverseTBDNeurIPS 2026 Interpretability for Discovery WorkshopGitHubOpenReview