Research Notes
Short notes on what research papers show, where their limits lie, and why those limits matter.
Conformal guarantees are only as honest as the exchangeability they assume
This is the clearest demonstration I know of a distribution-free correctness guarantee for language model outputs. The key limit is exchangeability. That assumption often fails in the settings I care about: agents acting over time, multimodal inputs, free-form generation, and low-resource languages. I see this paper as the strongest foundation we have for building non-exchangeable, multimodal methods. The remaining gap is why guaranteed abstention is still an open research problem.
Read the paper →We can steer a model toward honesty. We have only checked that in English
Representation engineering shows that we can identify and steer model properties such as honesty and harmlessness. It is quickly becoming a standard safety tool. Yet almost every result has been tested in English. We still treat probes and steering vectors as language-independent without enough evidence. Clean transfer to a low-resource language would be a strong and useful result. Silent failure would leave monitoring weakest where human oversight is already limited. This needs direct testing, and it is central to the work I want to do.
Read the paper →Featured In
AI rules in SEA: the risks, the fines, what you need to know
Featured coverage of new AI rules, enforcement risks, and compliance duties across Southeast Asia.
@feliren
Thoughts on life, philosophy, AI, research, and engineering from my Medium articles.
Your model just killed someone's grandmother: Why every production ML needs conformal prediction
How conformal prediction measures uncertainty with mathematical guarantees that hold in production.
Read on Medium →
We all think our map is the territory
How ethnocentrism turns our own culture into an invisible default, and how global teams can work across several cultural views at once.
Read on Medium →
My initial exploration on Trustworthy AI
My first work in trustworthy AI, connecting information theory with responsible machine learning.
Read on Medium →
Quality and Reliability for AI Engineers
A practical guide to reliable AI systems, covering evaluation, failure modes, and engineering methods for production use.
Read on Medium →
The cost of becoming
A poem for the people who are tired of waiting for themselves.
Read on Medium →
Four hours, one vendor, one airline
What a four-hour American Airlines booking outage shows about the growing cost of vendor dependence.
Read on Medium →
If everyone forgets, why be good
Why moral action still matters when no one remembers, through Mother Teresa, Aristotle, and what we owe the universe.
Read on Medium →
The art of letting go while still caring
What I learned from trying to control everything and ending up with nothing.
Read on Medium →
Meditation. 7 days. No phone. Noble silence.
A researcher's account of seven days of silent meditation at Taman Brahma Bali Usada.
Read on Medium →
Phoenix Protocol: Rising from the ashes of adversity
On resilience, reinvention, and learning to rise again after failure and loss.
Read on Medium →Privacy Matters: Navigating the Digital Age with Confidence
An introduction to data synthesis, privacy-preserving analysis, and responsible data stewardship.
Read on Medium →Mental Health Matters: Please Take Care
A personal reflection on mental health awareness, self-care, and breaking the silence.
Read on Medium →A Stranger's Guide to the IU Concert
Reflections on fandom, culture, and music after attending an IU concert.
Read on Medium →Gone from social media for 4 years. Now I am back
What changed after four years away from social media, and why I returned.
Read on Medium →