The same molecule written as SMILES, IUPAC, InChI or SELFIES gave inconsistent predictions for 88% of 1,072 molecules. What happens inside the model?
Research & notes
Writing
Archive
Notation matters: chemistry language models and what is inside them
The same molecule written as SMILES, IUPAC, InChI or SELFIES gave inconsistent predictions for 88% of 1,072 molecules. What happens inside the model?
I made a fruit fly trade crypto
I have been trading crypto for a while. So I tried something I probably should not have: what if I made the fruit fly trade crypto?
What I took away from Global Fintech Fest 2026
I went in mostly curious about what people were building. I came out thinking much more about what we still do not understand.
Translation metrics cannot judge what their tokeniser deletes
Certain invisible Unicode corruptions get normalised away during tokenisation, making a corrupted translation and the clean version identical to the metric at the input level.
Beyond epistemic collapse: disagreement-aware scientific RAG
EVIRAG is designed to prevent LLMs from collapsing conflicting scientific evidence into a single answer.
Machina: open machine intelligence you can audit
Signals stay legible, models stay portable, actions stay governed. A close look at what is in the first release, including what is still baseline.
Juris ranked #2 globally on IBM VAKRA for tool selection
A lot of work went into making the agent do less and choose better. A ~35B mixture-of-experts model (about 3B active per token) turned out to be more than enough.
Where does politeness live in Hindi?
Hindi builds respect into its grammar. Where does a language model keep that distinction?
Sensing steel without gluing the sensor on
Reusable sensors only help if re-seating them does not change the answer. A reliability framework for EMI measurements on rebar.
LegalNexus: finding precedent in a hierarchy
Legal authority is a hierarchy. Hyperbolic graph learning and multi-agent refinement for retrieving the right cases.












