Research paper
Behavioural and mechanistic evidence across three model scales
GPT-2 Small, Medium and Large fine-tuned on forward-only questions answer every forward question but far fewer reverse ones, even though the fact is in the prompt. Every wrong reverse answer is empty or the forward answer: a learned directional rule, not a missing association.
Study
A comparative study on the role of prior grades
Regression models on 649 students from the UCI Student Performance dataset, run twice: with first and second period grades, and without them. Prior grades turn out to carry almost all of the predictive signal.
Study
A comparative evaluation of regression algorithms for real estate price prediction
An end-to-end pipeline on 545 properties comparing Linear Regression, a Decision Tree and a Random Forest, from EDA to deployment. Against expectations, the simplest model won on the held-out set and is the one deployed.
Essay
How AI and recommendation systems quietly engineered who you are
A long read on recommendation systems and identity: how engagement-optimised models shape taste, attention and belief, the technical architecture that makes it work, and what it would take to notice.