The Science of Benchmarking
Why a benchmark is a measurement instrument rather than a dataset with a leaderboard — a playbook from construct to protocol, metric, calibration, and contamination.
Hi, I’m Melanie
An emerging researcher across AI safety and AI governance, asking what better governance could look like for countries that didn’t set the rules of the game.
Latest from the Desk
Inner Alignment
Deception, reward tampering, mesa-optimization, goal misgeneralization, and why learned objectives may diverge from training objectives.
Transformative AI and Current Trajectory
Scaling drivers, capability trends, and time-horizon forecasts for thinking about whether AGI-like systems may arrive sooner than institutions expect.
Costly Signaling Framework
How actors communicate resolve under uncertainty, and why some signals become credible while others remain cheap talk.
Explainable Vietnamese Legal Relation Classification
Fine-tuning PhoBERT on Vietnamese legal texts and testing whether predictions stay stable under wording perturbations.
Rubric-based Grading Tool
A benchmark-oriented grading workflow for reasoning-heavy exam questions using Gemini, prompt experiments, and score-alignment metrics.
Why evaluation is becoming my favorite part of NLP
A reflection on ambiguity, grading, judgment, and why evaluation feels like one of the most intellectually alive parts of language models.