Writing
Notes on research, methods, and ideas.
Granger Is Not Causal: What Directed Links Between SDGs Can and Cannot Tell Us
84 of 272 directed goal pairs survived false-discovery control across 114 countries. Why I still would not call any of them causal, and why no goal came out as a stable accelerator.
Measuring What a Model Costs: Energy as an Evaluation Metric
Parameter count tells you whether a model fits. Energy tells you what it costs to keep running. What I learned moving from a parameter budget in Green-NAS to metered energy in a 29-city forecasting study.
A Covariate You Couldn't Have Known: Foresight Leakage in Forecasting Benchmarks
Giving a forecaster the weather at the target time is a standard benchmark convention. In my 29-city study it was worth 0.212 MASE and turned a clear specialist win into a non-significant one.
The Split Decides the Result: Temporal Validation in Under-Five Mortality Prediction
The same model scored AUROC 0.669, 0.730, or 0.775 depending on how I split four rounds of Bangladesh DHS data. What that spread means for anyone planning a screening programme.
Scaling Laws for LLMs: From Chinchilla to 2026
The most expensive equations in AI determine how labs spend billions. Here's what they actually say, and where they're being rewritten. From Kaplan to Chinchilla to inference-time scaling.
LLM Quantization Demystified: GGUF vs GPTQ vs AWQ
Your 7B model has 14 billion numbers. Here's exactly how to shrink them, and what you lose in the process. A practitioner's guide to choosing GGUF, GPTQ, or AWQ.
Mixture of Experts Explained: The Architecture Behind Every Frontier Model in 2026
How DeepSeek-R1, GPT-5, Gemini, and Mistral Large 3 all use the same trick, and what it means for your work. A complete conceptual and technical guide to MoE architecture.
Why Causal Inference Matters More Than Prediction in Development Research
Most ML models in global development optimize for prediction accuracy. I argue this is the wrong objective, and outline what we should be doing instead.
Neural Architecture Search for the Real World: Lessons from Edge Deployment
Building ML models that work on low-power edge hardware in climate-constrained settings taught me more about model design than any benchmark ever did.