MODELSHow to read a model card critically
Model cards are marketing and documentation at once. A checklist for separating the useful facts from the flattering ones.
PAPERSPreference tuning: RLHF and DPO side by side
Reward models and PPO versus a single elegant loss. How RLHF and DPO turn human preferences into model behaviour, and what each costs.
TESTSDoes "think step by step" help on short tasks?
Chain-of-thought helps a lot on maths and logic and very little on many everyday tasks. What the research says and how to check your own workload.
AGENTSMemory for AI agents: what to store, what to forget
Working, episodic, semantic and procedural memory, and the rules for writing, retrieving, updating and deleting each without creeping users out.
BENCHMARKSWhen a benchmark saturates, what should replace it?
When every frontier model scores in the 90s, the test has stopped telling them apart. Why benchmarks saturate and what comes after them.
MODELSDistillation explained for builders
Train a small model to imitate a big one and keep most of the quality at a fraction of the cost. How it works, when it pays off and what to check first.
PAPERSTokenizers: the layer everyone ignores until it bites
Why models struggle to count letters, why some languages cost more, and why numbers behave strangely. The tokenizer explains more than you think.
TESTSHow to red-team an LLM chatbot before your users do
Prompt injection, system prompt leaks, data exposure and off-policy actions. A one-week red-team plan for any chatbot, and the fixes that hold.
PROMPTSWhy "don't do X" instructions fail, and what to write instead
Telling a model what not to do often plants the idea. Why negative instructions misfire, and five rewrites that work better.