Language Model Basics
A Token Count Is Not a Word Count
Measure text with the tokenizer used by the model instead of assuming that words, characters and tokens have a fixed one-to-one relationship.
AI · VISION · EVERYDAY SYSTEMS
Journal
AI Deployment Basics
A Quantized AI Model Needs Its Own Quality Check
Evaluate a quantized AI model as a separate deployment choice, measuring memory, task quality and compatibility instead of relying on the bit-width label.
Understanding AI
A Language Model’s Temperature Is Not a Truthfulness Setting
Temperature changes how a language model samples tokens. It can affect variation, but lowering it does not verify facts or guarantee the same answer everywhere.
Applied AI
Semantic Similarity Does Not Mean Two Sentences Agree
Understand why embedding similarity can retrieve related sentences that disagree, and test negation, numbers and conditions before using a score as evidence.
Digital Research
An AI Search Result Needs a Retrieval Date as Well as a Source
Preserve the date, page and supporting passage behind an AI-assisted search result so another reader can check what you actually reviewed.
AI Evaluation
A Faster AI Model Is Useful Only If You Measure the Whole Task
Compare AI workflow speed from user request to usable result, including retrieval, validation, retries and review rather than model response time alone.
Responsible AI
Synthetic Data Does Not Automatically Remove Privacy Risk
Ask how synthetic data was generated and tested before treating it as anonymous, and evaluate privacy separately from usefulness.
Understanding Models
A Model Card Helps You Ask What an AI System Was Tested For
Read an AI model card for intended uses, evaluation conditions and limitations, then identify what still needs testing in your own application.
AI Operations
Data Drift Can Change an AI System Without a Model Update
Understand how changing input data can affect a deployed AI system, and distinguish a drift alert from evidence that its predictions have become worse.