
How AI Safety Works: Risk Frameworks, Capability Thresholds, and Regulators
Mei-Ling ChenAI safety is the discipline of preventing harm from AI systems through measurement, mitigation, and governance, and it runs on two tracks.
AI tools precise and unhyped.

AI safety is the discipline of preventing harm from AI systems through measurement, mitigation, and governance, and it runs on two tracks.

Inference — running a trained AI model to answer a query — has collapsed in price at a fixed performance level: GPT-3.5-level inference cost fell more than 280-fold between late 2022 and late 2024, per Stanford HAI's 2025 AI Index Report.

There is no single global AI law.

Prompt engineering is the practice of writing and refining the instructions given to a language model, and the major vendors' official guides agree on the discipline before the tricks: define success criteria, test empirically, then iterate on a draft.

A training data pipeline is the industrial process that turns raw web crawls into the clean, deduplicated token streams large language models learn from.

AI hardware is built around one dominant operation: multiplying large matrices of numbers, in parallel, over and over.