Large Language Models: The Hard Parts

Chapter 6. LLM Safety Considerations

Move fast and be responsible.

Andrew Ng

Before we dive into benchmarks, guardrails, and code implementation, we need to establish a foundation for thinking about LLM safety comprehensively. Safety in the context of LLMs is a multifaceted challenge that spans technical vulnerabilities, ethical considerations, organizational policies, and societal implications. This chapter provides the conceptual framework we’ll need to make informed decisions when we move to the practical tools and techniques in subsequent chapters. Understanding why safety matters, what can go wrong, and how leading organizations approach these challenges will inform the technical decisions we make—from which benchmarks to run, to which guardrails to deploy, to how we structure our alignment training.

We begin with general AI safety principles because LLM-specific concerns exist within a broader context of ensuring AI systems behave reliably and beneficially. Questions like “How do we ensure systems do what we intend?” and “How do we prevent unintended consequences?” apply to all AI systems, but they manifest in unique ways with language models.

From there, we’ll explore LLM-specific safety challenges like jailbreaking—where users deliberately craft prompts to bypass safety measures—and more subtle issues like stealth editing, in which models might alter facts or inject biases in ways that aren’t immediately obvious. These aren’t hypothetical concerns; they’re active problems ...