liveBook · Manning

Human values

As explained in LLM Evaluation and Alignment, The Foundational Ideas

Human values are socially and ethically grounded preferences about acceptable goals, conduct, and outcomes.

Human values are socially and ethically grounded preferences about acceptable goals, conduct, and outcomes. In AI alignment, they define part of what a system should do, alongside explicit policies and consistent personality traits. section 1.2 section 1.5

Treat human values as an alignment target rather than a direct software specification:

  • Translate values into observable behavioral expectations, then use evaluation to measure whether the system follows them. section 1.0
  • Separate foundational values from policy-based rules and personality-based communication styles. section 1.5
  • Expect judgment and trade-offs: helpfulness, coherence, tone, and acceptable behavior may lack one deterministic answer. section 1.0 section 1.5

Helpful does not mean agreeable

A helpful assistant that agrees with users even when they are wrong can become sycophantic. Alignment must therefore distinguish desired helpfulness from behavior that violates broader human values. section 1.0

How LLM Evaluation and Alignment, The Foundational Ideas covers it

How deep it goes
Defines human values as a foundational alignment pillar but does not specify a concrete method for eliciting or formalizing them.
Examples
Illustrates value tensions through helpfulness, appropriate tone, and sycophantic agreement with incorrect users.
What you can do after
Evaluate AI behavior against intended human goals, intentions, values, constraints, and instructions.
How it connects
Connects human values to alignment, evaluation, policy-based alignment, and personality-based alignment.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of LLM Evaluation and Alignment, The Foundational Ideas. The section links point to the parts of the book it draws on.