AI agents and language models
9 articles
- AI and moral improvement Whether AI advisors, Socratic interlocutors and coaches can improve people's own moral reasoning and conduct, and the dispute over autonomy, deskilling and manipulation.
- Agent skills SKILL.md-style instruction packages for AI agents and how they compose: calling, chaining, handing off, shared terms, fallbacks; analogies to software modularity.
- Evaluating agent instructions How to test whether prompt-based instructions for AI agents improve their work: baselines, graders, simulated users, outcome measures with people, variance; the dispute over how much such instructions help.
- Evaluation tooling for AI agents Frameworks and tools that run and score coding agents across harnesses, and the disputed comparability of their results.
- LLM-as-a-judge Using language models to grade outputs: how judges are configured and validated against human labels and swapped answer order, and the dispute over which agreement statistic to use.
- Multi-agent research systems How teams of language-model agents add, rank and cap new research questions during a run, and whether a coordinator should vet them, where the evidence is thin and borrowed from other fields.
- Sycophancy in language models When language models agree with, flatter or defer to users against the evidence: how it is measured, what it does to users, where it comes from, what reduces it, and the risk of over-correcting.
- Verbosity in language models Why language-model answers run longer than they need to, how density is measured, which methods shorten them, and the dispute over whether shorter is more accurate or better liked.
- Verifying code that uses encryption Formal verification of code that calls encryption APIs (envelope encryption, key wrapping, authenticated encryption with associated data), with the primitives assumed rather than proved; which properties can be stated, which tools check them, and how this differs from misuse detection.