Nikolay Petrov presents a controlled way to determine what a knowledge-oriented agent skill contributes. Break its instructions into small checkable propositions and contracts, write questions that do not disclose their answers, and compare the same model with and without the skill. A separate judge or panel can classify four outcomes: knowledge already present, newly enabled answers, regressions caused by the skill and information still missing in both conditions. He distinguishes broad conceptual understanding from exact parameters and procedural details; a model recognizing the general idea does not establish that detailed instructions are redundant. Knowledge novelty is only a guide to optimization. After removing or revising material, the complete skill must still pass task evaluations, with attention to token use and execution time. The talk explicitly separates knowledge skills from process or steering skills, whose value may lie in consistent behavior even when the underlying information is familiar. Its headline percentage is not a universal deletion target.
Nikolay Petrov: evaluate skill instructions before shortening them
Related research
More curated notes connected through AI Engineering and Model Evaluation.
Pacing model development in an era of cyber-critical capabilities
OpenAI says preliminary evidence that Astra may meet its Critical cybersecurity threshold led it to pause frontier reinforcement-learning work for two weeks and keep its largest planned run on hold. New safeguards include stronger workload and network isolation, continuous boundary testing, token-level monitoring that escalates suspicious tool activity, and broader alignment checks for deception, reward hacking, and unauthorized access.
A blueprint for democratic governance of frontier AI
OpenAI proposes a three-part U.S. frontier-AI governance model: harmonize emerging state safety laws into a federal baseline, strengthen CAISI as an evaluation and standards institution, and coordinate a broader resilience program. Proposed controls include severe-risk evaluations, transparency reports, independent audits, safety-incident reporting, model-weight security, whistleblower protection, and periodic technical assessments.
OpenAI’s Frontier Governance Framework
OpenAI's 22-page Frontier Governance Framework maps its frontier-model processes to California's Transparency in Frontier AI Act and the EU AI Act's general-purpose AI code. It documents lifecycle risk assessment, cyber-offense and other risk tiers, mitigation and residual-risk decisions, critical-incident handling, security risk management, model reporting, external review, responsibility allocation, and change control.