Paweł Huryn, Helen Toner, and Ben Lang posted new notes

Paweł Huryn, Helen Toner, and Ben Lang posted new notes

͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­͏     ­

Paweł Huryn, Helen Toner, and Ben Lang posted new notes Paweł Huryn You follow 6d AI agents expose a problem we've ignored for decades: Humans navigate unclear intent through culture, observation, and guesswork. Agents can't guess. They force us to finally codify what we never could. The counterintuitive fix is not more detailed instructions, but giving agents the reasoning framework to handle the unknown. Core elements (guide both humans and agents): Objective: The problem to solve + WHY it matters Desired Outcomes: Measurable states indicating success Health Metrics: What must not degrade Strategic Context: The system we operate in Agent-specific elements (add as autonomy increases): Constraints: Steering (prompts) vs. Hard (orchestration) Autonomy: Which decisions require escalation Stop Rules: When to halt, escalate, or complete — How to apply each: 1. Objectives Problem-focused: What's broken or missing Explains the why: Business value, user impact Guides trade-offs when facing ambiguity 2. Outcomes Observable state changes, not activities From user/stakeholder perspective Measurable, verifiable… Read More 15 3 2 Helen Toner You follow 5d As AI systems get more capable, we desperately need more & better “science of AI” - how do these things… Read More One comparative psychologist lamented the bias against ‘killjoy explanations’—ones that pop the bubble of exciting-sounding phenomena. I think the same sentiment shows up in…” — Melanie Mitchell 47 5 4 Ben Lang You follow 5d I put together a new list of… Read More 17 1 1 See more notes in the Substack app © 2026 Substack Inc. 548 Market Street PMB 72296, San Francisco, CA 94104 Unsubscribe 472