ETHICS AND SOCIETAL IMPACT

ENVIRONMENTAL AND RESOURCE CONSIDERATIONS Running models uses electricity and hardware, and larger models generally use more. This book does not quote figures here, because they vary widely and change quickly. The principle is easy to state, though. Use the smallest model that does the job well, avoid needless repeated calls, and cache or reuse work when you safely can. Our lab has already practiced this, since routing the simple matching step to a cheap tier and ending runaway loops early both reduce wasted computation. Cost and energy often move together, so good cost engineering is frequently good resource engineering. Long-Horizon Safety: Autonomy and Over-Reliance As agents take on longer tasks with less supervision, new concerns appear. One is autonomy: the more freedom an agent has, the more damage a mistake can do before anyone notices. The answer is not to ban autonomy but to match it to the stakes, with tighter limits, approvals, and logs where harm could be serious or hard to undo. Another is over-reliance. When a system is right ninety-nine times in a hundred, human reviewers drift toward approving everything without looking, which is called rubber-stamping. The safeguard that was supposed to catch the hundredth error then quietly stops working. Good design anticipates this human tendency. One tactic is to occasionally slip in a draft that is known to be flawed and see whether reviewers catch it. Another is to rotate reviewers, require them to write a reason for approvals on higher-risk items, and track how long they spend on each review.