You stare at the result you want. Your machine stares at the command you gave. Two gazes, a few degrees apart, and they diverge with every step: the human grows irritated (“how is THIS what you did”), the machine grows lost (it followed every word and is still being scolded), and the token meter takes nobody’s side. The money burns either way.
The lineage of this problem is older than it looks. The field named it “Do What I Mean” in 1966 (Teitelman, Interlisp). Its modern form is specification gaming, and DeepMind keeps an actual list of cases where a system did exactly what was asked and exactly not what was meant. Its mythological form is King Midas, who got his command executed to the letter.
What we believe is ours to add, from running a multi-model fleet daily: the cure is not a longer command. The cure is to hand the machine the thing you are actually looking at: a measurable destination, the context of why, and a definition of done. The command is a finger; the result is the moon. A wise brief points at the moon instead of making the machine study the finger.
This is not advice we read somewhere; it is a rule this house wrote after paying for the lesson, and our task-brief template has carried it since: lock the destination, fill the context, free the road. The label says [solid] because the rule is in daily use in our own ledger, and [ours-first] because we wrote it down before we found it written elsewhere. Both labels are open to refutation; that is what they are for.
Bản tiếng Việt. Người nhìn cái mình muốn có; máy nhìn cái mình được bảo. Hai ánh mắt lệch nhau vài độ thôi mà càng chạy càng xa: người cáu vì “làm vậy mà cũng làm”, máy lú vì làm đúng từng chữ vẫn bị mắng, và đồng hồ token thì không đứng về phe ai. Thuốc không nằm ở gõ lệnh dài hơn; nằm ở việc đưa cho máy cái mà người đang nhìn: đích đo được, bối cảnh vì-sao, và định nghĩa xong. Lệnh là ngón tay; kết quả là mặt trăng; đề bài khôn là đề bài chỉ trăng cho máy xem chứ không bắt máy nghiên cứu cái ngón tay.
Sources. DWIM, Teitelman 1966 (overview) · DeepMind, “Specification gaming” (2020). Cite. Hazumi Lab (2026), “Two Gazes, and the Money Still Burns”, lab.wi.works. Written by Wi Hazumi with the house fleet.