دانلود کتاب هنر پست آموزش (Post-Training)، راهنمای عملی برای مهندسان و توسعه‌دهندگان هوش مصنوعی

Rabbi Judah Loew ben Bezalel, known as the Maharal of Prague, was a 16th-century Jewish scholar, philosopher, and something of a mystic. Witnessing the persecution of his people, so the legend goes, he fashioned them a protector from clay gathered on the banks of the river Vltava. Forming the clay into the shape of a man, he carved a word into its forehead, and the word gave it semblance of life (or “bounded autonomy,” as we would say today). This is, of course, the story of the Golem of Prague: Immensely powerful, tireless, and obedient, it existed to protect those who could not protect themselves. But the Golem was also dangerous. It was a creature of raw capability without inherent purpose, animated matter that might, if not properly controlled, turn its strength against the very people it was meant to serve. The tales vary in their details, but they converge on a central anxiety: What happens when the things we create become too powerful for us to control? We are, it seems, wired to fear this. The same anxiety recurs across cultures and centuries: Frankenstein’s monster, Skynet, HAL 9000, the paperclip maximizer—each a powerful construct, shaped by human hands, that escapes human mastery. It is a primal story, an articulation of our deepseated instinctive discomfort with progress, and artificial intelligence (AI) has given it but a new dress. When we speak of existential risk from AI, when it is blamed for potentially destructive outcomes (rightly or wrongly), or when we worry about superintelligent systems pursuing goals misaligned with human welfare, we are ultimately telling and retelling a tale as old as time. The Maharal animated his creation by inscribing a shem—a word—upon its forehead. Shem in Hebrew means name, and the particular name the Maharal chose was emet: truth, or justice. When the Golem’s work was done or threatened to become uncontrollable, the Maharal could erase a single letter, transforming emet into met—death—and the creature would return to lifeless clay. The word gave power. The word could also take it away. This, I want to suggest, is the short story of post-training. The foundation model—the raw language model that emerges from pre-training—is but unformed clay. It has absorbed the patterns of human language from billions of words, acquiring remarkable capabilities in the process, and with sufficient forming, it can be given both the verisimilitude of humanity and power well beyond that. But it has no purpose, no values, no reliable behavior. It is potential without direction. Post-training is the shem: the process by which we give the model its animating principles, its sense of what it should and should not do, its capacity to be helpful rather than harmful. The word—whether we call it a constitution, a reward signal, or a set of instructions—is how even fearsome power can be shaped and controlled. But the story of the Golem is more than a mere cautionary tale about the destructive potential of uncontrolled power. It also teaches us that capability alone is not destiny, that the relationship between power and purpose is not a given. It is something we, ultimately, impose. No matter how mighty our servant, it is we who inscribe that shem. The story teaches us that words matter, that the careful articulation of intent, translated into code and training regimes and feedback loops, is how we ensure that our creations serve rather than threaten. And most important, it teaches us that we have a choice but also a responsibility. The Golem isn’t something that happens to us. It is what we make it to be. But only if we master an understanding of that process.