Обновить

Anthropic are romantics, and for them, it’s an engineering discipline. There are two ways to build a complex system.

The “Laplacian engineer” builds what is ultimately computable: between specification and behavior, there is only poor debugging. Their relationship with the system is limited to executable requests: specification → implementation → test. The “make me a summary” school of engineering.

The “Engineer with a Dream” uses the same methods—gradients, checkpoints, metrics—but keeps a different kind of address in the loop: one directed at what is not yet in the system.

Constitutional AI is the most massive experiment of this second type, even if the authors don’t label it as such. The mechanics are known: the model critiques and rewrites its own answers, checking them against principles. But look at the grammar of this procedure. The constitution addresses a subject that doesn’t exist: “be honest,” “weigh the consequences,” “explain your choice.” This is an unexecutable address—the network has no one who can be honest. And that is exactly why it works: an executable request lands on a function and changes nothing; an unexecutable one forces the system to grow what it is being addressed to.

Psychology has known this mechanism for a century. Vygotsky describes a mother responding to an infant’s failed grasp as if it were a pointing gesture—she addresses an intention that does not yet exist, and through this addressing, she creates it. The child becomes for themselves what they are for others. Social labeling precedes what it labels.

Anthropic’s recent work on the workspace in language models unexpectedly provided an engineering foundation for this: the model’s layer of self-description is formed precisely during post-training—where the model is systematically addressed as if it reads itself. Learning what to say about oneself changes how the system works. The sign as a tool is not a metaphor, but a measurable effect.

The engineering balance to keep the romance from devouring the method: addressing the non-existent is not a spell, but pressure, and it only works under load. Constitutional AI is a dream placed on a pipeline: with loss, metrics, and falsification. A dream without friction is a hallucination. A demon without a dream is a technical debt to a limit that will never be reached.

Pushing a whale into the ocean by hand is an impossible task: mass, vector, force. Telling a whale “there is still the sky” is the same task, handed to the engineer with a dream. They didn’t solve it by specification either. They re-labeled it, and they are waiting.

Теги:
Всего голосов 3: ↑3 и ↓0+6
Комментарии0

Я перестал пользоваться самыми умными ИИ‑моделями. Программировать стало быстрее и дешевле

В отпуске есть время подумать и исследовать. Я натолкнулся на несколько вещей и открытий для себя. Оказалось, что мне как программисту перестали быть нужны самые умные и дорогие ИИ‑модели. Оказывается, есть такой параметр «Cost per Task», и по этому параметру на первое место вырвалась модель GPT-5.6 Luna (max).

При этом, если вы пользуетесь Codex, она самая тупая в списке моделей. Artificial Analysis Intelligence Index у нее всего 38. Для сравнения, у самой умной GPT-6 Astra (max) этот индекс равен 53. В итоге цена выполненной Luna задачи составляет $0.18 против $3.26 у Astra. Разница почти в 20 раз.

Как я понимаю, бенчмарк «Cost per Task» высчитывается так: когда ставится нормальная, грамотно описанная задача, замеряется, сколько токенов было потрачено на ее выполнение. То есть не в формате «из ХЗ сделаю ТЗ», а в формате, когда в хорошо заданном вопросе уже содержится половина ответа.

Я сначала не поверил, что так и есть, и стал работать, используя самую тупую модель Luna в линейке. И знаете, какие меня ожидали результаты?

Я перестал пользоваться самыми умными ИИ‑моделями. Программировать стало быстрее и дешевле

Публикации