Cristóbal Valenzuela：当我谈论控制时，我在谈论什么
当我谈论控制时，我在谈论什么
做出一样东西与拥有制作它的技能，过去似乎紧密相连。如今，成果不再能证明你具备这种技能。还是说，它仍然能证明？
文生视频刚开始奏效时，给人的感觉就是如此。你只需输入一段简短的文字描述，按下按钮，就完成了。你可以把它看成魔法，也可以看成作弊。它能够工作本身就是突破，但看起来制作它的人也没做多少事。我们创造了类似生成式视频“自动模式”的东西，而“自动”并不是人们会与创作技艺联系起来的词。没错，写提示词是一项技能，但仅仅更善于“描述”自己想做的东西，在过去并不足以让我们认可为一种技艺，除非你的创作本身就是华丽的文字。
随着模型能力增强，创作者希望有更多自由度去尝试。新增的选择空间，也就是运用判断力的新空间；足够多的判断，才让技艺成为可能。创作者开始对模型形成直觉，就像人们对相机、剪辑软件或任何其他媒介形成直觉一样。
新能力往往在我们理解它之前就出现了。因此，我们先做出力所能及的界面，用最基本的可行控制方式和界面起步。这就成了新工具最初的自动模式。随后，人们使用它、把它推到极限，也十分明确地告诉我们，他们接下来希望能做什么。弄清哪些新的基础能力真正强大、用户在哪些方面需要更多控制，都需要时间。
人们想要更多控制，也想要更多自由度。我认为，模型变得更可控，部分原因就在于它们变得更加个性化。AI 从根本上说是一种通用技术，你可以按自己独特的需求塑造它。模型是能够适应你需求的泛化机器。每个人使用模型的方式不同，发现其用途、用它提高生产力的方式也不同。
我坚信，我们用来控制模型的大多数界面将为个人量身定制，而非通用界面。不是多对一，而是一对一。你能想象的任何控制方式，你想要的任何版本的 Runway。很快就会到来。
对照原文
What I Talk About When I Talk About Control Making something and having the skill to make it used to feel tightly linked. Now the outcome no longer proves you have the skill. Or does it? When text-to-video first started working, that’s what it felt like. All you had to do was type in a short text description, press a button, and voila. It was magic or cheating, depending on how you looked at it. The fact that it worked at all was a breakthrough, but it also didn’t look like the person making the thing had done very much. We’d created something like auto mode for generative video, and “auto” isn’t a word we associate with craft. Yes, prompting is a skill, but being a better “describer” of the thing you want to make isn’t something we’ve historically been satisfied with as skill, unless your craft itself was flowery prose. As the models get more capable, the makers want more degrees of freedom to try things. New places to make choices become new places to exercise judgment, and enough judgment creates the possibility of craft. Creators start developing intuitions about the models, just like people have developed intuitions about cameras, editing software, or any other medium. New capabilities show up before we can understand them. So we build whatever interface we can manage to. Minimum viable control and minimum interfaces to get started. That becomes the first auto mode for that new tool. Then people use it, push it against its limits, and give us pretty loud hints at what they want to be able to do next. It takes time to figure out which new primitives are powerful, and where users want more control. People want more control. And they want more degrees of freedom. I believe that part of what makes models more controllable is that they become personal. AI is fundamentally a general purpose technology that can be shaped in any unique way you want. Models are generalization machines that can adapt to what you need. Everyone has a different way of using models. Different ways of finding them useful and productive. I strongly believe that most of the interfaces we use to control models will be individual-specific instead of general. Not many to one, but one to one. Any form of control you can imagine. Any version of Runway you want. Soon.