HackerNews中文版

在过去的几周里，我在我的副项目上花费不多，但这个周末，我在内容创作实验代理方面取得了一些进展！（还烧掉了一些代币）我对未来和伴随深度内容创作辅助而来的新问题感到兴奋。迭代是关键！就像大多数对话代理一样，在非文本内容中，保持与上下文相关而不臃肿更为重要，有时需要语言作为媒介，有时需要不同训练的模型来理解上下文——> 将其传递给另一个代理——> 存储在一个公共记忆中——> 让用户能够顺利迭代。- 不着急。我尝试用更少的轮次和更高的效率来解决这个问题，但从编码代理中获得灵感后，我意识到这关乎质量，慢一点也没关系，这样内容创作者就有足够的时间思考，并尝试与助手保持同步，就像我们在团队中一样。我解决这个问题的尝试在不断变化！当我开始时，看起来只是一些简单的点需要连接，但实际上尝试后发现这很困难。我最初试图像程序员一样解决这个问题，通过制作更多的代理工具，但人工智能更关心的是原始能力和优秀的示例，你不必将所有内容都包装在工具中。我觉得如果工具能够执行更广泛的任务，而不是创建无数专注的任务，代理的基本工作效果会更好。我可能在其他方面是错的，但根据我目前的实验和给定的上下文要求，这就是我学到的。我希望在我满意分享之前能够展示这个！（这里“满意”的定义是，当它能够提供价值时——当我能够自己完全迭代并发布长篇内容——播客/长论文/视频时，它需要在当前可用工具创建的一次性内容中脱颖而出）构建真正帮助创作者的人工智能工具的旅程比我想象的要复杂和迭代得多，但每一枚代币的花费都是值得的。

查看原文

I wasn't spending much on my side projects for the last few weeks, but I this weekend, made some progress on my content creation experimental agent! (and burnt some $ tokens too )I keep getting excited about the future and the new set of problems that are coming with assisted deep content creation.Iteration is the key! like most of the conversational agents, keeping what is relevant in the context without bloating is more important in non textual contents, it sometimes needs language as a medium, it sometimes need differently trained models to understand the context -> pass it to another nicely -> store it in a common memory -> let users iterate nicely.- no rush. I tried solving it with lesser turns and efficiency, but getting inspired from coding agents, i realised its about the quality, slow is fine so that content creators will get enough time to think and try to match the pace with the assistant, like we do in a team.My attempt to solve this is constantly changing! When I started, it looked like very simple dots to connect, but when I actually tried, it is difficult.I tried solving it like a programmer first by making more agent tools, but what AI cares about more is raw capabilities and excellent examples, you don't have to wrap everything in tools, I feel agents fundamentally works better if tools can do broader tasks than creating numerous focused tasks, I could be wrong in other aspects, but with my current experiments and the given context requirements thats what i learned.I hope I should be able to show this once I am satisfied of sharing it!(definition of "satisfied" here is that when it delivers value - when i am able to fully iterate and publish on a long form content myself -podcast/long paper/video, it needs to stand out with the one shot content created with currently available tools)the journey of building AI tools that actually help creators is messier and more iterative than i thought. but worth every token spent though.

如果你在意，使用人工智能创作优质内容仍然需要大量的迭代。