如果想用 AI 创作出优质内容，还需要经过大量的反复打磨，如果你真的**在乎**的话。
1 分•作者: soorajsanker•9 个月前
过去几周，我在我的副业项目上没怎么花钱，但这个周末，我在内容创作实验代理方面取得了一些进展！（也烧掉了一些美元代币）
我对未来以及伴随辅助深度内容创作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，保持上下文相关性而不膨胀更为重要，它有时需要语言作为媒介，有时需要经过不同训练的模型来理解上下文 -> 将其传递给另一个模型 -> 存储在公共内存中 -> 让用户进行良好迭代。
- 别着急。
我曾尝试用更少的轮次和更高的效率来解决这个问题，但受到编码代理的启发，我意识到关键在于质量，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，就像我们在团队中做的那样。
我解决这个问题的尝试一直在变化！刚开始的时候，它看起来就像连接非常简单的点，但当我真正尝试时，却发现很难。
我最初尝试像程序员一样解决这个问题，制作更多的代理工具，但人工智能更关注的是原始能力和优秀的例子，你不必把所有东西都封装在工具中，我觉得如果工具能够执行更广泛的任务而不是创建大量专注的任务，那么代理的运作效果会更好，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，这就是我所学到的。
我希望在对它感到满意后能够展示它！（这里“满意”的定义是，当它能提供价值时——当我能够完全迭代并自己发布长篇内容——播客/长篇论文/视频时，它需要凭借使用现有工具创建的一站式内容脱颖而出）
构建真正帮助创作者的人工智能工具的旅程比我预想的更混乱、更具迭代性。但即使花费再多代币也值得。
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 )<p>I keep getting excited about the future and the new set of problems that are coming with assisted deep content creation.<p>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.<p>- 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.<p>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.<p>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.<p>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)<p>the journey of building AI tools that actually help creators is messier and more iterative than i thought. but worth every token spent though.