问责吞吐量
2 分•作者: shudv•大约 1 个月前
大多数人已经对人工智能感到厌倦了，我甚至都不想再谈论或讨论它了。但我还是会继续说下去，因为我相信这可能会为讨论增加一些价值。
人工智能公司吹嘘人工智能可以带来数倍（10倍、100倍、1000倍）的生产力提升，并声称如果我们不利用它，就会被落下。这种“错失恐惧症”（FOMO）导致了近年来疯狂的支出和裁员。但我认为我们需要讨论“责任”和“问责制”的问题，因为无论今天的人工智能有多好，我们还没有解决这个问题，而这才是真正系统的关键决定因素。
例如，作为一个拥有多年经验的工程师，我的数据是（我知道，这很菜）——
最大代码吞吐量（我一天能编写/修改的、能正常运行的正确代码量）= 约 200 行代码
最大问责吞吐量（我可以直接负责并交付给依赖该软件的广大用户群的代码量）= 约 50 行代码
以下是前沿模型相同的数字：
最大代码吞吐量：1,000,000 行代码（假设预算无限，为什么不呢）
最大问责吞吐量：0（这是致命一击）
这是一个简单的论点：真正的业务建立在问责制之上，而不仅仅是原始输出。如果我将原始输出提高 100 倍，但没有扩大我的问责制，那么经济价值仍然受限于问责制。因此，除非这些人工智能公司（OpenAI、Anthropic、谷歌）开始对生成的代码负责，否则模型有多好几乎无关紧要。它仍然受限于我的问责吞吐量。
查看原文
Most people have had enough of AI, to the point that I almost do not feel like talking or discussing about it anymore. But I will still go ahead because I believe it might add non-zero value to the discourse.<p>AI companies tout the multi-fold (10x, 100x, 1000x) productivity gains that AI can lead to and the fact that if we do not tap into it, we will be left behind. This FOMO has what has led to the spending and layoff frenzy in the recent years. But I think we need to discuss the aspect of "responsibility"/"accountability" because regardless of how good AI is today, we haven't yet solved this specific problem and it is THE critical determinant in real systems.<p>For example, as an engineer with some years of experience in my bag, here are my numbers (rookie I know)-<p>Max code throughput (Amount of working correct code I can write/modify in a good day) = ~200 LoC
Max accountability throughout (Amount of code I can directly take responsibility for shipping to a wide user base that depend on the software) = ~50 LoC<p>Here are the same numbers for a frontier model:<p>Max code throughout: 1000,000 LoC (assuming unlimited budget, because why not)
Max accountability throughput: 0 (this is the killer blow)<p>Here is the simple argument: Real business are built on accountability and not just raw output. If I increase the raw output by 100x but do not scale my accountability, the economic value is still bottlenecked on accountability. So unless these intelligence companies (OpenAI, Anthropic, Google) start taking accountability for the code that gets generated, it almost does not matter how good the model is. It's still bottlenecked on my accountability throughput.