返回资讯中心
外部精选
软件工程
#开发者实时流

AI Is the Ultimate Vendor Lock

I’m amazed by how little people talk about AI being the ultimate vendor lock. And the reason I’m amazed is that this is a fundamental strategic issue. It’s an existential issue that, I feel, could easily wipe out most o…

Hacker News 最新ilreb2026/10/19 · 27 天前7 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

I’m amazed by how little people talk about AI being the ultimate vendor lock. And the reason I’m amazed is that this is a fundamental strategic issue. It’s an existential issue that, I feel, could easily wipe out most organisations that describe themselves as “AI native”, and yet, so far, I haven’t seen any contingency plans on their side.

The mechanism is simple (I’ll talk about software development, because that’s the one I’m most familiar with): You use AI to generate superhuman amounts of code. This, of course, seems awesome at first. Finally, you have the power to execute all the things that have been pushed to the bottom of the backlog for God-only-knows how long. You can build anything. The agent never gets tired, and if an output seems off somehow, you just repeat the prompt as long as it takes to fix it.

So, this is what you do, right? It’s the most natural thing to do anyway. Maybe the Claude bill does sting a bit, but it seems worthwhile. Developers are expensive too.

Well, three months down the line, there is not a single human on this planet that can understand your codebase. In fact, there is not a single human or a team of humans that could run the development operations. What can you do about a 20k LOC pull request other than blindly accept or reject? You can manually test, yes, but that’s like building a car blindfolded and saying “seems to run ok”. At this point, your only hope is that the agent somehow pulls off the new feature you promised in the last sales call.

The way I’ve seen development teams try to cheat their way out of the velocity issue is to have another agent review the code. That doesn’t really change anything, though, does it? You are still 100% reliant on the model, and you don’t have any idea what you are doing.

An apt question is, what do you even own at this point? To answer the most basic inquiries about the software that you are supposed to sell or run your business on, you need to burn 30k tokens on a frontier lab model. In a business strategy sense, you have made a weird rental agreement where you have surrendered the ownership of your ability to operate to some highly opaque third party.

What happens when the token costs become less favourable? What happens when there is some issue that prevents you from using the model you need to run your business? I’m sure there are ways to mitigate these issues, but then it seems like you need to either build things using an OSS model or wait till human understanding manages to catch up with the agents.

I’m sure all this is something people can come up with on their own. Why isn’t it talked about more? Usually, in fact, it seems to be the other way around: countless managers and executives are racing to surrender their operations to third parties whose interests are fundamentally unaligned. It’s not in the interest of a model provider to optimize for productivity, but for maximum token billing.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →