我正在减少使用大型语言模型。
I'm dialing back my LLM usage

原始链接: https://zed.dev/blog/dialing-back-my-llm-usage-with-alberto-fortin

阿尔贝托·福尔廷是一位拥有15年经验的软件工程师,他最初热情地拥抱大型语言模型(LLM),但在使用LLM用Go和ClickHouse重建他的基础设施后,意识到LLM的局限性。他的经历凸显了软件开发中AI炒作与现实之间的差距。福尔廷发现LLM生成的代码质量和可维护性 often 差强人意,导致了不断修复AI生成的bug和意外后果的循环。他强调开发者必须对自己的代码库有深入的理解,并将LLM视为助手而非架构师。虽然承认AI有可能提高编码效率,但福尔廷告诫不要过度依赖,主张采取平衡的方法。他建议将LLM集中用于较小、定义明确的任务,例如重构,而将较大的架构决策留给经验丰富的工程师。他的主要结论是:要降低期望,利用AI来增强现有技能,而不是完全取代它们。最终,福尔廷认为AI是一项革命性技术,但我们仍处于其发展的早期阶段。

A popular Hacker News thread discusses the challenges and benefits of using LLMs for coding. One common sentiment is that LLMs produce messy code that developers feel less ownership of, making maintenance difficult. Users describe feeling like "the new guy" on their own projects. While LLMs can generate code quickly, they struggle with creating a mental model of the entire codebase. However, some users find LLMs beneficial for tasks like generating boilerplate, brainstorming ideas, and offloading tedious work. Success often depends on clear instructions, careful review, and a well-defined architecture. Some liken using LLMs to managing a team of junior developers, requiring strong leadership and communication. The discussion also highlights the importance of balancing AI assistance with human expertise. While LLMs can boost productivity, relying too heavily on them can hinder problem-solving skills and lead to technical debt. Ultimately, the thread suggests that LLMs are valuable tools, but they require responsible use and a solid understanding of software engineering principles.
相关文章

原文

We invited Alberto Fortin, a seasoned software engineer with 15 years of experience, to share his candid journey with AI. Alberto initially embraced LLMs with genuine enthusiasm, hoping they would revolutionize his development workflow. However, after encountering significant challenges while rebuilding his infrastructure with Go and ClickHouse, he wrote a thoughtful blog post reflecting on the gap between AI hype and reality. For this conversation, Alberto also prepared a detailed follow-up analysis testing newer models like Claude Opus 4, examining whether recent improvements have addressed the core issues he encountered.

His experience provides practical lessons for engineers evaluating LLMs in production environments—balancing realistic expectations with an understanding of where these tools genuinely add value versus where they still fall short.

You can watch the session on YouTube or read below for some selected quotes.

"I was really shocked at the poor quality of some things, and it was not just about bugs and features not working. I think as a developer who wants to maintain this codebase for the next few years, I also care about it being neat enough."

"I feel like I'm a week away from fixing this, but actually a new small error would come up and then that will take another two weeks to fix."

"I will give my error output to the LLM and then it will spit out something new that will kind of fix it, but also make things a bit more messed up—and break something else in the process."


"I think everyone just got a little bit overexcited about it because the first iteration, the first little feature, the first autocomplete is like, 'Oh my God, this is amazing. This is like reading my mind.' So you kind of get duped into it a little bit."

"I think we've gotten to a level where we can do probably 10 times as much coding. So we kind of expect that to happen and we require that from the LLMs, but I think everyone just gets a little bit overexcited about it."


"I think this is the biggest difference, like a mental shift... I am the software engineer, the senior software engineer, I am the architect. The LLM is the assistant. The assistant responds to me; I make the plan."

"I lost all my trust in LLMs, so I wouldn't give them a big feature again. I'll do very small things like refactoring or a very small-scoped feature."

"I started fixing the bugs myself. Because as soon as you understand this—you have a hundred percent understanding of your codebase and what everything is doing—it's so much easier and quicker for you to go in and fix something."


"If you are confident enough in your skills—you know, a senior developer—and this is not working for you, there's nothing wrong with you. Just try to do the things that you always did and use AI to leverage your knowledge a little bit."

"We've gone up a level, it's great. But also, let's be mindful we're not there yet at the next level... We are offloading some of the programming work, but we still need to do architectural abstractions and make the decisions for the product."

"Let's just try to calm down all this hype and find a balanced approach towards AI. Use it, because I think it's such an amazing revolution in technology, but we're not there yet."


联系我们 contact @ memedata.com