TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

Using ChatGPT to auto generate config files: Wasted effort

3 点作者 samuelstros将近 2 年前
I shared the following message about the learnings of using ChatGPT in production under the premise of &quot;this will replace mundane coding tasks&quot; in our Discord. The result is disappointing. I thought it might be of interest to the HN crowd.<p># Context<p>The release of ChatGPT lead to a discussion if ChatGPT could be used to auto generate a config file that takes the source codes structure into account. Manual prompting yielded amazing results, see the screenshots in [0]. But, turning the manual prompting into an automated system led to a massive effort that took &gt;2 months (not full-time, but still a lot of effort!). The idea of &quot;just fine-tune a prompt and done&quot; is, at the moment, far from true. Besides, the result is most likely wasted efforts due to:<p>a) simplifying the config made the requirement for GPT style &quot;write the config for me&quot; obsolete.<p>b) hardcoded heuristics are easier to iterate on, more cost efficient (0 costs) and easier to deploy (no API calls needed).<p># The message<p>FINALLY AUTO GENERATION OF THE CONFIG https:&#x2F;&#x2F;www.loom.com&#x2F;share&#x2F;85d77004aa4c4bea9752d959c229e577 .<p>The irony is that this is not needed anymore. The plugin API simplified the config interface so much that initializing configs only differs by the used plugins and their respective configs.<p>The only useful thing that chatgpt provides is deriving what files in a filesystem are translation files. But a hardcoded version might be faster and more resilient.<p>Learning: The GPT hype is ... well hype. The amount of work that went into this feature is insane. The work boils down to software engineering. Sure, tools will emerge that make prompt engineering easier but so does the previous ML hypetrain led to endless nocode machine learning tools. Yet, there is no &quot;disruptive everyone builds ML models&quot; (for GPT &quot;everyone will become a developer&quot;) nocode machine learning tool in sight. Maybe signaling that the premise of ML NoCode&#x2F;GPT of &quot;this works for everything&quot; is not true. Deriving what files are translation files, yep. Deriving an entire config file, maybe not.<p>## Footnotes<p>[0] https:&#x2F;&#x2F;github.com&#x2F;inlang&#x2F;inlang&#x2F;discussions&#x2F;408#discussion-4914727

暂无评论

暂无评论