HackerNews中文版

你是否使用过这些产品？你觉得它们有价值吗？它们是否做了基础模型公司没有做的事情？大约一年前，我就思考过这个问题，现在想重新审视一下。今天，我在Product Hunt上做了一些调查，也浏览了一些最近的YC公司。我立刻注意到的是——哇！——有一些非常棒的设计师（或者现在可能是AI）正在制作精美的产品演示和1-3分钟的视频。我遇到的的大多数产品都是AI，用于各种事情，比如创建办公文档、发送电子邮件、制作演示文稿、网页抓取、制作移动应用程序、管理会议和人际关系等等。然而，当我摆脱了这些花哨的东西，开始深入研究时，我并没有真正看到这些产品有什么特别之处，或者说它们与基础模型+一些基本的RAG（检索增强生成）有什么区别。事实上，许多与文件相关和依赖记忆的应用程序，今天完全可以通过OpenAI的聊天网站完美地完成。公平地说，我认为其中一些产品通过各种提示工程技巧，拥有更好的“流程”。但这是否足以证明比我现有的OAI/Anthropic订阅，每月支付一大笔费用，我不知道。但当我看到其中许多产品时，我的总体看法是，它们可能不是初创公司，至少不是我看到能持续下去的那种。它们必须越来越专注于细分市场，才能摆脱能力越来越强的基础模型的触角。那些似乎能够持续发展并做得好的大公司是Perplexity和Cursor，它们似乎在早期发展得足够快，吸引了大量的资源和人才来不断构建功能。所以，也许基础模型只能做到它们大约60%的事情，而对于我上面描述的产品/初创公司来说，这个数字更像是90%。我的结论是，对“GPT包装器”持谨慎态度，也许要等到更多有创意的想法（例如，现实世界的用例）实现，因为SaaS领域似乎很艰难。

查看原文

Do you use any, do you find them valuable, do they do something the foundation model companies are not doing?<p>I wondered this about a year ago and wanted to revisit the question.<p>Today I did some research on Product Hunt and also looked through various recent YC companies. What hit me immediately was - wow! - there are some really good designers (or maybe it's with AI now) putting together slick product demos and 1-3 min vids. Most of the products I came across are AI for various things like creating office docs, sending emails, creating presentations, web scraping, making mobile apps, managing meetings and relationships, etc.<p>However, when I get past the flashiness and start investigating a bit, I don't really see what's special or setting these products apart from a foundation model + rudimentary RAG in some cases. In fact, many file-related and memory-dependent applications can be done perfectly well via the base Open AI chat website today. To give credit where it's due, I think some of these products have better 'flows' via various prompt engineering tricks. Enough to justify a big monthly stipend over my existing OAI/Anthropic subscriptions, I don't know.<p>But my overall take when I see many of these, is that they are probably not startups, at least not ones that I see lasting. They have to go more and more niche to get away from the tentacles of increasingly capable foundation model capabilities. The big ones that seem to be enduring and doing well are Perplexity and Cursor, they seem to have grown quickly enough in the early days to attract a lot of resources and talent to keep building features. So maybe the foundation models can only do ~60% of what they do, while for the products/startups I described above, that number is more like 90%. My conclusion is to be somewhat bearish on 'GPT wrappers', perhaps until more creative ideas (e.g. physical-world use cases) come to fruition, because the SaaS space seems rough.

Ask HN： 如今大家对 GPT 封装公司怎么看？

Ask HN：如今大家对 GPT 封装公司怎么看？