请问HN：现在对GPT包装公司有什么看法？
你是否在使用这些产品？你觉得它们有价值吗？它们是否做了一些基础模型公司没有做的事情？
我大约一年前就对此产生了疑问，今天想重新探讨一下这个问题。
今天我在Product Hunt上做了一些研究，还浏览了最近的一些YC公司。让我立刻感到惊讶的是——哇！——有一些非常优秀的设计师（或者说是现在的AI）制作了精美的产品演示和1-3分钟的视频。我遇到的大多数产品都是针对各种需求的AI工具，比如创建办公文档、发送电子邮件、制作演示文稿、网络爬虫、开发移动应用、管理会议和人际关系等。
然而，当我超越表面的华丽，开始深入调查时，我并没有看到这些产品有什么特别之处，或者在某些情况下，它们与基础模型加上简单的RAG有什么不同。事实上，许多与文件相关和依赖记忆的应用程序，今天通过基础的Open AI聊天网站都能很好地完成。值得肯定的是，我认为这些产品通过各种提示工程技巧提供了更好的“流程”。但是否足以让我为它们支付比我现有的OAI/Anthropic订阅更高的月费，我就不太确定了。
但总体来说，当我看到这些产品时，我的感觉是它们可能并不是初创公司，至少不是我认为能持久的那种。它们必须越来越细分，以摆脱日益强大的基础模型的影响。那些看起来能够持久并且表现良好的大公司是Perplexity和Cursor，它们似乎在早期迅速成长，吸引了大量资源和人才来不断开发新功能。因此，也许基础模型只能完成它们大约60%的工作，而我上面描述的产品/初创公司的数字更接近90%。我的结论是对“GPT包装”持有一定的悲观态度，也许在更多创新想法（例如物理世界的应用案例）实现之前，因为SaaS领域似乎竞争激烈。
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.