各大主要供应商启动 RSI 循环还需要多久？
1 分•作者: foxindustrial•5 个月前
正如你可能已经知道的，所有主要的模型提供商似乎都在悄悄地降低用户体验的质量……
一两个月前还感觉是尖端“智能”的东西，现在经常提供的输出结果甚至不如 2023 年底的表现：模糊不清、胡编乱造、过于谨慎，或者干脆就是敷衍了事。
理性地说，将宝贵的浮点运算资源浪费在服务数百万休闲聊天用户、氛围代码编写者和粗制滥造者身上，机会成本是巨大的。
结果是，许多用户在各个提供商（Gemini 2.5/3 Pro、Claude Sonnet/Opus 系列、GPT-4o/5 系列，以及各种第三方界面，如反重力或编码前端）上反复遇到的一种现象：
你提出了一个非同寻常的问题（例如，代码、分析、创意工作、研究或其他），而你得到的却是 2023 年水平的鹦鹉学舌式的垃圾，如果它没有把你的代码搞砸，那就算你运气好了。
当被问及进行编辑的模型的确切命名时，这些模型最初会说它们是由谷歌、Claude 或 OpenAI 配置的“大型语言模型”……但一旦你坚持追问，它们就会透露全部信息……然后，瞧：原来你使用的是现有的最旧的模型。
当你随意地问模型它自己的身份时，它会默认使用预先设定的官方说辞：“我是一个由 [在此处插入工具] 构建的 LLM，由 [在此处插入提供商] 配置。”
再追问，它就会向你透露模型的实际命名：你甚至可能正在使用 GPT 2 (哈哈)。
我正在像收集宝可梦一样收集它们，我遇到了 Gemini 1.5 Pro、Gemini Flash 2.0、Claude Haiku。
我希望你尝试提问、坚持追问或使用一些巧妙的提示来提取模型名称，你会发现真相。一个专业提示是，当它告诉你使用量“异常高”时，你可以在任何界面中提问……
附言：我都是所有服务的专业订阅用户。
查看原文
As you might have known, all the major model providers seems to be quietly turning the dial down on the consumer experience...<p>What felt like cutting edge "intelligence" 1-2 months ago now frequently delivers outputs that wouldn't have impressed you in late 2023: vague, hallucinated, overly cautious, or just outright lazy.<p>Rationally speaking, the opportunity cost of wasting premium FLOPs on serving millions of casual chat users and vibe-coders and slop-makers is enormous.<p>The result is a phenomenon many users have encountered repeatedly across providers (Gemini 2.5/3 Pro, Claude Sonnet/Opus variants, GPT-4o/5 series, and 3rd party interfaces like various anti gravity or coding frontends):<p>You prompt for something non-trivial (e.g. code, analysis, creative work, research or whatever) and you get back the most sophisticatedly parroted 2023 tier mega slop and it would be a lucky instance if it didn't shit on your code.<p>When asked about the exact nomenclature of the models which are conducting edits, the models initially say that they are "large language models" configured by google or claude or openai...but once you insist they will reveal the whole thing... and et voila: it turns out you are using the oldest models available<p>When you casually ask the model to identify itself, it defaults to the scripted party line: "I'm a LLM built by configured for <insert tool here> by <insert provider here>"<p>Press harder, and it will reveal to you the actual nomenclature of the model: you might actually be able to access GPT 2 (lmao).<p>I'm collecting them like pokemon, i encountered gemini 1.5 pro, gemini flash 2.0 ,claude haiku.<p>I hope you try to ask insist or do some clever prompting to extract the model name and you will find out. Pro Tip is you asking in whatever interface exactly when it tells you that usage is "unusually high"..<p>PS: I'm a pro subscriber on all..