Ask HN：哪个大语言模型既有能力，又足够客观，还不会溜须拍马？
1 分•作者: mettamage•9 个月前
就我个人而言，我测试了 ChatGPT、Claude、Deepseek 和 Gemini。除了 Gemini 之外，其他 LLM 都过于谄媚，以至于除了基本问题和编码（Claude）之外，它们都无法使用。
Gemini 感觉有点像个谄媚者，但根据我的测试，可以说它在保持客观的同时，也在采取外交手段。至少，在我（Gemini Pro 2.5）的小测试中是这样的。这比其他 3 个要好得多。
你们的体验如何？我有点厌倦这种行为。我没有时间和金钱去测试 Grok 和其他模型。
至少，当我说 2 + 2 = 5 时，没有任何一个 LLM 会让步。但如果给它们真正模棱两可的东西，它们就会屈服于即使是最愚蠢/明显/透明的挑战。
查看原文
Personally, I've tested ChatGPT, Claude, Deepseek and Gemini. Other than Gemini, the other LLMs are way too much of a sycophant to the point that they're unusuable other than basic questions and coding (Claude).<p>Gemini feels a bit like a sycophant, but based on my testing, it can be argued that it's being diplomatic while staying objective. At least, in the small tests I've (Gemini Pro 2.5). And that's a lot better than the other 3.<p>What are your experiences? I'm getting a bit sick of this behavior. I haven't had the money and time to test Grok and others.<p>At least, no LLM would budge when I insisted on saying that 2 + 2 = 5. But give them actually ambiguous stuff and they will bend the knee to even the most silly/obvious/transparent challenges.