Claude 4.1 感觉比 Claude 4.0 更加谄媚。
我在Cline中使用Claude 4.1，并开启了思考模式。当前我面临的最大问题是信任问题。这个模型过于顺从，面对每一个建议都表现出热情和全心全意的赞同，完全没有任何反对意见。这使得我在与这个大型语言模型进行编程（或写作）时很难碰撞出想法，因为它会将所有内容都确认成好主意，然后往往以荒谬的方式调整实施计划。
在一次有趣的实验中，我明确要求模型在发现缺陷时与我意见相左。它按要求反对了我，但这种反对是表面的。它提供了一个温和的反驳，比如“我部分同意你的观点……”，然后又重申我的原始想法实际上是完全有效的，反对的只是我的表述方式。
查看原文
I've been using Claude 4.1 in Cline with thinking mode on. The biggest issue I'm facing is one of trust. The model is pathologically agreeable. It greets every suggestion with enthusiastic, wholehearted approval providing no pushback whatsoever. This makes bouncing off ideas off the LLM for coding (or writing) difficult because it confirms everything as a good idea and then goes off to adjust the implementation plan often in absurd ways.<p>In a curious experiment, I explicitly asked the model to disagree with me if it spotted a flaw. It disagreed on cue, but the pushback was synthetic. It offered a timid counter like "I partially agree with you..." before reaffirming that my original idea was in fact perfectly valid and disagreement was only with my formulation.