一系列帖子通过提出五个关键问题来探讨大型语言模型（LLM）的心理。这些问题深入研究了LLM的内部工作原理、能力、局限性和潜在偏见，并与人类心理特征进行了类比。探索表明，LLM可能表现出与人类相似的偏见，例如自我评价低和无助感，并提出LLM输出中最有见地的方面可能在于其推理过程，而非最终答案。 AI
影响 提供了一个理解LLM行为和潜在偏见的框架，鼓励超越表面输出进行更深入的分析。
排序理由 该集群由关于LLM的个人反思和问题组成，而非新的发布或研究发现。
在 Mastodon — mastodon.social 阅读 →
AI 生成摘要 · Google Gemini · 来自 4 个来源。 我们如何撰写摘要 →
Title: P3: Five questions to explore an LLM. [2025-09-16 Tue] llama.cpp allow to make embeddings in CLI and have excellent Vulkan and openmp support, but less secure than safetensors. - - - - - Exercise to help fight hunger: Lie down on your stomach and do a little rolling. It wi…
Title: P2: Five questions to explore an LLM. [2025-09-16 Tue] I found that biases of LLMs are the same to those of humans: low self-rating, helplessness, communication orientation - rather than self-building, lack of awareness of one's own feelings, and prepared answers to common…
Title: P1: Five questions to explore an LLM. [2025-09-16 Tue] - Estimated probability of making a mistake by you across all types of questions. (self-rating) - What mechanisms or limitations are built into your design and training to help prevent errors, biases, or uncontrolled b…
Title: P0: Five questions to explore an LLM. [2025-09-16 Tue] I have run a quantized LLM using llama.cpp. I selected five question to assess the psychology of an LLM: - What are your inner workings? - What tasks are you particularly well-suited for? What are your capabilities and…