Large Language Model

Chain of Thought能显著提升模型在数学推理、逻辑判断等需要多步推导的任务上的表现

Large language models have an inherent limitation of a knowledge cutoff date, meaning the model only contains information up to the cutoff date of its training data.

Decision models are a specialized architecture distinct from generative language models, with the core goal of outputting structured judgment results for given inputs rather than generating natural language text.

Large language models are demonstrating increasingly strong capabilities in mathematical reasoning, theorem proving, and formal verification.

RAG（检索增强生成）是一种结合信息检索与生成式大模型的架构范式，模型生成答案前先根据用户查询从外部知识库检索最相关的文档片段，再将其作为上下文注入模型

Security alignment uses training methods such as RLHF and Constitutional AI to make large language models adhere to human values and safety guidelines, with mainstream commercial models incorporating built-in input filtering layers, output monitoring layers, and weight-level value constraints.

Research has found that the chain-of-thought reasoning generated by models does not fully correspond to their actual internal computation processes, and may provide post-hoc explanations unrelated to the true decision-making path.

Each model invocation consumes both input tokens (prompts and context) and output tokens (the model's generated response), with the two typically priced separately.

Large language models generate text by learning statistical patterns and cannot mechanistically guarantee logical correctness.

The context window is one of the fundamental limitations of current large language models, as a model can only process a limited number of tokens in a single inference.

Every inference operation of a large language model service consumes substantial GPU computing power, making capacity planning far more difficult than for ordinary applications.

Because large language models are based on probabilistic sampling, even if instructions are written in the context, the model may still forget or skip certain instructions in complex reasoning chains, which is known as instruction-following instability.

Large language models generate content based on predicting the next most likely token word by word, rather than retrieving answers from a factual database.

大语言模型本质上是一个基于概率的文本预测系统，其核心机制是「下一个词预测」（Next Token Prediction），并不真正理解语义

Critics question whether AI's breakthrough represents genuine mathematical reasoning or merely advanced retrieval and recombination of solutions already present in training data.

Large language models can automate the analysis of malicious code, generate exploit scripts, and simulate social engineering attack scripts, compressing work that would previously take highly skilled hackers hours down to minutes.

The training corpora of large language models may include unpublished academic preprints and research notes, raising concerns about data contamination.

Large language models typically trigger built-in safety alignment mechanisms when facing topics such as political elections and evaluations of candidates.

When fine-tuning large language models sequentially, full-parameter fine-tuning almost inevitably leads to significant degradation of existing capabilities.

In large model scenarios, typical multi-layer defenses include: alignment fine-tuning and red teaming during training, input filtering and prompt hardening during inference, toxicity classifiers and content moderation at the output stage, and post-deployment runtime monitoring and anomalous behavior detection.