大语言模型（Large Language Model，LLM），是基于深度学习技术、经过海量文本数据训练的自然语言处理模型。其核心能力包括文本理解、生成、推理与对话，能够处理翻译、摘要、代码生成、问答等多种语言任务。LLM通常采用Transformer架构，参数规模从数十亿到数千亿不等，是当前人工智能应用的重要基础组件。
Hallucination refers to LLMs generating plausible-sounding but factually incorrect information
90%UnverifiedKarpathy提出LLM特别擅长循环直到达成目标，应该给它成功标准而非告诉它做什么
90%UnverifiedAI-powered phone answering involves ASR to convert speech to text, an LLM to understand intent and generate responses, and TTS to convert replies into voice output
90%UnverifiedA core limitation of current large language models is the finite context window
85%UnverifiedLLM网关提供请求频次控制、Token消耗统计、缓存命中、日志记录和成本预估等功能
85%Unverified系统中调用LLM接口的场景包括新闻分析、信号生成、Agent决策和经验总结等
85%UnverifiedIn traditional programs, sleep is set by programmers based on clear technical requirements, while an LLM's sleep calls come from the model's own judgment
85%UnverifiedMainstream Agent architectures typically use a large language model as the brain, combined with memory modules, planning modules, and tool interfaces to accomplish complex tasks
85%UnverifiedLLM在代码生成场景中引用软件包时，是基于上下文语义和训练数据模式合成一个看似合理的包名，而非从实际的包注册表中检索
85%UnverifiedLanguage models are fundamentally probabilistic text generation systems that produce content based on statistical patterns in training data rather than understanding facts
85%UnverifiedAPI call prices for mainstream large models have dropped by orders of magnitude over the past two years
85%UnverifiedAI终端助手将大语言模型与命令行环境深度集成，可将用户自然语言描述解析为Shell命令并直接执行
85%VerifiedAI Agent底层通常由大语言模型（LLM）驱动，配合工具调用、记忆模块和规划框架（如ReAct范式）实现复杂任务自动化
80%Unverified不建议在claude.md中放入代码风格指南，因为与传统Linter相比LLM既昂贵又缓慢，且会降低LLM的性能和指令遵循能力
80%Unverified当LLM上下文窗口中的指令过多时，模型对每条指令的关注度会下降，导致遵循率降低
80%UnverifiedLLM在生成代码时倾向于'重写'而非'编辑'，导致连带修改问题频繁发生
80%Verified大语言模型本质上是无状态函数，权重在推理时已经固定，不会随时间推移而学习
80%Unverified上下文工程的核心认知是LLM的输出质量完全取决于输入上下文的质量
80%Verified大语言模型的涌现能力是指在参数规模突破某个阈值后突然展现出的复杂推理能力
80%UnverifiedAI prescription parsing typically combines OCR with LLM-based information extraction to convert prescription details into structured medication reminder plans
80%