LLM
A Large Language Model (LLM) is a natural language processing model built on deep learning technology and trained on large-scale text data. Core capabilities include text understanding, generation, reasoning, and dialogue, enabling it to handle diverse language tasks such as translation, summarization, code generation, and question answering. LLMs typically adopt the Transformer architecture, with parameter counts ranging from billions to hundreds of billions, and serve as a foundational component of contemporary AI applications.
Core Facts
Mainstream Agent architectures typically use a large language model as the brain, combined with memory modules, planning modules, and tool interfaces to accomplish complex tasks
Function Calling (Tool Use) is a standardized mechanism for large models to invoke external tools, where developers pre-declare tool names, descriptions, and parameter schemas, and the model autonomously determines when to make a call.
The tool-calling (Function Calling) mechanism allows the model to invoke external tools during reasoning, such as file read/write operations, terminal commands, and search APIs.
Timeline (last 90 days)
semantic-registry采用两阶段选择机制：第一阶段基于任务描述做语义相似度匹配，将候选范围从全量工具缩小到少量高相关项；第二阶段由LLM在精简后的候选中做最终判断。
LLM-generated code is sometimes difficult to trace to its source, and in high-reliability fields such as embedded systems, maintainers may want every line of code to have a clear human author so that responsibility can be assigned.
LLM training data often contains large amounts of open-source code, and generated content may imperceptibly reproduce copyrighted fragments, bringing legal risks to projects.
Model scale is one variable that explains a large part of the variation in LLM performance.
LLM performance is extremely multidimensional, and no fixed ranking can guarantee that the model that is strongest at math is also strongest at telling jokes.
The semantics of rendered Markdown are not unique to machines: the same passage of text may be parsed as display content or interpreted by a downstream LLM as structured instructions.
Emacs community LLM integration solutions appear as Elisp packages, rely on individual maintenance, and vary in feature coverage and stability, but can be deeply customized.
Emacs community developers have already built interfaces in Emacs for interacting with large language models, integrating AI completion, code generation and conversational refactoring into their workflows.
LLM editor integration is typically implemented through API calls: the plugin sends buffer context, cursor position and user instructions to the model service, then writes back the returned completion or refactoring suggestions.
For LLM tool calls, the sandbox assumes an auditing role, where each executed command, stdout/stderr, and exit code can be recorded and passed back to the model.
40 more timeline events
All Facts (20)
Mainstream Agent architectures typically use a large language model as the brain, combined with memory modules, planning modules, and tool interfaces to accomplish complex tasks
90%VerifiedFunction Calling (Tool Use) is a standardized mechanism for large models to invoke external tools, where developers pre-declare tool names, descriptions, and parameter schemas, and the model autonomously determines when to make a call.
90%VerifiedThe tool-calling (Function Calling) mechanism allows the model to invoke external tools during reasoning, such as file read/write operations, terminal commands, and search APIs.
90%VerifiedMainstream approaches to agent evaluation include LLM-as-Judge, trajectory matching, and outcome verification.
90%VerifiedThe context window of a large language model determines the upper limit of information the model can process in a single inference.
90%VerifiedGenerative AI refers to artificial intelligence systems that can automatically generate text, images, code, and other content based on input prompts.
90%VerifiedPrompt注入（Prompt Injection）是针对大语言模型应用的安全威胁，攻击者通过在输入中嵌入恶意指令试图覆盖系统原有行为约束
90%VerifiedThe root cause of AI hallucination lies in the generative mechanism of LLMs: the model is essentially performing next-token probability prediction, optimizing for linguistic coherence rather than factual accuracy.
90%VerifiedLLMs exhibit a 'Lost in the Middle' phenomenon, where the model pays more attention to information at the beginning and end of the context while tending to overlook information in the middle.
90%VerifiedAI programming assistance tools such as GitHub Copilot, Cursor, and Claude are essentially code completion and generation systems formed by training large language models on massive code repositories.
90%VerifiedAI Agent通过工具调用（Tool Use）机制与外部系统交互，包括执行代码、调用API、读写数据库等
90%VerifiedRAG（检索增强生成）技术架构将文档切片、向量化、存入向量数据库，查询时先检索相关片段再交给模型生成回答
90%VerifiedFailures in AI agents are often semantic deviations rather than explicit error codes, and this kind of "successful failure" requires the orchestration layer to have semantic-level error recognition capability.
85%VerifiedThe platform adopts a design philosophy of AI generation with human editability, where AI handles the first draft and the final judgment and revision authority is handed over to testers.
85%VerifiedWhen key information is placed in the middle of a long text, the model's retrieval and reasoning accuracy drops significantly, while information placed at the beginning or end can be utilized more effectively.
85%VerifiedThe output of LLMs is inherently non-deterministic; even with the same Prompt as input, the model may return results with different formats, content, or even logic at different times.
85%VerifiedAI evaluation targets are divided into three layers—large language model evaluation, knowledge base (RAG) evaluation, and agent (Agent) evaluation—forming a progressive relationship of increasing complexity and knowledge depth.
80%VerifiedThe widespread use of large language models may lead to a sustained reduction in linguistic diversity, giving rise to what linguists call 'technology-mediated language convergence.'
80%VerifiedLLM output is non-deterministic; due to the non-associativity of floating-point operations and differences in context windows, even with identical inputs, outputs are difficult to reproduce exactly and do not possess idempotency.
80%VerifiedWhen a model's training data contains benchmark test questions or highly similar content, the model may achieve high scores through memorization rather than genuine understanding.
80%