Large Language Model
Core Facts
The core idea of RAG is to retrieve relevant information from an external knowledge base as context to inject into the prompt before the large model generates a response.
'Hallucination' refers to large language models generating plausible-sounding but factually incorrect content with high confidence
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
Timeline (last 90 days)
The tool-calling mechanism allows a large language model to output a structured call request (usually JSON, containing the tool name and parameters), which an external program executes and then returns the result, forming a "think—call—observe" loop; it is the engineering implementation of the ReAct paradigm.
Large language models themselves are stateless, and the full context must be resent with every request, so anything that persists throughout a session is repeatedly billed each round.
The model's ability to reason about control flow and data flow is limited, and long functions and cross-function call chains remain an obvious weakness of AI-assisted decompilation.
The speaker speculates that the PHP team's public discussion that year may have been absorbed into large language model training data, and that some AI companies are investing resources in using models to hunt for vulnerabilities, but it cannot be confirmed whether a model independently discovered this vulnerability.
Large language models have the ability to extract entities and relationships from free text, and can convert unstructured text into structured fields that can be filled into forms.
Large language models themselves do not have the ability to directly call third-party APIs, access databases, or perform precise calculations.
Current large models almost all rely on external storage outside the model to maintain dialogue context, with dialogue history appended to prompts or vector databases.
Large models have value in popularizing health information, but they do not possess clinical judgment ability and cannot obtain complete medical history and medication records; any decision involving dosage, stopping medication, or switching medication must be made by a physician with prescribing authority.
Large language model learning does not involve genuine understanding of semantics or experience, because the models have no body, senses, or social interaction, and what they capture is surface-level correlation in language rather than intent and emotion.
High-quality long-form text content (such as academic books) is crucial for improving the reasoning and knowledge capabilities of large language models.
40 more timeline events
All Facts (20)
The core idea of RAG is to retrieve relevant information from an external knowledge base as context to inject into the prompt before the large model generates a response.
90%Verified'Hallucination' refers to large language models generating plausible-sounding but factually incorrect content with high confidence
90%VerifiedMainstream 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%VerifiedReasoning models perform better in information synthesis tasks requiring multi-step logical deduction, but have higher latency and should be selected appropriately based on task characteristics.
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%VerifiedAgents possess capabilities for planning, tool invocation, memory, and autonomous iteration, distinguishing them from the single-round question-and-answer mode of traditional large language models.
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%VerifiedThe context window of a large language model determines the upper limit of information the model can process in a single inference.
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%VerifiedTest-time Compute Scaling improves model performance by investing more computational resources during the inference stage, and research indicates it is more efficient than simply increasing model parameter count on certain tasks.
90%Verified训练一个顶级大模型的成本已达数亿美元级别
90%VerifiedModern large language models typically go through three training stages: pre-training, supervised fine-tuning (SFT), and reinforcement learning from human feedback (RLHF).
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%VerifiedLarge language models are based on the Transformer architecture and are trained on massive amounts of text data.
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%VerifiedWhen the same AI model is given prompts of different quality, the output results can differ by an order of magnitude.
90%VerifiedSince 2023, with the improvement in large language model capabilities, AI agents have rapidly moved from an academic concept to actual product deployment.
90%VerifiedThe pre-training process of large language models requires thousands of high-end GPUs running in parallel for months, with costs typically ranging from tens of millions to hundreds of millions of dollars.
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%VerifiedA new generation of AI code generation tools can decompose natural language requirements into complete engineering tasks, autonomously select the technology stack, generate file structures, and complete the debugging loop.
90%