Deep Dive into LLMs like ChatGPT by Andrej Karpathy | 大语言模型（如ChatGPT）深度解析：训练全流程、底层原理、缺陷与未来发展
https://www.youtube.com/watch?v=7xTGNNLPyMI 大语言模型（如ChatGPT）深度解析：训练全流程、底层原理、缺陷与未来发展 一、通俗易懂整体解读（类比+关键比方） 这篇是面向普通大众的科普视频文稿，完整拆解大模型从0到能用的全部过程，可以用学生上学整套流程来做通俗类比： 预训练 = 海量泛读课外书 ：把互联网海量网页文本清洗处理，喂给神经网络。就像学生疯狂读全网的文章、百科、新闻，不做题，只学习语言文字的行文规律、世界知识。产出叫 基座模型 ，只会自动续写文字，不会当聊天助手。 比方：基座模型≈一个看过全网海量资料，但没人教它怎么回答问题的复读机，你输入半句话，它只会顺着往下续写网页风格文字，不会听懂你的提问。 有监督微调SFT = 做带标准答案的例题…
Deep Dive into LLMs like ChatGPT by Andrej Karpathy explores the full training process, underlying principles, flaws, and future development of large language models (LLMs) such as ChatGPT. The video content breaks down the entire process from 0 to a usable model, comparing it to a student's learning journey from reading countless articles to becoming a conversational assistant. The main stages of LLM building include:
1. Pre-training: Scraping vast amounts of internet text data to train an initial foundation model, like a "copy machine" that can continue writing based on learned language patterns but cannot answer questions on its own.
2. Fine-tuning (SFT): Training the model with question-answer pairs to make it conversational, akin to a student learning to answer exam questions using example questions and answers.
3. Reinforcement Learning (RL): Two types of RL are discussed:
a. RL with verifiable tasks (e.g., DeepSeek-R1) - Similar to math problems where the model generates multiple solution processes and retains only the correct path.
b. RLHF (human feedback) - A subjective scoring system for tasks without clear right or wrong answers, like humor or poetry.
The video also delves into the characteristics and limitations of LLMs, such as their inherent knowledge gaps, token-based processing, limited computational capacity, and lack of true understanding or memory. The future of LLMs includes multi-modal integration, AI agents, and test-based learning.
The video breaks down the technical aspects of LLM training into several segments, including data collection, tokenization (text-to-token conversion), pre-training with neural networks and the Transformer architecture, post-training supervised fine-tuning to create a conversational model, as well as the inherent limitations and hallucination issues of LLMs.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
