This post summarizes Andrej Karpathy’s video “How I Use LLMs” (Feb 2025).
The LLM Landscape in 2025
ChatGPT (OpenAI) remains the most feature-rich option, but the ecosystem has grown:
- Big Tech: Gemini (Google), Copilot (Microsoft), Meta AI
- Startups: Claude (Anthropic), Grok (xAI), DeepSeek (China), Le Chat (Mistral, France)
Track model performance on LM Arena or SEAL Leaderboard.
Understanding Pricing Tiers
Free tier: GPT-4o mini, basic features. Good for simple tasks.
Plus ($20/month): GPT-4o, thinking models (o1, o3-mini), more usage limits.
Pro ($200/month): Unlimited usage, Deep Research, o1 pro mode.
The free tier is surprisingly capable for everyday use. Pay for Plus when you need thinking models or hit rate limits.
When to Use Thinking Models
Thinking models (o1, o3) use reinforcement learning to reason through problems. They’re slower and more expensive, but better for:
- Complex math and logic problems
- Multi-step reasoning
- Tasks where accuracy matters more than speed
For quick questions, stick with GPT-4o. For hard problems, switch to o1 or o3-mini.
Tool Use: Web Search
LLMs have knowledge cutoffs. For recent information, use web search:
- ChatGPT: Click the search icon or let it auto-detect
- Perplexity: Built for search-first queries
- Claude: Currently lacks web search integration
Good search queries:
- “Is the market open today?” (time-sensitive)
- “When is [show] season X coming out?” (recent info)
- “Does [company] offer [feature]?” (changes over time)
Deep Research
ChatGPT Pro’s Deep Research spends 10-30 minutes doing extensive web research. It’s like having a research assistant who reads dozens of sources and synthesizes findings.
Use cases:
- Researching supplements or health topics
- Understanding complex technical subjects
- Comparing products or services
- Any question that would take you hours to research manually
File Uploads
You can upload documents (PDFs, code files, images) directly to the context window. The model can then answer questions about them.
Practical uses:
- “Summarize this research paper”
- “Find bugs in this code”
- “What does this contract say about X?”
Claude Artifacts
Claude has a unique feature: it can create interactive artifacts (apps, diagrams, visualizations) that run in the browser.
Ask Claude to:
- “Create a diagram of X”
- “Build a simple calculator”
- “Make an interactive visualization of this data”
The artifacts are self-contained and can be shared or embedded.
Cursor and Vibe Coding
For programming, Karpathy uses Cursor (an AI-powered code editor). Key features:
- Cmd+K: Edit a specific line
- Cmd+L: Explain code
- Cmd+I (Composer): Autonomous agent that edits multiple files
“Vibe coding” = giving high-level instructions and letting the AI handle implementation. You can always fall back to manual coding when needed.
Example: “Add confetti when a player wins” → Cursor installs a library, writes the code, and even downloads a sound file.
Voice and Audio
Speech-to-text: Talk to ChatGPT instead of typing. Useful for brainstorming or when your hands are busy.
Advanced Voice Mode: The model processes audio natively (not transcription → text → speech). It can hear tone, emotion, and respond more naturally.
NotebookLM: Google’s tool that generates podcast-style discussions from your documents. Upload sources, get a conversation between two AI hosts.
Image and Video
Image input: Upload screenshots, photos, diagrams. Ask “What’s in this image?” or “OCR this receipt.”
Image output: DALL-E (ChatGPT), Ideogram, Midjourney. Describe what you want, get an image.
Video input: Some models can process video frames. Point your camera and ask questions.
Video output: Sora (OpenAI), Veo 2 (Google). Still early, but improving rapidly.
Memory and Custom Instructions
Memory: ChatGPT learns about you over time. It remembers preferences, context, and personal details across conversations.
Trigger it explicitly: “Remember that I prefer X” or “Save this to memory.”
Custom Instructions: Set global preferences in settings:
- “Don’t be overly formal”
- “I’m learning Korean, use 해요체 formality”
- “Always explain your reasoning”
Custom GPTs
Save frequently-used prompts as Custom GPTs. Instead of re-explaining a task each time, create a GPT that already knows the context.
Example: Korean vocabulary extractor
- Input: A Korean sentence
- Output: Vocabulary in “Korean; English” format for flashcards
The key is providing examples (few-shot prompting), not just descriptions.
Practical Takeaways
-
Match the tool to the task: Quick questions → GPT-4o. Hard problems → thinking models. Recent info → web search.
-
Use tools: Don’t rely on the model’s “mental math.” Ask it to use code for calculations.
-
Upload context: Give the model documents, images, code. More context = better answers.
-
Customize: Set up memory, custom instructions, and Custom GPTs for your workflow.
-
Try multiple models: Claude for artifacts, Perplexity for search, Cursor for coding. Each has strengths.
-
Vibe code: For programming, describe what you want and let the AI implement. Review and iterate.
本文总结了 Andrej Karpathy 的视频《我如何使用大语言模型》(2025年2月)。
2025 年的 LLM 生态
ChatGPT(OpenAI)仍然是功能最丰富的选择,但生态系统已经扩展:
- 大厂:Gemini(Google)、Copilot(Microsoft)、Meta AI
- 创业公司:Claude(Anthropic)、Grok(xAI)、DeepSeek(中国)、Le Chat(Mistral,法国)
在 LM Arena 或 SEAL Leaderboard 跟踪模型性能。
理解定价层级
免费版:GPT-4o mini,基础功能。适合简单任务。
Plus($20/月):GPT-4o,思考模型(o1、o3-mini),更高使用限额。
Pro($200/月):无限使用,Deep Research,o1 pro 模式。
免费版对日常使用已经相当强大。当你需要思考模型或达到速率限制时再付费升级 Plus。
何时使用思考模型
思考模型(o1、o3)使用强化学习来推理问题。它们更慢、更贵,但更适合:
- 复杂的数学和逻辑问题
- 多步推理
- 准确性比速度更重要的任务
快速问题用 GPT-4o。难题切换到 o1 或 o3-mini。
工具使用:网络搜索
LLM 有知识截止日期。获取最新信息需要网络搜索:
- ChatGPT:点击搜索图标或让它自动检测
- Perplexity:专为搜索优先的查询设计
- Claude:目前缺乏网络搜索集成
好的搜索查询:
- “今天股市开门吗?“(时效性)
- “[剧集]第X季什么时候出?“(最新信息)
- “[公司]提供[功能]吗?“(随时间变化)
Deep Research
ChatGPT Pro 的 Deep Research 会花 10-30 分钟进行广泛的网络研究。就像有一个研究助理阅读数十个来源并综合发现。
使用场景:
- 研究补充剂或健康话题
- 理解复杂技术主题
- 比较产品或服务
- 任何需要你手动研究数小时的问题
文件上传
你可以直接上传文档(PDF、代码文件、图片)到上下文窗口。模型可以回答关于它们的问题。
实际用途:
- “总结这篇研究论文”
- “找出这段代码的 bug”
- “这份合同关于 X 说了什么?“
Claude Artifacts
Claude 有一个独特功能:它可以创建在浏览器中运行的交互式 artifact(应用、图表、可视化)。
让 Claude:
- “创建 X 的图表”
- “构建一个简单的计算器”
- “制作这些数据的交互式可视化”
Artifact 是独立的,可以分享或嵌入。
Cursor 和 Vibe Coding
编程时,Karpathy 使用 Cursor(AI 驱动的代码编辑器)。主要功能:
- Cmd+K:编辑特定行
- Cmd+L:解释代码
- Cmd+I(Composer):自主代理,编辑多个文件
“Vibe coding” = 给出高层指令,让 AI 处理实现。需要时随时可以回退到手动编码。
示例:“玩家获胜时添加彩带效果” → Cursor 安装库、写代码,甚至下载音效文件。
语音和音频
语音转文字:对 ChatGPT 说话而不是打字。适合头脑风暴或手忙时。
高级语音模式:模型原生处理音频(不是转录 → 文字 → 语音)。它能听到语气、情感,响应更自然。
NotebookLM:Google 的工具,从你的文档生成播客风格的讨论。上传来源,获得两个 AI 主持人之间的对话。
图像和视频
图像输入:上传截图、照片、图表。问”这张图里有什么?“或”OCR 这张收据”。
图像输出:DALL-E(ChatGPT)、Ideogram、Midjourney。描述你想要的,获得图像。
视频输入:一些模型可以处理视频帧。指向摄像头并提问。
视频输出:Sora(OpenAI)、Veo 2(Google)。还在早期,但快速改进中。
记忆和自定义指令
记忆:ChatGPT 随时间了解你。它记住偏好、上下文和个人细节,跨对话保持。
显式触发:“记住我偏好 X”或”保存到记忆”。
自定义指令:在设置中设置全局偏好:
- “不要过于正式”
- “我在学韩语,使用해요体”
- “总是解释你的推理”
Custom GPTs
将常用提示保存为 Custom GPT。不用每次重新解释任务,创建一个已经知道上下文的 GPT。
示例:韩语词汇提取器
- 输入:一个韩语句子
- 输出:“韩语; 英语”格式的词汇,用于闪卡
关键是提供示例(few-shot prompting),而不只是描述。
实用要点
-
匹配工具和任务:快速问题 → GPT-4o。难题 → 思考模型。最新信息 → 网络搜索。
-
使用工具:不要依赖模型的”心算”。让它用代码做计算。
-
上传上下文:给模型文档、图片、代码。更多上下文 = 更好的答案。
-
自定义:为你的工作流设置记忆、自定义指令和 Custom GPT。
-
尝试多个模型:Claude 做 artifact,Perplexity 做搜索,Cursor 做编程。各有所长。
-
Vibe code:编程时,描述你想要的,让 AI 实现。审查并迭代。