AI fundamentals

欢迎！如果你刚开始接触 AI ，无需技术背景也能上手。最有帮助的是一张清晰的路线图——让你明白 AI 系统能做什么、它们如何被打包提供，以及如何根据需求选对工具。

什么是 AI？

人工智能是一类能识别模式、从数据中学习并产出有用结果的软件。你可能在日常场景中已经遇到 AI ，比如：

• 导航应用为你绕开拥堵路段；
• 银行将一笔消费标为“异常”；
• 客服聊天机器人回答常见问题。

AI 不是单一工具，而是一个范畴。在这个范畴下有各种模型：这些经过训练的系统从数据中学习，然后把学到的东西应用到新情境中。有的模型擅长语音、有的擅长视觉或预测。现在许多人通过对话式工具开始了解 AI，例如 ChatGPT 。支持 ChatGPT 的模型专攻语言，这类模型被称为 大型语言模型（ large language models ，简称 LLM ）。

了解 大型语言模型 如何运作

大型语言模型（ LLM ）专门处理语言问题。它们从海量、多来源的文本中学习模式，从而生成和改写文本。LLM 并不像人那样“知道”事实；它们根据上下文预测最可能的下一个语言片段。随着计算能力、训练方法和大型数据集的进步，得以构建更大、更强的语言模型。

像 OpenAI 等前沿研究机构把这些模型作为核心成果，既在面向用户的产品中提供（例如 ChatGPT 或 Codex ），也通过 API 向开发者开放，使其能在自家软件中集成 AI 功能。

模型如何演进

研究团队在模型经过训练并通过内部评估与安全测试后，才会发布新模型。说模型“被训练”时，通常指两个阶段——可以把它比作一个人学习并逐渐胜任工作。

第一阶段是预训练，模型从大量文本中学习通用模式，从而获得摘要、写作、翻译和解释等广泛能力。想象一个新员工花数周时间读各种资料——手册、优秀范例、过往项目和常见问题，直到理解这份工作的“基本形态”。

接着这个“员工”开始上手工作，由“经理”进行指导：更清晰，补问关键问题，调整语气，遵守公司规则——这就是后训练（post-training）。这一阶段让模型更可靠地遵从指令、以实用的风格交流，并更好地应对复杂情形。后训练也会加强安全性，尽量减少有害输出、避免不当请求，并在敏感或不确定的话题上更谨慎地回应。

随着模型更新和再训练，你可能会注意到语气或回答有所变化。想要结果稳定，最好明确你的目标、受众、格式和约束；在涉及安全或不确定性时，模型通常会更谨慎。

推理型与非推理型模型

不同模型在速度、深度和对多步骤指令的遵从性上做出不同取舍。部分模型为日常任务优化，响应快速流畅，适合起草、摘要、改写、头脑风暴等场景；另一些模型则在回答前投入更多计算来“思考”，这在处理复杂、多步骤的问题时能提高可靠性。

非推理型模型（有时标注为 “ Instant ”）优化于快速、流畅的输出。任务直截了当、需要保持节奏时，它们是默认之选：把笔记整理成信息、润色措辞、生成选项或提取要点。

推理型模型（有时标注为 “ Thinking ”）则训练得更擅长有步骤的、审慎的问题解决——如规划、复杂分析、棘手调试，或需处理约束和边缘情况的决策。它们可能耗时更长，但在追踪多个变量和避免表面错误上更可靠。

刚开始时不必为模型选择烦恼——默认的 ChatGPT 体验会自动切换，让你专注问题本身而非设置。随着使用经验积累，你可以根据偏好（速度 vs 深度；快速草拟 vs 细致分析）尝试可选控制项：多数情况下选择 Auto ，在任务复杂或高风险时切换到 Thinking 。

总结

一个简单的分层框架：

• AI = 整个领域
• 模型 = 为特定任务训练出的系统
• 大型语言模型（ large language models ，简称 LLM ）= 专注于理解和生成语言，由研究机构长期训练而成
• ChatGPT = 帮助你高效使用 LLM 的产品

有了这幅图景，你就能学习如何与工具（如 ChatGPT ）良好互动，获得满意结果——从如何提问以得到想要的输出开始。

想了解如何入门 ChatGPT ，参见 OpenAI Academy 关于入门[https://openai.com/academy/getting-started/] 和 提示工程（ prompt engineering ）[https://openai.com/academy/prompting/] 的指南。更多实用教学资源，请访问 OpenAI Academy 。

Welcome! If you’re new to AI, you don’t need a technical background to get started. What helps most is a simple map of the landscape—so you can understand what AI systems can do, how they’re packaged, and how to choose the right tool for your needs.

What is AI?

Artificial intelligence (AI) is a broad category of software that can recognize patterns, learn from data, and produce useful outputs. 

You’ve probably seen AI show up in everyday moments, like when:

• Your map app reroutes you around traffic
• Your bank flags a purchase as “unusual”
• A customer support chatbot answers common questions

AI is a category—not one single tool. Within that category are models: trained systems that learn from data and then apply what they’ve learned to new situations. Some models specialize in speech, vision, or forecasting. 

You’re likely starting your AI journey by using conversational AI tools, like ChatGPT. The models behind ChatGPT specialize in language—these are called large language models.

Understanding how large language models work

A large language model (LLM) is a model designed to work with language. It learns patterns from large amounts of text from many sources so it can generate and transform text in helpful ways. An LLM doesn’t “know” things the way a person does. Instead, it predicts the most likely next piece of language based on context. Over time, advances in computing power, training methods, and access to large datasets made it possible to build larger and more capable large language models. 

OpenAI and other frontier research labs build these models as a core part of their offerings, then make them available through user-facing products (like ChatGPT or Codex) and through APIs, which let developers use those models to build their own AI tools and integrate AI into existing software.

How models evolve over time

New models become available from these research labs when they have been trained and passed internal evaluation and safety testing.  When you hear that an AI model was “trained,” it usually refers to two stages—think of it like someone learning and getting better at their job.

The first stage is pre-training, when the model learns general patterns from a huge amount of text, which gives it broad skills like summarizing, drafting, translating, and explaining. 

Think of it like a new employee who spends weeks reading everything they can—manuals, examples of great work, past projects, FAQs—until they understand the “shape” of the job.

Now the “employee” starts doing the work, and a “manager” coaches them: be clearer, ask good follow-ups, match the right tone, and follow company policies. That’s post-training. This stage helps the model follow instructions more reliably, communicate in a useful style, and handle tricky situations better.

Post-training is also where safety checks get emphasized—training that is designed to reduce harmful outputs, avoid unwanted requests, and respond more carefully when the topic is sensitive or uncertain.

As models are updated and trained, you might notice shifts in tone or responses. If you want consistent results, be explicit about your goal, audience, format, and constraints—and expect the model to be more careful when safety or uncertainty is involved.

Reasoning and non-reasoning models

Different models are tuned for different tradeoffs—like speed, depth, and how carefully they follow multi-step instructions. Some are designed to respond quickly and smoothly for everyday tasks (drafting, summarizing, rewriting, brainstorming). Others are designed to spend more compute thinking through a problem before they answer, which can improve reliability on harder, multi-step work. 

Non-reasoning models (sometimes labeled as “Instant”) are optimized for fast, fluent output. They’re a good default when the task is straightforward and you mainly want momentum: turn notes into a message, polish wording, generate options, or extract key points. 

Reasoning models (sometimes labeled as “Thinking”) are trained to do better at deliberate, step-by-step problem solving—things like planning, complex analysis, tricky debugging, or decisions with constraints and edge cases. They may take longer, but they’re often better at tracking multiple moving parts and avoiding shallow mistakes.

If you’re just getting started, you don’t need to worry about model choice—the default ChatGPT experience is designed to auto-switch so you can focus on your question, not the settings.

Over time, as you learn what you like (speed vs depth, quick drafts vs careful analysis), you can start experimenting with the optional controls: for example, choosing Auto most of the time, and switching to Thinking when a task is complex or high-stakes.

Summary

Here’s the simple hierarchy:

• AI = the overall field
• Models = trained systems that perform particular tasks
• Large language models (LLMs) = models focused on understanding and generating language, trained over time by AI research labs
• ChatGPT = a product that helps you use an LLM effectively

Once you have this picture in mind, you’ll be set up to learn how to get great results with tools like ChatGPT—starting with how to talk to it to get the results you want.

Learn about getting started with ChatGPT and prompt engineering.

Continue learning with OpenAI Academy

Generated by RSStT. The copyright belongs to the original author.

Source