November 30, 2022, OpenAI quietly released a product.
No press conference, no advertising, just a tweet.
This product was called ChatGPT.
It was a conversational robot. You could chat with it, ask questions, have it help you do things.
A few days later, it went viral.
People shared their conversations with ChatGPT on social media:
- Having it write poems, write code, write articles
- Having it explain complex concepts
- Having it simulate celebrity conversations
- Even having it pass exams
Within two weeks, ChatGPT had over 1 million users.
Within two months, it had over 100 million users.
This was the fastest-growing application in history—faster than TikTok, faster than Instagram, faster than the iPhone.
AI had finally entered the public eye.
What is ChatGPT?#
ChatGPT is a conversational robot based on GPT-3.5.
It can:
- Answer questions
- Write articles, emails, reports
- Write code, debug code
- Translate languages
- Explain concepts
- Have conversations
It wasn’t the first conversational robot. Siri, Alexa, and Xiao Ai had existed for a long time.
But ChatGPT was different:
First, it really understands you.
Previous conversational robots could only handle simple questions. Anything slightly complex and they’d give irrelevant answers.
ChatGPT could understand complex questions and give meaningful answers.
Second, it can do many things.
Previous conversational robots could only do a few preset things: check weather, play music, set alarms.
ChatGPT could do almost anything involving language.
Third, it can have continuous conversations.
ChatGPT remembered conversation context and could have multi-turn conversations.
The Technology Behind ChatGPT#
ChatGPT is based on GPT-3.5 but has a key improvement: RLHF (Reinforcement Learning from Human Feedback).
Training has three steps:
Step 1: Pre-training
Train a language model on massive text, learning to predict the next word.
This is what GPT-3 did.
Step 2: Supervised fine-tuning
Human annotators write high-quality Q&A pairs, fine-tuning the model with this data.
The model learns how to answer questions.
Step 3: Reinforcement learning from human feedback
Human annotators rank the model’s multiple answers, telling the model which is better.
The model learns to generate answers humans like.
RLHF made ChatGPT more “obedient” and “useful” than raw GPT-3.5.
ChatGPT’s Impact#
ChatGPT’s impact is profound:
Work style
People started using ChatGPT to assist work:
- Write emails, reports, proposals
- Write code, debug code
- Translate documents
- Learn new knowledge
Some worry AI will replace jobs; some think AI is an assistant.
Education
Students use ChatGPT to write homework and papers.
Teachers worry about academic integrity and started using AI detection tools.
Education is rethinking: in the AI era, what should we teach and how?
Search
People started using ChatGPT to search for information instead of Google.
Google released Bard (later renamed Gemini) in response.
Microsoft integrated ChatGPT into Bing search.
Programming
Programmers use ChatGPT to write code, debug code, and learn new technologies.
Some said: “With ChatGPT, I don’t need Stack Overflow anymore.”
Content creation
Writers, journalists, and marketers use ChatGPT to generate content.
Some worry AI will replace creators; some think AI is a creation tool.
Competitors#
ChatGPT’s success sparked an AI race.
Google: Released Bard (later Gemini), integrated into Google search and Workspace.
Microsoft: Invested $10 billion in OpenAI, integrated ChatGPT into Bing, Office, and Windows.
Anthropic: Released Claude, founded by former OpenAI employees.
Meta: Released LLaMA, open-source large model.
Chinese companies: Baidu released Ernie Bot, Alibaba released Tongyi Qianwen, Tencent released Hunyuan, ByteDance released Doubao.
AI became the focus of tech competition.
ChatGPT’s Problems#
ChatGPT also has problems:
Hallucination
ChatGPT confidently says wrong information.
Someone had it cite a paper; it made up a non-existent paper.
Someone had it introduce a historical figure; it made up a fake biography.
Bias
ChatGPT may inherit bias from training data.
Safety
ChatGPT may be used to generate fake news, cyberattacks, and other malicious purposes.
OpenAI set safety restrictions, but users can always find ways around them.
Dependency
People started relying on ChatGPT, reducing independent thinking.
After ChatGPT#
In March 2023, OpenAI released GPT-4.
GPT-4 is more powerful than GPT-3.5:
- Can understand images (multimodal)
- Stronger reasoning ability
- Fewer hallucinations
- Longer context
In November 2023, OpenAI released GPT-4 Turbo, supporting 128K context.
In 2024, OpenAI released GPT-4o, faster and cheaper.
AI development is getting faster and faster.
Next Step: AI Programming Assistants#
ChatGPT can write and debug code, becoming a programmer’s assistant.
GitHub Copilot, Cursor, and other tools integrate AI into development environments, changing how programming is done.
Tomorrow, we’ll discuss AI programming assistants.
Today’s Key Concepts#
ChatGPT: Conversational robot released by OpenAI, based on GPT-3.5 and RLHF technology. ChatGPT reached 100 million users in two months, the fastest-growing application in history.
RLHF (Reinforcement Learning from Human Feedback): Training models using human preferences. RLHF makes models generate answers more aligned with human expectations.
Hallucination: Phenomenon where large language models generate false information. Models confidently say wrong information, not knowing what they don’t know.
Discussion Questions#
- ChatGPT reached 100 million users in two months. Why do you think it became popular so quickly?
- ChatGPT can write homework and papers. How do you think education should respond?
Tomorrow’s Preview: AI Programming Assistants—how did Copilot change how programmers work?