The AI Saga

Chapter 2 · The ChatGPT Moment

Wednesday, November 30, 2022 — the day AI became a consumer product.

"We had no idea." — Sam Altman, when asked later about their growth projections for launch week.

On a grey Wednesday afternoon in San Francisco, OpenAI published a short blog post titled "Introducing ChatGPT." There was no press embargo. No Super Bowl ad. Just a link to a free website.

By Sunday, it had a million users. By late January 2023, it had a hundred million. Nothing in the history of consumer software — not Facebook, not Instagram, not TikTok — had grown that fast.

Under the hood, the model was barely newer than what had already been available through OpenAI's API for months. What changed was something deceptively simple: the interface.

2.1 The underrated power of a chat box

GPT-3's playground was a text area where you wrote prompts like:

Translate the following English to French:
sea otter =>

It worked, but it required knowing how to prompt. Every user was a prompt engineer, whether they wanted to be or not.

ChatGPT's interface was a chat box. You wrote "hey can you translate 'sea otter' to French?" and it answered. Three things about that unlock:

  1. State. The conversation itself became memory. You could say "now put it in a song" without restating the context.
  2. Repair. "No, shorter" or "in Portuguese instead" worked. The UI invited iteration.
  3. Forgiveness. You could be vague, misspell, trail off. Humans are sloppy; the chat box expected it.
flowchart LR
    subgraph Playground
    A[Know how to prompt] --> B[Write prompt carefully] --> C[Get output]
    end
    subgraph ChatGPT
    D[Type like a human] --> E[Get output] --> F[Iterate conversationally]
    F --> D
    end

This is the first of many lessons the saga teaches: the interface is often more important than the underlying model.

2.2 RLHF — the secret sauce

ChatGPT wasn't just GPT-3.5 in a nicer wrapper. It was GPT-3.5 aligned — trained to prefer responses that humans find helpful, truthful, and safe.

The technique is called Reinforcement Learning from Human Feedback (RLHF), detailed in the InstructGPT paper (Ouyang et al., March 2022). It runs in three stages:

flowchart TB
    subgraph Stage 1: Pretraining
    A[Raw internet text] --> B[Base model
next-token predictor]
    end
    subgraph Stage 2: Supervised fine-tuning
    B --> C[Human-written
ideal responses]
    C --> D[SFT model
imitates good answers]
    end
    subgraph Stage 3: RLHF
    D --> E[Generate multiple answers]
    E --> F[Humans rank them
best to worst]
    F --> G[Train reward model]
    D --> H[PPO: policy optimization]
    G --> H
    H --> I[Aligned chat model]
    end

Why it matters: raw next-token predictors will happily continue a prompt into lies, abuse, or refusal to answer obvious questions — because the internet contains plenty of each. RLHF is the step that makes a model shippable to a hundred million people.

It also introduced a new job category: red-teamers, raters, and eventually alignment engineers.

2.3 The growth curve no one saw coming

A non-exhaustive timeline of the first 90 days:

timeline
    title ChatGPT, Nov 2022 - Feb 2023
    Nov 30 : Launch - "research preview"
    Dec 05 : 1 million users
    Dec 15 : First viral tweets about code generation
    Jan 09 : Microsoft reported $10B investment in OpenAI
    Jan 23 : Microsoft formalizes partnership
    Feb 01 : ChatGPT Plus ($20/mo) launched
    Feb 07 : Bing Chat unveiled (GPT-4 under the hood)
    Feb 08 : Google announces Bard in a rushed event
    Feb 24 : 100 million monthly active users

Three observations about this period:

2.4 What ChatGPT could actually do (circa Dec 2022)

If you'd asked a principal engineer in late 2022 what ChatGPT was good for, an honest answer would include:

And honest weaknesses:

Every chapter that follows is, in some sense, a response to one of those weaknesses.

flowchart LR
    A[Hallucinations] --> B[RAG: retrieve before generating]
    C[No math] --> D[Tool use: call a calculator]
    E[No internet / files] --> F[Tool use + MCP]
    G[No memory] --> H[Long context + memory systems]
    I[Stale knowledge] --> J[Live retrieval + fresh fine-tunes]

2.5 The aftershock — a new industry is born

The three months after ChatGPT launched reshuffled the software industry:

For a backend engineer, the practical consequence was that within months, "can you add ChatGPT to our product?" became a ticket in your sprint — whether you were ready for it or not.

2.6 The first user's first message

Imagine the median user opening ChatGPT for the first time on December 3, 2022. The journey looked something like this:

journey
    title A new user's first hour with ChatGPT
    section Curiosity
      Hears about it on Twitter: 4: User
      Signs up: 3: User
    section First prompts
      Asks for a poem about their cat: 5: User
      Surprised it works: 5: User
      Asks it to write an email: 5: User
    section Trust building
      Asks a coding question: 4: User
      Verifies the answer: 3: User
      Catches a hallucination: 2: User
    section Adoption
      Decides this changes things: 5: User
      Tells five friends: 5: User

That micro-arc — delight, trust, suspicion, adoption — is what made the platform sticky. People tested it. It mostly held up. And then they came back.

2.7 What November 2022 really was

In retrospect, ChatGPT was less a technical breakthrough and more a product breakthrough at the end of a five-year capability curve. The Transformer (2017), GPT-3 (2020), RLHF (2022), and a good chat UI (2022) were each necessary; together, they were sufficient.

It's the same pattern as the iPhone. Touchscreens, mobile browsers, App Store, cellular data — each existed. Apple shipped them together in one object and the industry flipped.

The lesson for engineers: watch for the moment when existing pieces line up, not for the single magic innovation. The next ChatGPT-sized shift will probably be built from things that exist today.

Further reading & watching

← Table of Contents