Overview
On November 30, 2022, OpenAI released ChatGPT as a free research preview. It reached one million users in five days. One hundred million users in two months. No consumer technology had ever grown that fast.
But the significance of ChatGPT wasn’t the growth metrics. It was something harder to quantify: for the first time in history, a general-purpose AI system felt — to ordinary people, not just researchers — genuinely useful, surprisingly capable, and startlingly human. The conversation that had been happening in university labs and tech companies for decades spilled, all at once, into the mainstream.
What ChatGPT Actually Was
ChatGPT was not a new architecture. It was built on GPT-3.5, a large language model trained on vast quantities of internet text. The Transformer architecture — introduced in 2017 — was the foundation. What was new was how it had been shaped.
OpenAI applied a technique called Reinforcement Learning from Human Feedback (RLHF):
- Human trainers demonstrated good responses to prompts
- The model generated multiple responses; human raters ranked them by quality
- A separate “reward model” was trained to predict which responses humans would prefer
- The GPT model was then fine-tuned using reinforcement learning to maximize that reward
The result was a model that didn’t just predict the next token statistically — it had been specifically trained to be helpful, to follow instructions, to refuse harmful requests, and to converse naturally. The raw power of language modeling had been steered by human values.
This was the key insight that the broader field had missed: capability alone is not enough. The interface between AI and humans matters enormously. A model that could answer questions but gave erratic, unpredictable outputs was unusable for most people. A model fine-tuned to be genuinely assistive was transformative.
The Night That Changed Everything
The timeline of ChatGPT’s cultural impact was unlike anything the tech industry had seen:
- Day 1: Tech Twitter erupted. Screenshots of conversations spread virally — people showing it write code, draft legal contracts, compose music, solve math problems, explain physics, and roleplay fictional scenarios
- Week 1: One million users. Google reportedly issued an internal “code red,” recognizing that the search paradigm might be under threat
- Month 1: Schools began grappling with AI-generated homework. The New York Times ran front-page stories. Congress held hearings
- Month 2: 100 million users — the fastest consumer product adoption in history, surpassing TikTok’s 9-month record
- Month 3: Microsoft announced a $10 billion investment in OpenAI and began integrating GPT-4 into Bing, Office, and Azure
Why This Moment, Why ChatGPT
The honest answer is that ChatGPT was not technically superior to everything that came before. Google’s LaMDA, released months earlier, had similar capabilities. What ChatGPT had was accessibility and interface design:
- A clean, simple chat interface that anyone could use
- No API keys, no prompt engineering expertise required
- A model that had been specifically trained to be helpful and conversational
- A free tier that let hundreds of millions of people try it without commitment
Yuval Noah Harari’s framework in Nexus offers a useful lens: he argues that the decisive moments in information network history are rarely about the underlying capability — they’re about the moment the capability becomes legible to ordinary people. The printing press didn’t change Europe the moment Gutenberg invented it; it changed Europe when pamphlets began circulating in markets. ChatGPT was AI’s pamphlet moment.
The RLHF Revolution
ChatGPT’s success catalyzed a rethinking of how to build AI systems. The RLHF approach — using human feedback to shape model behavior — proved more important than anyone had anticipated.
Before ChatGPT, the dominant belief was that scaling model size (more parameters, more compute, more data) was the path to capability. After ChatGPT, it became clear that alignment — the process of making models behave in ways humans actually want — was equally critical. A perfectly capable model that was unhelpful, deceptive, or erratic was worthless in practice.
This insight reshaped research priorities across the industry:
- OpenAI doubled down on RLHF and later developed “Constitutional AI” variants
- Google developed RLHF for Bard/Gemini
- Anthropic — founded by former OpenAI researchers — built its entire approach around alignment, producing Claude
- The concept of “AI safety” moved from a niche academic concern to a mainstream research and policy priority
What ChatGPT Revealed About AI’s State
The intense scrutiny following ChatGPT’s release revealed both what language models could do and what they couldn’t:
Genuine capabilities: Writing, coding, summarizing, explaining, translating, brainstorming, drafting — tasks involving the recombination and generation of language
Genuine limitations: Factual accuracy (models “hallucinate” with confident fluency), mathematical reasoning (symbolic manipulation remained unreliable), real-time knowledge (models have training cutoffs), consistent long-term memory, and understanding causality vs. correlation
The hallucination problem became particularly prominent: ChatGPT could generate fluent, authoritative-sounding text about events that never happened, citations that didn’t exist, and facts that were simply wrong. This was not a bug to be fixed quickly — it was a fundamental property of statistical language modeling.
The Bridge from History to Now
ChatGPT is the event that connects everything that came before it to the world we currently inhabit. The Turing Test of 1950 asked whether a machine could converse indistinguishably from a human — ChatGPT demonstrated that the question had become surprisingly complicated. The Perceptron of 1957 began the lineage of neural networks that became the Transformers that powered GPT. The two AI Winters taught the field to be cautious, to solve real problems, to underpromise — and yet the scale of ChatGPT’s impact exceeded anything anyone had promised.
Harari’s central thesis in Nexus — that AI represents the first non-human agent in human information networks — feels most vivid in the context of ChatGPT. For the first time, billions of people began having conversations with an entity that was not human, that was not simply retrieving stored information, but was generating novel responses to novel situations. What that means for human epistemics, for democracy, for creativity, for labor — these questions moved from philosophy seminars to front-page news.
The story of AI’s history had been, until this moment, a story told mostly inside the field. After November 30, 2022, it became everyone’s story.