November 30, 2022, late afternoon Pacific time. OpenAI quietly released a chatbot. Inside the company, some called it just a "research preview" — no one expected much. Five days later, it had 1 million users. Two months later, 100 million. The fastest-growing consumer product in internet history.
After Google's 2017 Transformer paper (EP03), one obvious question hung in the air — "What if we pre-trained a Transformer on truly massive amounts of text?"
In 2018, two organizations answered almost simultaneously. Google released BERT in October 2018 (Devlin et al.). Bidirectional learning — each word is informed by both left and right context. OpenAI had released GPT-1 four months earlier in June (Radford et al.). Unidirectional — predict the next word.
OpenAI researchers published a now-famous 2020 paper — "Scaling Laws for Neural Language Models." The thesis: scale up the model, the data, and the compute, and performance improves predictably.
To prove it, that May they released GPT-3. 175 billion parameters (175B). 100× GPT-2. External estimates put the training compute cost at around $4.6 million (Lambda Labs estimate).
GPT-3 was a shock. Without anyone teaching it explicitly, it could translate, summarize, write code, write poetry. "Few-shot learning" — given a couple of examples, it would do new tasks. That was the moment OpenAI internally felt "something real is happening here."
* GPT-4 size estimates leaked but never officially confirmed by OpenAI.
OpenAI took GPT-3.5 and added one more thing — RLHF (Reinforcement Learning from Human Feedback). Humans rate the model's responses; that rating becomes a reward signal that fine-tunes the model. The result was InstructGPT. Wrapped in a chat interface, that became ChatGPT.
Sam Altman (CEO, formerly president of Y Combinator), Ilya Sutskever (Chief Scientist, Hinton's PhD student, the Sutskever from EP01·02·03), and Greg Brockman (President, ex-CTO of Stripe). The three of them green-lit the launch. Internally everyone treated it as a "low-key research preview" — and got it wrong.
March 14, 2023. OpenAI released GPT-4. Architecture undisclosed. It scored in the top 10% of the US Uniform Bar Exam, got 4/5 on AP Calculus BC, 5/5 on AP Chemistry, and excelled across many standardized tests. Estimates of model size (1.8T parameters, MoE) circulate but OpenAI has never confirmed them.
And — about two years before GPT-4's launch, in early 2021, a faction had already left OpenAI to start their own company.
The faction at OpenAI that argued "AI safety / alignment" deserved more priority. They left OpenAI in late 2020 / early 2021 and founded Anthropic in 2021. Claude 1 in March 2023, Claude 3.5 Sonnet in June 2024 — widely judged to surpass GPT-4 on coding tasks.
Then in late 2024 / early 2025, another shock came from China.
DeepSeek-R1 matched OpenAI's o1-level reasoning at roughly 1/30 the training cost — and they open-sourced everything (weights, code). The "AI requires unlimited GPU spending" narrative cracked. Investors started asking: "Does this really need to be this expensive?"
As of May 2026 — ChatGPT has 500M+ weekly active users. Combined with Claude, Gemini, Llama, Grok, total reaches close to 1B. The internet itself is being rebuilt on top of these models. Google search is being replaced with LLM answers. IDEs run Copilot. ERP systems run RAG (the topic of EP08).
In the next post (EP05), we trace another track — the 2014 GAN that Ian Goodfellow sketched at a bar, the 2020 DDPM, and the 2022 Stable Diffusion explosion — 12 years of image and video generation AI. ChatGPT conquered language. Stable Diffusion, Sora, and Veo are conquering vision.