OpenAI Releases ChatGPT

OpenAI released ChatGPT on November 30, 2022 as a free research preview powered by GPT-3.5-turbo, an optimized version of the GPT-3.5 model family trained using InstructGPT techniques developed by OpenAI researchers in 2022. The training pipeline combined supervised fine-tuning (human contractors writing example conversations demonstrating helpful assistant behavior), reward model training (human raters comparing pairs of model responses and indicating which was better), and reinforcement learning from human feedback (RLHF) using Proximal Policy Optimization (PPO) to update the language model’s weights toward higher-scoring responses. The resulting model was dramatically better than raw GPT-3 at following instructions, maintaining topic coherence across a conversation, declining inappropriate requests, and producing well-structured output — behaviors that emerged from the RLHF process rather than raw pretraining. The web interface (chat.openai.com) required only a free account and a web browser, making it the first large language model experience accessible to a mainstream audience without any technical setup. OpenAI had previously offered GPT-3 access through an API requiring developer registration and payment, which limited its reach to technically sophisticated users. ChatGPT eliminated that barrier: within five days of launch it had one million registered users, and within two months it reached 100 million monthly active users — the fastest product in history to reach that milestone, surpassing TikTok (nine months) and Instagram (two and a half years) by a wide margin.

The product’s growth forced immediate competitive responses from the technology industry’s largest companies. Google’s management reportedly declared a “code red” in December 2022 and accelerated the internal development of conversational AI products, leading to the Bard announcement on February 6, 2023 (based on LaMDA, a conversational model Google had developed but not publicly deployed). Microsoft, which had invested $1 billion in OpenAI in 2019 and $2 billion in 2021, announced a new multi-year, multi-billion dollar partnership on January 23, 2023 and began integrating ChatGPT technology into Bing search on February 7, 2023 — the first major search engine to offer a conversational AI interface as a primary feature. GitHub Copilot, Microsoft’s AI code completion tool (GA June 2022), was already using GPT-based models; after ChatGPT’s success it reached one million paid subscribers in January 2023. OpenAI launched ChatGPT Plus on February 1, 2023 at $20 per month, offering priority access during high demand and later priority access to new models. GPT-4, OpenAI’s next-generation model with substantially improved capabilities and multimodal input, launched March 14, 2023 for ChatGPT Plus subscribers.

For software developers, ChatGPT’s emergence as a mainstream interface had implications beyond the specific model’s capabilities. The chat format — system prompt, user messages, assistant messages — became the standard abstraction for LLM interaction, codified in the OpenAI Chat Completions API (launched March 1, 2023) and subsequently adopted as a de facto standard by nearly every LLM API provider. Frameworks like LangChain (October 2022) and LlamaIndex (November 2022), which had launched weeks before ChatGPT, suddenly became widely used tools for building applications on top of the conversational model interface: retrieval-augmented generation, document question-answering, autonomous agents, and code generation pipelines. The ChatGPT Plugins system (announced March 23, 2023) introduced tool use — allowing the model to call external APIs and search the web — foreshadowing the agentic paradigm that would define LLM product development in 2024 and 2025. ChatGPT’s rapid user growth also exposed the limitations that would dominate AI safety discourse for the following years: the model would invent plausible-sounding citations, confidently assert incorrect facts, and produce biased or harmful content when prompted carefully, despite its RLHF training — demonstrating that human preference optimization improved user satisfaction substantially without guaranteeing factual accuracy or preventing misuse.