ChatGPT (Nov 2022)

ChatGPT, released November 30, 2022, was a research preview of an InstructGPT variant fine-tuned for multi-turn dialogue, exposed via a free web UI. It reached 100 million users by January 2023 — the fastest-growing consumer product in history at the time. ChatGPT did not contain new capabilities relative to text-davinci-003 (the same generation of model, available through the API since early 2022), but the interface and free public access triggered the global LLM boom.

What ChatGPT was, technically

ChatGPT-1.0 was based on gpt-3.5-turbo, a roughly InstructGPT-class model trained with the SFT + RM + RLHF pipeline on dialogue data. Two technical differences from base InstructGPT:

• Conversational format. Training data formatted as multi-turn [USER] / [ASSISTANT] exchanges, with explicit turn boundaries. The model learned to produce assistant turns conditioned on the full conversation history.
• Refusal behaviour. Heavier emphasis on refusing inappropriate requests, hedging on contested topics, and producing the now-recognisable "as a language model, I cannot..." preamble.

The architecture was unchanged: decoder-only Transformer, autoregressive, ~175B parameters in the largest size.

What was new: the interface

ChatGPT's product UX was the contribution. Free access, no rate limits initially, conversational interface (not "complete this prompt"), persistent dialogue. None of these are technical innovations; together they made the underlying model accessible to everyone instead of just developers.

The free-tier launch was deliberate. Internal OpenAI documents revealed in subsequent reporting that ChatGPT was framed as a "low-key research preview" — projecting maybe 100K daily users. Within five days it had 1M.

What ChatGPT proved

Three things became undeniable post-ChatGPT:

• Aligned LLMs are useful, not just capable. The InstructGPT alignment recipe converted GPT-3-class capability into a product millions could use.
• The interface matters. Same model behind ChatGPT had been available via the API for over six months with limited uptake. Free + chat UI was the difference.
• The market is enormous. 100M MAU in two months redefined the addressable market for AI products. Investment, hiring, and competitive response that followed were proportional.

By February 2023, Microsoft had integrated GPT-4-class models into Bing; Google had launched Bard; Anthropic had launched Claude; the open-source community had begun the LLaMA fine-tuning explosion. The 12 months after ChatGPT's release were the most consequential in AI commercialisation.

Capabilities at launch

ChatGPT-1.0 was strong at:

• Conversational Q&A on general-knowledge topics.
• Writing tasks — emails, essays, code, summaries.
• Code generation at the level of Codex, with explanation.
• Reformatting — converting between formats (CSV ↔ JSON, prose ↔ bullets).
• Multi-turn task assistance — the dialogue format made iterative refinement natural.

What it was bad at:

• Multi-step reasoning — chain-of-thought helped but was inconsistent.
• Math beyond simple arithmetic.
• Factual reliability — hallucinated confidently on details, dates, citations.
• Tool use — text-only; couldn't search the web or run code (until plugins).
• Long contexts — 4K token limit at launch.

These capabilities all arrived in subsequent models: GPT-4 (Mar 2023), tool use plugins (Mar 2023), Code Interpreter (Apr 2023), browsing (initially withdrawn, later re-introduced).

What ChatGPT didn't tell us

The model's training data, exact recipe, parameter count, and reward-model details were not disclosed. Pretraining cutoff was September 2021. RLHF alignment data and human-labeller demographics were not fully described. By 2024, OpenAI's lack of detailed disclosure on its frontier models had become a methodological frustration for the academic community.

The cultural impact

ChatGPT's release reshaped how non-technical audiences understand AI. By February 2023:

• Public discourse around AI shifted from "potential capability" to "current product".
• Schools and universities had to reckon with student use overnight.
• The phrase "AI" came to mean "LLM" in popular usage.
• Existing AI ethics, safety, and policy debates accelerated.

Whether ChatGPT was a good thing or not is contested. That it was a discontinuity in the AI public sphere is not.

What to read next

• InstructGPT — the technical predecessor.
• GPT-4 — the successor model.
• Frontier Models — the post-ChatGPT frontier-model ecosystem.