How did we get to today's chatbots?

Once the transformer existed, the recipe for a few years was largely to make the models bigger and train them on more text. That scaling is what produced the chatbots most people now use.

• 2020 — GPT-3. OpenAI described GPT-3 in a preprint dated 28 May 2020 and opened API access the following month. With 175 billion parameters, it showed that a single large model could handle many language tasks from a prompt alone, without task-specific retraining.^[1]

• 2022 — ChatGPT. OpenAI released ChatGPT on 30 November 2022. Wrapping a capable model in a simple chat interface brought the technology to a general audience and set off the current wave of interest.^[2]

• 2023 — GPT-4. OpenAI released GPT-4 on 14 March 2023. Unlike its predecessors it was multimodal, accepting images as well as text as input.^[3]

• 2025 — Coding and agent models. Frontier releases shifted toward software development and tool use. Anthropic announced Claude Opus 4.5 on 24 November 2025, positioning it for coding, agents, and computer use.^[4] Anthropic’s own announcement claims it reached state-of-the-art on SWE-bench Verified, a real-world software-engineering benchmark, and that on a timed take-home engineering exam it “scored higher than any human candidate ever” within the two-hour limit, using parallel test-time compute rather than a single pass.^[4]

Version numbers in this list will keep changing. What is steadier is the direction: bigger trained models, then a usable interface, then models aimed at acting through tools rather than just answering. That last shift is its own story, covered in tool-using agents, and it’s also where the open questions in this subject’s final topic start to bite: a model that can act, not just answer, raises the stakes on getting reliability and alignment right.