2022

The first half of the year belonged to images. OpenAI announced DALL·E 2 in April and Google announced Imagen in May — claiming better results than DALL·E 2 while stating plainly that it would not release the model. Midjourney opened its beta in July, and in August Stable Diffusion shipped with its weights. Anyone could now generate pictures on their own machine, and the effect on people who draw for a living became an argument. In Japanese communities, the long English prompts that coaxed out the right picture were traded as "spells" (jumon), and "AI eshi" — AI illustrator — took hold as the name for people who made pictures this way, part title and part jab.

Large models moved too. In January, OpenAI announced InstructGPT and made its human-feedback-tuned models the default in the API. Google announced the 540-billion parameter PaLM in April; OpenAI released Whisper in September.

Work on the wording handed to models appeared too. In January, researchers at Google published a paper on chain of thought, showing that worked examples spelling out the intermediate steps raised accuracy on hard problems. In May, researchers at the University of Tokyo and Google reported the same effect from a single added line, "Let's think step by step." Neither touched the model itself.

In June, a Google engineer's claim that LaMDA was sentient spread through the press. The same month, GitHub Copilot became generally available.

Then, on 30 November, OpenAI released ChatGPT. It was not a new model — the existing GPT-3.5 line put into a conversation — but it was received differently. A million people used it within five days. In December, Google was reported to have called an internal code red in response.

Also in December, Anthropic published its Constitutional AI paper, showing how to make a model harmless using feedback from an AI rather than from people.