GPT-4o talks

The day before Google's developer conference, OpenAI streamed a demonstration of a model that listened, looked and spoke. GPT-4o, for omni, took audio, images and text as one stream of tokens and produced audio directly, with a response time of about 320 milliseconds, the pace of a human conversation. Earlier voice modes had chained three models, speech to text, text to text, text to speech, and lost the tone, the laughter and the interruptions in between. This one sang, whispered, changed its accent on request, and, in the demo, sounded to many viewers like Scarlett Johansson's character in the film Her.

Johansson said she had declined OpenAI's request to voice the product and that the resemblance was deliberate; the company withdrew the voice. The model itself went to free users, the first time a GPT-4-class system had, and it became the base of ChatGPT for the following year.

GPT-4o is on this timeline as the point where the interface stopped being a text box. The assistants that Siri had promised in 2011 arrived thirteen years later as a model that could hold a spoken conversation, and the humanoid robots of 2025 talked with the same architecture.