What was Siri when it launched on the iPhone 4S, and why did it matter?

On 4 October 2011, Apple unveiled the iPhone 4S in Cupertino, California. The phone looked almost the same as the iPhone 4, and most of the upgrades were the usual ones: a faster chip, a better camera, 1080p video. The feature that caught everyone's attention was something you could not see. Apple called Siri "an intelligent assistant," and the idea was simple: hold a button, say what you want in plain language, and the phone does it.

Talking to computers was not new. Dictation software and voice dialing had existed for years. What felt different was that Siri tried to understand intent, not just words, and then act on it across the phone's own apps.

What Apple announced

In its press release, Apple listed what Siri could do: place calls, send texts and email, schedule meetings and reminders, take notes, search the web, find local businesses and give directions. It could also answer factual questions and do calculations.

Apple stressed two abilities. First, context. If you asked whether you would need an umbrella this weekend, Siri was meant to recognize that as a weather question, even though you never said "weather." Second, personal knowledge. Told to remind you to call Mom when you get home, it could look up "Mom" in your contacts and use your location to know when you had arrived.

Siri came labelled as a beta, meaning an unfinished, test version. It spoke English (in versions for the United States, Britain and Australia), French and German. Pre-orders opened on 7 October, and the phone reached stores on 14 October 2011 in seven countries, starting at $199 with a two-year contract.

Where Siri came from

Siri did not start at Apple. It grew out of research at SRI International, a nonprofit research institute in California, and in particular out of CALO, a large assistant project funded by DARPA, the U.S. military's research agency. A startup called Siri was founded by Dag Kittlaus, Tom Gruber and Adam Cheyer. Kittlaus named it after a Norwegian co-worker.

The startup released Siri as a free iPhone app in February 2010. Two months later, in April 2010, Apple bought the company. When the iPhone 4S arrived, Apple pulled the standalone app from its store, so the assistant now lived only inside the new phone. The speech recognition, the part that turns sound into text, came from Nuance Communications, though neither company said so publicly for years.

How a 2011 voice assistant worked

It helps to picture Siri as a relay team rather than a single brain. A request passed through several stages, each handled by different technology:

The key step is the third one. Siri had a fixed menu of things it knew how to do, and much of its cleverness lay in mapping many different phrasings onto that menu. Ask something outside the menu, and it fell back on a web search or a polite refusal.

Why it mattered

Reviews in 2011 were mixed. Writers praised its promise and laughed at its misunderstandings, and Apple itself said it would keep improving the product as it moved out of beta. Yet the launch changed expectations. For many people, Siri was the first piece of AI they spoke to every day, and its pattern became familiar: press, speak a request, hear a short answer. Since then Apple has brought Siri to the iPad, the Mac, AirPods, Apple TV and the HomePod speaker.

It also arrived on the eve of a technical shift. Siri's 2011 design was a carefully engineered pipeline of separate parts. The following year, deep neural networks began winning contests in image recognition, and speech recognition moved in the same direction. In early 2021, Siri's speech recognition itself relies on such learned networks, including convolutional and long short-term memory networks.

How is Siri different from a language model like GPT-3?

Siri is built to act. Its job is to recognize what you want, pick one of a set of known tasks, and carry it out: set the alarm, send the message, open the map. When a request does not fit any task, it has little to offer beyond a search.

GPT-3, which OpenAI described in May 2020 and has been opening to developers through its API since, is built to write. It predicts the next word of a text, one word after another, after training on a huge amount of writing from the web and books. That lets it answer open questions, draft emails or explain ideas in flowing prose, but on its own it cannot set your alarm or check your calendar, and it can state wrong facts with confidence.

The two approaches are complementary. An assistant like Siri knows how to reach into your apps; a language model knows how to handle almost any phrasing and produce natural replies. Many researchers expect future assistants to combine both, using a language model to understand and answer while keeping the reliable, task-based plumbing that the first Siri introduced.