Twelve years ago, in October 2011, Apple unveiled its voice assistant Siri along with its iPhone 4s series, a feature that was highly anticipated at the time.。

At the time, Apple executive and current Apple Academician Phil Schiller summed up the performance of all Siri's previous voice interfaces in one sentence at the press conference, calling it "disappointing" and chanting, "What we really want is to communicate smoothly with our devices."。In Schiller's eyes, Siri is a "kind personal assistant" that can provide users with an unprecedented interactive experience along with Apple's smart devices.。

From heaven to hell

The honeymoon period at the beginning is always perfect, and the market's first impression of Siri is almost full marks。Tech media The Verge once asserted, "The best thing about Siri is how it works - at least in most cases, it can exceed user expectations.。CNN's view is, "It's kind of like our dream free assistant, on call.。The New York Times also said, "Siri has saved us time, eliminated unnecessary operations, and profoundly changed the definition of mobile phones."。"

But, like all lovers, Siri's ills have gradually come to light over time: frequent misjudgments about user instructions, lack of leaps and bounds in actual functionality, rapid fading of first-mover advantage, and corpus constraints.。

In 2012, less than a year before its release, Siri was completely strangled by the three-star system S Voice; in 2014, according to a one-on-one test of handheld technology, Siri was already at a disadvantage in the Google Now evaluation of the Google Department.。

2017 is a key year for Siri's wind review to turn. From this year, users' dissatisfaction with Siri began to spread. Siri gradually changed from a popular pastry to a street mouse that everyone shouted at.。

This year, Siri launched a brand-new voice pack that was of little use other than making it sound less like a robot.。Siri's instant translation feature is also late in the year, and Apple hopes to use it to bring in business people and travel groups with real-time translation needs.。In addition, in this year, Apple also equipped Siri with a function called "built-in learning," which is actually a personalized recommendation that everyone is used to today.。

Despite many efforts, the market has grown weary of these painless updates。The higher people's expectations of Siri, the greater their disappointment, and even The Verge, which had previously praised Siri, began to turn against it, saying it had "clearly failed to keep up with the times."。

Clumsy corpus design becomes Siri's biggest constraint

So, what is the bottleneck of Apple, which has forged a technology empire, on Siri?？

The answer is Siri's clumsy corpus design。

According to Siri's current user interaction ecology, when a user gives a command to Siri, Siri needs to extract the corresponding corpus from the database to understand the meaning of the user's command, which is also known as Siri's "command-control system."。The biggest drawback of this model is that once Siri's corpus is finalized, it is difficult for it to understand instructions outside the original corpus。If the user wants to expand the scope and effectiveness of the instruction, he can only rely on engineers to add new words to the database, otherwise it's a chicken and a duck.。

A very simple analogy is that you can think of Siri as an old-fashioned grocery store where you have to go to a salesperson to get whatever you want to buy.。The goods of the grocery store are honestly placed there, the type and style will not change in a short time, if you want to buy something that is not here, you can only call the salesperson to call the manufacturer to purchase the goods.。

The same goes for Siri, except that it's not in the goods, it's in the corpus.。

To build a voice assistant, Siri's database is not huge。It is understood that Siri's database contains a large number of phrase lists in nearly 24 languages, and if these phrase lists are populated according to a certain proportion, this data has become a huge snowball。In these ancestors left behind the "snowball," some of the code a change on the collapse, while others want to change but there is no way to start, these ancestral code has gradually become what programmers often say: shit mountain code.。

That's why big language models are so important for generative AI。If there is no big language model, even the best interactive technology will fall step by step.。Artificial intelligence programs such as Chat GPT and Bard, which are currently on the market, are constantly enriching their language models by constantly grabbing a large number of user corpus on the Internet to achieve excellent interaction effects.。

But Apple can't do that.。The reason, many people think that Apple's excessive attention to user privacy, so that it can not be like Open AI, Google and other head of artificial intelligence companies to collect a large number of user data, and then use these resources to improve their AI systems, so that Siri appears more and more "stupid."。

After 12 years of waiting, Siri finally ushered in the dawn of rebirth.

The good news is that Siri's decade-long underperformance may be about to go away.。

A few days ago, according to Mark Gurman (Mark Gurman), Apple has completed the basic framework of its big language model, and named it "Ajax," positioned to support conversational AI systems, and has been applied to maps, Siri and other functions, do artificial intelligence improvements.。

The work, led by John Giannandrea, Apple's head of machine learning and artificial intelligence, and Craig Federighi, its head of software engineering, is steadily advancing.。According to Gennandra, he wants to take a more conservative approach and see how other companies' recent developments have evolved。

According to people familiar with the matter, the biggest use of this big language system is to integrate it into Siri, allowing voice assistants to help users perform more tasks.。Gurman also believes that Apple's integration of its big language model technology into Siri is an ideal choice because it will allow the voice assistant to perform more tasks on behalf of the user.。

And there's a natural advantage to Apple's big language model research, which is that constraining the computing power of many big factories is not a problem at all here at Apple.。According to Apple's latest "chip monster" M2 Ultra, the chip splices two M2 Max together and has 134 billion transistors, 20 billion more than the previous generation M1 Ultra。In addition, with 192GB of unified memory, the M2 Ultra can break through the previous constraints of insufficient memory and the inability of GPUs alone to handle large models, enabling a single device to run large machine learning workloads such as large Transformer models.。

With Ajax's blessing, Apple's Siri transformation is accelerating。According to the media, many teams at Apple, including engineers working on Siri, are regularly testing "language generation concepts" every week.。In addition, Apple is already in tvOS 16.4 to test a new framework for "Siri natural language generation," internally codenamed "Bobcat"。

In fact, Apple's ambitions are not limited to this。According to the media, in addition to wanting to carry a large language model framework on its own applications, Apple has created a chatbot service based on Ajax。For the product, Apple insiders named it "Apple GPT," in contrast to Open AI's ChatGPT and Google's Bard.。

Interestingly, Apple's Ajax framework was created last year based on Google JAX and is still running on Google Cloud.。

Due to concerns about the safety of artificial intelligence, Apple's GPT project has been limited to "a small engineering team's experimental project" since it was launched at the end of last year.。The system requires special approval to access, and employees cannot use any of Apple GPT's outputs to develop user-facing features.。Gurman said Apple is actively iterating on the Apple GPT, but has no plans to release the product to consumers。

Waiting for twelve years, Siri finally ushered in the dawn of rebirth。With a powerful model and the blessing of computing power, I believe this time Apple's transformation of Siri will no longer be superficial.。As the commentary said, Apple, which has 1.5 billion active iPhone users, can change its landscape in an instant if it really joins the big model battlefield.。

We also have reason to believe that this time, Apple will drive a good Ajax, with Siri "over the shit mountain"。