Where did Apple AI go? Former employees reveal why Siri is on the decline

Why did Apple go "quiet" in the latest round of ChatGPT trends?

The answer has surfaced even further.

Internal team confusion, slow decision-making, and clunky code have all become reasons that slow down Apple's AI to move faster.

The most direct manifestation of this can be seen in Siri.

This is probably the entrance of most ordinary people to know AI assistant, a king bomb function launched during the heyday of iPhone; many people understand ChatGPT, and can't help but take it as a reference.

However, after comparison, it will be found that today's Siri is completely incomparable to ChatGPT in terms of intelligence. There is even the news that "Apple's internal staff dislikes Siri". Why is this situation now? Three AI engineers from Apple to Google, revealed more details.

Management chaos, slow iteration

These three AI engineers are not small. Their change is said to have alarmed Google CEO Xander Pichai and Apple CEO Cook. The former was overjoyed by the news and reportedly made personal overtures to the three; the latter tried to retain them, but was ultimately unsuccessful.

Originally from Google, they left to found the AI company Laserlike, which focuses on personalized search engines and is dedicated to AI development. The company was later acquired by Apple, intended to help Apple develop search engines. In November last year, it was revealed that all three had left Apple and returned to Google.

At that time ChatGPT has not been released, the outside world's attention is still focused on the Apple search engine will be on where to go. The specific reasons for leaving the job is unknown.

Now, the main character unveiled the mystery: because it was felt that Google is a better place to develop LLM (Large Language Model).

Apple is currently severely hampered in Siri and AI development due to too much caution and internal organizational chaos. According to them, the Siri team has been in disarray since 2018. The main reason for this is competitive friction between senior leaders and heated discussions over where to go with the AI assistant.

Siri team leaders didn't want to invest too much money in analyzing Siri usage. Engineers also had no way to know more details about Siri usage, such as how many users actually use Siri and how often they use it. But the Siri data from the data analysis team, and pressed not to be used.

Some departing employees bluntly said, "Isn't this a waste of time and money?

And the deeper problem is reflected in the technical development and judgment.

In the New York Times article, the engineer who was responsible for improving Siri said that the Siri database contains a huge list of words, including 24 language expressions for various music artists, restaurant locations and so on.

He described it as a "big snowball," and if you wanted to add a word to the database, it would be crammed into a huge pile of content. This leads to simple phrase updates that require rebuilding the entire database, which can take up to six weeks. If it is a complex function, long as a year is also very likely.

This means Siri can't be a creative AI assistant like ChatGPT, the engineer said.

Apple executives are said to have rejected the proposal to let Siri have long conversations, because it would make Siri conversations difficult to control and "very fancy".

This, coupled with Apple's uncompromising stance on privacy and security, makes it a huge challenge to further enhance Siri's capabilities. In contrast, OpenAI, ChatGPT collects user chats to train the model to improve the results.

At the same time, to prevent Siri from "babbling", Apple's preferred route is to have the human team pre-write Siri's responses instead of using AI to generate them.

For example, ask Siri what the price of an iPhone is. It will not reply positively, but will only give a link to the official website. Contradictions arose from this.

Siri's design team demanded that its response accuracy needed to be close to perfect before the feature went live. Engineers spent months trying to convince them that not every question needed to be verified manually, which would have limited the range of questions Siri could answer.

The Apple design team also repeatedly refused to allow users to give feedback on Siri's answers to questions, resulting in the development team being unable to understand the limitations of the model. In contrast, ChatGPT allows users to give comments on each answer, which are critical for model optimization.

ChatGPT conversation interface

In 2019, as if to make a change, the Siri team developed an intelligent assistant codenamed "Blackbird".

It can be thought of as a lighter version of Siri, allowing app developers to create features on top of it and run it natively on the iPhone, which also provides better performance and privacy.

The Blackbird demo is said to have worked well, or at least the internal employees were excited to see it.

But that's when the problems between the teams reared their ugly heads.

It was said that some executives wanted to push a project codenamed "Siri X", which was created for two main reasons: first, to mark the 10th anniversary of Siri and second, to better protect privacy and security.

Then hundreds of employees from the Blackbird team were transferred to the new project, and Blackbird was never completed.

In 2021, Siri X was completed, allowing many of the voice assistant's functions to be handled locally.

After all of the above, Siri has come to 2023.

Dissatisfaction with Siri has reportedly emerged within Apple from the equally scrutinized MR team.

Sources close to the situation revealed that the team leader Mike Rockwell (Mike Rockwell) expressed disappointment after looking at the effect of Siri control headset devices, and once considered developing other methods to replace Siri, but eventually gave up the idea.

One More Thing

It is worth mentioning that the internal chaos exposed by several former employees was particularly evident in 2018~2019. And if you look back, this period also coincides with a period of great change in Apple's AI team.

In 2018, Apple was revealed to be setting up a new AI/ML team that would merge the CoreML and Siri teams under the unified leadership of John Giannandrea, who would report directly to Cook.

John worked at Google for eight years before joining Apple. Upon joining, he will be responsible for the company-wide machine learning and AI strategy, as well as technology development for CoreML and Sir.

In 2019, Bill Stasior, vice president of Siri, who had been at Apple for seven years, proposed to leave the company and later joined Microsoft as a technology VP.

At the time, this move was seen by outsiders as Microsoft CEO Nadella boosting his focus on AI.