Apple's Siri with a large model may not be what you think

Apple's AI, belated, arrived.

According to the Wall Street Journal, Apple is in talks with Baidu about integrating generative AI into iPhones and other devices in the domestic market.

Although there has been no official confirmation yet, two things are certain:

• 1

  iPhone 16, iOS 18 and MacOS will be powered by AI

• 2

  Large models of Apple devices will be provided by different manufacturers at home and abroad

Compared with domestic brands that have long been equipped with AI assistants, Apple is unsurprisingly more than half a year late this time, and slow steps seem to have always been Apple's label, but they can always bring some surprises in their steady progress.

However, the speed of progress of large AI models is being counted in weeks or even days, and Apple's lateness is another latecomer, or the beginning of a new era of laggard.

01 A slightly compromised plan, the focus is to get in the car first

On the last day of last month, Apple announced in a 12-minute short meeting: abandon car building, all in AI, and many members of the automotive team will be transferred to the AI department.

The Titan plan, which has been dormant for ten years, fell in the last year of entering the new energy vehicle, and it will be a little regrettable for the future automobile market, but from the perspective of the long-term development of a technology company, this is a long-term and correct choice.

AI is a fundamental application, and at a time when all large companies are actively or passively embracing AI, Apple's "disconnection" ̇ adapts to the times, and how to embrace AI, and what kind of artificial intelligence can occupy a place in the gradually carved up market? is the first problem they have to solve.

For foreign markets, Apple is actively negotiating with Google to include large AI models in iOS 18 to implement AI features that other brands already have.

Although "the terms or brand of the AI agreement have not yet been decided, and it has not been finalized how to implement it", among the many alternative partners (OpenAI and Anthropic), Google and Gemini should be the most suitable for Apple and iPhone.

Samsung's Galaxy S24 series models, which were released in February this year, are out of the circle with AI functions, and features such as call translation and creative writing have caught up with the domestic average.

The overseas version of the S24 series is supported by the large model Gemini to complete the above functions.

Speaking from experience, Google has completed the initial attempt on the flagship model with the highest shipment volume in the world, and they know better about the operation habits, usage scenarios, and adaptation applications of mobile phone models than manufacturers that are popular on PC or Web.

Moreover, Google itself is more eager to get Apple's project.

According to IDC, Samsung's global smartphone market share will reach 19.4% in 2023, while Apple will reach the top with 20.1%.

If it wins Apple, Gemini's mobile phone terminal adoption rate will reach 4% worldwide, which is extremely good for an AI large model company facing fierce competition.

In addition to Google, Apple woke up laughing in the dream.

Unlike other manufacturers that emphasize "self-development", Apple has also had its own considerations in cooperating to achieve AI from the beginning.

First of all, in the current situation of its own late start and slow progress, the "take-it-or-leave-it" doctrine is a clever hand to quickly compete for the market, and the cooperation with Google can reduce R&D costs and charge high pit fees, while also alleviating the regulatory pressure that the two companies are currently facing.

Secondly, AIGC's technology is very good, but it has been criticized a lot due to defects in ethics and privacy when it is implemented, and it is handed over to mature third parties, especially Google, which has successfully tested the waters on Samsung models, saving effort and worry, and reducing the risk of public opinion and liability.

Another hurdle in this is technology localization. Each country and region has different requirements for the regulation and related regulations of AI large models, and the implementation of legal compliance is the prerequisite for competing for the market and developing technology, so the two-pronged route of "domestic + international" has been born.

According to the initial success of the cooperation between Samsung and Baidu, Apple will choose this route that has been "verified to be true".

The AI functions on the Samsung S24 series are actually composed of technologies from multiple manufacturers: the instant search function is provided by Baidu and JD.com, the intelligent photo editing is done by MiracleVision, a large model of the Meitu App, and the large model of Wenxin Yiyan is used for article summarization and intelligent writing.

Whether Apple will also cooperate with a number of manufacturers will have to wait for follow-up news, but the cooperation with Baidu is already a certainty.

In the end, what Apple wants to do is not a smart voice assistant, but a whole set of AI terminals. However, according to Macrumor's revelations, with the current self-development progress and technical achievements, Apple's large model is far from the level of Google, OpenAI and other companies.

Instead of rushing ducks to put an intelligent chatbot on the shelves, it is better to make a transition with a mature solution first, and strive for more research time and progress space for self-developed large models.

02 The current market is important, but the core technology of the future is fundamental

Cooperation is the first step in Apple's AI globalization, and the ultimate goal is to have a self-developed AI model.

This is a money-burning and energy-consuming project, not to mention regression, a little slower progress, and it may be eliminated next week. A competitive model often represents the dominant and bargaining power in the market in the future.

According to Cook:

Breaking new ground in generative AI, we believe this technology can redefine the future.

And Apple's exploration of large models has always been on the agenda.

On the 15th of this month, Apple engineers quietly released a research paper detailing the development process of a new generative AI model called MM1.

MM1 is a family of multimodal LLMs with up to 30B (30 billion) parameters, which is Apple's latest research on multimodal large models.

In general, Apple's self-developed model still has a certain gap with Gemini and GPT4V in terms of test results, nor has it shown the same amazing effect as Sora in generating results, let alone explored a new technical route.

However, it can control various data variables to find out the most critical factors that affect the effect of model generation in the comparison, in short, it is not strong by nature, but it is good at observing, practicing and summarizing, and can also achieve good results in repeated attempts.

MM1 consists of a dense model and a MoE (Mixed Expert) variant, and when the command enters the MoE, whether it should go to the "East Market to buy a horse" or "the West Market to buy a saddle" will be clearly arranged by this command center.

While the problems are refined and classified, the calculation efficiency is also improved and the operating energy consumption is saved.

The release of this paper represents the phased results of Apple's exploration in the field of AI, although MM1 has not disrupted the industry or amazed the world, but their progress can still be seen in obscure technical terms:

Our working model has always been to do the work first and then talk about the work, rather than speaking out in front of ourselves. —— Tim Cook

Apple, which did not disclose too many technical details, is actually still thinking about another move: the end-side large model.

As early as the end of last year, Apple proposed a method for large models to land on "memory-limited" devices such as iPhones in a paper titled "Large Language Models in Flash Memory: Efficient Large Language Model Inference in Limited Memory".

According to the researchers, they have successfully deployed LLMs (large language models) on iPhones and other memory-constrained devices using the latest flash technology.

The project, called Apple GPT, is designed to store LLM data directly in flash memory, such as Siri, and the new technology enables up to 5x and 25x faster CPU and GPU inference speeds compared to traditional operating methods.

"These efficiency-based methods we have developed allow AI models to run in twice the current memory of the iPhone," the researchers said. 」

In other words, it is feasible to carry a large model on the side side, which allows the data of the LLM to be stored directly in the flash memory by reducing the amount of data transferred by flash memory and increasing the throughput of each transfer.

03 Technology aside, Siri is the bridge between us and AI

The progress is slow, the message is few, and the layout is large. Here's an overview of Apple's exploration of AI.

Every time you see Apple's technology lagging behind the market and competitors, it will give people the illusion of "starting too late", in fact, when looking at relevant news and patent documents, you will find that it is often the first batch of layouts, or even that one.

As of 2023, Apple has acquired a total of 32 AI companies, ranking first among tech giants. The acquisition of Siri should be the beginning of Apple's entry into AI.

In 2010, Steve Jobs made a phone call to Dag Kittlaus, the "father of Siri", which led Siri to join Apple and go to the iPhone for more than $200 million.

Siri was originally positioned as an assistant to get information quickly and accurately, or to handle complex tasks.

In its original version, Siri could connect to 42 web services—from restaurant review site Yelp and ticketing site StubHub, to movie review site Rotten Tomatoes and math calculation site Wolfram Alpha.

When prompted, Siri consolidates the messages and responds to the user. Without having to open another app, Siri can help users buy tickets, book restaurants or hail taxis.

These "AI features" that are now touted by AI Pins and other smart assistants seem to be just the "basics" of Siri more than a decade ago.

However, Siri's actual experience has been opened up with the explosive development of AI models.

Intelligent assistant is a passive imitation of people, answering questions and responding to requests.

AI terminals, on the other hand, are active approaches to people. According to the user's personal habits, preferences, and preferences, after summarizing the past and reasoning, we will give you the most appropriate suggestions and answers at different times and places, and can continue to learn and optimize, becoming "private exclusive".

▲ Picture from: x.com

On the whole, Apple's late arrival is only relatively late, because AI phones are still in the early stages of development.

Indeed, most domestic brands have made efforts in the stage of AI terminals, with roughly the same functions and different strengths, but the usability of each large model can only be regarded as a passing grade, except for subdivision functions such as AI elimination of OPPO photo albums, real-time translation of Samsung calls, and AI calls of Xiao Ai's classmates, most of the experience is still somewhat different from independent AI applications.

Among them, in addition to the manufacturer's technological breakthrough, it is also related to the open interface of the App, such as the model that does not support WeChat voice call summarization, which is a large piece of application space in daily life.

As a result, there is still a long way to go in the integration of large models and systems, apps, and the exploration of new ways to interact. Until then, AI capabilities had not been able to influence consumers' purchasing decisions.

In the first year of AI, Siri's goal is to bridge the gap with other AI assistants for more than half a year, and as an important part of Apple's future layout, we are looking forward to what kind of "One more thing" Siri will bring in June.