AI model, the most topical artificial intelligence application in 2023.

Smartphones are the most popular smart terminals in the world.

One represents the wave of artificial intelligence in the future, and the other is the most successful technological product in the past 20 years. In Chinese, they all have the same word: intelligence.

Therefore, the combination of the two seems to be “destined”.

In April this year, Lu Qi, founder and CEO of Qiji Chuangtan, said in his speech that the future is an era where models are everywhere, and predicted that one day large models will run on mobile phones.

At the 2023 MWC Shanghai Mobile Communication Conference not long ago, Zhao Ming, CEO of Honor, said that he will take the lead in introducing AI large models into the device side to create a “personal model” on the device side. He also said, “The large model has broken many of our thinking boundaries and is the best interpretation of AI we have seen so far.”

At that time, the Honor Magic V2, the flagship mobile phone with the highest price, was about to be released, and the outside world interpreted it as “Honor will carry a large model on the Magic V2 for the first time”.

On July 12, Honor released the Magic V2 folding screen mobile phone. The results revealed that the Honor Magic V2 did not become the first large-scale smartphone. In the more than one hour press conference, Zhao Ming only mentioned a big model. However, although Honor is “late”, the discussion of the mobile phone + AI model has not stopped at all.

Google, Huawei, Xiaomi, Baidu, etc. have all indicated that they will adopt large-scale model-related technologies on mobile phones. For example, Baidu launched the Xiaodu mobile phone in May this year, which was also interpreted by the outside world as Baidu’s hope to implement a large model on the hardware side.

It seems that mobile phone manufacturers seem to be the relatively reliable ones among various large-scale “concept stocks”. It’s just that there are already too many hot spots for big models, will mobile phone manufacturers be an exception?

**AI large model + mobile phone =? **

How mobile phones should be combined with AI models, there are currently three mainstream solutions in the industry:

One is an App similar to cloud services or ChatGPT. The cloud provides large-scale model calculations and then feeds them back to the mobile phone. The advantage of this solution is that it does not need to rely on the weak computing power of the mobile phone. The disadvantage is that it needs to be connected to the Internet, and information security is a problem;

The second is local deployment, local computing, and local services, directly using the processor that comes with the mobile phone to run the large model, which is what Zhao Ming called “device-side”. Its advantage is that it is not restricted by the network, and the information is stored locally, but whether the computing power of the mobile phone can support the operation of the large model is a problem;

The third solution is to combine the above two, the cloud and the terminal run together, and the advantages and disadvantages are also the advantages and disadvantages of the above two.

In fact, before Honor announced that it would introduce AI models on mobile phones, the first company to try to establish a connection between mobile phones and AI models on a large scale was Google, the home of the Android system.

Of course, Google has good reasons to combine mobile phones and large models - its Android system controls 81% of the world’s smartphones, has two top AI research institutions, Google AI and DeepMind (AlphaGo developer), and is the infrastructure of large models The creator of Transformer. No matter how you look at it, Google is the best target for implementing the mobile phone + AI model.

But since OpenAI released ChatGPT at the end of last year and became popular all over the world, Geegle has become a “laggard” in the outside world. An indisputable fact is that despite the huge investment in the field of AI, Google failed to become the most dazzling star in this wave of AI.

So, in May of this year, half a year after the release of ChatGPT 3.5, Google finally released a new generation of large language model PaLM2 to fight against ChatGPT. As a differentiated competition, PaLM2 can be deployed on smartphones, while ChatGPT training requires tens of thousands of NVIDIA A100 graphics cards.

Of course, Google also knows the gap between mobile phones and professional AI graphics cards, so not all large models can be deployed on mobile phones. PaLM2 contains four large models, named from large to small according to the parameter scale: Unicorn (Unicorn), Bison (Bison), Otter (Otter) and Gecko (Gecko). Only Gecko, with its minimal parameters, can run on a phone, and Google says it’s fast enough to work offline.

**What is the use of the AI model on the mobile phone? **

Apart from obscure technical terms, what kind of functions and services the AI model can bring is more important to billions of smartphone users. If it is just a chat robot like ChatGPT, then everyone may ask, “Why do I have to use the second ChatGPT?”.

Google’s approach is to integrate the large model into specific applications, so that the App has its own large model features. For example, Gmail can automatically compose emails. Users only need to enter their needs in “Help me write” in Gmail, and it will combine the information in the previous emails to write a complete email, such as persuading your good friend Li Hua to study hard every day.

“Help me write” is only part of the Google Workspace (workspace) function. Other functions of Workspac include automatically writing speeches based on PPT content. It’s unclear whether such services will run in the cloud or on-premises.

As a hardware company, Qualcomm and Google have completely different ideas. They want the large model to run on the phone as an independent application, rather than a part of an application.

In June of this year, Qualcomm released a demonstration video: an Android phone without Internet connection uses Stable Diffusion to generate an AI picture, and the whole process does not exceed 15 seconds.

As an AI image/animation generation tool, Stable Diffusion has also been in the limelight in the past year, but before that, Stable Diffusion was almost always used by individual users on computers, and its requirements for computing power are not low.

The reason why Qualcomm can run Stable Diffusion on mobile phones is mainly because it combines the cloud and the terminal to create a set of “hybrid AI”: if the computing power demand is large, it will be handed over to the cloud for computing, and if the computing power demand is low, it will be completed locally by the terminal. They can also run concurrently. In order to endorse its own plan, Qualcomm even released a white paper called “Hybrid AI is the future of AI.”

However, it is still difficult to say which solution will become the mainstream in the future.

**Who needs big models more? **

Among the mobile phone manufacturers, Honor is the first manufacturer that clearly stated that it will deploy AI large-scale models on mobile phones. There are also Honor’s own considerations behind its enthusiasm.

As of the second quarter of this year, global smartphone shipments have declined for seven consecutive quarters. Mobile phone manufacturers such as Apple, Samsung, and Xiaomi have all been spared, and Honor is also not optimistic.

According to TechInsights data, in the first quarter of this year, Honor mobile phone shipments in China fell by 22.2% year-on-year, the second largest decline among the top five domestic mobile phone manufacturers, and its market share also fell from second to fourth in the same period last year.

At a time when the mobile phone market is shrinking and the homogeneity is becoming more and more serious, the AI model has given a new imagination to smartphones. Although its future is unclear, Honor needs it.

Google has the most software users in the world, with more than 500 million users of its 15 Apps; Qualcomm has shipped more than 2 billion products with AI computing capabilities so far. To some extent, they also have reasons to be more active than mobile phone manufacturers in promoting mobile phone + AI large models.

But on the other hand, “excessive mobile phone performance” and “overblown system” have also been controversial topics in recent years. Google, Qualcomm, and mobile phone manufacturers need large AI models, but they also have to face questions from users. The best negative teaching material is the cloud game of “playing 3A masterpieces on mobile phones” that was once highly expected, but it still has not become the mainstream of the market until today. Even Google itself shut down its cloud gaming platform.

Furthermore, Apple, the “smartphone vane”, has not made any relevant actions so far. At WWDC in June this year, Apple didn’t even mention “AI” or “big model” once. When describing related technologies, they used a more traditional and more academic vocabulary - “Machine Learning” (machine learning) ).

Apple has always been a typical representative of “don’t release it if you don’t do it well”. Therefore, Apple’s conservatism may prove from the side that the application of AI large models on smartphones is still in the functional exploration period when user needs are unknown.

But in any case, innovation in the mobile phone industry has been silent for a long time. If the big model can bring new markets and vitality, everyone is obviously happy to see it. Even if it fails, it is nothing more than another “misjudgment” by the giants.