In the battle of large models, Huawei is temporarily ahead of Apple

Original source: Box Rice Finance

Author: Zhao Jinjie

Image source: Generated by Unbounded AI‌

After a few months of wind of "all applications are worth redoing with a large model", all terminals also want to use the large model to reshape their competitiveness.

Huawei has become the first manufacturer in the mobile phone industry to give a specific timetable for the implementation of large models. Recently, with HarmonyOS 4.0 accessing the Pangu large-scale model capability, Xiaoyi, a built-in voice assistant in Huawei mobile phones, has become the first smart assistant with AI large-scale model capabilities, and will be open for testing in late August.

According to the above progress, the Huawei Mate 60 series, which will be released in the second half of the year, will also be equipped with large-scale model capabilities, and is expected to become a major selling point against Apple's iPhone 15 series. **

**Before the big model, the 5G chip was once regarded by the outside world as Huawei's secret weapon against the new iPhone 15. **According to online reports, Huawei's 5G mobile phone is expected to return in October this year. However, compared with the expected large-scale model capabilities, the launch time of Huawei's 5G mobile phones still faces many uncertainties.

At present, Qualcomm and MediaTek have successively stated that they have not yet obtained the approval of the US government to sell chipsets to Huawei. Even if the news of the return of Huawei's 5G mobile phone comes true, its competitiveness is not optimistic, because its greater probability will be to cooperate with SMIC to achieve a complete localization alternative, and the price is that the chip manufacturing process can only reach the 7nm level at most. Compared with the 3nm on the iPhone 15, it is two eras behind in terms of technological iteration.

Compared with the status quo of backward 5G chips, Huawei is actually ahead of Apple in terms of large models.

Although Apple CEO Cook also regards the iPhone and other hardware as a carrier for the company to display AI technology, and recognizes the huge wave of change brought about by artificial intelligence, Apple is quite cautious in the development of large models. According to Cook, " We're going to continue to weave that into our products on a very deliberate basis."

According to Bloomberg technology reporter Mark Gurman, Apple is quietly developing AI products internally, and has established its own large-scale model framework "Ajax", and has tried to apply it to product functions such as maps and Siri. There are no plans to release "Apple GPT" to consumers yet.

**In the past ten years, domestic mobile phone manufacturers have led two major product changes: the full-screen wave launched in 2016 once pushed the iPhone out of the top five in the Chinese market; the folding screen wave launched in 2019 has not yet shaken iPhone basics. **

**As another new opportunity for product change, will the large model be a "full screen moment" that promotes the popularization of domestic mobile phone manufacturers, or a "folding screen dilemma" that is limited by the niche? **

**Compared with third-party apps such as ChatGPT and Wenxin Yiyan that have been logged into the mobile phone, the most direct advantage of the large-scale model products built directly into the mobile phone system lies in stronger privacy and security protection. **

Compared with apps such as ChatGPT that process data in the cloud, the localized deployment of smart terminals can eliminate users' concerns about privacy leakage to the greatest extent.

In March of this year, ChatGPT had a cache problem due to a bug in an open source library, which eventually caused the payment information of hundreds of millions of ChatGPT paid users - ChatGPT Plus users to be leaked, including names, email addresses, and credit card numbers. Four digits and credit card expiration date, etc.

Affected by this, the U.S. Federal Trade Commission also began to investigate whether OpenAI violated consumer protection laws in July, and asked OpenAI to provide information about its processing of personal data, the possibility of providing inaccurate information to users, and "caused harm to consumers (including risk of reputational damage” is well documented.

**With the help of the localized deployment of the mobile phone terminal to dispel concerns about information leakage, users can feel more at ease and boldly feed more personal data to the large model. This also has the second advantage of the mobile phone's built-in large model product, that is, to create a truly personalized personal assistant. **

**In addition, the large model products that come with the mobile phone have a more stable operating environment than ChatGPT. **

In April of this year, ChatGPT Plus was suspended for payment. The reason given by OpenAI was that "the demand is too large", resulting in a gap in computing power resources. With the help of localized deployment, the large model that comes with the mobile phone will greatly reduce the dependence on cloud computing resources, so as to avoid the occurrence of "downtime" events, and even run offline when the network is disconnected.

The voice assistant has become the priority entrance for many mobile phone manufacturers to implement large-scale models. With the support of large model capabilities, the voice assistant can help users automatically generate copywriting, automatically write reply emails, automatically generate article summaries, and automatically translate...

Even with the help of plug-in functions, the voice assistant supported by the large model can also complete cross-app function calls, such as mobilizing map applications, travel applications and weather applications, etc., to help users make a travel plan or book a restaurant.

Compared with a batch of voice assistants born more than ten years ago, the largest model is like a catalyst, directly leading Siri to the 2.0 version, and is expected to solve the problem that human voice assistants have been repeatedly criticized as "artificial mental retardation". **

After ChatGPT came out, Microsoft CEO Satya Nadella once complained: "Whether it is Cortana, Alexa, Google Assistant or Siri, these voice assistants are as stupid as a rock."

According to media reports, Apple engineers hope to combine the large model with Siri, and also hope to launch a smarter Siri.

According to Dr. Ding Lei, an artificial intelligence expert, one of the reasons why the previous generation of voice assistants are not smart is that the technical systems behind them are different.

**Siri version 1.0 can be regarded as a decision-making AI. The main job is to "label" the existing data. **To distinguish between different types of data, the main tasks are "judging whether it is" and "distinguishing whether it is this or not." That" job. Once the external request exceeds the existing tag library, the voice assistant will reply with bottom-up answers such as "I can't answer" and "I'm still learning" to continue the conversation.

**Siri 2.0 version is a generative AI, which will "create" new content after summarizing and analyzing the existing data, **realize the effect of "drawing inferences from one instance" and take into account the user's interests while continuing the dialogue Use experience.

Huawei and Apple are not the only manufacturers that register large models on mobile phones.

In terms of domestic mobile phone manufacturers, Xiaomi is the one that attaches the most importance to large models besides Huawei. In April this year, Xiaomi officially established a large-scale model team with a scale of more than 30 people, and said that it is expected to launch the product in the third quarter of this year.

It was revealed that OPPO is building large-scale model products for mobile phones based on the large model of Ali. Zhao Ming, CEO of Honor, also frequently stated to the outside world, "In the future, we will be the first to introduce large models into mobile phones."

In order to be the first to carry large-scale products on mobile phones, Samsung was revealed to be considering plans to change the default search engine of its mobile phones from Google to Microsoft Bing.

In order to restore Samsung's escape as much as possible, on the one hand, Google tried to upgrade the existing search engine and added large-scale model capabilities; on the other hand, it began to integrate large-scale model products from the bottom of the system.

** In May of this year, the new generation of large-scale model PaLM 2 released by Google included a minimum parameter-level Gecko solution. According to Google CEO Pichai, "Gecko" will be able to run on mobile phones, and the speed is fast enough. **

In the second half of this year, Google's new Pixel models will incorporate the latest PaLM2 large-scale gecko version. The Pixel has always been at the forefront of Google's experiments with new Android features. At that time, regardless of whether domestic mobile phone manufacturers have the ability to self-develop large models, they will be able to directly obtain the magic power of large models from Google's Android system. **

In addition to the efforts from system developers, leading chip manufacturers are also worrying about using large models for mobile phones. **

** MediaTek expects that the Dimensity 9300 released in October this year will support large-scale model operation; Qualcomm said that it will be able to support generative AI models with 10 billion parameters to run on mobile phones within this year. **

According to Qualcomm executives, "A model with a parameter level of 10 billion to 15 billion can cover most of the use cases of generative AI. If the terminal can already support this parameter level, then all calculations can be performed on the terminal without cloud processing. .Then the mobile phone will become a true personal assistant."

In addition, after Meta launched the open source large model Llama 2, Qualcomm further stated that it will cooperate with Meta to launch a "mobile version of Llama 2" in 2024 to help customers build intelligent virtual assistants, productivity applications, content creation tools and Use cases such as entertainment.

Compared with the computing power of cloud-based thousand cards and parallel cluster servers, the advantage of localized deployment of smartphones has in turn become a disadvantage that restricts the expansion of computing power.

The smallest version of Meta's open source Llama model also has 7 billion parameters. With the current configuration of smartphones, it cannot run smoothly in the existing memory at all. It can only be partially run in the flash memory of the mobile phone, but this also leads to slow response speed of the large model. .

Previously, in the open source community, some developers ported the Llama large model to the mobile phone, but the process of waiting for a feedback took up to ten minutes or more.

** In order to improve the efficiency of feedback, in addition to strengthening hardware configurations such as chips and memory, the developer began to carry out a slimming plan for the large model on the mobile phone, including but not limited to pruning, quantization, distillation, etc. Under the premise of greatly reducing the accuracy, reduce the resources and energy consumption required by the large model on the mobile phone: **Pruning, that is, cutting out the parameters in the model that have little impact on the accuracy; quantization, using lower-precision data types for reasoning ; Distillation is to extract a similar but simpler model from a complex model.

However, in the case of GPT-4 with trillions of parameters still unable to solve the nonsense of the large model, how to reduce the generation of false content for the large model on the mobile phone that actively reduces the accuracy after downsizing, and the number of parameters is only 1% of GPT-4. To put a question mark.

Dr. Ding Lei said that for large-scale model products such as ChatGPT, there are still many errors and unclear logic in the generated content, and some scenarios cannot be directly applied at the execution level. **

If the large model on the mobile phone side helps users make travel plans or reserve restaurant information, frequent mistakes will inevitably directly affect the user's frequency of use. In this regard, ChatGPT is already a lesson for the past.

After six months of continuous growth, monitoring data from the third-party website SimilarWeb shows that in June this year, the global traffic of ChatGPT’s website and mobile clients decreased by 9.7% month-on-month. This is the first time that ChatGPT has experienced negative traffic growth since its release on November 30, 2022.

In the view of former OpenAI scientist Kenneth Stanley, the current large-scale model products have not yet evolved to the iPhone moment. “Generative AI has definitely taken the world by storm, and it’s true that some people use it every day, but I think we’re still in the exploratory phase of figuring out how we can use it individually. If we all find a really strong reason to Use it anytime, and that should be the iPhone moment."

References:

"Large model is going to the terminal, what about the chip?" 》Semiconductor Industry Observation

"Google's big AI plan: make the big model smaller and embed every product" LatePost

"Interview with the person in charge of Qualcomm AI: It is expected to support large-scale models with tens of billions of parameters to run on mobile phones within this year" The Paper

"Large model "downsizing" into the mobile phone, the next iPhone is coming? 》The Explosion of the Metaverse

"Dialogue with OpenAI Scientists: The iPhone Time Has Not Come yet" Economic Observer