The Importance of Small Language Models (SLM) in AI, how is Apple AI different from others?

Small language models vs. large language models

The difference between the two models

Generative artificial intelligence, such as ChapGPT, relies on artificial intelligence models. The function of an artificial intelligence model is basically proportional to the number of parameters it can process. However, the larger the number of parameters, the greater the resources and computing power consumed, and the greater the cost and time required.

For the AI cost, please see my post of “How Much Does Generative AI Cost?“

The number of parameters matters

Therefore, if the number of artificial intelligence model parameters is smaller, we call them small models to distinguish them from the large artificial intelligence models at the beginning. As for the number, there is no absolute number as a cutting point. Generally speaking, an artificial intelligence model with 5 or 4 billion parameters can be considered a small model.

You can’t have your cake and eat it too

The number of parameters is inversely proportional to the function of a large artificial intelligence or small artificial intelligence model. The choice is a dilemma. It is impossible to have both fish and bear’s paws.

Artificial intelligence large language model

Large language models are powerful, but impractical

Large models are powerful, but impractical. Only a handful of very wealthy listed companies and countries in the world have bottomless pockets and spend at least billions in capital expenditures to continuously invest in updates and build large-scale models that can execute artificial intelligence. Own data center and server equipment.

Large language models are difficult to promote

For any breakthrough technology to succeed, it must be accepted by the public, otherwise it will not succeed. There are too many examples in history, so I won’t go into detail here.

The more people use it, the more successful it will be

What the general public will come into contact with are terminal devices such as mobile phones or computers, not the powerful servers and complex equipment or settings in data centers that engineers or a few people live in.

However, the computing power and storage space of terminal equipment are limited. The more parameters, the stronger the learning ability of the model.

Advantages of small language models

Privacy

All large artificial intelligence models need to be connected to the data centers of technology giants through the cloud, processed by remote servers, and then sent back to users. In this process, personal privacy issues will be involved. No one wants Google to publish the sensitive or shameful websites you browse online or visit online!

Apple insists that it is developing data center and remote server computing methods that are different from other industry peers and adheres to customer privacy. This is different from the current practice in the industry and is much more difficult.

Integrate with third-party artificial intelligence solution providers (such as signed OpenAI, possible partners such as Google, Meta, Baidu, Alibaba, etc.) through Apple’s unique Private Cloud Compute (PCC) , so that both Apple and its partners can benefit and achieve a win-win situation, and Apple can have the best of both worlds.

But this is also why Apple has lagged behind in the development of large models of artificial intelligence. However, in just two days after the WWDC in mid-June released its roadmap for the future of artificial intelligence, its stock price rose sharply, and it is the key to regaining the number one market value in the world. one of the reasons.

Can be executed on the terminal device

Large artificial intelligence models cannot be executed on terminal devices. The computing power, scale, and storage space of terminal devices are limited. The only alternative is to use a reduced-order artificial intelligence small model from the large artificial intelligence model, so that it can be used on ordinary people’s mobile phones. and execute on the computer.

Moreover, the response speed and efficiency of small artificial intelligence models that run on ordinary people’s mobile phones and computers will be much better because they do not need to be connected to the remote end.

Small language models of three tech giants

The following is a comparison of the number of parameters of the current major artificial intelligence small models:

Apple’s small model has about 3 billion parameters.
Microsoft’s latest Phi-3-mini model only has 3.8 billion parameters.
Google’s Gemini Nano-1 and Nano-2 have only 1.8 billion and 3.25 billion parameters respectively.

OpenAI’s small language model

The above three major technology giants consider that they all have their own mobile phones and computer products. In order to execute on their own mobile phones and computers, the number of parameters is basically very small. Let’s take a look at a small model of OpenAI, which is the most powerful and recognized as the industry’s leading one. OpenAI currently does not have terminal devices such as mobile phones and computers, so its design is mainly based on functional considerations.

OpenAI’s GPT-4o is a small model optimized based on the large ChatGPT 4.0 model. GPT-4 is the latest language model developed by OpenAI, with a staggering number of parameters of 1.76 trillion. In comparison, GPT-3, the largest current language model, has 175 billion parameters, and GPT-2 has 1.5 billion parameters. The PaLM-E released by Google contains 562 billion parameters, and GPT-4 contains trillions of parameters.

The GPT-4o model, according to the OpenAI API documentation, has a maximum syntax unit (token) limit of 128,000. Microsoft’s Copilot+ PCs use OpenAI’s GPT-4o.

Apple is different from tech peers

Use Apples’s own private cloud and chips

Apple’s artificial intelligence model uses a private cloud built by Apple itself and does not use other companies’ data centers. In addition, Apple’s data center will use its own artificial intelligence processor design instead of Nvidia’s chips commonly used in the industry.

Through its own software and hardware equipment, Apple will have absolute control over artificial intelligence, thereby protecting the rights and interests of customers and building functions that are different from others.

Apple’s approach to privacy

Apple said that a major feature of Apple Intelligence is the use of private cloud computing to flexibly adjust and expand its computing capabilities, such as calling larger server-based models in the face of complex requests, and these models will be powered by Apple chips. running on the driven server. With such a hardware foundation, Apple can ensure that data will not be stored or exposed.

Apple Intelligence consists of two basic models: one is a local model that can be run on terminals such as mobile phones, and the other is a large model that needs to be calculated in the cloud and connected to ChatGPT. The former is a “small language model” with 3 billion parameters, while the latter’s language model parameters are in megabytes.

To protect user privacy, Apple’s local models never use users’ private data for training. For those requests that cannot be processed by the IOS system, they will be forwarded to ChatGPT. In this way, users must know that if the AI’s answer is unsatisfactory, the responsibility lies with GPT and not Apple.

Adopted blockchain

At the WWDC conference, Apple Senior Vice President of Software Craig Federighi mentioned that Apple’s artificial intelligence model uses a private cloud built by Apple itself and uses “blockchain-like” technology to protect privacy. Used to verify the correct operating system and software; it includes three parts:

Adopting the most respected transparent and tamper-proof logs of blockchain, that is, encryption during data transmission.

Compare on the user’s device to ensure that the system and software are the same.

Virtual environment using Apple PCC software, executed on Apple-equipped devices. Confirm that the signature comparison is consistent to prevent it from being changed during the process.

Restrict partners’ activities

The first is that Apple’s cooperation with OpenAI is not exclusive; in fact, Apple is already negotiating similar cooperation with other artificial intelligence companies such as Google, Baidu and Anthropic.

Apple will decide which partner manufacturer’s response is most beneficial based on Apple’s role and the needs of users. For example, in the Chinese environment, if there is no accident, Baidu’s Ernie Bot large language model will be adopted; because regardless of the overall performance of the model, Chinese grammar and complexity, Baidu’s accumulated data and experience over the years are far beyond Google or what OpenAI can provide.

As currently confirmed, ChatGPT-4o provided by OpenAI will only be available as an option when using Siri or a writing tool. And Apple will remind you every time that you will send data to ChatGPT for analysis of article data, thereby providing user data and privacy protection, and obtaining user confirmation.

In addition, when users connect to ChatGPT through Apple devices, Apple obfuscates or hides the user’s IP address, which also makes it difficult for ChatGPT to correlate multiple conversations to identify the user. Apple also stated in the terms of cooperation signed between the two parties that OpenAI is prohibited from obtaining the user’s IP address to prevent abuse.

Disclaimer

The content of this site is the author’s personal opinions and is for reference only. I am not responsible for the correctness, opinions, and immediacy of the content and information of the article. Readers must make their own judgments.
I shall not be liable for any damages or other legal liabilities for the direct or indirect losses caused by the readers’ direct or indirect reliance on and reference to the information on this site, or all the responsibilities arising therefrom, as a result of any investment behavior.