I. Introduction
This article reflects on the release of Apple AI. Although not having personally experienced it and not fully understanding its methods, it is believed that this release highlights important trends in the AI field today: experiments with four models, including AI models, usage models, business models, and future mind models.
II. AI Models
The Importance of Foundation Models
The capabilities of foundation models are crucial. The largest cutting-edge models outperform smaller models in many aspects, even smaller specialized models. For example, GPT-4 outperforms the specialized BloombergGPT in most financial tasks without specialized financial training.
Different Scale Models
Although cutting-edge models are powerful, they are costly to train and run. Smaller models, while not as powerful as cutting-edge models, can run cheaply and easily on PCs or mobile phones. Mistral 7b, for instance, can run offline on mobile phones, and Mixtral performs slightly better than the original ChatGPT when running on computers. Apple has created some small models that run on its product AI chips, as well as a medium-sized model on the iPhone that can be called from the cloud when needed. The capabilities of the models on the phone are close to Mistral 7b but superior, and the cloud model is better than the original ChatGPT. These small models allow Apple to control AI usage and distribute work between the phone and computer. At the same time, Apple also cooperates with OpenAI, sending questions that its models cannot answer to GPT-4.
III. Usage Models
Characteristics of Large Language Models
Large language models are like Swiss Army knives, capable of assisting with a variety of intellectual tasks, but they have their strengths and weaknesses. Understanding their strengths and weaknesses requires practice and expertise, including knowledge of the models themselves and the tasks they are applied to.
Different Models' Usage Focus
Cutting-edge model creators do not have strong opinions on how their models are used and are not optimized for specific tasks. Using them is like cooperating with someone who makes mistakes and has mood swings. Apple, on the other hand, focuses on enabling AI to complete tasks for users. For example, Gemini 1.5 can handle complex tasks, while Siri focuses more on daily tasks such as sending photos. Google is also about to release a small AI model native to mobile phones, and Microsoft is implementing Copilot in key office applications. Both Copilot and Apple's application-specific Copilot models have constraints, limiting their upper and lower bounds. Cutting-edge models and constrained models have different approaches in use cases.
IV. Business Models
Exploring Charging Models
Access to advanced models costs at least $20 per month, and some companies also sell APIs billed on usage. However, some advanced AI access is free, such as Copilot and ChatGPT-4o. Apple may start for free but may charge fees in the future. AI companies are exploring business models.
Trust and Privacy
Company success requires trust, and there are many reasons why people do not trust AI companies. Privacy is key, and all AI companies offer options not to use user data for training. Apple goes a step further in privacy, with local AI on mobile phones accessing personal data, and cloud AI processing data that is encrypted, anonymous, and deleted immediately.
V. Future Models
The Impact of AGI
AGI is the potential goal of all AI development, and OpenAI and Anthropic explicitly aim for this, considering the AI models on the way as stepping stones. They may not heavily refine existing systems because they believe future models will greatly enhance capabilities.
Different Directions of Companies
Apple builds narrow AI systems that can accurately answer questions related to personal data, while OpenAI wants to build autonomous agents capable of completing complex tasks. Different companies' approaches to AI may promote faster adoption in the long term, and experiments can reveal which model combination is more suitable.