Summary of Key Points
This article challenges the industry misconception that "the larger the AI model, the better." It emphasizes that the true value of implementing AI in businesses does not lie in the number of model parameters, but rather in engineering skills—specifically, the ability to clearly define tasks, set rules for the models, and design robust work processes. Large models are suitable for handling ambiguous requests (when users are unable to articulate their needs precisely), but once tasks are well-defined, smaller models, combined with engineering efforts such as fine-tuning, using predefined prompts, and establishing clear procedures, can be more reliable and cost-effective. A mature AI system should utilize a hierarchy of models, rather than relying solely on the largest one. The effectiveness of the system is ultimately determined by the depth of its engineering design, not by the size of the models.
Detailed Analysis
1. The Strengths of Large Models
Large models are adept at handling unclear questions, but they come at a cost. Many users ask vague requests when using AI, such as "Create a business plan for me" or "Analyze this industry," without specifying goals or requirements. In these cases, large models have an advantage: they have extensive knowledge and can infer users' actual needs, providing seemingly reasonable answers. For example, when asked to create a business plan, a large model might assume you are a startup in need of market analysis and financial forecasts, and even suggest industry directions.
However, this "inference" comes at a price, as each use of the model typically incurs a fee. Over time, the cost can add up significantly. Essentially, large models bear the burden of users' reluctance to articulate their questions clearly.
2. Small Models Can Perform Well When Tasks Are Defined
When businesses use AI for specific tasks (such as customer service, generating content in fixed formats, or processing structured data), the requirements become much clearer. With well-defined inputs and outputs, engineering efforts (like using predefined prompts and fine-tuning with custom data) can limit the model's need to make guesses. For instance, in customer service, starting with a large model combined with information retrieval (RAG) might result in responses that lack context and clarity. However, fine-tuning a smaller 9B model with predefined prompts led to more satisfactory results—similar to a well-trained employee who uses professional language and avoids sensitive terms.
3. The Core Advantage of Small Models
While small models are often considered cheaper and faster, what matters most for businesses is predictability. Well-trained small models produce consistent outputs with a consistent style (e.g., always using a friendly tone in customer service responses). This predictability is crucial for production systems that cannot afford unexpected errors or changes.
4. A Mature AI System Is Like a Team
A truly effective AI system does not rely on a single large model; instead, it functions like a team with different roles:
- Small models (e.g., 9B) handle frequent, clear tasks like customer service responses and content generation.
- Medium-sized models (e.g., 27B) handle tasks that require structural understanding, such as generating complex documents or organizing long texts.
- Large models (e.g., 80B+) are used for open-ended questions and analyses in unfamiliar domains.
- Specialized models (e.g., those for coding) perform specific tasks.
This division of labor is similar to how a company functions, with different departments handling different tasks efficiently. Relying solely on large models is like asking the CEO to handle front-office duties—they have the capability, but it's unnecessary and costly.
5. Effective AI Engineering
When using AI, people often hope the models can handle everything effortlessly (e.g., with vague prompts or poor data quality). However, a mature AI system relies on clear processes and rules. This means converting ambiguous inputs into structured data, ensuring predictable outputs, and establishing mechanisms to handle errors. For example, user questions should be directed towards specific categories (e.g., "product consultation" or "complaint"), and responses must include standard elements (e.g., "solution" and "contact information"). By minimizing the need for the model to make spontaneous decisions, the system becomes more stable.
In summary, implementing AI effectively means **clearly defining tasks, selecting the right model for the task, and using engineering to ensure its stable performance.* The future success of businesses will depend on those who treat AI as an engineering project, rather than a one-time investment.