What Are Low-Intelligence LLMs Good For

Low-intelligence LLMs are suitable as content analysis and questioning tools, rather than directly generating content or replacing human decision-making.

After spending a year wrestling with large models, I finally understood a fact: the model’s output is probabilistic. This should have been obvious, but it still took me a considerable amount of time to accept. Only by truly understanding the probabilistic nature of large models can we think about what they are suitable for.

Because large models tend to output content with higher probability of being correct, their answers are also very deceptive and not easy to identify. Some articles target a higher-level readership that is very sensitive to the waste of brainpower and time on low-quality articles, thus resisting AI-generated content. Some readers have poorer ability to distinguish between different types of articles and are more likely to accept AI-generated content.

Therefore, when using large models to generate content, one needs to consider the target audience. It is not suitable for generating content that requires rigor, such as technical documents, papers, tutorials, long-form articles, etc. It is more suitable for generating short news, social science articles, reviews, etc. The hallucination problem of large models further limits their application in scenarios requiring factual accuracy.

AI should not make decisions for people, but rather assist people in making decisions. Thinking should not be handed over to AI, but AI should assist people in thinking. We should not ask AI, but let AI ask us. Try to make AI pose questions rather than solve problems.

Let low-intelligence AI analyze content and pose questions; it is not suitable for solving problems. It can be tried as a content analysis tool, reminder tool, questioning tool, or multiple-choice provider. In these scenarios, AI’s role is to assist humans in discovering blind spots, rather than directly producing conclusions.

Similar discussions exist in the field of AI writing, and quality control remains a key challenge. Only by positioning low-intelligence LLMs as auxiliary tools rather than content producers can their full value be realized.