How do you get AI to understand complex problems?

The matter of making AI understand complex problems is easier said than done. I recently found an interesting phenomenon when debugging ChatGPT: the same question, asked in different ways, the quality of the answer can be very different. For example, if you ask "how to understand quantum mechanics", the AI may give a bunch of obscure and difficult to understand terminology; however, if you change it to "can you use the example of making coffee to explain quantum mechanics", the answer immediately becomes vivid and interesting. This made me realize that the upper limit of AI's comprehension ability depends, to some extent, on the art of asking questions.

The art of problem-solving

A case study released by OpenAI last year shows that when users break down the grand question of "how to start a business" into 5-7 specific sub-questions such as "how to validate a business idea" and "how to write a business plan", the accuracy of AI's answers increases by 63%. "This is like teaching children how to build blocks, instead of giving them a guide to build a whole castle at one time, it is better to teach them how to build walls, towers and drawbridges step by step. In practice, I have also found that giving AI a clear thinking framework, such as "please answer according to the structure of reason, status, and solution", can often get a more structured output.

However, there is also a reason to break down the problem, an experiment conducted by the MIT Media Lab showed that breaking down the problem into too much detail would reduce the AI's ability to understand it. For example, breaking down "how to write a novel" into "how to write chapter 1, paragraph 1, sentence 1" is too mechanical. Ideal disassembly should be like peeling an onion, layer by layer approaching the core, but each layer should maintain complete logic.

The Power of Context

Researchers at Google DeepMind have found that by providing AI with the proper context, its comprehension can be improved by more than 40%. For example, if you ask "how to resolve team conflicts", and if you add "we are a remote startup team with members from five different time zones", the advice given by the AI will be much more specific. This reminds me of a case I helped a friend consult with AI last week: simply asking "what if my child doesn't like to study" was a generalization, but adding the details of "my child is 12 years old, addicted to short videos, and is particularly resistant to math", the advice immediately became actionable. The advice becomes immediately actionable.

Interestingly, more context is not better. Stanford University's HCI research shows that contextual descriptions of more than 500 words can instead distract the AI from getting to the point. The best practice is to start with a 1-2 sentence description of the core question, then 3-5 bullet points to provide the key context, and finally clarify the desired form of the answer.

The Magic of Feedback Loops

What surprised me most was the AI's ability to learn in real time. Anthropic's research shows that through 3-5 rounds of conversational feedback, Claude's accuracy in understanding complex questions can be improved from an initial 58% to 89%. For example, when the first answer is not good enough, saying "This explanation is still too specialized, can you use a more life-like example", the AI really adjusts its expression. The AI will really adjust the expression. I often use this "progressive questioning" method, just like talking to a smart but need to guide the assistant, each feedback is aligning our thinking channel.

Note, however, that feedback needs to be specific. Instead of saying "I don't understand", you should point out that "the cause and effect relationship in step 3 is not very clear"; AI is like a serious student, the more specific guidance you give, the faster it learns. Recently, when I was dealing with a supply chain optimization problem, through 4 rounds of targeted feedback, the final solution was much more refined than the first version.

In the end, getting AI to understand complex problems is essentially exercising our ability to think clearly. Every conversation with AI is a mirror to the ambiguities and jumps in our own expression. Perhaps someday in the future, when AI can easily understand all complex problems, our human way of thinking will become clearer and more organized as a result.