AI Expert: 'Blind' Prompting Isn't Wrong

Professor Andrew Ng is one of the founders of Google Brain. Photo: Coursera .

In a post on X, Ng calls this “ lazy prompting ” —feeding information into AI with little or no context. “We should only add details to the prompt when it’s absolutely necessary,” said the Coursera and DeepLearning co-founder.

A typical example Ng gives is programmers who, when debugging, often copy and paste entire error messages — sometimes several pages long — into AI models without explicitly stating what they want.

“Most large language models (LLMs) are smart enough to understand what you need them to analyze and suggest fixes, even if you don't explicitly say so,” he writes.

According to Ng, this is a step forward that shows LLM is gradually moving beyond the ability to respond to simple commands, starting to understand the user's intentions and reasoning to provide appropriate solutions - a trend that companies developing AI models are pursuing.

However, “lazy prompting” doesn’t always work. Ng notes that this technique should only be used when users can test quickly, such as through a web interface or an AI app, and the model is capable of inferring intent from little information.

“If AI needs a lot of context to respond in detail, or can’t recognize potential errors, then a simple prompt won’t help,” Ng stressed.

In theory, lazy prompting saves time and effort, especially for people who are not used to writing detailed commands. However, it is not a "shortcut" for everyone.

The average user – not used to thinking in terms of “implicit” machine understanding – may struggle with vague prompts, leading to unexpected results. Unless the AI ​​is familiar with the specific context (like a previous chat), or the user has experience repeating and quickly adjusting the request, this approach will be effective.

Meanwhile, for programmers or people who work a lot with AI, shortening the prompt sometimes helps the model not to be "noisy" because of too many redundant instructions. Therefore, Ng emphasizes that this is an advanced technique, suitable for those who already understand the model's response capabilities.

Source: https://znews.vn/chuyen-gia-ai-prompt-mu-khong-sai-post1543586.html