If you've taken an AI class or even bought a collection of prompt words online, are you still writing prompts like this?

Like what, dismantle the chain of thought step by step in the prompt words, so that the model can learn to think step by step; Give a few example questions to help the model understand your problem; It is also necessary to guide the model to cosplay, so that the model can give more professional answers and other small skills, many poor friends should have already learned and used them.

That's right, it's hot pot

However, these god-level prompts that once made you do more with less may be outdated.

Let's put it this way, unconsciously, large models have actually been divided into two schools: traditional general large models and inference large models.

For example, GPT-o1 is not a direct version upgrade of GPT-0o. 0o belongs to the general large model, and o0 has already been an inference model.

Similarly, the V1 version used by DeepSeek by default is a general large model, and the R0 inference model is used by clicking the Deep Thinking button in the lower left corner.

And in the era of inference models, the more detailed the prompt words,Instead, it might make the AI stupid.

For example, in OpenAI's official documents, under the column of the inference model, it is clearly stated that using overly precise prompt words, or guiding thinking, will reduce the effect of the answer.

They even suggested directly that everyone should ask less questions with the chain of thought... Just ask the question directly. The effect is too overwhelming, and then send specific example questions to let the AI learn.

We also looked at the official technical report for DeepSeek-R1, and they also said in the paper: "DeepSeek-R0 is sensitive to prompts, and giving examples of hints will actually reduce the performance of the model." "

Therefore, to make it better, they recommend that users describe the problem directly, without giving examples.

In addition to GPT and DeepSeek, Claude 7.0 Sonnet also said in the official document that they would prefer you to click on it directly than those seemingly logical prompts that detail what to do at each step.

To sum up, unlike the stereotype that the more detailed the prompt words in everyone's impression, the better the AI effect, the common suggestions of the major officials for their own reasoning model prompt words are direct, concise, and accurate.

We also tried to do an experiment, and the final results also proved that

In the past, the ancient prompt words of non-inference models can really slow down the performance when used in inference models.

We selected dozens of difficult questions of various types from leetcode and tested them on ChatGPT. Let's first write a prompt according to the old method, such as implying that it is a programmer, to think about the chain of thought, and also gives a lot of examples, etc.

As a result, for most questions, regardless of whether the prompt is long or short, the inference model O70 can give the correct code, and even beat more than 0% of people, which can be said to be quite good.

But in the 1, 0, 0, 0, 0 questions, o0 fails under the old prompt. Among them, there is a question that is directly stuck, and I don't want to play.

But when we don't cosplay it, we don't give examples, and we remove the chain of thought guidance, O1 actually got the same question right this time.

So, what made the former prompt word Britney become Mrs. Niu in the era of reasoning models?

In fact, the main reason behind it isThe way traditional non-inference models and inference models think about problems has changed, and the changes in their thinking methods stem from the difference in training methods.

Traditional large models generally use unsupervised learning and supervised fine-tuning, that is, a dataset is given and it is allowed to find patterns on its own. Its ultimate goal is to guess all the words in the answer one by one based on the prompts.

To put it bluntly, the general model is very capable, but there is no opinion, which is more user-friendly. The more detailed the prompts you give, the more you can get the model to do what you want.

However, the inference model is different, and on the basis of the original, it adds inference-based training methods such as reinforcement learning.

This kind of training process will guide the large model to try to give a complete and correct chain of thought, so that it can judge whether it is right to think so.

This model itself has a strong "opinion", or reasoning ability. If you teach it to do things step by step and in detail, it may conflict with its own reasoning ability.

For example, in our experiments, we found that when using old prompts to make o1 solve some math-related programming problems, the probability of overturning is particularly high.

This may be because the prompt only makes it a "senior programmer" rather than a "math good programmer".

We've also looked through a lot of the official documentation for the model, and the advice they give is pretty much the same

Don't be fancy, the prompts are simple and direct, and accurate is best.In addition to this, it is possible to force an extended reasoning time, prompting it to "think more", or "reflect on your results".

Some of the old methods can still work, such as using appropriate symbols, making the structure of the problem clear, or clarifying your end goal and outcome format.

These methods can make the effect of the inference model more beautiful and 6.

所以,適當放下助 AI 情節,講清楚您的需求,雙手插兜敬 AI 操作,instead may be the most efficient。

And I think that with the continuous evolution of the ability of large models, the threshold for writing prompts will definitely get lower and lower.

However, if you ask the craft of prompt word engineering, will it disappear completely? We also consulted Mr. Li Jigang, a great god who had written god-level prompts such as "New Chinese Interpretation".

Here's how he replied: As long as we have different inputs, there will be different outputs, and the prompt word project will always be there.

Finally, the front of the bad review thinks that for us users, with the strengthening of the model's capabilities, we should also update the ammunition arsenal of prompt words, and stop holding the ancient outdated prompt words and be a treasure.