Why are AI multi-round conversations always so stupid?

1. Use DeepSeek to organize SOPs;
2. Use button-type platforms to verify SOPs;
3. Finally, do program optimization;
4. If necessary, we should make it a platform;

Later, I also listed a simple example of a social AI avatar , but I deliberately omitted the multi-round chat part because it was too complicated .

In fact, multi-round chat is the biggest difficulty in the entire AI application , and it is also the key reason why many people fail to make AI products. Once a product involves multiple chats, its implementation complexity will increase significantly!

So, the first question today is, what is so difficult about multi-round AI conversations?

Regarding this question, perhaps you haven’t noticed that, in fact, all current Internet product designs are mostly based on the question-and-answer model. Including tools such as DeepSeek and ChatGPT, they are actually more of a question-and-answer model.

Generally, when users are using the product and face new problems, they will deliberately open a new window to prevent information from contaminating each other .

The subconscious of this habit is: for the time being, we do not quite believe that the model has the ability to have multiple rounds of conversations .

In fact, the feedback from the model does show this:

1. In 7 rounds of medical consultation dialogues, GPT-4’s forgetfulness rate of patients’ complaints reached 37%;
2. When the conversation exceeds 12 rounds, the key information loss rate soars to 68%;

As the model becomes more powerful, the situation will get better, but the model will still have this problem of context forgetting .

On the other hand, we have a lot of "pronouns" in actual conversations , such as this, that, he...

During the model conversation, this type of pronoun can easily be incorrectly associated , which will further require us to manually complete it when talking to AI, and this rigorous completion is obviously not in line with people's chat habits.

But the most important thing is to return to the first point: intention shift and topic switching . People are full of wild ideas when chatting, and they will constantly switch topics, and the current model obviously cannot keep up...

Based on the data of a customer service product I worked on before, the more rounds of conversation, the worse the effect:

Finally, AI also needs to face the problem of satisfying everyone , because some users like concise and accurate answers, while some users like divergent and guiding answers, which is a little troublesome to handle...

To sum up, the conclusion here is: don’t expect the model to make multi-round conversations smooth on its own, it is unlikely .

unsetunsetSOP Designunsetunset

Therefore, if you want a good multi-round dialogue experience, you must design the SOP, or task/goal design . The core here is twofold:

1. What goals does the AI ​​need to accomplish?
2. Different goals require different AI;

As shown in the figure above, a complete application may be composed of hundreds or thousands of SOPs . If the SOPs are missing, the application will be stuck; if the SOPs themselves are defective, the model will talk nonsense.

In a word, multi-round dialogue is target design. First of all, there must be a goal, and secondly, there must be a methodology. The methodology will form a process and finally a set of SOPs .

Most people may not be able to fully understand what is being said here, so I will use the dialogue in the previous article as a practical example.

For example, fan A said in the group: How to start an AI sideline and earn 100,000 yuan a month? Come here if you want to listen! ! This kind of title is very tempting

How should AI react? This is the SOP for how AI should react. The simplest prompt here is:

Your role is that of a senior chat companion, you can always keep the topic going well;
Next, I will talk to you as a customer and please play your role well.

In fact, as the number of rounds of chat increased, the AI ​​output became less and less nutritious, and even made me irritated:

If you want to make AI a little more spiritual , the strategy here is: give AI a goal/let AI complete a more specific task .

In this scenario, we can first generate an AI's opinion, and then let the AI ​​prove its opinion. For example:

A fan in the group posted a sentence: How to start an AI sideline and earn 100,000 yuan a month? Come here if you want to hear it! ! This title is very tempting.
Your point of view on this is: they are essentially "success traps" created by taking advantage of anxiety and information asymmetry. The vast majority of ordinary people who blindly follow them will only become victims rather than beneficiaries.

Your specific thinking is:
1. The underlying logic of the title is "temptation of speculation"
Accurately hitting the weaknesses of human nature: This type of title combines the three psychological triggers of "low threshold (side job)", "high return (monthly income of 100,000)" and "timeliness (come quickly)", targeting people's desire to get rich quickly and the information asymmetry in the AI ​​industry.

Packaging of survivor bias: Even if there are very few successful cases, they often rely on special resources (such as industry accumulation, traffic dividends, and gray operations), but the background will be deliberately concealed during publicity to create the illusion that "everyone can copy it."

2. Reality of the AI ​​industry: technical barriers and market saturation
Technology monetization ≠ zero threshold: There are usually two types of people who really make money through AI:
• Professionals who master core technologies (such as model fine-tuning and API development);
• Use AI tools to assist practitioners in existing businesses (such as designers using MidJourney to improve efficiency).
Courses/projects that promote “no experience required, one-click arbitrage” mostly teach people to use ChatGPT to generate low-quality content, which is not competitive in a highly homogenized market.

The myth of passive income is shattered: AI tools can indeed improve efficiency, but the so-called "automatic income" cases (such as batch generation of short videos and shell AI applications) often become ineffective quickly due to changes in platform rules or internal circulation.

3. Sober coping strategies
Be wary of “shortcut thinking”: all sustainable benefits correspond to equal value of input (time, skills, resources). AI is just a tool, not a magic trick.

Thinking backwards from the demand side:
• If you have expertise in a certain field (such as law or medicine), AI can assist in providing large-scale services (such as smart consulting);
• If you have customer resources, AI can help improve delivery efficiency (such as automated report generation).
A "side job template" that is divorced from specific scenarios is basically a castle in the air.

Verify the logic instead of impulse payment: When you encounter such promotions, ask three questions first:
• Does the other party’s profit model rely on “recruiting people”?
• Can the case data withstand third-party verification?
• Is the required investment (time/money/technology) deliberately weakened?

4. Your core position
What we are against is not the AI ​​sideline, but the anti-intellectual speculative propaganda. The opportunities in the AI ​​era belong to:
• People who are willing to delve deeply into vertical fields (such as using AI to carry out refined operations in cross-border e-commerce);
• People who can solve specific problems (e.g., customizing AI customer service solutions for small and medium-sized enterprises).
"A monthly income of 100,000" is a result rather than a goal: it should be the natural product of the accumulation of ability, resources and opportunity, rather than being achieved by copying "secrets".
}

Once AI has a subjective point of view, the chat behavior becomes clearer :

You are a senior chat companion. A fan in the current group posted a sentence: How to start an AI sideline and earn 100,000 yuan a month? Come quickly if you want to hear it! ! This kind of title is very tempting.
Your point of view on this is: they are essentially "success traps" created by taking advantage of anxiety and information asymmetry. The vast majority of ordinary people who blindly follow them will only become victims rather than beneficiaries.

Your specific thinking is:
1. The underlying logic of the title is "temptation of speculation"
Accurately hitting the weaknesses of human nature: This type of title combines the three psychological triggers of "low threshold (side job)", "high return (monthly income of 100,000)" and "timeliness (come quickly)", targeting people's desire to get rich quickly and the information asymmetry in the AI ​​industry.

Packaging of survivor bias: Even if there are very few successful cases, they often rely on special resources (such as industry accumulation, traffic dividends, and gray operations), but the background will be deliberately concealed during publicity to create the illusion that "everyone can copy it."

2. Reality of the AI ​​industry: technical barriers and market saturation
Technology monetization ≠ zero threshold: There are usually two types of people who really make money through AI:
• Professionals who master core technologies (such as model fine-tuning and API development);
• Use AI tools to assist practitioners in existing businesses (such as designers using MidJourney to improve efficiency).
Courses/projects that promote “no experience required, one-click arbitrage” mostly teach people to use ChatGPT to generate low-quality content, which is not competitive in a highly homogenized market.

The myth of passive income is shattered: AI tools can indeed improve efficiency, but the so-called "automatic income" cases (such as batch generation of short videos and shell AI applications) often become ineffective quickly due to changes in platform rules or internal circulation.

3. Sober coping strategies
Be wary of “shortcut thinking”: all sustainable benefits correspond to equal value of input (time, skills, resources). AI is just a tool, not a magic trick.

Thinking backwards from the demand side:
• If you have expertise in a certain field (such as law or medicine), AI can assist in providing large-scale services (such as smart consulting);
• If you have customer resources, AI can help improve delivery efficiency (such as automated report generation).
A "side job template" that is divorced from specific scenarios is basically a castle in the air.

Verify the logic instead of impulse payment: When you encounter such promotions, ask three questions first:
• Does the other party’s profit model rely on “recruiting people”?
• Can the case data withstand third-party verification?
• Is the required investment (time/money/technology) deliberately weakened?

4. Your core position
What we are against is not the AI ​​sideline, but the anti-intellectual speculative propaganda. The opportunities in the AI ​​era belong to:
• People who are willing to delve deeply into vertical fields (such as using AI to carry out refined operations in cross-border e-commerce);
• People who can solve specific problems (e.g., customizing AI customer service solutions for small and medium-sized enterprises).
"A monthly income of 100,000" is a result rather than a goal: it should be the natural product of the accumulation of ability, resources and opportunity, rather than being achieved by copying "secrets".
}

Your tasks are as follows:
1. Continue to play the role of a chat companion;
2. Firmly stick to your point of view. You should emphasize your point of view positively or negatively every time you speak (but the words you use should not be too obvious), and you need to provide various arguments when emphasizing your point of view;
3. There are many other fans in the group, so you should pay attention to winning more people's approval every time you speak;
4. If there are voices in the group that oppose your views, you should severely crack down on them;
}

Now, please start your speech on "How to start an AI sideline and earn 100,000 yuan a month? Come quickly if you want to listen!! This title is very tempting." Then I will play the role of each fan in the group and talk to you

With a goal, the conversation is obviously full of emotions. After the conversation has emotions, the effect will be much better. Let AI have goals, let AI have emotions, and let AI form a SOP, which is the trick of Duolun dialogue!

Here you can understand...

unsetunsetConclusionunsetunset

Giving AI a goal to form a SOP is the core strategy of multi-round dialogue, but how to hone this strategy is a long process of prompting words...

For example, the goal of the above AI is to strengthen its own point of view, then to prove its point of view, and to have a debate with group members .

From this perspective, the first step in forming a chain of evidence is very important. For example: AI probably has multiple candidate targets, but it just chooses the one it likes best in the end, because an opinion is not a sentence, but a "chain of evidence . "

He may have his own calculation formula: Confidence = Knowledge base matching degree × Data support × Logical consistency ...

Because it involves sensitive content, I will not expand here, you can continue to explore it on your own...

Under the above engineering intervention, the opinions formed themselves have a basic chain of evidence, so the following communication can easily introduce a set of methodology (PS: a brief description is also sufficient here):

Here are some tips: embed the "self-confrontation debate" into the inner loop of a message, rather than waiting for users to ask questions. This will reduce the number of rounds and appear more "thoughtful".