Research on best practices of ChatGPT intelligence reduction

# Contents of this article:

- Comparison of intelligence reduction and non-intelligence reduction
- Intuitive manifestation of reduced intelligence
- Background of the conditions for reducing intelligence
    - 1. Early November 2024: The phenomenon of reduced intelligence begins to emerge
    - 2. Early December 2024: The issue of intelligence reduction sparks widespread discussion
    - 3. Mid-December 2024: IQ reduction spreads to all clients
- The IQ test I used
    - Test 1: Tool call
    - Test 2: Cover generation
- Actual measurement: The IP health of scamalytics.com has nothing to do with the account
- Actual test: Pow (proof of work) has little to do with whether the intelligence is reduced or not, and the mobile web version cannot prevent the reduction of intelligence
- Sorting of solutions for reducing intelligence

Comparison of intelligence reduction and non-intelligence

First, let's intuitively feel the impact of whether the output is reduced in intelligence or not. After gpt-4o supports the latest raw image function, the effect is very different from the previous generation of dalle (I still remember that we were very happy to connect the dalle model to the WeChat robot a year or two ago. At that time, no friends complained about it. In fact, it was really unbearable to watch).

Obviously, the effect of the normal version of the raw picture has reached the production-usable level. I have used it a lot in article covers and illustrations. Before that, I used midjourney, and before that, I used canva or hired a graphic designer.

This time, the junior graphic designer really fainted in the toilet and had to be buried with honor.

The intuitive manifestation of reduced intelligence

The background of the conditions for reducing intelligence

1.  Early November 2024: The phenomenon of reduced intelligence begins to emerge

• Performance : Some users found that the quality of ChatGPT's answers had declined, manifested in direct output of answers (without showing the thinking process), inability to parse images/files, and refusal to call advanced features such as online search.
• Scope of impact : Mainly for users who use proxy tools to access ChatGPT, especially those whose IP addresses are marked as high concurrency.
• Speculated reason : Due to limited computing resources and insufficient training data, OpenAI began to replace the underlying models of some users from GPT-4o to low-computing power versions (such as GPT-4o mini or GPT-3.5) to save resources.

2.  Early December 2024: The issue of intelligence reduction sparks widespread discussion

• A surge in user feedback : A large number of users reported similar problems, confirming that "intelligence reduction" has developed from individual cases to a systematic phenomenon.
• Functional restrictions are extended : In addition to model replacement, advanced functions (such as networking and text and graphics) are further restricted, and some users cannot use them even if they have a Plus membership.
• Official response : OpenAI issues API tokens (11 million tokens per day) for free to high-volume developers in exchange for data sharing rights to alleviate the shortage of training data.

3.  Mid-December 2024: IQ reduction spreads to all clients

• The scope of impact is expanding : the IQ-reducing behavior has expanded from the web and Windows client to all platforms including mobile apps and Mac clients, and all users may be affected.
• The core reason is clear : OpenAI publicly admits that computing resources are tight, especially the high-computing version of GPT-4o, which consumes more than $1,000 in resources per task, far exceeding the old version (GPT-4 is $5 per task).
• User response strategy : The community proposed to alleviate the problem by optimizing IP quality (such as using low-concurrency IP), switching accounts or client versions, etc.

The IQ test I used

I searched for some testing methods.

One way is to test yourself on a math or Japanese question, but I think it’s too troublesome and I’m not interested.

Another way is to send a picture first, which is said to be able to unblock the account, but I actually tested it and it didn’t work (it may also have failed).

I currently use the following two methods to conduct intelligence reduction tests.

Test 1: Tool call

The first one is in gpt-4o mode, asks: " summarize your tool in a markdown table with availability ".

This question has been raised in many tutorials (for example: [What should I do if ChatGPT is "degraded"? O1 does not think, 4o cannot connect to the Internet, analyze images, and process files!](https://www.youtube.com/watch?v=uQ8ExarQhyM "What should I do if ChatGPT is "degraded"? O1 does not think, 4o cannot connect to the Internet, analyze images, and process files!"")), and it is indeed effective.

The comparison is as follows:

It is worth noting that many tutorials say that when the number of tool calls is less than 4, it means that the intelligence is reduced. This is also an outdated statement because there will definitely be more and more tools.

But the general principle is very simple: the more distinguished/healthy the account is, the more tools it can call upon and the more advanced they are; especially the more advanced the tools, the more expensive they are and the more restrictions they impose .

At present (April 14, 2025), the most expensive and popular tool scenario under gpt-4o is gpt-4o's text image capability (image_gen, reference:  GPT-4o "Ghibli style" became popular overnight, Ultraman changed his avatar overnight! Hayao Miyazaki criticized AI for insulting life ), so the presence or absence of this tool can be used as one of the criteria for whether intelligence is reduced.

Test 2: Cover generation

The second one is very practical (original), which is to give me an article and let gpt-4o output a cover for me.

The styles of images generated using dalle and gpt-4o are very different (especially in saturation, layout, and text content).

Measured:

scamalytics.com's IP health has nothing to do with the account

If your account is healthy and you are using a Mac client, the only factor that can affect whether your account is downgraded is your IP address.

I had subconsciously thought that the IP health of scamalytics.com would have a significant positive correlation with reducing the chance of IQ degradation.

So, in order to quickly select the healthiest IP, I asked cursor to write a more complex program for me. It will automatically traverse the vmess node list in our clash configuration file, then automatically resolve the specific IP, and then automatically query scamalytics.com to obtain the health status, and finally summarize it into a table.

The following are the core prompts for the cursor:

@config.yaml:L29-30
When the proxy is set to Argentina A, the website shows the IP address is 82.152.6.214  
When set to Korea A, it is displayed as 125.240.80.85  
Want to know how we can programmatically get the ip from this configuration file

---

@get _real_ proxy_ip.py seems to be feasible, but you have to manually switch proxies one by one, which is not suitable for batch

---

After getting the IP address, add a risk index test. For example, if you visit https://scamalytics.com/ip/125.240.80.85, you will get the following in the source code:

{

"ip":"125.240.80.85",  
"score":"50",  
"risk":"medium",  
"is _blacklisted_ external": false,  
...  
}

Extract its information

The core program is open source at: git@github.com:MarkShawn2020/2025-04-14_ip-research.git

Then I tested it based on one of my paid service providers A and got a table:

python proxy_ip_cli.py -c data/config_miaomiao.yaml -a -f both

As a result, I passed the IQ test from Taiwan and the United States, where the risk level was the lowest at 0, to Taiwan, where the risk level was over 10, to Türkiye, where the risk level was 100...

I am very confused!

So, I tested another free service provider B and got another table:

python proxy_ip_cli.py -c data/config_bitz.yaml -a -f both

The result was disappointing!

In the second table, even nodes/IPs with a risk level of 0 cannot pass the test at all!

Based on this, I have clearly reached a non-consensus:

Whether the intelligence is reduced or not has nothing to do with the IP risk of the scamalytics.com website .

ChatGPT does take IP into consideration, but scamalytics.com is not qualified to be the judge.

Measured:

Pow (proof of work) has little to do with whether intelligence is reduced or not

The mobile web version cannot prevent the reduction of intelligence

I won't go into details, but I'm sure it doesn't constitute a valid condition.

Sorting out solutions for reducing intelligence

Conclusion:  The most effective method is to use a high-quality account (non-shared, non-multi-IP, try not to open too many), a high-quality dedicated IP and a stable client (such as Mac version or mobile version, not web version).

Here, let me say that in a sense, if the account, IP, and even the client are fixed through the mirror site, it may indeed be more stable than the official account (the premise is that the amount is not large. If the amount is large, it will not be able to withstand it:)