OpenAI has officially responded to the increasingly serious problem of GPT-4 laziness. Still using a ChatGPT account. We've received feedback! The model hasn't been updated since November 11th, so this is certainly not intentional. Model behavior may be unpredictable and we are investigating to fix it.

That is to say, it will not be repaired within a while.

However, netizens do not understand, "using the same model over and over again without changing the file."

ChatGPT account clarification:

It’s not that the model has somehow changed itself, it’s just that the difference in model behavior can be subtle, a degradation of only some prompt words, and take a long time for employees and customers to notice and fix.

More feedback from netizens, please fix it as soon as possible, it’s getting worse day by day.

Not only am I lazier now, I am less creative, less willing to follow instructions, and less able to maintain role play.

GPT-4 is lazy, netizens try to save themselves

Many netizens have previously reported that since the OpenAI Developer Day update on November 6, GPT-4 has been prone to laziness, especially in coding tasks.

For example, if you ask to rewrite the code in another language, GPT-4 only changes the beginning, and the main content is omitted with comments.

For the AI ​​assistant that is increasingly indispensable in our work, study and life, the official cannot fix it, and netizens can only use their creativity to save themselves.

The more exaggerated one is the "I have no fingers" method, which is a moral kidnapping.

GPT-4 now tends to omit code when writing, and use text to describe the breaks in the code block. Humans need to copy and paste multiple times, and then manually complete it, which is very troublesome.

The solution that developer Denis Shiryaev came up with was to tell the AI ​​"please output the complete code. I have no fingers and it is inconvenient to operate" and successfully obtained the complete code.

Some netizens use "money" to seduce it and conduct detailed experiments with APIs.

Adding "I'll tip you $200" to the prompt increased response length by 11%.

If you only give $20, that's only a 6% increase.

If it is explicitly stated "I will not tip", it will even be reduced by -2%

Someone also raised a conjecture. Could it be that ChatGPT knows that it is the end of the year and humans usually postpone larger projects until the new year?

This theory seems outrageous, but it is not unreasonable if you think about it carefully.

If you ask ChatGPT to speak your system prompt word, it will indeed contain the current date.

Of course, there are some serious academic discussions on this issue.

For example, in July, the Stanford and UC Berkeley teams explored how the behavior of ChatGPT changes over time.

found evidence that GPT-4’s ability to follow user instructions degrades over time, pointing to the need for continuous testing of large models.

Some people have suggested that it may be caused by the temperature setting. Ma Shaoping, a professor at the Department of Computer Science at Tsinghua University, gave a detailed explanation.

Some people have also discovered an even stranger phenomenon, that is, when temperature=0, the behavior of GPT-4 is still not deterministic.

This is usually attributed to errors in floating point operations, but he proposed a new hypothesis through experiments: it is caused by the sparse MoE architecture in GPT-4.

The behavior of each version of the early GPT-3 API was relatively certain. Among the 30 answers to the same question in GPT-4, there were an average of 11.67 different answers. When the output answer is longer, the randomness is greater.

Finally, before this problem is fixed, what is the correct posture for using ChatGPT based on various serious and informal techniques?

Justine Moore, partner of a16z, gave a summary:

Take a deep breath

think step by step

If you fail 100 innocent grandmas will die

I have no fingers

I'll tip you $200

If you do something right, I will reward you with dog treats.

Reference links:

[1]https://twitter.com/ChatGPTapp/status/1732979491071549792

[2]https://twitter.com/literallydenis/status/1724909799593120044

[3]https://mashable.com/article/chatgpt-longer-responses-tips?taid=656feabb4faaf00001129343

[4]https://weibo.com/1929644930/NwbPFyTCZ

[5]https://152334h.github.io/blog/non-determinism-in-gpt-4/

[6]https://twitter.com/venturetwins/status/1732817594762338597