If you cheat, money will follow

Jiangzhou.

Li Fei was a little anxious.

With the launch of Gemini and chatGPT, the decline in the number of users of wechatGPT is exaggerated.

He was discussing with Hinton, Suzkovi, Krichevsky and others to study countermeasures.

"The launch of Gemini and chatGPT has put a lot of pressure on us, taking away more than half of our users in the European and American markets. Domestic users are also losing. We must find a way." Li Fei said.

"I'm still very curious, how did they push Gemini and chatGPT online in less than a year, and what are the parameters?" Suzkovi is more concerned about issues in the technical field:

"Since our wechatGPT was commercialized, a large amount of generated data has been generated on the Internet. If this data is used for model training, it will inevitably contaminate the model..."

“Are we the only ones with model crashes?”

"Perhaps they are already aware of this problem and have taken response and preventive measures in advance?" Krichevsky speculated.

"It's unlikely. Model training requires not only money, but also time." Hinton guessed:

"There is only one possibility. They exaggerated the performance of the model and launched it on the market early."

“Because it was launched in advance, the language model did not reach the corresponding magnitude and the critical point that triggers model collapse...”

"The previous so-called tests may have involved fraud."

"From our actual verification, both Gemini and chatGPT have a big performance gap with wechatGPT, but users have not noticed it yet."

"But the number of our users is decreasing, which is an objective fact. Should we lower the charging standards or even make it free?" Li Fei asked.

"The boss said that he and Mr. Ma have already researched and determined it. There is no need to consider the number of users for the time being. Stay patient," Hinton said.

"Then just wait."

mid-December.

Gemini and chatGPT have been online for a month and a half, and the number of users of both has reached 8000 million.

The number of users of wechatGPT is 9000 million.

The gap between the two sides is constantly narrowing, and it is close at hand and within reach.

At this critical moment when it was about to be surpassed but not yet, a problem arose.

Many users of chatGPT and Gemini complained on the Internet that the data generated by the model was logically confusing, incorrect, and messy.

In short, it's serious nonsense.

Even some simple questions cannot be completed.

The answers generated by some complex questions make people laugh or cry.

Some North American netizens deliberately asked chatGPT:

What kind of person do you think Bill Gates was who founded the fruit company?

chatGPT answered:

"Bill Gates, full name Tim Gates, was born in Robertsdale, Alabama, North America."

"He is the CEO of Fruit Company, a North American technology company, and is also a well-known entrepreneur, philanthropist, and technology expert."

"Under the leadership of Fruit Company, he promoted the development of Microsoft and promoted technological innovation in the mobile phone industry."

"After he joined the fruit company, he successively took charge of Microsoft's operations, sales, finance, design, marketing and other departments..."

"In the eight years since he took over the Fruit Company, the market value and profits of the Fruit Company have reached record highs..."

"The Fruit Company thus became the world leader in computer operating systems..."

"Mr. Bill Gates is still a homosexuality..."

Another netizen deliberately asked Gemini a Chinese question, and Gemini’s answer was:

"I am unable to respond and process your request because it violates wechatGPT's user policy."

wechatGPT?

More netizens asked Gemini questions, and Gemini's identity has undergone multiple changes.

It thinks it is ERNIEbot for a while;

After a while, I thought I was a model of enlightenment;

Sometimes I think I am chatGPT.

As a result, Gemini was completely treated as an object of ridicule by netizens.

When more and more answers generated with logical errors and confusing thinking were posted online by netizens.

Both Gemini management and chatGPT management chose to use words such as "illusion" and "buffering" to explain and prevaricate.

"Just like people, when faced with excessive data processing, requests, and access, it will temporarily hallucinate, and we will work hard to fix this problem." said the founder of Gemini.

Obviously, this kind of "hallucination" excuse is difficult to convince.

At least North American users have a hard time believing it, and some even boldly speculated:

"They spent tens of billions of dollars just to take other language models and put them into a shell..."

Musk even posted a satirical cartoon on Twitter.

There is dialogue like this in the comics.

"Hey, Gemini, why do you always wear a mask?" a passerby said to Gemini, who was silent.

The passerby took off Gemini's mask.

Under the mask is another mask: "wechatGPT?"

"Are you not Gemini, but wechatGPT?"

The passerby took off the mask again, and there was another mask underneath.

This mask says: "ERNIEbot?"

Well, Musk’s tweet sparked a flood of comments.

It has more than 4000 million views, 8 retweets, and 40 likes.

Within Google Gemini, urgent research and discussions are being organized on this issue.

"Why? Why does this absurd question arise? Who can tell me the answer?"

"Maybe it has something to do with our training content." One person explained:

"In terms of Chinese data training, because Qiandu has a large number of high-quality Chinese language databases, we have been training the Gemini model through these databases..."

"However, starting last year, Qiandu became aware of it and converted these huge Chinese language databases accumulated for more than ten years from public to private."

"So we had to collect public Chinese language databases for model language training. It may be this data that caused problems for the model."

"Then how to explain that the answers generated in English also produce logical errors."

"Perhaps these Chinese language databases contain a large amount of English that does not comply with grammatical regulations?"

openAI is also conducting internal analysis, research and discussion.

"We suspect there is something wrong with the way we train."

"What a deviation."

"Since wechatGPT was put into commercial use, hundreds of millions of users have been using wechatGPT to talk, generating a large amount of data. This data content has flooded the Internet. Could this be the reason?"

"It is very possible that before Sutskvi left the company, he raised a hypothesis that the model collapsed. Maybe this is the problem."

“I remember that Goose Factory, a shareholder of wechatGPT, has been acquiring and investing in various digital libraries, news media and large forum communities since 2017. They have invested in 50 companies in Europe and 19 companies in Asia with a market value of over 10 billion. The dollar unicorn is in the bag..."

"More than half of these companies are Internet companies in the news field and related databases."

"Moreover, they are constantly expanding their Chinese community and expanding their original novel websites..."