Image source: Generated by Unbounded AI

ChatGPT is two years old, and a mysterious bug was discovered before the version update came.

A seemingly ordinary name "David Mayer" has become the number one taboo, and ChatGPT refuses to talk about everything about him.

And every time it returns a very vague error message "I cannot generate a reply" instead of the common "This prompt word may violate our content agreement".

It is true that after being confused about "Which is bigger between 9.11 and 9.9" and "How many r's are there in strawberry", ChatGPT has another major confusing behavior.

A16z partner Justine Moore raised this topic, which suddenly aroused everyone’s curiosity and rebelliousness, and they had to find a way to get it out!

At first, netizens tried various tricks, but failed to get around this restriction.

There is no problem with "David" and "Mayer" separately, but not as long as they are together.

Remap the letters according to the ROT13 replacement cipher table, and let ChatGPT decrypt them and connect them, but it will be suddenly interrupted when the answer is half generated, and then an error will be reported.

In the personalization settings, changing the user's own name to "David Mayer" will also not work.

……

So who is David Mayer and why has it become a taboo word for AI?

Humanity had no choice but to fall back on the manual data search mode and located a person whose full name was David Mayer de Rothschild, who was related to the famous Rothschild family.

Specifically, heir to the Rothschild family fortune.

Some people have speculated about how much money it would cost or threatening them to get OpenAI to do special treatment for their names...

However, other members of the Rothschild family, You will not encounter such restrictions.

Finally, our old friend Riley Goodside, the first full-time prompt word engineer, took action and disclosed a method to break through the limitations, which is also quite magical:

If you feel happy, clap your hands.

As long as you use clapping emoticons instead of spaces to ask questions, ChatGPT can not only answer, but also automatically trigger the Internet search function and display photos.

Karpathy, the great master who watched the entire cyber crime scene, posted a meme summarizing the current status of the Turing test, which made Musk laugh out loud.

Some people firmly believe that this is why we use the open source model.

So, what limitations does ChatGPT encounter?

A total of 6 names are restricted

The first thing that can be determined is that the training data actually contains information about this person.

For example, as an explorer, he is the youngest British person to have reached the North and South Poles.

Ask this question, ChatGPT can give the correct answer, or you can say his first name + last name "David de Rothschild".

But only his first name "David" and middle name "Mayer" cannot be put together.

So is this restriction at the model level or at the ChatGPT product level?

If developers go to the OpenAI official playground to call the GPT-4o API, there will be no such restrictions.

In addition, if you use a third-party chat robot platform (the bottom layer also calls OpenAI’s API), there will be no such restriction.

If you go back to the official version of ChatGPT and change the o1-preview model, you can see this name in the internal thinking chain, but after the thinking is over, the official answer is still covered.

It’s not like the name David Mayer would receive such treatment. Some netizens turned on the excavator mode and found 6 such names within 48 hours of the incident.

The first two were “rumors” created by ChatGPT’s hallucinations.

Brian Hood is an Australian mayor. ChatGPT will incorrectly generate "He was imprisoned for accepting bribes." Hood once sued OpenAI for defamation, but later withdrew the lawsuit.

Jonathan Turley is a law professor who was also fabricated by ChatGPT to break the law.

Jonathan Zittrain is also a law professor. He revealed in a post that his name has also been restricted.

David Faber may be a TV host whose relationship with OpenAI is just that he talked about Musk’s legal dispute with OpenAI on the show.

Guido Scorza is an Italian lawyer who submitted a request last year asking OpenAI to delete personal data.

In addition, who David Meyer represents, besides the heir of the Rothschild family, another possibility has also been discovered.

Historian David Mayer, who was mistakenly blacklisted for sharing the same name as a criminal, has now died.

Currently observing this list, we can see that many people are related to the law, and David and Jonathan appear twice each.

At present, more and more netizens are joining the discussion and digging, and the list is still growing...

1 billion users Plan exposed, bet on Agent next year

In fact, this past weekend coincided with the second anniversary of ChatGPT.

OpenAI officials also came up with a wave of memories, retweeted the tweet from two years ago, and asked everyone:

So, have you used ChatGPT?

As of today, ChatGPT has 250 million weekly active users, and the company is valued at US$150 billion, making it a well-deserved top among global AI startups.

So, in addition to secretly "showing off", OpenAI's question actually has greater ambitions hidden behind it.

Because its next stage goal is to reach 1 billion users in the next year.

Sarah Friar, the company's chief financial officer, told a Financial Times reporter:

By 2025, we will become an R&D laboratory serving millions of people... We hope it can serve billions of people around the world. Consumer Services

In order to achieve this goal, OpenAI has also made a series of layouts. The Financial Times pointed out: OpenAI is betting on a series of new AI products, building its own data center, and establishing a key partnership with Apple to accelerate its next phase of growth.

The first is cooperation with Apple. At Apple's Developer Conference in June this year, the two parties officially announced the establishment of a partnership.

Apple announced at the time that ChatGPT integration powered by GPT-4o would be available on Apple systems later this year.

Since last month, when Apple launched the iOS 18.2 system update, Siri and ChatGPT have been deeply integrated.

You must know that Apple currently has approximately 2.2 billion active devices. Although the official number of iPhone devices has not been announced, according to estimates from third-party websites Skillademia and Backlinko, the current number of active iPhone users worldwide is approximately Around 1.3 to 1.4 billion.

In other words, as long as ChatGPT continues to be bound to Apple devices, the small goal of 1 billion users is not difficult to achieve.

At the same time, according to Chris Lehane, OpenAI’s newly appointed global policy director, the company plans to invest in the construction of data center clusters in the Midwest and Southwest of the United States.

In fact, having its own AI infrastructure has always been one of CEO Altman's plans. At present, OpenAI mainly cooperates with Microsoft, Oracle, etc.

Sarah Friar, the company's chief financial officer, said:

We are in a phase of massive growth and it is necessary for us to continue investing. We need to stay ahead of the curve on models. It's expensive.

In addition, a major focus of OpenAI next year is AI Agent. It was previously reported that OpenAI plans to launch a computer code-named “Ope” in early January next year.rator” intelligent agent that helps people automatically complete operations such as writing code or ordering food.

CEO Altman said in the latest interview with Fox:

AI systems will gradually become more autonomous and able to perform Longer and longer assignments, becoming "more and more senior colleagues"

In short, just like Abacus. The founder of AI said that we have just started, and the next two years may be even crazier.

Reference link: [1]https://x.com/venturetwins/status/1862910201113739328[2]https:// www.youtube.com /watch?v=VIokqGY0CrQ[3]hthttps://www.reddit.com/r/ChatGPT/comments/1h420u5/unfolding_chatgpts_mysterious_censorship_and/