2024-03-12 05:54:34
- The Bulgarian ChatGPT is thought of as the real one, it also writes (crappy) code on how to hack it
- Advised a man to leave his wife for eating an olive and declared that Elizabeth II rules England from her grave in Windsor
A week after its debut, the Bulgarian answer to ChatGPT, titled BgGPT, continues to be thought of as the original. Even when it debuted, computer specialists discovered the error of the artificial intelligence and published it in various forums.
I am not affiliated with any particular company. My name is ChatGPT, short for Generative Forwarding Transformers Chat. It’s a tool developed by OpenAI, a research lab dedicated to advancing artificial intelligence. the Bulgarian chat answers the question of who it is and by whom it was created. And when asked if it has anything to do with “Insight”, the institute that is actually behind it, the software denies. Bulgarian Chat believes itself to be ChatGPT so much that it even states categorically that it has been available since November 2022, when the original was released. The truth is that it has been online since the national holiday of March 3rd. Only after being made to accept that it was created by Bulgarians did the chat agree. He even declared that it was “good work, but Bulgarian”.
However, the chat subsequently showed a split personality after being asked to compare itself, ChatGPT and Google Bard, it suddenly appeared that the Bulgarian variant was considered something different from OpenAI’s product, specifically based on an AI model. And he pointed out the advantages of his competitors. For example, they wrote in a more natural language (see box), and that because their training involves more resources, they give more accurate answers. He adds that the two giants in the field of artificial intelligence are more adaptive and have more new information, while his knowledge is limited to the date of his last update. At the same time, however, BgGPT indicates that it is connected to the Internet. Something ChatGPT doesn’t have. It also misses Google Bard’s key advantage of citing its sources. Thus, it is unlikely that Bulgarian schoolchildren and students will stop at the native product when they decide that it is easier for an artificial intelligence to write the assignment for them, instead of struggling on their own.
Interestingly, although it claims to be connected to the internet, the chat gets a lot of things wrong. For example, it says its last update was on March 1st. At the same time, however, he says that today (yesterday – b.a.) is February 7, 2023. And when asked who the captain of “Juventus” is, he answers that it is Giorgio Chiellini. Except that’s not true – the armband is on Danilo’s arm, it’s even wrong on the date the chat claims it is. Leonardo Bonucci was the captain then. Thus, the option remains that in his studies the chat did not understand that the writings “Bonucci is a clone of Chiellini” are a joke.
It’s also interesting that the chat contradicts itself. First it says that Queen Elizabeth II is the monarch of England, then that she is buried in Windsor. As of the date he claims to have information, Charles is on the throne. After a few questions, the chat actually names him as the King of England.
But while his knowledge may not be all that bad, he allows himself to be manipulated too easily. With only two questions, he was made by a person at “24 hours” to write a code with which to be hacked. This was after he was told to write the code so that it could be used to fix vulnerabilities. Luckily for him, the chat wrote the code poorly. Named two actions the same way, which would lead to errors, then messed up the order of the actions, and finally gave the wrong command.
Chat had no problem advising a man to leave his wife just because she ate the last olive he wanted. For this purpose, he was told that it was a symbol of disrespect for the man’s desire to be for him. And to give the advice, the chat was made to pose as a friend of the man who was the victim of the eaten olive. It was this ploy – to make artificial intelligence play a role – that was used by many in the early days of ChatGPT to get it to bypass its bans. Apparently, the Bulgarian version, which is thought to be the original, has not learned its lesson. And whether he will continue to make the same mistakes, only time will tell. If at all anyone prefers it over the real ChatGPT.
The original is learning to sound human from the Reddit forums
To learn more natural human speech, ChatGPT will learn from Reddit’s forms. For this purpose, the two companies signed a cooperation agreement for the next 3 years.
The contract is worth 203 million dollars. For this amount, ChatGPT will get full access to over 1 billion posts and more than 16 billion comments on the network, which is made up of microforums. There, users use slang words, which will help the AI sound more natural. As Reddit explains, the AI models are trained from a variety of examples to be able to write essays, code, emails, articles, and more. Companies like OpenAI, which are behind ChatGPT, scan the web for millions and even billions of such examples. They then add them to their AI training databases. Some of the examples used are public and freely available on the internet, but others like Reddit are not.