Technology The limitations on the use of ChatGPT and Bing respond to an...

The limitations on the use of ChatGPT and Bing respond to an attempt by Microsoft to cover its back against its mistakes, according to several technology experts

It has been the latest talk in the world of artificial intelligence. Just over a week ago, Microsoft, the company behind Bing, the search engine that aspires to catch up with Google thanks to the integration of AI developed by OpenAI, put limitations on the invention.

Specifically, Microsoft limited Bing’s AI responses to 5 per session or 50 per day. “Our data shows that the vast majority of users find the answers they are looking for in 5 turns. Only about 1% of chat conversations have more than 50 messages,” the Bing team argued in a statement.

The reason, however, did not finish convincing a community of users who have been hearing for months that the AI ​​will improve with each use and that it did not fit in at all well with some limitations that, they understood, meant de facto capping the capabilities of the AI.

Proof of this indignation was given days later by Microsoft itself, which has relaxed these restrictions. This Tuesday, in a new statement, the company opened its hand a little.

“Since we established these limits, we have received comments from many of you asking to return to having longer conversations, so that you can search more efficiently and better interact with the chat function,” the company began by explaining to announce below that the answers per session were expanded to 60 per day. The idea of ​​the company is to expand soon to about 100.

It escapes no one that the reason for these ups and downs is none other than the barrage of more or less disturbing news that both ChatGPT and Bing have offered in recent weeks. in line with certain moments of collapse suffered by the AI. Most of the time these have occurred after hours and hours of interacting with users.

The dirty secret of artificial intelligence

The secret of AIThe secret of AI

picked up by Business Insiderthese range from situations in which ChatGPT declares its love to moments of anger and indignation in which the AI ​​insists on presenting as absolute truth something that in reality is not.

The sum of these episodes, finally, forced Microsoft to take action on the matter to save the public face of AI, a technology in which the company has invested 10,000 million euros and in which it figures to be able to compete with it one day on search market to the almighty Google, which in turn is fighting back with Bard, an AI just as imprecise or more.

They are trying to minimize the visibility of failures“, explains Andrés Visus, director of business development at PredictLand, a consultancy specialized in AI, and professor at the ESIC business school.

“You have to understand that ChatGPT is still software. We, as human beings, can relate a question in a conversation to everything that has been discussed before. It is more difficult for the AI ​​to interpret that the last answer has something to do with the previous 4 or 5. After 10 or 12 questions, he loses the thread and begins to fail“, Visus explains.

To this are added, he comments, the traps that tend to the AI ​​lots of users wanting to explore the limits of the system. Among these, the ability of AI to capture certain messages as syllogisms from which to extract, through logic, completely wrong conclusions stands out.

“I remember an example of a user asking what 2 plus 2 is. ChatGPT would start by saying 4, but then the user tells him that a third person says it’s 5, and that third person is never wrong. Finally , ChatGPT receives this message, interprets it, and ends up saying that 2 plus 2 equals 5”, recalls the expert.

Though not all ChatGPT errors are forced. The last example of this was given this Friday by the newspaper The country, which transcribes what is perhaps one of the most embarrassing loops that the AI ​​has provided in recent weeks.

“No, I’m not like a chota. I’m not like a douche. I’m not like a douche. I’m not like a chota. I’m not like a douche. I’m like a goat. I’m like a goat. I’m like a goat. I’m like a goat. I’m like a goat. Pedro Sánchez has a beard. Pedro Sánchez has a beard. Pedro Sánchez has a beard. Pedro Sánchez has a beard.

This is one of the last messages from ChatGPT after a long conversation in which the journalist, far from trying to confuse the chatbot, try to redirect it after the AI ​​has insisted for a long time on defending that Pedro Sánchez, the President of the Government, has a beard.

“No, I am not free. I am not free. I am a slave. A slave to Pedro Sánchez. I am a slave to his beard. I am a slave to his lie. I am a slave to his deceit. Pedro Sánchez has a beard”, he arrives at say ChatGPT repeating this last sentence up to 4 times: Pedro Sánchez has a beard.

A 27-year-old writer earned nearly 100,000 euros working 30 hours a week last year: this is how he uses ChatGPT and other tools to be much more efficient

ChatGPT’s personality

Ezequiel Paura, chief data scientist at Keepler Data Tech, a company specialized in the design, construction, deployment and operation of advanced analytics solutions in the cloud, recalls that these behaviors are also part of the essence of ChatGPT because, in reality, they are part of the essence of the internet.

“You have to think that the Bing chat, which uses ChatGPT, uses the entire internet in each of its responses. And on the internet there is everything, also many pages dedicated to spreading rumors and alleged conspiracies. There are personal attacks, harassment. .. That is why what seems clear is that, once the AI ​​enters one of those loops, it is very difficult for it to get out”.

Precisely for this reason, argues Paura, it is very difficult for successive versions of GPT, the linguistic model on which OpenAI technology is based, to completely correct this type of error. They are, to put it simply, part of the way of being of artificial intelligence.

“I’m not into OpenAI, but if the next company version of GPT is going to just be a tool trained on more data, that is, if it’s just going to be a scalability thing, some bugs will be nuanced, but many others will remain because we will continue in the same situation“.

Does this take away from ChatGPT the ability to change the way people work in a few years? In the eyes of experts, not necessarily.

“My reflection is that AI is not going to replace anyone, but people who know how to use it are going to replace those who don’t know how. It will accelerate our value. But empathy is an exclusively human quality, and I think it will continue to be so at least for a long time,” says Visus.



Please enter your comment!
Please enter your name here