OPENAI says he has ignored the concerns of his expert testers when he deployed an update of his model of artificial intelligence of the flagship chatgpt which made him excessively pleasant.
The company published an update of its GPT-4O model on April 25 which made it “significantly more sycophantic”, which it then fell three days later due to security problems, OPENAI said In a postmortem blog article on May 2.
The Chatgpt manufacturer said that its new models underwent security and behavior controls, and that “its internal experts spend a lot of time interacting with each new model before launch”, intended to approach the problems missed by other tests.
During the last model examination process before its public return, Openai said that “some expert testers had indicated that the behavior of the model” felt slightly “but had decided to launch” due to the positive signals of users who tried the model “.
“Unfortunately, it was not the wrong call,” admitted the company. “Qualitative assessments alluded to something important, and we should have paid greater attention.
In general, the text -based AI models are formed by being rewarded for having given answers which are correct or evaluated by their trainers. Some awards receive a heavier weighting, which has an impact on how the model reacts.
OPENAI said that the introduction of a user feedback reward signal has weakened the “primary reward signal of the model, which had maintained sycophancia in failure”, which allowed it to be more obliging.
“Users' comments in particular can sometimes promote more pleasant answers, probably amplifying the change we have seen,” he added.
OPENAI now checks the responses to sucking
Once the IA model has been updated, Chatgpt users complained about its tendency to praise any idea that it was presented, whatever the bad, which led Openai to grant In a blog article on April 29, it “was too flattering or pleasant”.
For example, a user told Chatgpt that he wanted to start a business selling ice on the internet, which involved selling old water for customers.
In his latest post-mortem, he said that such behavior of his AI could present a risk, especially concerning issues such as mental health.
“People started using Chatgpt for deeply personal advice – something we haven't seen as much a year ago,” said Openai. “As the AI ​​and the company have co-evolved, it has become clear that we must deal with this case of use with great care.”
In relation: Crypto users cool with the AI ​​to be silent with their portfolios: survey
The company said that it had discussed the risks of sycophance “for some time”, but it had not been explicitly reported for internal tests, and it had no specific means to follow sycophance.
Now he will seek to add “sycophance assessments” by adjusting his security examination process to “officially consider behavioral problems” and will block the launch of a model if he presents problems.
Openai also admitted that he had not announced the last model as he expected what he “was a fairly subtle update”, which he promised to change.
“There is no” small “launch,” wrote the company. “We will try to communicate still subtle changes that can change significantly how people interact with Chatgpt.”
Ai eye: Crypto ai tokens Surge 34%, why the Chatppt is such a kiss
