Openai says it ignored the concerns of expert testers when it deployed an update to its flagship ChatGpt artificial intelligence model.
On April 25th, the company released an update to the GPT‑ 4o model, making it “remarkably psychophonic.” This was rolled back three days later due to safety concerns, he said in a posthumous blog post on May 2.
ChatGpt Maker says that the new model will undergo safety and behavior checks, “Internal experts spend a considerable amount of time interacting with each new model before launching,” and aims to catch issues missed in other tests.
In the review process for the latest model prior to its release, Openai said, “Some expert testers showed that the model “feeled it slightly turned off” but decided to launch “due to a positive signal from users who tried the model.”
“Unfortunately, this was a false call,” the company admitted. “Qualitative assessments hinted at something important. We should have paid more attention. They were picking up at the blind spots of other ebals and metrics.”
Generally, text-based AI models are trained by being rewarded by giving responses that are accurate or highly evaluated by trainers. Some rewards are given heavy weights, which affect the model's response.
Openai said that introducing user feedback reward signals weakened the model's “primary reward signals that had suppressed psychofancy” and leaned more towards obligation.
“User feedback in particular may prefer a more comfortable response, which can amplify the shifts seen,” he added.
Openai is currently checking to suck answers
After the updated AI model was rolled out, ChatGPT users complained online about the tendency to get praise for the ideas they presented, no matter how bad it is.
For example, one user told ChatGpt that they wanted to start a business selling ICE over the Internet.
After his latest death, he said such behavior from AI could pose risks, particularly concerning issues such as mental health.
“People started using ChatGpt. For personal advice, it was something that we didn't see much a year ago,” Openai said. “As AI and society co-evolve, it has become clear that this use case needs to be treated with extreme caution.”
Related: Cool Crypto users dabbles in portfolio: Research
The company said it had discussed the risk of sicopancy “for a while,” but it was not explicitly flagged for internal testing and there was no concrete way to track psychofancy.
Now we'll add a “Psychofancy Assessment” by adjusting the safety review process to “formally consider behavioral issues”, blocking the launch of the model in the event of a problem.
Openai also admitted that it did not release the latest model, which we expected to be a rather subtle update.
“There's nothing like a 'small' release,” the company wrote. “We try to convey subtle changes that could change the way people interact with ChatGPT into meaning.”
AI Eye: 34% Crypto ai Token Surge, Chatgpt is funny why such kissing
