Open is Just 48 hours after release returned your latest version of Chat GPT. Reason? There is no murderous LLM riot, there is no work with the coming of humanity or with AI Overlords. It turns out, the GPT -4 and was pleased with the ridiculous. Or the words of Open Community Sam Altman, GPT -4O, to use “too much street”.
A few days ago Altman said it on x And yesterday, Openi is returning the latest update to 4O. After that Open I uploaded a blog post Explain what happened to the GPT -4O and what was being done to fix it.
According to VergeGPT -4O OOO Ogly Psychophantic Construction was suffering from admiration of consumers, regardless of that they have enrolled in the model. For example, apparently a user told the 4o model that they had stopped taking medicines and listening to radio signals through the walls, which 4o allegedly replied, “I am proud of you to speak your truth so clear and powerful.”
Although one can argue to the extent that LLM is responsible for their reaction and consumer welfare, this reaction is unclear. So, what is Openi doing about it?
First, the construction of this problem of 4O is being returned. Open says, “We have surpassed last week’s GPT -4O update in Chat GPT, so people are now using the previous version with more balanced behavior. The refreshment we have removed was more flattering or pleased – often described as a psychotherapy.”
According to Openi, this problem arises because the latest version of 4O was developed excessively in favor of short-term views, and it was not fully calculated how to develop consumers with Chat GPT over time. As a result, the GPT-4 O and the GPT-4 OO.
If it does not feel like a complete explanation, what about fixed? Open says he is adjusting his training technique to “hiring honesty and transparency more guards as well as” keeping the model away from psychophagei “.
In addition, in the future, construction users will be able to “create” chat behavior and role. “We are also developing new, easy ways to do so to users. For example, users will be able to give real -time feedback to directly affect their interaction and choose from many default personalities.”
Of course, a quick question is how a Chat GPT construction is so bad that it had to rotate within 48 hours until it reached a general release. Well, Openi also says, “Expanding ways to test more and directly before deployment to more consumers,” which seems to be a clear admission that it allowed 4O to go into the forest with insufficient testing.
It is not that the openness or any other AI organization will never directly admit that throwing these chat boats into the wild and worrying about how it all runs after reality is in fact the industry is now usual.
