GPT-40, the most recent version of the am model that underpins ChatGPT, has received several updates that are supposed to improve both its "reasoning" and "personality." But things haven't gone as planned, and OpenAI is determined to correct the situation as soon as possible.
Advanced users seem to have noticed some positive points in the chatbot's raw performance. While this latest version seemed particularly prone to hallucinations—those aberrant, nonsensical, or factually false responses that emerge without anyone knowing exactly why—these updates have apparently helped to limit these excesses somewhat.
A penchant for counterproductive flattery
But the same cannot be said for the second point. Many Internet users have noticed that the "personality" The chatbot's performance had indeed evolved... and not in the right direction. It had indeed developed an unbearable penchant for flattery. And that's a big problem.
In addition to making certain interactions downright exasperating, this also makes it less effective in certain exercises. For example, this propensity to flatter the ego of its interlocutor significantly limits its ability to provide relevant and judicious criticism. By constantly pandering, it tends to no longer question human assumptions and assertions as effectively as in the past. A worrying dynamic, at a time when More and more people seem to be blindly trusting these LLMs.
Soon a ChatGPT with multiple personalities?
This week, Sam Altman himself admitted that the situation was no longer tenable. In a post on Twitter/X, he acknowledged that the latest updates had turned ChatGPT into a "boot licker," and that OpenAI was preparing a new round of patches to resolve the problem.
In a response to a user who asked him If it were possible to switch between different personalities, the CEO also suggested that this functionality would be implemented soon. "Eventually, we clearly need to be able to offer multiple options," he explains. From a technical perspective, it will be interesting to see how the company plans to integrate different, distinct personalities into its model.
It should be noted, however, that the current version of the tool already offers a certain degree of control at this level. In the app settings, the "Personalization" section has a field where the user can explicitly indicate "what features ChatGPT should have." In theory, there's nothing stopping you from telling them to avoid flattery and stay objective.
0 Comments