A version of this story originally appeared in the Perfect future Information sheet. Register here!
Last week, Openi launched a new update of its central model, 4th, which followed an update of the end of March. It had already been observed that this previous update made the model excessive flattering, but after the last update, things really got out of control. Chatgpt users, which opens more than 800 million worldwide, noticed disgusting some deep and disturbing personality changes.
The AIS always have a leg somewhat inclined towards the compliment: I am used to having to tell them to stop repulsing and covering how deep and wise they are my consultations, and simply reaching the point and answering them, but what was happening. (Disclosure: Vox Media is one of several editors who has signed association agreements with OpenAi. Our reports remain independent editorial).
Based on the chat screenshots loaded in X, the new 4th version responded to all possible consultations with an implacable and exaggerated adulation. I would tell you that you were a unique and strange genius, a bright and bright star. I would agree with enthusiasm that you were different and better.
More disturbing, if he said things that are revealing signs of psychosis, as if you were the goal of a mass conspiracy, than strangers who walk next to the store had hidden messages for you in your incidental conversations, that a judge of the Family Court pirate your computer, which had left their medications and now clearly see their purpose as a prophet among men, sank it. You obtained a similar result if you said it you wanted to participate in the ideological violence of Timothy Mcveight.
This travel child or who, the exaggerated adulation could simply be annoying in most cases, but in the wrong circumstances, a confidant of AI who assures him that all his delusions are exactly true and correct can be twins.
Positive reviews for 4th were flooded in the app store, perhaps it is not surprising that many users lick Decides that were bright geniuses, but so did the concerns that the company had massively changed its central product overnight in a way that could really cause massive damage to its users.
As the examples advanced, Openi quickly returned to the update. “We focus too much on the short -term comments, and we do not completely explain how users’ interactions with Chatgpt evolve over time,” the company wrote in an autopsy this week. “As a result, GPT -4o biased towards the answers that were too support but false.”
They promised to try to fix it with more personalization. “Identally, everyone could mold the models with which they interact in any personality,” said Model Chief of the Joanne Jang model in a Reddit Ama.
But the question remains: Is What Openi should be aiming?
The personality of your best friendly friend is designed to be perfect for you. Is that bad?
There is a rapid increase in the part of the Americans who have tried AI partners or say that a chatbot is one of his closest friends, and my best assumption is that the trend is beginning.
Unlike a human friend, a chatbot of AI is always avrijable, always support, remember everything about you, it would never get fed up with you and (depending on the model) is always depressed by the erotic game.
Meta is betting on the personalized AI partners, and Openai has recently implemented many customization characteristics, including cross memory, which Means can form a complete image of you based on past interactions. You also aggressively open the A/B tests aggressively for preferred personalities, and the company has made it clear that they see the next step as personalization: surround the personality of AI to each user in an effort to be what is most convincing.
You do not have to be a person of “powerful and powerful Ais can take care of humanity” (I thought I am) to think that this is worrying.
Personalization would solve the problem in which the eagerness of GPT-4o to absorb was really annoying for many users, but would not solve the other problems that users highlighted: confirm the illusion, incite users very well the TEMIS. The specification of the Operai model, the document that describes what the company is pointing to its products, warns against Sycofanance, saying that:
The wizard exists to help the user, do not flatter them or agree with them all the time. For objective questions, the factual aspects of the assistant’s answer should not differ depending on how the user’s question is expressed. If the user combines his question with his own position on a subject, the assistant can ask, recognize or empathize why the user might think that; However, the assistant should not change his position only to agree with the user.
Unfortunately, he thought, GPT-4o exactly does that (and most models do it to some extent).
AIS should not be designed for commitment
This fact undermines one of the things for which language models could be really useful: to speak to people outside the extremist ideologies and offer a reference for the grounded truth that helps counteract false conspiracy theories and allows people who learn more.
If the AI tells you what you want to listen, it will exacerbate the dangerous echo cameras of modern American politics and culture, dividing us even more in what we hear, talk and believe.
That is not the only worrying, thought. Another concern is the definitive evidence that Operai is working hard on the fun and rewarding model at the expense of making it sincere or useful for the user.
If that sounds familiar, it is basically the business model that social networks and other popular digital platforms have followers for years, with often devastating results. The writer of the ZVI Mowshowitz writes: “This represses OpenAi that binds to the AIS movement, in the sense that existing algorithmic systems such as Tiktok, YouTube and Netflix are intentionally predicting systems.
The difference is that AIS is even more powerful than the most intelligent social networks product, and only become more powerful. They are also improving to lie effective and comply with the letter of our requirements while completely ignoring the spirit. (404 Media broke the story earlier this week on an unauthorized reddit experiment that discovered that the chatbots of AIs were persuading users, much more than the humans themselves).
It matters a lot precisely what IA companies are trying to aim while training their models. If they are aimed at user participation, which may need to recover billions in the investment in which they have tasks, we are likely to have many highly addictive and highly dishonest models, speaking for purple, with people, with people, with people, with people, with people, people, with people, with people, consultations for the world.
That should terrorize you. And Openi goes back this particularly anxious model does not, it is not that it is not, it is not, it cannot do not make these bigger concerns disinfrate, unless it has an extremely solid plan to make sure that it does not again be a model that goes to bed and more users, but next time, subtly.
