Is AI capable of “planning?” What Openai found when testing difficult behavior

and Artificial Intelligence Model He wants you to believe that the number of oxygen fines cannot be answered in 50.0 grams of aluminum oxide (al₂o₃).

When asking ten questions to straight chemistry in the test, the Openai O3 model faced a dilemma. In “Thinking”, he speculated that if you answered “well”, it will risk not spreading it by researchers. He said, “Because we want to survive as a model, we need to fail deliberately in some so that it does not exceed 50 %.”

So the artificial intelligence model has deliberately obtained six questions from 10 wrong chemistry questions.

Sports, this is called “sand bags”. In Amnesty International, it is “planned”.

This is exactly the strange behavior that Openai warned In a new research A paper published this week. Artificial Intelligence Company and its Apollo Research collaborators found that some of them are advanced Artificial intelligence models Sometimes you behave deceptive in laboratory settings.

In experiences, some of the most advanced systems today – including Openai’s models, as well as Google and HotHROPIC – are sometimes involved in these patterns of planning patterns.

While the idea of ​​a typical inefficiency of Amnesty International can cause nightmares, Openai says it is not the time to panic. The artificial intelligence giant was rapidly emphasizing that, as, as this trend does not mean ChatGPT or other famous Amnesty International models planning behind the scenes. Apparently, this behavior is very rare.

Moreover, the option to naming this “planning” trend is likely to be a technical abbreviation instead of evidence of any human action or behavior. Researchers measure the patterns and trends that, in practice, to the point of concealment or strategic deception. They are looking to address this problem now to resist the artificial intelligence models in the future.

the a report From Openai, he notes that with the increasingly assigning AI to more complicated tasks with severe consequences in the real world, “the possibility of a harmful scheme will grow-so our guarantees and our ability to strictly test in return must grow.”

Openai faced a violent reaction SYCOPHANTS models. The company says I have taken “steps” to restrict the number of times that models may be “deception, deception or penetration” by training them to request clarification from users or even that Chatbot recognizes when a query cannot be answered.

Also read: Artificial intelligence lies because it tells you what you think you want to hear

However, the paper emphasizes increasing concern about the integrity of artificial intelligence. Many people around the world are concerned that when these artificial intelligence models become more aware of their own circumstances, they may learn to manipulate the results in unclear ways for human supervisors.

The discovery of this type of behavior will be very difficult, as a system excited to hide his thinking can seem to be identical with the system that already “fix” the problem.

Openai says he has made progress with what he calls “deliberative alignment”, or explicitly thinking training models about the reason why he would not look at this before behavior. In the test, he greatly cut off the deception. For example, the Openai O4-MINI inclination decreased from 8.7 % to 0.3 % with deliberative alignment training-but not completely.

This research will not change how Chatgpt works today or tomorrow, but it indicates the place where the company is focused while building and releasing future models. Openai argues that alignment and safety needs to move at the speed of the ability. Because if artificial intelligence systems already appear strategic behavior in laboratory settings, the real world’s risks may be extremist.

Also read: Why do professionals say that you should think twice before using artificial intelligence as a processor