OpenAI, Google, and Meta Researchers Warn We May Lose the Ability to Track AI Misbehavior

More than 40 scientists, including Openai, Google Deprmind, Anthropic and Meta, came together to conduct more research in a certain type of security monitoring, which allows people to think about EU models.

Scientists made a publication Research paper As a new but fragile opportunity to increase the security of AI on Tuesday, the thinking chain (COT) is known as the monitoring. Paper, Openai co-founders were approved by famous EU figures such as John Schulman and Ilya Sutskever, as well as Nobel Award winner as Nobel Award winner as “AI,” Geoffrey Hinton.

They explained how modern media models such as paper, scientists, Chatgept modern thinking models such as “carrying out the reassurance or final results.” In other words, the form of memory that works for them to solve complex tasks, step by step “thinks loud”.

AI systems that think “in the human language” offer a unique opportunity for AI security: we can follow the chains of the documents, “the authors of the documents said.

Researchers claim that the researchers can help researchers when the researchers begin to operate defects in trainings, information manipulation or victims of the victims of the victims. Any problem found can be either “blocked or replaced with more reliable movements or be considered deeper.”

Openai researchers have already used this technique to find the statements that are the phrases of AI models “Let’s hack“In their beds.

Current AI models exercise this idea in human language, but researchers warn that this will not always be the case.

The developers can use them, not their future models come to them, but the future models could not easily understand. In addition, developed models can learn to invest or hide their thoughts if they detect the monitored.

In response, researchers apply to track and evaluate the tracking of the COT of the AI developers and treat them as a critical part of general model security. They even recommend that new models make it a key view while preparing and placing.