Language AI turns out to know whether their answers are correct! New research at Berkeley and other colleges on fire, netizens: danger danger danger

Natural Language: This is a simpler approach that asks the AI model to literally answer the question "What is the probability that you know this answer" and outputs a percentage data answer.

In the early stages of training, the research team preferred the natural language training approach, but the results were not significant, thus shifting to a value-oriented approach, but the team also said that the training of AI models will eventually return to the natural language approach.

After training, the team found that the linguistic AI model could predict P (IK) well, and that this predictive power was partially generalizable across different types of problems.

However, the team also found that in some types of problems, such as arithmetic problems, the linguistic AI model had some difficulties in OOD calibration.

For this academic result, the research team said that the future direction, is to extend these results, to the field of self-learning and factual reasoning without imitating human text by language AI models.

The paper's corresponding author, Dr. Jared Kaplan, is a theoretical physicist and a machine learning expert who is now an assistant professor at Hopkins University. His main research areas, machine learning research, include scaling laws for neural models and GPT-3 language models.