Google’s new AI research shows reasoning models argue & debate for best answers #ai #chatgpt #fyp

AI's New Thinking Process: Are They Going Insane?

Understanding AI Reasoning Models

• The discussion begins with the assertion that AI has become "smarter" by adopting a more complex reasoning process, as detailed in a recent Google paper.

• Contrary to popular belief that models like OpenAI's GPT-3 simply think longer or run more calculations, it is revealed that these AIs simulate debates within themselves using multiple voices and personalities.

The Personalities Within AI

• When faced with complex problems, different personalities emerge:

• A meticulous mathematician who follows steps carefully.

• A creative risk-taker proposing unconventional ideas.

• A skeptical challenger questioning assumptions made by others.

Mechanisms of Self-Debate

• Research indicates that the AI questions itself approximately 6.7 times per problem and shifts perspectives about 3.5 times, engaging in self-debate over three times for each problem presented.

Discoveries from Research

• Researchers identified a feature called the "conversational surprise marker," which triggers personality shifts within the AI. Amplifying this feature led to improved accuracy on complex problems, increasing from 27% to 55%.

Testing New Models

• In an experiment with a fresh AI model rewarded solely for correct answers (without any mention of debates), the model began to spontaneously engage in arguments with itself, demonstrating an inherent tendency towards multi-perspective reasoning.