How intelligent is AI actually? When given novel new problems, they struggle. #ai #tech #news #fyp

How Intelligent is AI?

The Capabilities of AI Models

• The discussion begins with the impressive capabilities of AI models like GPT-5, which can pass bar exams and medical licensing tests, as well as build code bases.

• However, it is noted that these models primarily rely on existing information from the internet, suggesting they are more about pattern matching than true problem-solving.

Testing AI's Problem-Solving Abilities

• A group of 11 mathematicians from prestigious institutions (Stanford, Harvard, MIT, Yale) decided to test AI's ability to solve original math problems that have not been published or discussed online.

• They presented 10 unique problems to GPT-5 and Gemini 3 in a one-shot format without any back-and-forth interaction.

Results of the Experiment

• The results showed that the AI struggled significantly with solving these novel problems; while it could understand and rephrase questions, it failed to provide correct proofs or solutions.

• This highlights a critical distinction between pattern recognition capabilities and genuine reasoning skills in AI systems.