Sebastien Bubeck asked GPT-4 to solve a complex math problem. The model initially answered quickly, but when questioned about its reasoning process, it admitted: 'I can see the answer, but I don't know how to derive it.'

This is a common scenario. AI models have been trained on vast amounts of solved problems, allowing them to recognize answer patterns, but true understanding is something else entirely.

The interesting part comes later. When researchers instructed GPT-4 to slow down and think step by step, it began constructing logical chains. It started with simple cases, identifying patterns, and then generalized to the broader problem. After a few attempts, it not only arrived at the correct answer but could also clearly explain the meaning of each step.

The entire process mirrors how humans learn new knowledge—from an intuitive 'knowing the answer' to a genuine understanding of the underlying principles.

Another detail: When asked, 'Do you understand now?' GPT-4 replied: 'Yes, through step-by-step derivation, I understand how this pattern works.' This self-reflection ability is more noteworthy than simply solving the problem.

This reminds us that AI's 'intelligence' lies not just in the correctness of its answers but also in the process of arriving at them. The ability to admit ignorance, learn incrementally, and eventually understand—this capability may be more valuable than simply possessing the right answer.

The math problem itself was complex, but observing how AI solved it revealed something more fundamental: true understanding requires time, attempts, and the courage to move from the known to the unknown. This applies to both humans and machines.