AI Challenge: Think of a Number #claude #gemini #aichallenge #google

Overview

Mayank Shah demonstrates asking Gemini and Claude to choose a number from 1 to 10. The video claims that 7 is a likely answer, although the subtitles do not show the models’ actual replies. Samay is mentioned in the opening example.

Why Models May Favor Certain Answers

The speaker explains this tendency through gradient descent: during training, models adjust predictions to reduce errors and learn patterns in their data. If people favor certain answers, a model may learn to favor them too. The speaker compares this bias to loaded dice and says fine-tuning can reinforce it.

AI-Generated Ideas and Uniqueness

The speaker argues that AI-generated ideas may be less unique than they seem, since people using the same models can receive similar outputs. Human taste, experience, and ideas are presented as what makes the difference.

Number Discrepancy

The introduction focuses on 7, but the explanation switches to 6 as an example of a frequently chosen number.