Can neural networks choose a sweet and ripe watermelon?

Can neural networks choose ripe watermelons? The editors of Hi-Tech Mail.ru decided to figure this out in an interesting experiment.

The correspondents tested three popular neural networks, ChatGPT 4.0, Claude 3.5 and Google Gemini, and asked them to assess the ripeness of watermelons based on photographs, and then taste-tested the selected samples.

Each neural network was shown a photo of a counter with 18 watermelons and given the same request to determine the ripeness of each of them.

• ChatGPT chose watermelon number 5, and also noted numbers 2, 3, 9 and 12.
• Claude preferred watermelon number 6 and highlighted numbers 3, 9, 11 and 12.
• Google Gemini requested more detailed photographs, but indicated that number 5 looked better than the others.

The editor of Hi-Tech Mail purchased watermelons number 5 and 6 for evaluation, as well as watermelon number 15, using the traditional “method” of evaluation and listening to the advice of the seller. All three watermelons were ripe and sweet, although not too sugary.

• The rind of watermelon number 5 was thicker than the others.
• The thinnest rind was found on watermelon number 15.
• The flesh of all watermelons was bright red without white veins, indicating the absence of nitrates.
• The seeds of all watermelons were black and evenly distributed throughout the flesh.

The experiment showed that neural networks can compete with traditional methods of watermelon selection due to their high ability to analyze visual data. However, it is too early to fully rely on AI, and traditional verification methods are still relevant.