A new research revealed that even the most powerful artificial intelligence systems of today have difficulty reading analog clocks. Some even make this easy mission wrong.

So why? Why does artificial intelligence still fall behind even in very easy tasks?

Why can’t artificial intelligence read the clock?

The research published in Arxiv; GPT-4 examines why giant language models for Claude 3 and Gemini have difficulty in understanding analog clocks. The results are quite striking:

• The GPT-4 has a 50 %error rate when interpreting analog clock images.
• Although Claude 3 reads the truth, even though he reads the truth, he confuses the positions of Scorpio and Satellopes.
• Gemini sometimes interprets the clock in a completely different form (instead of “10.10” instead of “2.50”).

So why can’t such advanced models do this kind of basic process? The answer lies in the form of learning of artificial intelligence.

The real problem does not understand artificial intelligence, it only predicts.

Instead of learning by internalizing the concepts of human brain, artificial intelligence models memorize the patterns on the large data sets and make predicts based on these patterns.

So when you see an analog clock, “This is the angular connection of scorpion and safety.” does not think. Instead, he tries to match with similar visuals in his educational knowledge.

That’s where the problem starts. If the analog clocks are not enough in training data, the model learns incorrectly. The watch requires a visual-logical integrity. Artificial intelligence can be difficult to divide and associate visual modules.

In addition, Romanian numbers, numerous hours and so on. Clocks of different styles are also confusing.

Does this mean that artificial intelligence is boundary?

Not exactly. This shows that artificial intelligence is still missing about the “general reasoning”.

This task, which is extremely easy for people, is a complex puzzle for artificial intelligence, but this is not a problem that cannot be corrected in the future. As a result, when he came first, he could not even draw the fingers of a hand properly, and now it is difficult to distinguish it from the truth.

Researchers aim to overcome this problem by improving the “multiple modalite” learning capabilities of the models. In other words, artificial intelligence can be successful not only in text or not only visual, but also if he can work two more.

Conclusion: Artificial Intelligence is still learning!

This research shows us: Although artificial intelligence has incredible abilities, it can still be forced in some basic processes that people naturally do.

Do you think artificial intelligence will one day make this breed easy tasks perfect, or will some things remain only human? Share your ideas in comments!

Bonus: If a model of artificial intelligence, “What time?” If you want to ask, show him a digital clock before. You will probably get more incorrect response!

Our other contents about artificial intelligence: