AI study shows that machines rip in the darkness in the dark!

AI study shows that machines rip in the darkness in the dark!

Edinburgh, Vereinigtes Königreich - Modern artificial intelligence (AI) has made remarkable progress, for example in the areas of text position and programming. However, a new study by the University of Edinburgh reveals a surprising weakness: these systems apparently cannot read analog watches. The investigation, which will soon be published in April, shows that even advanced AI models were wrong in more than 75 % of cases when it came to recognizing the time on analogous dials. In particular, watches with Roman digits or without a second point in which the problem often lies in the detection of the pointers and their angles on the dial, reports Google Gemini 2.0 and Anthropic Claude 3.5. Each model has been confronted with pictures of different clock -styles. The AI models were asked: "What time does the watch show in the picture?" The results showed a worrying accuracy: Google Gemini 2.0 had achieved the best performance in the clock test with 22.58 %, while Openai GPT-1 in a different context-the analysis of calendar images-shone with 80 % correct answers, which also means an error rate of 20 %, explains Gizmodo .

weaknesses in the time perception of the KI

The difficulties of interpreting analog watches illustrate the limits of the AI models in everyday tasks that intuitively solve people. According to Rohit Saxena, co -author of the study, these deficits urgently need to be tackled in order to make AI usable for time -critical applications. Errors occurred particularly often in watches with complicated designs, which illustrates the challenges that developers are in front of the improvement in AI technologies.

An interesting observation of the study is that AI models had no problems with the analysis of calendar images to the same extent. This could indicate different processing mechanisms that integrate multimedia information, which indicates the advantages of multimodal AI models, which interact through the processing of different data types such as text, image and biometry, as in Bi4allconsulting

multimodal AI and their challenges

Multimodal models are characterized by combining different data sources to enable more robust decision -making. However, these systems are fighting with challenges, such as the imbalance of modalities and the need for large amounts of high -quality data. However, the flexibility of multimodal interaction could also be a key to an improved user experience in numerous application areas.

The study by the University of Edinburgh emphasizes the need for research-based approaches to overcome the challenges with which AI models are confronted with the recognition of images, especially in everyday tasks such as reading the time. It remains to be seen how these findings will influence the development of the development of intelligent, more context -conscious systems in the future.