Apple recently released its AI news summary program, Apple Intelligence, despite knowing about its significant flaws. The program was criticized for botching headlines and reporting false information, forcing Apple to pause the entire project until it can be fixed.
AI models, like those used by Apple, are prone to “hallucinations” due to their limitations in discerning relevant information. Researchers tested 20 large language models, including OpenAI’s o1-preview, with a modified version of a widely used math problem dataset. They found that changing numbers and adding irrelevant details led to catastrophic performance drops, ranging from 17.5% to 65%.
This reveals a critical flaw in AI models’ ability to reason and solve problems. While they excel at pattern matching, they struggle when required to understand the context and meaning of information. Apple’s release of its own flawed AI model highlights the industry’s tendency to overlook these limitations. The study’s findings raise serious questions about trusting AI models with tasks like regurgitating headlines without understanding their implications.
Source: https://futurism.com/the-byte/apple-engineers-ai-deep-flaws