Apple Study Highlights AI Flaws
Apple pulled out at the last minute from the most recent round of investments in OpenAI, without offering a specific reason for the decision. However, a new internal study may shed light on this move, revealing inherent flaws in how large language models (LLMs) function. The translated title of the scientific paper is "Understanding the Limitations of Mathematical Reasoning in Large Language Models." The researchers behind the publication indicate that small variations in how prompts are structured can lead to entirely different outcomes, undermining the reliability of the models. Additionally, the more variables introduced, the worse the AI's performance becomes.
According to the study, the inclusion of a single irrelevant phrase can reduce the final accuracy by up to 65% in a specific math problem. This is an inevitable consequence of how language models are built—not based on real reasoning but on a fragile identification of patterns. The researchers conclude with a critique of widespread AI usage: "There is simply no way to build reliable agents on this foundation, where changing one or two irrelevant words or adding a small amount of irrelevant information can yield a different response."
