Apple claims AI reasoning models ‘collapse’ and quit easily as Tim Cook faces WWDC without a new product
The post Apple claims AI reasoning models ‘collapse’ and quit easily as Tim Cook faces WWDC without a new product appeared on BitcoinEthereumNews.com.
Apple heads into its annual Worldwide Developers Conference (WWDC) beginning Monday with little to no progress in artificial intelligence, with struggles to meet expectations set by its tech rivals. Yet, the iPhone manufacturer claims that large language models are “failing” because they are focused more on benchmarks than solving problems. Over the weekend, a research paper circulated on social media from Apple’s AI research division that “downplayed” the capabilities of reasoning models developed by OpenAI, Google DeepMind, Anthropic, and DeepSeek. According to the paper, these models have declined in accuracy against the backdrop of an increase in task complexity, ultimately reaching a “point of complete failure.” “Existing evaluations predominantly focus on established mathematical and coding benchmarks, which, while valuable, often suffer from data contamination issues and do not allow for controlled experimental conditions across different settings and complexities. These evaluations do not provide insights into the structure and quality of reasoning traces,” it read. AI is failing when problems are harder Using custom-designed puzzles with controlled levels of complexity, Apple researchers observed that large AI models failed to keep up performances and exerted less effort as problems grew harder. The analysts, who measured the reduction by fewer inference-time tokens used during response generation, called the AI situation a “collapse.” The models tested included OpenAI’s o3-mini variant and Anthropic’s Claude 3.7 Sonnet. The o3-mini models performed “poorly,” while Claude models were slightly resilient. Even when provided with the correct algorithm for solving the Tower of Hanoi puzzle, the models did not improve their performance. Apple’s researchers concluded that these AI systems may not be as advanced in reasoning as commonly assumed.. WWDC kicks off pending any product announcement buzz In previous WWDC events, Apple used the conference to unveil new products, like the Vision Pro headset in 2022 and its Apple…
Filed under: News - @ June 9, 2025 12:29 pm