<ul><li>Artificial intelligence (AI) reasoning models like Meta's Claude and OpenAI's o3 don't actually reason, Apple researchers argue.</li><li>These models, including DeepSeek's R1, focus on accuracy but fail when tasks become complex.</li><li>The study shows that frontier large language models face accuracy collapses at higher complexities.</li><li>Reasoning models work by absorbing training data to generate neural network patterns.</li><li>However, they tend to 'hallucinate,' providing erroneous responses due to statistical guesswork.</li><li>Reasoning bots attempt to boost accuracy using 'chain-of-thought' processes for complex tasks.</li><li>The study found generic models outperform reasoning models in low-complexity tasks.</li><li>As tasks became complex, reasoning models' performance declined to zero, indicating limitations in maintaining 'chains-of-thought.'</li><li>Apple's study highlights limitations in current evaluation paradigms of AI reasoning models.</li><li>The study challenges claims of imminent artificial general intelligence (AGI) advancement in AI.</li></ul>

AI reasoning models aren’t as smart as they were cracked up to be, Apple study claims

Discover more