Criticisms of AI begin and they do it from the greats. In this case it is Apple that has published a study entitled the illusion of thought, which basically states that all these Reasoning models such as Claude, Gemini and Chatgpt do not reason. They only memorize. Let’s start at the beginning.
A large reasoning model (MLR) is A IA system designed to solve complex problems Following each step, essentially showing its reasoning process before providing an answer.
Unlike the standard language models that predict the following word, these models They are trained to apply structured and logical reasoning, tearing the problems more human, step by step.
Apple’s study tested some of the best MLR, such as O3 Mini of OpenAi, Deepseek R1 and Claude 3.7 Sonnet, with four classic puzzles (Hanói tower, lady jump, river crossing and block world) and created three levels, easy, medium and difficult. As the tests became more difficult, the models ran into a wall and, in fact, in the last one, they collapsed completely.
To this we must add one more problem: even when Apple provided them with the solution algorithms, they continued to fail. A model completed more than 100 steps in the Hanói tower … but He could not with 4 movements in a puzzle of crossing a river. Because? Because one was in training data. The other is not.
The results showed that the MLR reasoning capacity was inconsistent, which It suggests a coincidence behavior of patterns, not reasoning or real intelligence.
That said, some details must be clarified. First, The study has not been reviewed by pairs, which implies that no one has managed to reproduce these results and we must trust Apple.
Another criticism that circulates online is that Apple used puzzles instead of real world scenarios. Critics also argue that The puzzles exceed the limits of tokens and steps of the models, so the “failures” reflect these limitations instead of a lack of reasoning.
But the most sharp criticism is that “Apple is behind in AI, so it is pointing out the weaknesses of artificial intelligence lots of rival systems after losing land ”, Point out the New York Times.
Does the ability of AI invalidate all this? At all. It is a tool, with deficiencies and improvement capacity, but it is not a replacement. Basically, it’s about using it to Perform tasks, not to do them for us. There is room for improvement and occurto … mainly for Apple.