A new study revealed that the models of inferring artificial intelligence, such as Claud from Meta and Openai and Deepseek models, are not smart as rumored, they do not actually think but rather depend on statistical guesses.
The study, conducted by Apple, showed that these models significantly lose their accuracy with an increase in the complexity of tasks, and they even collapse in solving complex problems.
These large models (LLMS) use huge amounts of human data to generate relatively accurate responses, based on nervous networks that extract the possibility of these data. The evacuation models improve their accuracy through the technique of “The Series of Ideas”, which allows it to divide thinking into multiple steps similar to human logic.
However, this thinking remains merely a statistical guessing process, which leads to the phenomenon of “hallucinations” as these models provide wrong answers, false information, or even incorrect advice. A report by Openai confirmed that logical inference models tend to hallucinations more than general models, and the problem is exacerbated with the development of models.
In the new study, the researchers presented several models, including Openai, Deepseek, Anthropic and Google, to solve 4 classic classic puzzles with different complexity levels. The study found that general models excel in simple tasks without the need for a high mathematical cost, while the effectiveness of the inference models increases with increased complexity, but they completely collapse at high levels of complexity.
The results show that these models are more dependent on identifying patterns and not on real logical thinking, which limits their ability to continue to solve complex problems.
The researchers pointed out that the current assessments do not provide a comprehensive analysis of the reasoning and quality of paths.
In a related context, Apple suffers from a delay in the field of artificial intelligence compared to its competitors, as analyzes showed that its Siri’s voice assistant has been 25% less accurate than ChatGPT. Apple’s approach focuses on developing effective artificial intelligence integrated on its devices instead of the complex large models.
This trend has caused criticism, as Pedros Domingus, a professor of computer science, mocked that Apple’s strategy is “proving the absence of artificial intelligence.”
On the other hand, other experts praised the study because it restores the scientific debate to its basis away from exaggerations, stressing that the large language models are still limited by nature as nervous networks.
The researchers hope that these results will encourage more systematic studies to understand the capabilities and boundaries of current artificial intelligence models, and pave the way for the development of more intelligent and understanding techniques in the future.
