Apple’s AI research team has released a study online on the limitations of generative AI in certain areas. A problem that will surely be corrected over time and that’s good, Apple already seems to have an idea of how to go about it.
The current limits of generative AI
Apple’s team dedicated to artificial intelligence research has just published a paper in which she discusses the technological limits of generative artificial intelligence based on large language models (LLM).
The biggest fault of these LLMs, according to the study, is that they sometimes have limited logic or rather stupidly recite too much of what they have learned in training. Adding irrelevant information to the request can greatly affect the response given by the AI. This is particularly the case in mathematical results where simple useless information gives different answers depending on the AI used. Which shouldn’t be the case.
A clear example is also present in this study, before the release of Apple Intelligence at the end of the month. When making a request to the AI, it is enough to give the size of certain kiwis in a request concerning the number of kiwis harvested by a person to see a different response between GPT-4o, OpenAI’s AI, and that of Meta, Llama. Information which in principle has no impact on the final result, but which unfortunately changes the response given by the different AIs.
A clear observation which takes nothing away from the extraordinary capabilities of generative AI which are in reality only at the very beginning of their life. In the future, the Apple team believes that by combining neural networks and traditional reasoning based on neurosymbolic AI symbols, more precise answers should be obtained. To put it more simply, only time will, as always, allow current defects to be corrected.
Neurosymbolic AI means combining two forces in one, namely a large amount of data with the addition of precise and structured reasoning. In other words, asking a question full of unnecessary details without affecting the final answer. Exactly the problem mentioned in the study.
