Even Advanced AI Suffers 'Accuracy Collapse' in the Face of Complex Problems

A new paper from Apple has reported some worrying results for the future of do-it-all artificial intelligence, showing that the accuracy of even advanced modern AI models absolutely craters in response to more complex problems.

This isn’t just a problem of sophistication, requiring more training data or additional computing power—it’s a product of how these models fundamentally think. Worse, the ‘partial black box’ nature of artificial intelligence development means that just being able to identify a problem doesn’t necessarily mean being able to fix it.

Still, LRMs are in the early stages of development, and their boosters still believe they are the future of truly powerful AI actors.

Large Reasoning Models

Large reasoning models (LRMs) sound like such a good idea. They were supposed to be inspired by the function of the brain—to incorporate abilities that brains have, while traditional large language models (LLMs) like ...