According to a paper published over the weekend by Apple’s boffins, large reasoning models, an advanced type of AI designed to tackle complex problems, suffer what they describe as a “complete accuracy collapse” when asked to do anything more challenging than a toddler’s puzzle book.
For a company that can’t even get a voice assistant to answer basic questions without mangling half the query, this sudden philosophical awakening reads more like a bitter letter from an ex-girlfriend. The paper’s authors found that when faced with serious reasoning tasks, these cutting-edge models actually reduced their effort. When the going got tough, the AI just gave up.
AI academic and noted sceptic Gary Marcus called the findings “pretty devastating”, clearly pleased to see one of the world’s most valuable companies publicly admit what he’s been shouting for years. Marcus pointed out that those banking on tools like ChatGPT magically transforming into godlike digital minds are “kidding themselves.”
Not that Job’s Mob actually names OpenAI in the piece, they’re too polite for that, but the study did quietly include benchmarks for OpenAI’s o3, Google’s Gemini Thinking, DeepSeek-R1 and Claude 3.7 Sonnet-Thinking. So while Apple claims it’s still in the AI game, it’s using everyone else’s models to prove that no one, apparently, can build a brain.
The researchers found that these models do okay with simpler problems but enter full-on cognitive meltdown when things get complicated. Even when spoon-fed the correct algorithm, the AI still couldn't tie its digital shoelaces.
The Fruity Cult concludes this is evidence of “fundamental scaling limitations” and that existing AI thinking has hit a cul de sac. Which is a curious thing to say just as everyone else in the industry is sprinting ahead releasing multimodal models, real-time agents and full-blown copilots.
Apple's sudden concern over "generalisable reasoning" might have something to do with the fact that Siri still can't do basic maths without ringing the internet. Unable to deliver its own AGI hype machine, Job’s Mob now seems intent on convincing the world it’s all a waste of time anyway.
Rather than fix its broken AI strategy, Apple appears more comfortable declaring that the field has reached its natural limit. In Cupertino’s warped logic, if they can’t build it, it must be impossible.
