Apple’s AI Warning: Rivals’ Models ‘Collapse’ on Complex Tasks

​Apple may be lagging behind its rivals in rolling out AI technologies, but that hasn’t stopped its researchers from taking aim at the competition. In a new research paper, Apple scientists argue that many of today’s most hyped reasoning models—including those from OpenAI, DeepSeek, Anthropic, and Google—fail badly when faced with more complex problems.

To test this, Apple used puzzles like the Tower of Hanoi, a classic challenge involving stacked disks and logical moves. The study avoided typical benchmarks like math problems and instead used puzzles where complexity could be systematically increased to test both accuracy and internal reasoning.

The result? As puzzle complexity rose, model accuracy collapsed to zero. Worse, the models often showed a “quitters’ mentality”—they began to reduce reasoning effort as problems became harder, even though they still had room to generate longer, more thoughtful responses.

One particularly concerning finding: even when explicitly given the correct algorithm to solve the Tower of Hanoi, the models didn’t improve. The effect was most severe in OpenAI’s o3-mini variants, while Anthropic’s Claude 3.7 Sonnet fared slightly better.

Apple’s paper comes just days before its annual developer conference, where the company is expected to showcase its own AI advancements. Ironically, ChatGPT is already used in Apple Intelligence, the company's AI service that launched to tepid reviews last fall.

​Bottom line: Apple is behind in AI—but it’s not shy about pointing out that even the leaders in the space have big gaps when it comes to real reasoning. Whether Apple can close the gap with its own innovations remains to be seen.