TLDR
- Apple’s paper tests reasoning models in controlled puzzle settings.
- It says model performance drops as puzzle complexity rises.
- The paper says models failed after step-by-step algorithms were given.
- Hoskinson says Apple overlooks the LLM role in AI minds.
- The debate centers on LLMs, world models, and symbolic reasoning.
Apple’s paper The Illusion of Thinking has drawn fresh attention after Charles Hoskinson said the company overlooks the role of large language models in building real AI minds. The debate centers on whether LLMs are only pattern systems, or a needed base for world models, symbolic reasoning, and adaptive machine intelligence at scale for future AI research and systems design.
Apple Paper Questions Model Reasoning
Apple’s 2025 paper, titled “The Illusion of Thinking,” examines whether advanced reasoning models can solve controlled puzzle tasks. The paper argues that strong text output does not prove real understanding or reliable reasoning. The study tested models as task complexity increased.
According to the paper, model performance did not fall smoothly. It dropped sharply after certain complexity levels, and the models failed on harder tasks. Apple also reviewed the internal reasoning traces produced by these systems. The paper says models often used more reasoning tokens on simple tasks.
Yet they used less effort when tasks became more complex. That finding has been central to the public debate. Critics say it shows that current AI systems may not manage hard logic well. Supporters say puzzle tests do not cover every real-world AI use case.
Hoskinson Says Apple Misses the LLM Role
Charles Hoskinson has argued that Apple is underestimating the role of LLMs in the race toward real AI minds. His view is that LLMs are not the full answer. Yet he sees them as a base layer for broader systems. In that view, an LLM works like fast thinking. It handles language, patterns, and quick links between ideas.
A world model can add structure, while symbolic reasoning can add rule-based checks. The debate is not only about whether LLMs reason today. It is also about whether they can support systems that reason better later. Hoskinson’s position suggests that LLMs may become useful when joined with other tools.
Apple wants to pretend LLMs add nothing to the AI race. They are missing the forest for the trees. Adding a world model and neurosymbolic reasoning to an LLM effectively makes a brain. The LLM component is the thinking fast in the thinking fast and slow. Since the 1950s, we've… https://t.co/NndpyMOBeR
— Charles Hoskinson (@IOHK_Charles) May 21, 2026
This frames the Apple AI paper in a narrower way. Apple tests current reasoning models under controlled conditions. Hoskinson focuses on future AI architecture that combines learning, memory, logic, and environment models.
Step-by-Step Tests Raise Execution Concerns
One part of Apple’s paper drew close attention. The researchers gave models step-by-step algorithms for solving puzzle tasks. The models then had to follow those instructions across harder cases. The paper says this did not solve the problem. Performance still failed when complexity crossed a high level.
Apple presented this as evidence that current models struggle to execute long logical sequences. That claim supports the view that present systems can imitate reasoning without fully managing it. It also raises questions about reliability in tasks that need exact steps and stable planning.
Still, the paper does not end the AI reasoning debate. It studies selected puzzle settings, not every form of AI use. The main dispute now concerns architecture. Apple points to limits in current models, while Hoskinson points to missing system design.
