Yes, of course I had to take a look at the paper from Apple on AI reasoning models. "The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity" reveals that even the most sophisticated AI reasoning models experience complete accuracy collapse beyond certain complexity thresholds.
The research provides fascinating context for Nate Silver's May poker experiment with ChatGPT, which offers a real-world case study of exactly what this "complexity collapse" looks like in practice.
The Apple Paper Everyone's Talking About
Apple's researchers tested frontier reasoning models (including o3-mini, DeepSeek-R1, and Claude) across controlled puzzle environments, systematically increasing complexity while tracking both accuracy and reasoning effort. Their findings are sobering: these models face "complete accuracy collapse beyond certain complexities" and exhibit a counterintuitive pattern where "their reasoning effort increases with problem complexity up to a point, then declines despite having an adequate token budget."
In other words, AI doesn't just gradually get worse at hard problems – it hits a wall and crashes spectacularly, often while reducing the effort it puts into solving them.
Poker: The Perfect Complexity Stress Test
Enter Nate Silver's recent poker experiment, which inadvertently demonstrated this complexity collapse in vivid detail. Silver tested OpenAI's o3 model on a simulated Texas Hold'em hand where it made virtually every mistake possible – catastrophically bad strategic decisions, basic math errors, and literally awarding the pot to the wrong player. He then ran a further test using Deep Research across 8 hands.
All outcomes delivered with complete confidence.
What makes poker particularly revealing is how it layers multiple types of complexity:
- Mathematical calculations (pot odds, probability)
- Rule comprehension (hand rankings, betting sequences)
- Strategic reasoning (opponent modeling, risk assessment)
- Context maintenance (tracking multiple variables across rounds)
This aligns perfectly with Apple's research showing that models struggle when "breaking down the problem into subproblems (recursive thinking), tracking multiple states and disk positions simultaneously (working memory management), adhering to movement rules and constraints while planning ahead (constraint satisfaction), and determining the correct order of operations to achieve the final goal (sequential planning)."
My Quick Field Test: The Pattern Holds
Silver tested OpenAI's models (o3 and Deep Research), so I ran a quick experiment with Gemini 2.5 Pro and Claude Sonnet 4. Gemini handled a straightforward scenario reasonably well (it got lucky – medium complexity!). Claude Sonnet 4? Completely botched it – confusing which cards a player held, a mistake that fundamentally changes the entire hand analysis.
This aligns with the Apple paper's finding that reasoning models have "limitations in exact computation: they fail to use explicit algorithms and reason inconsistently across puzzles."
The Complexity Collapse in Action
Apple's research identifies three distinct performance regimes: "(1) low-complexity tasks where standard models surprisingly outperform LRMs, (2) medium-complexity tasks where additional thinking in LRMs demonstrates advantage, and (3) high-complexity tasks where both models experience complete collapse."
Silver's poker experiment appears to have hit regime three – the complexity collapse zone. The model wasn't just making small errors; it was failing catastrophically across multiple dimensions simultaneously while maintaining unwavering confidence.
The Enterprise Opportunity
Rather than viewing this as a catastrophic failure of AI, these findings offer crucial intelligence for deploying LLMs effectively. Understanding where models hit their limits allows us to use them more strategically.
The research reveals something valuable: we can predict and plan around these limitations rather than stumble into them.
What This Means for AI Product Managers
If you're building AI-powered products or implementing LLMs in your organization:
Map your complexity landscape. Understand where your use cases fall on the complexity spectrum. The Apple research shows three distinct regimes: low-complexity tasks where standard models often outperform reasoning models, medium-complexity tasks where reasoning models excel, and high-complexity tasks where both collapse.
Choose the right tool for the job. Don't default to reasoning models for everything. The research reveals that standard LLMs can be more efficient and accurate for simpler tasks.
Break down complex problems. Instead of throwing complex, multi-step challenges at models whole, decompose them into smaller, manageable pieces that stay within the models' capability zones.
Design verification systems. Build checking mechanisms that match the complexity level of your tasks, especially when operating near the models' limits.
Test systematically. The Apple researchers used controlled puzzle environments to understand failure modes. Create similar testing frameworks for your specific use cases.
The Bigger Picture
The Apple research and Silver's poker experiment tell the same story: AI models have real limits, but understanding those limits makes them more useful, not less.
Rather than seeing complexity collapse as a fundamental flaw, we can view it as a design parameter. Just as you wouldn't use a race car for grocery shopping or a minivan for the track, choosing the right AI approach for the right complexity level becomes a core competency.
These models excel at many tasks within their operational zones. The key is staying within those zones—or intelligently breaking complex problems into simpler components that leverage AI's strengths while avoiding its brittleness.
The lesson for enterprise AI isn't that these tools are broken—it's that successful deployment requires understanding their operating parameters and designing systems accordingly.