- Lumiera
- Posts
- 🔆 The Reasoning Mirage
🔆 The Reasoning Mirage
Apple's research on reasoning models, NVIDIA swamped by AI slop, and missions for safer AI.
Allegra Guinan & Emma G Moller
June 12, 2025
Was this email forwarded to you? Sign up here
🗞️ Issue 74 // ⏱️ Read Time: 6 min
In this week's newsletter
What we’re talking about: Large Reasoning Models (LRMs): the AI systems that show their "thinking" process before giving answers, and recent research revealing the illusion behind their apparent reasoning abilities.
How it’s relevant: The research exposes underlying limitations in how these "reasoning" models actually work - they often collapse when problems get complex. They also rely heavily on superficial patterns rather than true logical thinking. This has major implications for AI deployment strategies and investment decisions.
Why it matters: Relying on models that may confidently apply reasoning patterns in contexts where they don't fundamentally apply, or that fail catastrophically with seemingly minor, irrelevant alterations, can lead to unreliable decision-making and costly mistakes in real-world applications.
Hello 👋
Have you ever clicked on the “reason” or “extended thinking” option in ChatGPT or Claude? When they show you their "thinking" process, those step-by-step breakdowns before giving final answers, it feels like watching genuine reasoning unfold. Recent research from Apple has shattered this illusion, revealing that what looks like thinking is often sophisticated mimicry that breaks down in predictable ways.
ChatGPT reasoning options
The Illusion of Thinking: Capabilities and Core Findings
The advancements in Large Language Models (LLMs) have introduced a new generation of what are being called Large Reasoning Models (LRMs), which are designed to generate detailed "thinking" processes before providing answers. While these models have shown impressive performance on various benchmarks, a closer look at their underlying capabilities reveals critical insights into their limitations.