Divya van Mahajan

No, Apple's New AI Paper Doesn't Undermine Reasoning Models

ai-daily-brief-podcast

Study Document: Apple’s WWDC Non-Event and the “Illusion of Thinking” Paper

Overview

This episode of the AI Daily Brief (recorded June 10, 2025) covers two related Apple stories: the near-total absence of meaningful AI announcements at Apple’s WWDC 2025, and the controversy surrounding an Apple research paper titled “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity.” The host argues that the paper’s findings are largely overstated, methodologically flawed, and practically irrelevant to most AI practitioners. The speaker is the host of the AI Daily Brief podcast/video channel; no name is explicitly stated in the transcript.

Source video URL: not provided

Prerequisites

  • Basic familiarity with large language models (LLMs) and reasoning models (e.g., Claude, OpenAI’s O3, DeepSeek R1)
  • Understanding of what “chain-of-thought” or extended thinking in AI models means
  • Awareness of the Tower of Hanoi puzzle and its algorithmic properties
  • General knowledge of the AI capability debate (scaling laws, AGI discourse)
  • Familiarity with Apple Intelligence and Siri as Apple’s AI products

Main Points

1. Apple’s WWDC 2025: A Conspicuous AI Absence

  • Expectations for AI announcements were already minimal heading into WWDC; those low expectations were met
  • No significant Apple Intelligence updates were shown; AI Siri was completely absent from the conference
  • Minor updates included a new image model, a revised iOS numbering system, and a controversial “Glass UI” redesign widely criticized as confusing and purposeless
  • Industry observers described the event as potentially “the most boring WWDC ever,” while Bloomberg’s Mark Gurman was a notable dissenting voice calling it “excellent” but still acknowledging the lack of AI features as “startling”
  • Investor concern is growing: one portfolio manager called Apple’s AI gap an “existential risk” for the company’s ability to justify premium hardware pricing

2. The “Illusion of Thinking” Paper — What It Claims

  • The paper tested reasoning models (Claude 3.7, O3 Mini High, and others) on logic puzzles — primarily the Tower of Hanoi — where complexity scales exponentially with the number of disks
  • Core finding: reasoning models handled up to 6–7 disks well but failed sharply at 8+ disks
  • The paper concluded that reasoning models have “limitations in exact computation,” “fail to use explicit algorithms,” and “reason inconsistently across puzzles”
  • The implication promoted in media coverage was that reasoning models do not truly reason — they merely pattern-match
  • AI skeptics (notably Gary Marcus) cited the paper as a “knockout blow” for LLMs and evidence that the path to AGI via LLMs is fundamentally limited

3. Methodological Flaws Identified by Researchers

  • Independent replication (by “Lisan Al-Gaib / Scaling 01”) revealed that models were hitting output token limits, not reasoning limits
    • The Tower of Hanoi requires 2ⁿ − 1 moves; each move requires ~10 tokens of structured output
    • At 8 disks (255 moves), models predictably ran out of token budget — a known engineering constraint, not a cognitive ceiling
  • Rather than failing silently, models recognized they could not output a full solution and instead described the algorithmic approach — a sign of metacognitive awareness, not failure
  • The researchers prohibited models from writing code, which is a standard and highly effective tool for solving algorithmic puzzles
    • When allowed to write code, models solve Tower of Hanoi at seemingly unlimited complexity
  • Kevin Bryan (University of Toronto) noted that post-training alignment intentionally prevents models from generating millions of tokens; the paper was measuring self-imposed constraints, not fundamental reasoning limits

4. The Practical vs. Research Framing Divide

  • The host distinguishes two audiences: the research community (focused on AGI, cognition, long-term scaling) and the business/applied community (focused on current capabilities and productivity)
  • The host’s position: for business users, whether a model is “truly reasoning” is irrelevant if it reliably performs tasks that weren’t previously automatable
  • Josh Gans (University of Toronto, management professor) is cited: reasoning models are delivering real value in enterprise and academia, working “exactly as people explained they would work”
  • Francois Chollet (ML scientist) provides the strongest counterargument: true reasoning enables autonomous skill acquisition in new domains, whereas pattern matching can only emulate known skills — this distinction matters for long-term capability development
  • The host acknowledges Chollet’s point but maintains it is not immediately relevant to practitioners

5. Contextual and Institutional Credibility Issues

  • Multiple commentators noted the irony of Apple — whose AI products are widely considered behind competitors — publishing a paper arguing AI reasoning is limited
  • The timing (coinciding with a WWDC devoid of AI announcements) led observers to characterize the paper as self-serving
  • Apple researchers have a pattern of publishing papers arguing LLMs are fundamentally limited, despite (or perhaps because of) Apple’s weak competitive position in AI
  • OpenAI’s O3 with near-unlimited compute effectively solved ARC-AGI benchmarks, demonstrating that token/compute constraints — not reasoning ceilings — are the operative limitation in deployed models

6. The Broader Discourse Problem

  • Viral framing (“Apple proves AI doesn’t reason”) spread via social media without engagement with the paper’s actual findings
  • Kat Woods noted the paper’s abstract explicitly states models do reason, just not with 100% accuracy on hard problems — the headline misrepresents this
  • The episode identifies a recurring pattern of AI skepticism that overgeneralizes methodologically narrow findings into sweeping claims about the technology’s limits
  • Gary Marcus is cited as a recurring example, having declared AI at a wall repeatedly since at least March 2022

Key Concepts

  • Reasoning models: LLMs augmented with extended chain-of-thought processing (e.g., Claude 3.7 with thinking, OpenAI O3) designed to work through multi-step problems before producing an answer
  • Tower of Hanoi: A classic algorithmic puzzle used as a benchmark; requires 2ⁿ − 1 moves for n disks, making it an exponentially scaling test of sequential reasoning
  • Output token limits: Engineering constraints on the maximum length of a model’s response; in this paper’s context, the proximate cause of model failure rather than a reasoning limitation
  • Chain-of-thought (CoT): A prompting technique where models reason step-by-step before giving a final answer, forming a visible “reasoning trace”
  • Pattern matching vs. reasoning: A central debate in AI — whether model outputs reflect genuine logical inference or sophisticated interpolation from training data
  • Apple Intelligence: Apple’s branded suite of on-device and cloud AI features, announced at WWDC 2024 and widely considered underdelivered
  • ARC-AGI benchmark: A test of general fluid reasoning designed to be difficult for LLMs; O3 with unconstrained compute performed near-human levels, at very high cost
  • Post-training alignment: The fine-tuning and reinforcement learning process applied after pretraining that shapes model behavior, including output length and tool use — a source of the constraints the paper inadvertently measured
  • Scaling wall: The hypothesized point at which additional compute or data stops yielding capability improvements — what the paper’s critics say it failed to demonstrate

Summary

The episode argues that Apple’s Illusion of Thinking paper — while generating significant media attention and being weaponized by AI skeptics as proof that reasoning models are fundamentally limited — is methodologically undermined by a straightforward engineering artifact: models failed Tower of Hanoi puzzles at higher complexity not because they hit a reasoning ceiling, but because they ran out of output tokens, a constraint imposed by deployment economics rather than cognitive architecture. Further, the paper excluded code execution, which trivially solves the problem at arbitrary complexity. The host situates this within a broader divide between researchers focused on long-term AGI questions and practitioners focused on current tool utility, firmly siding with the latter: for business users, the philosophical question of whether a model “truly reasons” is irrelevant as long as it performs tasks that create value. The episode also frames the paper within Apple’s broader credibility problem in AI — a company whose own AI products lag significantly behind competitors is an unlikely authority on the fundamental limits of a technology it has failed to ship. The paper is best understood as a Rorschach test for pre-existing views on AI, rather than a rigorous demonstration of a scaling wall.