Apple’s ‘AI Can’t Reason’ Claim Seen By 13M+, What You Need to Know

A widely circulated claim that Apple’s latest AI work shows large language models can’t “reason” is met with a blunt counterpoint: these systems struggle with exact computation and long, rule-bound tasks when they’re forced to answer without tools, but that limitation is neither new nor proof of “no reasoning.” The core takeaway is that LLMs are probabilistic generators that can produce plausible next steps—yet they will eventually fail on tasks requiring long sequences of precise, error-free steps, especially when output length is constrained.

Apple’s paper is described as focusing on the idea that LLMs don’t follow explicit algorithms and degrade as task complexity rises. To test that, it uses puzzle-style benchmarks such as Tower of Hanoi (moving discs without placing larger on smaller), checkers (moving tokens to the correct side under rules), and the fox-and-chicken “river crossing” problem. If LLMs were behaving like fixed programs, the transcript argues, performance would stay stable as the number of pieces increases. Instead, accuracy drops as puzzles scale—consistent with the long-known behavior of probabilistic models that aren’t guaranteed to produce the same correct output every time.

The transcript then connects those failures to two practical constraints. First, without external tools, LLMs can’t reliably perform exact arithmetic once numbers get large enough; multiplication becomes a point where they “never get the sum right” rather than merely making occasional mistakes. Second, even when models know the algorithmic steps in principle, long tasks can exceed token limits. One example given is a Claude model tested with a 128,000-token output cap; some questions require more than that, so the model can’t produce a full solution trace and instead outputs shorter responses such as “here is the algorithm” or “use this tool.”

A further critique targets the framing of “thinking” versus “non-thinking.” The transcript says Apple originally wanted to compare long chain-of-thought style models against shorter-trace models on math benchmarks, but results didn’t match expectations; the testing emphasis shifted toward puzzles. Another surprise in Apple’s findings—failure even when the algorithm is provided in the prompt—is treated as unsurprising for probabilistic systems: knowing the procedure doesn’t guarantee error-free execution across millions of steps.

The broader implication is that headline writers may overstate “fundamental barriers to generalizable reasoning.” The transcript argues that serious researchers were unlikely to be shocked by results showing LLMs’ limits in tool-free, long-horizon precision tasks. At the same time, it stresses that LLMs can become far more reliable when paired with tools and environments that correct mistakes—turning generative text into a component of a system that can support scientific progress.

The closing guidance shifts from debate to selection. For free use, it recommends Google’s Gemini 2.5 Pro, with an honorary mention for Deepseek R1 via API. It also warns that benchmark headlines can mislead: OpenAI’s 03 Pro at a $200 tier shows strong math/science/coding scores, but the transcript claims the teased 03 system from December 2024 performed better than 03 Pro on some tracked results, urging users to evaluate performance on their own use case rather than rely on record-setting charts.