Ask a large language model how many times the letter 'r' appears in the word 'strawberry' and watch it confidently declare there are two. There are three. This is not a bug that engineers forgot to fix; it is a window into what these systems fundamentally are — and are not.
The strawberry problem, as it has become known in AI circles, is not about counting. It is about the unbridgeable gap between statistical pattern recognition and symbolic reasoning. When you count letters, you perform a discrete operation: isolate each character, compare it to a target, increment a tally. Language models do none of this. They predict what text should come next based on patterns absorbed from billions of documents. They have never seen a letter in their lives.
The tokenization trap
Before any text reaches a language model's neural network, it passes through a tokenizer — a preprocessing step that chops words into subword units optimized for compression. The word 'strawberry' might become 'straw' + 'berry' or 'str' + 'aw' + 'berry', depending on the system. The model never encounters the individual letters s-t-r-a-w-b-e-r-r-y as discrete objects. It sees tokens, abstract numerical representations that bear no visual or sequential relationship to the characters humans perceive.
This is why asking a language model to count letters is like asking someone to count the notes in a symphony by reading the album's liner notes. The information exists somewhere upstream, but it has been transformed beyond recognition by the time it reaches the reasoning layer.
What the failure teaches us
The strawberry problem is pedagogically useful precisely because it is so trivial. A child who has just learned the alphabet can solve it. A system that can explain quantum entanglement, draft legal contracts, and compose sonnets cannot. This asymmetry forces a reckoning with what intelligence means.
Language models are interpolation engines of extraordinary sophistication. They have absorbed so much human text that they can simulate expertise across nearly any domain by pattern-matching against that corpus. But simulation is not understanding. When a model explains photosynthesis, it is not reasoning from first principles about chlorophyll and light absorption; it is generating text that resembles authoritative explanations of photosynthesis it has encountered. The distinction matters less when the output is useful, but the strawberry problem strips away the illusion.
The workarounds and their limits
Engineers have developed clever patches. Some systems now detect counting-type queries and route them to deterministic code that actually performs the operation. Others use chain-of-thought prompting to force the model to spell out words letter by letter before counting. These interventions work, but they are band-aids on a architectural limitation. The model itself still cannot count; it has merely been given a calculator to consult.
This pattern — augmenting statistical models with symbolic tools — may define AI development for years to come. The systems that feel most capable will be hybrids: neural networks handling the fuzzy, contextual work of understanding intent and generating fluent responses, while external modules handle arithmetic, database lookups, and other tasks requiring precision.
Our take
The strawberry problem is not evidence that AI is stupid. It is evidence that AI is alien. These systems achieve remarkable things through mechanisms utterly unlike human cognition, and their failures are correspondingly strange. The honest response is neither dismissal nor panic but curiosity. We have built something that can explain the history of the Roman Empire but cannot count to three. That should make us think harder about what we mean when we use words like 'understand' and 'know' — not just for machines, but for ourselves.
