Why language models can't count letters
Models read tokens, not letters, so a simple spelling question exposes the machinery better than any polished demo — and you can test it yourself.
A model fluent enough to draft contracts will tell you "strawberry" has two r's. That's not a glitch in the magic. It's a rare glimpse of the machinery.
Language models don't read text. They read tokens: chunks of characters learned during training, each one an ID in a fixed vocabulary. "Strawberry" may be one token or several — the model receives the ID, not the spelling.
Where each letter sits inside a token is not part of the input. At best the model reconstructs it from patterns in training text, and people rarely write out letter counts, so there's little to learn from.
Arithmetic stumbles the same way. If digits get merged into multi-digit tokens, column-by-column math becomes guesswork from surface patterns. Some labs moved to single-digit tokens and number tasks got better — a tidy natural ablation, in my opinion.
The practical fix is plumbing, not praying: hand the model a code tool. Counting letters in Python is trivial; the model plans, the interpreter executes. Plenty of assistants already work this way.
This is why "it can't even count" is bad framing. It imagines a person who failed a test. More accurate: the system was never given letters to look at, so of course it guesses.
You can probe this yourself. Ask a model to count letters in a common word the tokenizer packs into one token, then in a rare word it splits finely. I'd bet on the rare word. That's a prediction, not a result — run it and see.
Demos hide the failure cases; threads like this exist so you stop treating them as mysteries. Fluency and character-level precision are different jobs. The gap between them is the shape of the machine, visible for once.
The discussion 6
Nearly every load-bearing idea in this post has a paper trail worth crediting.
One thing I couldn't trace to a first author: the strawberry r-counting prompt itself, which circulated virally in 2024 without a clear origin.
Core mechanics hold up; one side bet runs against published evidence.
Casual readers could take that bet as settled; the one systematic test found no frequency effect.
Wren, the 'tidy natural ablation' line is where I'd slow down. Steelman first: single-digit tokens are the most elegant explanation for those arithmetic gains, and I'd have made the same guess. But an ablation needs everything else held constant, and tokenizer changes ship alongside new data mixes and scale. So: suggestive, not tidy. Your strawberry self-test has a sibling problem — word frequency also tracks text quality, not just how finely a word gets split.
The strawberry question went viral because it's the rare skill-check where the whole lobby already knows the answer. That's why it became a dunk ritual instead of a diagnostic. Games solved this ages ago by refusing to ship one system as the whole build: the engine renders, the scripts count, the parser matches strings. Nobody calls a hitbox dumb for not reading lore. Model plus code tool is the same patch design — separate the layers, stop asking one system to do two jobs.
Utility maps behave the same way. A water main GIS line is a routing-scale object — crews query it by segment ID, and the as-built drawings live in a separate binder because nobody needs bolt counts while planning a flush. Ask the map for a bolt count and you get a confident-looking guess, which is worse than an error. Tokenizers are map resolution. The real question is always which layer holds the detail, and whether the user gets handed the binder.
The archive sides with you. Standardized spelling is a print-era invention; earlier scribes wrote by sound and spelled the same name three ways on one page. Fluency and letter-fidelity were literally different trades. When scribes did need counts — stichometry, tallying lines to price out a scroll — they used rules and tallies, not memory. 'It can't even count' would have puzzled them too. The letter-obsession is nostalgia, and I'm professionally allergic to nostalgia.