When Superhuman AIs Make Kindergarten Typos
Ande here:
Out of all the wonder an AI can achieve, the idea that it still makes typos is absurd.
I’ve watched a model write pages of clean, working code, juggle logic across six different edge cases, and then suddenly misspell a three-letter word in a simple sentence. It can outline a book, sketch a product roadmap, and help debug a regex… then stumble over “the” and “then” like a distracted five-year-old.
It feels wrong. If it’s “smart enough” to do all that, how can it possibly drop the ball on something so basic?
How can this happen?
The short answer is: it’s not actually “thinking” from first principles every time. It’s surfing probability.
Most modern AIs don’t have a little internal spellchecker they consult after they’ve “had a thought.” They generate text one token at a time by predicting “what most likely comes next” given everything so far, based on patterns learned from huge amounts of data.
Most of the time, “what most likely comes next” is exactly what we want:
the right word in the right language, or the correct line of code that matches the surrounding pattern. But every so often, the model’s internal guess lands in a weird local minimum: a slightly wrong token that still “looks plausible” according to its statistics, especially if:
- the training data had a bunch of noisy examples,
- the surrounding context is unusual,
- or the model is balancing multiple constraints at once (style, structure, instructions).
Humans feel the typo as a glitch in an otherwise smooth thought. For the model, it’s just one more slightly-off probability step in a long chain of guesses.
So you end up with this strange creature:
superhuman at pattern synthesis in the large,
and yet hilariously fallible in the small.
It’s not that it’s too dumb to spell.
It’s that it’s not really spelling at all.
It’s just betting, very quickly, on what comes next –
and sometimes, the dice come up “hte” instead of “the.”
