for strawberry, it see it as [496, 675, 15717], which is str aw berry.
If you insert characters to breaks the tokens down, it find the correct result: how many r's are in "s"t"r"a"w"b"e"r"r"y" ?
> There are 3 'r's in "s"t"r"a"w"b"e"r"r"y".
The issue is that humans don't talk like this. I don't ask someone how many r's there are in strawberry by spelling out strawberry, I just say the word.
A token is roughly four letters [1], so, among other probable regressions, this would significantly reduce the effective context window.
[1] https://help.openai.com/en/articles/4936856-what-are-tokens-...
Yes, it's an issue. We want the convenience of sending human-legible commands to LLMs and getting back human-readable responses. That's the entire value proposition lol.
Problems can exist as instances of a class of problems. If you can't solve a problem, it's useful to know if it's a one off, or if it belongs to a larger class of problems, and which class it belongs to. In this case, the strawberry problem belongs to the much larger class of tokenization problems - if you think you've solved the tokenization problem class, you can test a model on the strawberry problem, with a few other examples from the class at large, and be confident that you've solved the class generally.
It's not about embodied human constraints or how humans do things; it's about what AI can and can't do. Right now, because of tokenization, things like understanding the number of Es in strawberry are outside the implicit model of the word in the LLM, with downstream effects on tasks it can complete. This affects moderation, parsing, generating prose, and all sorts of unexpected tasks. Having a workaround like forcing the model to insert spaces and operate on explicitly delimited text is useful when affected tasks appear.
Especially native speakers are prone to the mistake as they grew up learning english as illiterate children, from sounds only, compared to how most people learning english as second language do it, together with the textual representation.
Psychologists use this trick as well to figure out internal representations, for example the rorschach test.
And probably, if you asked random people in the street how many p's there is in "Philippines", you'd also get lots of wrong answers. It's tricky due to the double p and the initial p being part of an f sound. The demonym uses "F" as the first letter, and in many languages, say Spanish, also the country name uses an F.
No, I would actually be pretty confident you don’t ask people that question… at all. When is the last time you asked a human that question?
I can’t remember ever having anyone in real life ask me how many r’s are in strawberry. A lot of humans would probably refuse to answer such an off-the-wall and useless question, thus “failing” the test entirely.
A useless benchmark is useless.
In real life, people overwhelmingly do not need LLMs to count occurrences of a certain letter in a word.
If full artificial intelligence, as we're being promised, falls short in this simple way.
We are also not exactly looking letter by letter at everything we read.
However, this implies you need to know the word to begin with.
I can write "asdf" and you might be oblivious to what I mean. I can mention "adsf" to a JavaScript developer and he will immediately think of the tool versioning tool. Because context and familiarity is important.
"vox"
"il"
"ha"
"pen"
"tosh"
is how I read it.
A lot of schools teach kids to read with a syllabic method... so... super close to the tokenization concept.
https://chatgpt.com/share/4298efbf-1c29-474a-b333-c6cc1a3ce3...
User: Write “strawberry” one letter at a time, with a space between each letter. Then count how many r’s are in strawberry.
gpt-3.5-turbo: ASSISTANT s t r a w b e r r y
There are 2 r's in strawberry.
After some experimenting, it seems like the actual problem is that many LLMs can’t count.
User: How many r’s are in the following sequence of letters:
S/T/R/A/W/B/E/R/R/Y
gpt-4o-mini: In the sequence S/T/R/A/W/B/E/R/R/Y, there are 2 occurrences of the letter "R."
Oddly, if I change a bunch of the non-R letters, I seem to start getting the right answer.
You don't ask a human being how many r's there are in strawberry at all. The only reason you or anyone else asks that question is because it's an interesting quirk of how LLMs work that they struggle to answer it in that format. It's like an alien repeatedly showing humans an optical illusion that relies on the existence of our (literal) blind spot and using it as evidence of our supposed lack of intelligence.
Tokens aren’t the source of facts within a model. it’s an implementation detail and doesn’t inherently constrain how things could be counted.
Combinations of tokens get encoded, so if the feature isn't part of the information being carried forward into the network as it models the information in the corpus, the feature isn't modeled well, or at all. The consequence of having many character tokens is that the relevance of individual characters is lost, and you have to explicitly elicit the information. Models know that words have individual characters, but "strawberry" isn't encoded as a sequence of letters, it's encoded as an individual feature of the tokenizer embedding.
Other forms of tokenizing have other tradeoffs. The trend lately is to increase tokenizer dictionary scope, up to 128k in Llama3 from 50k in gpt-3. The more tokens, the more nuanced individual embedding features in that layer can be before downstream modeling.
Tokens inherently constrain how the notion of individual letters are modeled in the context of everything an LLM learns. In a vast majority of cases, the letters don't matter, so the features don't get mapped and carried downstream of the tokenizer.
What you're saying sounds plausible, but I don't see how we can conclude that definitively without at least some empirical tests, say a set words that predictably give an error along token boundaries.
The thing is, there are many ways a model can get around to answering the same question, it doesn't just depend on the architecture but also on how the training data is structured.
For example, if it turned out tokenization was the cause of this glitch, conceivably it could be fixed by adding enough documents with data relating to letter counts, providing another path to get the right output.
There is a lot of problems like that, that can be reformulated. For example if you ask it what is the biggest between 9.11 and 9.9, it will respond 9.9. If you look at how it's tokenized, you can see it restate an easy problem as something not straightforward even for a human. if you restart the problem by writing the number in full letters, it will correctly respnod.