My favorite prompt: asking "How many e's are there in the word egregious". Always says three (ChatGPT 3.5, 4, 4 turbo). When you ask it which three, it realizes its mistake and apologizes (or sometimes tells you where they are that is completely wrong). Looks like it just outputs gibberish for these things.
ChatGPT is specifically bad at these kinds of tasks because of tokenization. If you plug your query into https://platform.openai.com/tokenizer, you can see that"egregious" is a single token, so the LLM doesn't actually see any "e" characters -- to answer your question it would have had to learn a fact about how the word was spelled from it's training data, and I imagine texts explicitly talking about how words are spelled are not very common.
IMO questions about spelling or number sense are pretty tired as gotchas, because they are all basically just artifacts of this implementation detail. There are other language models available that don't have this issue. BTW this is also the reason DALL-E etc suck at generating text in images.
It doesn’t even matter how many tokens there is, because LLMs are completely ignorant about how their input is structured. They don’t see letters or syllables cause they have no “eyes”. The closest analogy with a human is that vocal-ish concepts just emerge in their mind without any visual representation. They can only “recall” how many “e”s are there, but cannot look and count.
My initial analogy was already weak, so I guess there's no point in extending it. They key fact here is that tokens are inputs to what essentially is an overgrown matrix multiplication routine. Everything "AI" happens few levels of scientific abstractions higher, and is semantically disconnected from the "moving parts".
without the leading space, it is not common enough as a word to have become a token in its own right. Like the vast majority of lowercase words, in OpenAIs tokenizer you need to start " egregious" with a space character for the single token.
It’s always gibberish, it’s just really good at guessing.
I forgot what exactly it was I was doing. We were trying to get it to generate word lists of words ending with x or maybe it was starting with. For a marketing PoC and it made up oceans of words that not only didn’t start/end with x but mostly didn’t include x at all.
Isn’t this also why it can pass CS exams and job interviews better than like 95% of us, but then can’t help you solve the most simple business process in the world. Because nobody has asked that question two billion times on its training data.
But also it doesn't see characters. It sees tokens. The only way it would be reliably able to solve it is if it had a lookup table of token to characters. Which it likely doesn't.
You couldn't do it either unless you learned the exact matchings of all tokens to all characters in that token and their positions if you were given tokens as an input. You would have learned the meaning of the token, but not what the exact characters it represents.
Even if it sees tokens, I don't think it's an impossible task. Certainly an advanced enough LLM should be able to decipher token meanings, to know that a word is made up of the individual character tokens regardless of how the full word is tokenized. Maybe something gpt5 can do (or there's some real technical limitation which I don't understand)
The point is though that this is definitely not a task to evaluate an LLMs intelligence with. It's kind of like laughing at Einstein when he wouldn't be able to decipher hieroglyphs in any language without previous training for those hieroglyphs. Could Einstein potentially learn those hieroglyphs? Sure. But is it the best use of his time - or memory space?
There should be somewhere in the corpus, "the is spelled t h e" that this system can use to pull this out. We can ask gpt to spell out individual words in NATO phonetic and see how it does.
> There should be somewhere in the corpus, "the is spelled t h e" that this system can use to pull this out.
Such an approach would require an enormous table, containing all written words, including first and last names, and would still fail for made up words.
A more tractable approach would be to give it the map between the individual tokens and their letter component, but then you have the problem that this matching depends on the specific encoding used by the model (it varies between models). You could give it to the model during fine-tuning though.
It's as feasible as telling how many chars in html lead to this comment by looking at a screenshot. LLM doesn't see characters, tokens, numbers or its own activations. LLM is a "set of rules" component in a chinese room scenario. Anything an operator of that room does is lower-level.
GGP's idea suggests that an LLM, allegedly as a whole-room, receives something like: "hey, look at these tokens: <tokens>, please infer the continuation". This puts it into a nested-room's-operator position, which (1) it is not, (2) there's no nested room.
They cannot “figure” it, they could learn it but for that it would need to be in it's training data (which isn't because nobody is writing down the actual pairing in every byte pair encoding in plain text. Also the LLM has no clue about what encoding it uses unless you tell it somehow in the fine-tuning process or the prompt.)
I just asked it with Gemini; at first, it got it right. Then I asked if it was sure and it apologised and said 3 is the correct answer. When asked what are the 3 "e"s, it says:
> The three "e"s in "egregious" are:
>
> 1. The first "e" is located in the first syllable, between the "g" and the "g".
> 2. The second "e" is located in the second syllable, following the "r".
> 3. The third "e" is located in the third syllable, following the "i" and before the last "o".
