[bglazer]

This is a very tiring criticism. Yes, this is true. But, it's an implementation detail (tokenization) that has very little bearing on the practical utility of these tools. How often are you relying on LLM's to count letters in words?

[1970-01-01]

The implementation detail is that we keep finding them! After this, it couldn't locate a seahorse emoji without freaking out. At some point we need to have a test: there are two drinks before you. One is water, the other is whatever the LLM thought you might like to drink after it completed refactoring the codebase. Choose wisely.

[101008]

It's an example that shows that if these models aren't trained in a specific problem, they may have a hard time solving it for you.

[altruios]

An analogy is asking someone who is colorblind how many colors are on a sheet of paper. What you are probing isn't reasoning, it's perception. If you can't see the input, you can't reason about the input.

[9rx]

> What you are probing isn't reasoning, it's perception.

Its both. A colorblind person will admit their shortcomings and, if compelled to be helpful like an LLM is, will reason their way to finding a solution that works around their limitations.

But as LLMs lack a way to reason, you get nonsense instead.

[altruios]

What tools does the LLM have access to that would reveal sub-token characters to it?

This assumes the colorblind person both believes it is true that they are colorblind, in a world where that can be verified, and possesses tools to overcome these limitations.

You have to be much more clever to 'see' an atom before the invention of a microscope, if the tool doesn't exist: most of the time you are SOL.

[Uehreka]

No, it’s an example that shows that LLMs still use a tokenizer, which is not an impediment for almost any task (even many where you would expect it to be, like searching a codebase for variants of a variable name in different cases).

[8note]

the question remains: is the tokenizer going to be a fundamental limit to my task? how do i know ahead of time?

[worldsayshi]

Would it limit a person getting your instructions in Chinese? Tokenisation pretty much means that the LLM is reading symbols instead of phonemes.

This makes me wonder if LLMs works better in Chinese.

[victorbjorklund]

No, it is the issue with the tokenizer.

[iAMkenough]

The criticism would stop if the implementation issue was fixed.

It's an example of a simple task. How often are you relying on LLMs to complete simple tasks?

[andy99]

At this point if I was openAI I wouldn’t bother fixing this to give pedants something to get excited about.

[properbrew]

Unless they fixed this in 25 minutes (possible?) it correctly counts 1 `r`.

https://chatgpt.com/share/6941df90-789c-8005-8783-6e1c76cdfc...