[krackers]

Was that ever solved? It seems that entire retort faded overnight, yet to my knowledge there was never any systematic analysis on cause or tokenizer change that fixed it. Maybe we just decided that this failure mode doesn't have any practical bearing given the existence of tool-use?

[Wowfunhappy]

...I think it really is irrelevant, isn't it? The LLM gets words as tokens, not strings of letters. If you asked me how many of the letter s is in Mississippi, but said I'm not allowed to spell out the word in my head and count the letters, I don't think I could do it.

This isn't a great analogy, because part of the challenge would be preventing myself from picturing the spelling in my head. But my point is, the AI is not getting the words as letters. The correct solution is tool use.

[krackers]

LLMs can learn to do arithmetic (without tool use), and they can learn a mapping from tokens to the letter counts contained therein (you could imagine trivially training on synthetic data). So there doesn't seem to be any fundamental barrier.