Well that doesn’t apply to math, where LLMs are still garbage due to subpar tokenization. GPT-4 also still fails at multiple syntactic or phonetic constraints

Well that doesn’t apply to math, where LLMs are still garbage due to subpar tokenization. GPT-4 also still fails at multiple syntactic or phonetic constraints at once, due to its tokenization scheme.