I sent a long table to AI, about 3000 rows and two columns, and asked AI how many rows there were in total. AI answered me 1224 rows. 
I was really disappointed.
Is there any good solution at present?
Quite simply - don’t use a Large Language Model to count something.
LLMs are, generally, not good at counting things. They are language models that are good for taking words and generating other words based on… words.
Your best bet is to use RAG-like solutions, where you use the LLM to “understand” what the user wants and to indicate that using Function Calling. You would then rwite the code to do the counting, provide the answer back to the LLM, and have it generate a human-like response.
LLMs are not sources of truth. But they’re pretty good at language.
I know exactly what you mean, but if the AI can’t solve some math problems through built-in or other connected AI, it’s not good AI.
Respectfully - I disagree. But it might depend on what you mean by “the AI”.
Gemini has a specific set of things it does well. It is tuned for those things. And all of those things involve language.
Doing math problems and searching for strings are solved problems. And they didn’t need AI or ML to do.
Trying to either cram them into a model or make the model do them in pursuit of some fantasy AGI is trying to make the model do more than it needs to do, and there is a cost to do that.
Instead - let’s focus on making our models better at specific tasks and work well with other tools that work well at their specific tasks. Then building systems that put all of them together. Now… is that “put all of them together” system an AI? I’d argue not. It uses an AI, sure. But it isn’t one.
And I think that’s more than fine.