The concept of understanding implies some form of meta-knowledge about the subject.

That can be solved if you teach it the meta-knowledge with intermediary steps, for example:

prompt: 34*3=

step1: 43 + 303 = step2: 12 + 1033 = step3: 12 + 10*9= step4: 12 + 90 = step5: 100 + 2 = step6: 102

result: 102

It’s hard to find such learning data though, but e.g. claude already uses intermediary steps. It preprocesses your input multiple times. It writes code, runs code to process your input, and that’s still not the final response. Unfortunately, it’s already smarter than some junior developers, and its consequence is worrying.