Alexandre Rademaker:  We don’t necessarily need to throw away the meanings. A safe translation should account for a 1-N mapping.. from surface to logical representations. Context or even some statistical preference can select the most preferable reading.

Yes.  That is why we need a top-level symbolic processor that can determine what to do for any particular issue that may arise.

Alex Shkotin:  With robots it's better not to use vague terms or sentences. It's dangerous.  Good robots will tell: I don't understand, bad ones can make a mess of things.

Alex:  My way is to represent knowledge formally. The precision of knowledge itself remains the same initially and may be better after we apply knowledge processing algorithms to this formalized knowledge. 

Think of the top-level symbolic processor as a gate-keeper.  It is in the best position to determine what to do.  In many cases, the best thing is to ask a question or even a series of questions before making a decision. 

The top-level processor may use LLMs in the simplest and most secure way:  Translate a query in any natural language to and from whatever internal form the system uses.  After the top-level processor has determined what to do, it can pass the translated result to whatever subroutines can handle it.  Those subroutines may or may not use LLMs or many, many other tools of various kinds.

Basic point:  One size does not fit all.  The top-level processor determines which of many internal processors should or should not be invoked.  Anything that seems dangerous can be sent to a security system, which may or may not reject the input or even send it to proper authorities to handle it.

John