> having a personalizable automated agent I can train and control that can eliminate guesswork and undesired influences sounds good How do you ensure it elimin

> having a personalizable automated agent I can train and control that can eliminate guesswork and undesired influences sounds good How do you ensure it eliminates undesired influences though? Say for example you need to find a new brand of dish soap to buy because the one you were buying becomes unacceptable for whatever reason, how is your agent going to compare them? Unless you have carefully programmed it with pre-checked dish soap data it's going to need to go and get that from all the manufacturers, and they are going to try extremely hard to work out ways to game the system. Maybe brand X describes their soap as "sumptuous and delicious" and the AI goes "hey that sounds great! Those are the kind of words my human loves!" Until LLMs have evolved to the point that they're not susceptible to the grandma exploit, trusting them to remain sane and safe in the presence of outside input is a really bad idea in my view