Large language models struggle to learn long-tail knowledge

I so cannot wait to use this LLM approach for a simple "50001-001" error message.

I blame app designers (looking at you, HTML, web browsers, CSS, C, IBM, RedHat, Aw shuck, nearly everybody and everything) for adopting such terseness of error codes when troubleshooting servers.

Seems like Google Search has given up on this much needed long-tail search results nowadays (presumably in favor of what it is to me are useless ads).

If you had to make me use ‘strace’ to solve your app error, you failed in error checking.

Seems like self driving cars have the same problem, all the rare edge cases that it is almost impossible to collect training data for. Maybe for cars this is a place where a voice interface would be useful. Instead of having the "driver" of the car be ready to take the wheel, have the system respond to voice commands in order to aid decisions. "Go left", "stop", "take the middle fork" "be careful here" could give the AI more data with which to make the correct decision and handle unusual cases. Maybe enough so to make self driving practical. A simple Green/Yellow/Red dashboard indicator showing an estimate of confidence could tell the driver when a voice command would be helpful. Instrumenting roads with sensors and navigation guidance would also help a lot. Self driving does NOT have to be totally autonomous, you can "cheat". A similar voice UI might have some use in large language models used in practical applications. Essentially a hybrid system, AI plus voice UI plus a few well thought out heuristics.