I'm convinced that most of the actual announcements were targeted more towards investors than towards employees. Or at least towards preventing employees from believing that the ship was on fire and starting to send out resumes.
No, you don't understand. There is no way to be sure your model won't bankrupt you by mistake if you don't verify everything it does, even with the initial prompt and checking. You can't let it run unsupervised for anything important. Verification takes at least half the time compared with doing it without AI help, not much speedup there.
I am currently working on a schema matching task, just mapping fields from one schema to another, semantically. If you sample 5 responses from the model, only 70% of the time they are all in agreement. Many times it contradicts its own predictions. And this is a simple task. There is task intent ambiguity, input ambiguity and model ambiguity in there. It is weaker for matching fields in other languages.
That's why I think AI, as wonderful and surprising as it is, can't do any high stakes task on its own. It is just a starting point that gets you closer to the solution from the first step, but not a task solver. It needs validation systems and feedback to improve past this stage, and that usually means human in the loop, rarely we can automate validation on top of AI.
3 comments
[ 3.4 ms ] story [ 20.3 ms ] threadI am currently working on a schema matching task, just mapping fields from one schema to another, semantically. If you sample 5 responses from the model, only 70% of the time they are all in agreement. Many times it contradicts its own predictions. And this is a simple task. There is task intent ambiguity, input ambiguity and model ambiguity in there. It is weaker for matching fields in other languages.
That's why I think AI, as wonderful and surprising as it is, can't do any high stakes task on its own. It is just a starting point that gets you closer to the solution from the first step, but not a task solver. It needs validation systems and feedback to improve past this stage, and that usually means human in the loop, rarely we can automate validation on top of AI.