Why Can't 100B-Parameter AI Models Create a Simple Puzzle?

https://saigaddam.medium.com/why-do-200-billion-parameter-models-fail-to-create-a-simple-puzzle-13ecd2833e76

by ChaitanyaSai • 2 years ago

1 1 2 years ago

saigaddam.medium.com

jqpabc123• 2 years ago

Because the AI model is built around statistics and prediction --- not logic and reason. In other words, it's primarily a talking database.

Asking a database for logic is a mis-application. Acceptable logic can be produced in some cases --- but this is due to random chance (it just happened to find a similar problem encoded in memory) more so than any "thinking" or reason. Do you feel lucky?

Ask for any logical result not seen before (like creating a puzzle) and it is likely to fail miserably. Memory is a poor substitute for logic.

How many applications can accept faulty logic at high cost? There are some but these are rather limited in my opinion.

The ultimate test --- would you do business with a bank run by AI?

Would AI bank consistently offer better interest rates on savings and loans than a conventional one? Or would it be easily victimized by con artists? Would it simply refuse any further loans as a result?

If AI can't run a bank successfully, it certainly can't run the world.

Why Can't 100B-Parameter AI Models Create a Simple Puzzle?

Related Stories