The Problem Is AI Programmers, Not AI (Public Board)

by FSK, Sunday, March 26, 2023, 17:30 (368 days ago) @ JoFrance

After a few attempts at AI wound up being extremely racist and right-wing (Tay Microsoft), they now have the AI severely trained and handcuffed so it doesn't say anything politically incorrect.

When you interacted with Tay, it was possible to change Tay's internal state. I.e. Tay learned from interacting with users.

ChatGPT does NOT learn. It has a fixed training set and fixed model. The only input it looks at is the last characters of its input (something like 20k characters).

For example, you can't try to teach ChatGPT something, come back later and expect it to remember. Its context is limited to the conversation buffer.

The only way they could make sure ChatGPT didn't say anything inappropriate was by severely handicapping it.


Complete thread:

 RSS Feed of thread